xingcheng
/
disco-diffusion
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
151 Commits
25 Branches
1 Tag
12 MiB
 
 
 
Tag: Branch: Tree: 49d3b2da7d
Branches Tags
${ item.name }
Create tag ${ searchTerm }
Create branch ${ searchTerm }
from '49d3b2da7d'
${ noResults }
disco-diffusion/Disco_Diffusion.ipynb

4006 lines
211 KiB
Raw Blame History

{
"cells": [
{
"cell_type": "markdown",
"metadata": {
"id": "view-in-github",
"colab_type": "text"
},
"source": [
"<a href=\"https://colab.research.google.com/github/alembics/disco-diffusion/blob/main/Disco_Diffusion.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "TitleTop"
},
"source": [
"# Disco Diffusion v5.3.999 - Short-lived inclusions\n",
"\n",
"Disco Diffusion - http://discodiffusion.com/ , https://github.com/alembics/disco-diffusion\n",
"\n",
"In case of confusion, Disco is the name of this notebook edit. The diffusion model in use is Katherine Crowson's fine-tuned 512x512 model\n",
"\n",
"For issues, join the [Disco Diffusion Discord](https://discord.gg/msEZBy4HxA) or message us on twitter at [@somnai_dreams](https://twitter.com/somnai_dreams) or [@gandamu](https://twitter.com/gandamu_ml)"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "CreditsChTop"
},
"source": [
"### Credits & Changelog \u2b07\ufe0f"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "Credits"
},
"source": [
"#### Credits\n",
"\n",
"Original notebook by Katherine Crowson (https://github.com/crowsonkb, https://twitter.com/RiversHaveWings). It uses either OpenAI's 256x256 unconditional ImageNet or Katherine Crowson's fine-tuned 512x512 diffusion model (https://github.com/openai/guided-diffusion), together with CLIP (https://github.com/openai/CLIP) to connect text prompts with images.\n",
"\n",
"Modified by Daniel Russell (https://github.com/russelldc, https://twitter.com/danielrussruss) to include (hopefully) optimal params for quick generations in 15-100 timesteps rather than 1000, as well as more robust augmentations.\n",
"\n",
"Further improvements from Dango233 and nshepperd helped improve the quality of diffusion in general, and especially so for shorter runs like this notebook aims to achieve.\n",
"\n",
"Vark added code to load in multiple Clip models at once, which all prompts are evaluated against, which may greatly improve accuracy.\n",
"\n",
"The latest zoom, pan, rotation, and keyframes features were taken from Chigozie Nri's VQGAN Zoom Notebook (https://github.com/chigozienri, https://twitter.com/chigozienri)\n",
"\n",
"Advanced DangoCutn Cutout method is also from Dango223.\n",
"\n",
"--\n",
"\n",
"Disco:\n",
"\n",
"Somnai (https://twitter.com/Somnai_dreams) added Diffusion Animation techniques, QoL improvements and various implementations of tech and techniques, mostly listed in the changelog below.\n",
"\n",
"3D animation implementation added by Adam Letts (https://twitter.com/gandamu_ml) in collaboration with Somnai. Creation of disco.py and ongoing maintenance.\n",
"\n",
"Turbo feature by Chris Allen (https://twitter.com/zippy731)\n",
"\n",
"Improvements to ability to run on local systems, Windows support, and dependency installation by HostsServer (https://twitter.com/HostsServer)\n",
"\n",
"VR Mode by Tom Mason (https://twitter.com/nin_artificial)\n",
"\n",
"Horizontal and Vertical symmetry functionality by nshepperd. Symmetry transformation_steps by huemin (https://twitter.com/huemin_art). Symmetry integration into Disco Diffusion by Dmitrii Tochilkin (https://twitter.com/cut_pow).\n",
"\n",
"Warp and custom model support by Alex Spirin (https://twitter.com/devdef)."
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "LicenseTop"
},
"source": [
"#### License"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "License"
},
"source": [
"Licensed under the MIT License\n",
"\n",
"Copyright (c) 2021 Katherine Crowson \n",
"\n",
"Permission is hereby granted, free of charge, to any person obtaining a copy\n",
"of this software and associated documentation files (the \"Software\"), to deal\n",
"in the Software without restriction, including without limitation the rights\n",
"to use, copy, modify, merge, publish, distribute, sublicense, and/or sell\n",
"copies of the Software, and to permit persons to whom the Software is\n",
"furnished to do so, subject to the following conditions:\n",
"\n",
"The above copyright notice and this permission notice shall be included in\n",
"all copies or substantial portions of the Software.\n",
"\n",
"THE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR\n",
"IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,\n",
"FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE\n",
"AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER\n",
"LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,\n",
"OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN\n",
"THE SOFTWARE.\n",
"\n",
"--\n",
"\n",
"MIT License\n",
"\n",
"Copyright (c) 2019 Intel ISL (Intel Intelligent Systems Lab)\n",
"\n",
"Permission is hereby granted, free of charge, to any person obtaining a copy\n",
"of this software and associated documentation files (the \"Software\"), to deal\n",
"in the Software without restriction, including without limitation the rights\n",
"to use, copy, modify, merge, publish, distribute, sublicense, and/or sell\n",
"copies of the Software, and to permit persons to whom the Software is\n",
"furnished to do so, subject to the following conditions:\n",
"\n",
"The above copyright notice and this permission notice shall be included in all\n",
"copies or substantial portions of the Software.\n",
"\n",
"THE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR\n",
"IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,\n",
"FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE\n",
"AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER\n",
"LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,\n",
"OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE\n",
"SOFTWARE.\n",
"\n",
"--\n",
"\n",
"Licensed under the MIT License\n",
"\n",
"Copyright (c) 2021 Maxwell Ingham\n",
"\n",
"Copyright (c) 2022 Adam Letts \n",
"\n",
"Copyright (c) 2022 Alex Spirin\n",
"\n",
"Permission is hereby granted, free of charge, to any person obtaining a copy\n",
"of this software and associated documentation files (the \"Software\"), to deal\n",
"in the Software without restriction, including without limitation the rights\n",
"to use, copy, modify, merge, publish, distribute, sublicense, and/or sell\n",
"copies of the Software, and to permit persons to whom the Software is\n",
"furnished to do so, subject to the following conditions:\n",
"\n",
"The above copyright notice and this permission notice shall be included in\n",
"all copies or substantial portions of the Software.\n",
"\n",
"THE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR\n",
"IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,\n",
"FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE\n",
"AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER\n",
"LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,\n",
"OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN\n",
"THE SOFTWARE.\n",
"\n",
"--\n",
"flow-related - https://github.com/NVIDIA/flownet2-pytorch/blob/master/LICENSE\n",
"--\n",
"Copyright 2017 NVIDIA CORPORATION\n",
"\n",
"Licensed under the Apache License, Version 2.0 (the \"License\");\n",
"you may not use this file except in compliance with the License.\n",
"You may obtain a copy of the License at\n",
"\n",
" http://www.apache.org/licenses/LICENSE-2.0\n",
"\n",
"Unless required by applicable law or agreed to in writing, software\n",
"distributed under the License is distributed on an \"AS IS\" BASIS,\n",
"WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n",
"See the License for the specific language governing permissions and\n",
"limitations under the License.\n",
"\n",
"--\n",
"consistencyChecker - https://github.com/cysmith/neural-style-tf\n",
"--\n",
" GNU GENERAL PUBLIC LICENSE\n",
" Version 3, 29 June 2007\n",
"\n",
" Copyright (C) 2007 Free Software Foundation, Inc. <http://fsf.org/>\n",
" Everyone is permitted to copy and distribute verbatim copies\n",
" of this license document, but changing it is not allowed.\n",
"\n",
" Preamble\n",
"\n",
" The GNU General Public License is a free, copyleft license for\n",
"software and other kinds of works.\n",
"\n",
" The licenses for most software and other practical works are designed\n",
"to take away your freedom to share and change the works. By contrast,\n",
"the GNU General Public License is intended to guarantee your freedom to\n",
"share and change all versions of a program--to make sure it remains free\n",
"software for all its users. We, the Free Software Foundation, use the\n",
"GNU General Public License for most of our software; it applies also to\n",
"any other work released this way by its authors. You can apply it to\n",
"your programs, too.\n",
"\n",
" When we speak of free software, we are referring to freedom, not\n",
"price. Our General Public Licenses are designed to make sure that you\n",
"have the freedom to distribute copies of free software (and charge for\n",
"them if you wish), that you receive source code or can get it if you\n",
"want it, that you can change the software or use pieces of it in new\n",
"free programs, and that you know you can do these things.\n",
"\n",
" To protect your rights, we need to prevent others from denying you\n",
"these rights or asking you to surrender the rights. Therefore, you have\n",
"certain responsibilities if you distribute copies of the software, or if\n",
"you modify it: responsibilities to respect the freedom of others.\n",
"\n",
" For example, if you distribute copies of such a program, whether\n",
"gratis or for a fee, you must pass on to the recipients the same\n",
"freedoms that you received. You must make sure that they, too, receive\n",
"or can get the source code. And you must show them these terms so they\n",
"know their rights.\n",
"\n",
" Developers that use the GNU GPL protect your rights with two steps:\n",
"(1) assert copyright on the software, and (2) offer you this License\n",
"giving you legal permission to copy, distribute and/or modify it.\n",
"\n",
" For the developers' and authors' protection, the GPL clearly explains\n",
"that there is no warranty for this free software. For both users' and\n",
"authors' sake, the GPL requires that modified versions be marked as\n",
"changed, so that their problems will not be attributed erroneously to\n",
"authors of previous versions.\n",
"\n",
" Some devices are designed to deny users access to install or run\n",
"modified versions of the software inside them, although the manufacturer\n",
"can do so. This is fundamentally incompatible with the aim of\n",
"protecting users' freedom to change the software. The systematic\n",
"pattern of such abuse occurs in the area of products for individuals to\n",
"use, which is precisely where it is most unacceptable. Therefore, we\n",
"have designed this version of the GPL to prohibit the practice for those\n",
"products. If such problems arise substantially in other domains, we\n",
"stand ready to extend this provision to those domains in future versions\n",
"of the GPL, as needed to protect the freedom of users.\n",
"\n",
" Finally, every program is threatened constantly by software patents.\n",
"States should not allow patents to restrict development and use of\n",
"software on general-purpose computers, but in those that do, we wish to\n",
"avoid the special danger that patents applied to a free program could\n",
"make it effectively proprietary. To prevent this, the GPL assures that\n",
"patents cannot be used to render the program non-free.\n",
"\n",
" The precise terms and conditions for copying, distribution and\n",
"modification follow.\n",
"\n",
" TERMS AND CONDITIONS\n",
"\n",
" 0. Definitions.\n",
"\n",
" \"This License\" refers to version 3 of the GNU General Public License.\n",
"\n",
" \"Copyright\" also means copyright-like laws that apply to other kinds of\n",
"works, such as semiconductor masks.\n",
"\n",
" \"The Program\" refers to any copyrightable work licensed under this\n",
"License. Each licensee is addressed as \"you\". \"Licensees\" and\n",
"\"recipients\" may be individuals or organizations.\n",
"\n",
" To \"modify\" a work means to copy from or adapt all or part of the work\n",
"in a fashion requiring copyright permission, other than the making of an\n",
"exact copy. The resulting work is called a \"modified version\" of the\n",
"earlier work or a work \"based on\" the earlier work.\n",
"\n",
" A \"covered work\" means either the unmodified Program or a work based\n",
"on the Program.\n",
"\n",
" To \"propagate\" a work means to do anything with it that, without\n",
"permission, would make you directly or secondarily liable for\n",
"infringement under applicable copyright law, except executing it on a\n",
"computer or modifying a private copy. Propagation includes copying,\n",
"distribution (with or without modification), making available to the\n",
"public, and in some countries other activities as well.\n",
"\n",
" To \"convey\" a work means any kind of propagation that enables other\n",
"parties to make or receive copies. Mere interaction with a user through\n",
"a computer network, with no transfer of a copy, is not conveying.\n",
"\n",
" An interactive user interface displays \"Appropriate Legal Notices\"\n",
"to the extent that it includes a convenient and prominently visible\n",
"feature that (1) displays an appropriate copyright notice, and (2)\n",
"tells the user that there is no warranty for the work (except to the\n",
"extent that warranties are provided), that licensees may convey the\n",
"work under this License, and how to view a copy of this License. If\n",
"the interface presents a list of user commands or options, such as a\n",
"menu, a prominent item in the list meets this criterion.\n",
"\n",
" 1. Source Code.\n",
"\n",
" The \"source code\" for a work means the preferred form of the work\n",
"for making modifications to it. \"Object code\" means any non-source\n",
"form of a work.\n",
"\n",
" A \"Standard Interface\" means an interface that either is an official\n",
"standard defined by a recognized standards body, or, in the case of\n",
"interfaces specified for a particular programming language, one that\n",
"is widely used among developers working in that language.\n",
"\n",
" The \"System Libraries\" of an executable work include anything, other\n",
"than the work as a whole, that (a) is included in the normal form of\n",
"packaging a Major Component, but which is not part of that Major\n",
"Component, and (b) serves only to enable use of the work with that\n",
"Major Component, or to implement a Standard Interface for which an\n",
"implementation is available to the public in source code form. A\n",
"\"Major Component\", in this context, means a major essential component\n",
"(kernel, window system, and so on) of the specific operating system\n",
"(if any) on which the executable work runs, or a compiler used to\n",
"produce the work, or an object code interpreter used to run it.\n",
"\n",
" The \"Corresponding Source\" for a work in object code form means all\n",
"the source code needed to generate, install, and (for an executable\n",
"work) run the object code and to modify the work, including scripts to\n",
"control those activities. However, it does not include the work's\n",
"System Libraries, or general-purpose tools or generally available free\n",
"programs which are used unmodified in performing those activities but\n",
"which are not part of the work. For example, Corresponding Source\n",
"includes interface definition files associated with source files for\n",
"the work, and the source code for shared libraries and dynamically\n",
"linked subprograms that the work is specifically designed to require,\n",
"such as by intimate data communication or control flow between those\n",
"subprograms and other parts of the work.\n",
"\n",
" The Corresponding Source need not include anything that users\n",
"can regenerate automatically from other parts of the Corresponding\n",
"Source.\n",
"\n",
" The Corresponding Source for a work in source code form is that\n",
"same work.\n",
"\n",
" 2. Basic Permissions.\n",
"\n",
" All rights granted under this License are granted for the term of\n",
"copyright on the Program, and are irrevocable provided the stated\n",
"conditions are met. This License explicitly affirms your unlimited\n",
"permission to run the unmodified Program. The output from running a\n",
"covered work is covered by this License only if the output, given its\n",
"content, constitutes a covered work. This License acknowledges your\n",
"rights of fair use or other equivalent, as provided by copyright law.\n",
"\n",
" You may make, run and propagate covered works that you do not\n",
"convey, without conditions so long as your license otherwise remains\n",
"in force. You may convey covered works to others for the sole purpose\n",
"of having them make modifications exclusively for you, or provide you\n",
"with facilities for running those works, provided that you comply with\n",
"the terms of this License in conveying all material for which you do\n",
"not control copyright. Those thus making or running the covered works\n",
"for you must do so exclusively on your behalf, under your direction\n",
"and control, on terms that prohibit them from making any copies of\n",
"your copyrighted material outside their relationship with you.\n",
"\n",
" Conveying under any other circumstances is permitted solely under\n",
"the conditions stated below. Sublicensing is not allowed; section 10\n",
"makes it unnecessary.\n",
"\n",
" 3. Protecting Users' Legal Rights From Anti-Circumvention Law.\n",
"\n",
" No covered work shall be deemed part of an effective technological\n",
"measure under any applicable law fulfilling obligations under article\n",
"11 of the WIPO copyright treaty adopted on 20 December 1996, or\n",
"similar laws prohibiting or restricting circumvention of such\n",
"measures.\n",
"\n",
" When you convey a covered work, you waive any legal power to forbid\n",
"circumvention of technological measures to the extent such circumvention\n",
"is effected by exercising rights under this License with respect to\n",
"the covered work, and you disclaim any intention to limit operation or\n",
"modification of the work as a means of enforcing, against the work's\n",
"users, your or third parties' legal rights to forbid circumvention of\n",
"technological measures.\n",
"\n",
" 4. Conveying Verbatim Copies.\n",
"\n",
" You may convey verbatim copies of the Program's source code as you\n",
"receive it, in any medium, provided that you conspicuously and\n",
"appropriately publish on each copy an appropriate copyright notice;\n",
"keep intact all notices stating that this License and any\n",
"non-permissive terms added in accord with section 7 apply to the code;\n",
"keep intact all notices of the absence of any warranty; and give all\n",
"recipients a copy of this License along with the Program.\n",
"\n",
" You may charge any price or no price for each copy that you convey,\n",
"and you may offer support or warranty protection for a fee.\n",
"\n",
" 5. Conveying Modified Source Versions.\n",
"\n",
" You may convey a work based on the Program, or the modifications to\n",
"produce it from the Program, in the form of source code under the\n",
"terms of section 4, provided that you also meet all of these conditions:\n",
"\n",
" a) The work must carry prominent notices stating that you modified\n",
" it, and giving a relevant date.\n",
"\n",
" b) The work must carry prominent notices stating that it is\n",
" released under this License and any conditions added under section\n",
" 7. This requirement modifies the requirement in section 4 to\n",
" \"keep intact all notices\".\n",
"\n",
" c) You must license the entire work, as a whole, under this\n",
" License to anyone who comes into possession of a copy. This\n",
" License will therefore apply, along with any applicable section 7\n",
" additional terms, to the whole of the work, and all its parts,\n",
" regardless of how they are packaged. This License gives no\n",
" permission to license the work in any other way, but it does not\n",
" invalidate such permission if you have separately received it.\n",
"\n",
" d) If the work has interactive user interfaces, each must display\n",
" Appropriate Legal Notices; however, if the Program has interactive\n",
" interfaces that do not display Appropriate Legal Notices, your\n",
" work need not make them do so.\n",
"\n",
" A compilation of a covered work with other separate and independent\n",
"works, which are not by their nature extensions of the covered work,\n",
"and which are not combined with it such as to form a larger program,\n",
"in or on a volume of a storage or distribution medium, is called an\n",
"\"aggregate\" if the compilation and its resulting copyright are not\n",
"used to limit the access or legal rights of the compilation's users\n",
"beyond what the individual works permit. Inclusion of a covered work\n",
"in an aggregate does not cause this License to apply to the other\n",
"parts of the aggregate.\n",
"\n",
" 6. Conveying Non-Source Forms.\n",
"\n",
" You may convey a covered work in object code form under the terms\n",
"of sections 4 and 5, provided that you also convey the\n",
"machine-readable Corresponding Source under the terms of this License,\n",
"in one of these ways:\n",
"\n",
" a) Convey the object code in, or embodied in, a physical product\n",
" (including a physical distribution medium), accompanied by the\n",
" Corresponding Source fixed on a durable physical medium\n",
" customarily used for software interchange.\n",
"\n",
" b) Convey the object code in, or embodied in, a physical product\n",
" (including a physical distribution medium), accompanied by a\n",
" written offer, valid for at least three years and valid for as\n",
" long as you offer spare parts or customer support for that product\n",
" model, to give anyone who possesses the object code either (1) a\n",
" copy of the Corresponding Source for all the software in the\n",
" product that is covered by this License, on a durable physical\n",
" medium customarily used for software interchange, for a price no\n",
" more than your reasonable cost of physically performing this\n",
" conveying of source, or (2) access to copy the\n",
" Corresponding Source from a network server at no charge.\n",
"\n",
" c) Convey individual copies of the object code with a copy of the\n",
" written offer to provide the Corresponding Source. This\n",
" alternative is allowed only occasionally and noncommercially, and\n",
" only if you received the object code with such an offer, in accord\n",
" with subsection 6b.\n",
"\n",
" d) Convey the object code by offering access from a designated\n",
" place (gratis or for a charge), and offer equivalent access to the\n",
" Corresponding Source in the same way through the same place at no\n",
" further charge. You need not require recipients to copy the\n",
" Corresponding Source along with the object code. If the place to\n",
" copy the object code is a network server, the Corresponding Source\n",
" may be on a different server (operated by you or a third party)\n",
" that supports equivalent copying facilities, provided you maintain\n",
" clear directions next to the object code saying where to find the\n",
" Corresponding Source. Regardless of what server hosts the\n",
" Corresponding Source, you remain obligated to ensure that it is\n",
" available for as long as needed to satisfy these requirements.\n",
"\n",
" e) Convey the object code using peer-to-peer transmission, provided\n",
" you inform other peers where the object code and Corresponding\n",
" Source of the work are being offered to the general public at no\n",
" charge under subsection 6d.\n",
"\n",
" A separable portion of the object code, whose source code is excluded\n",
"from the Corresponding Source as a System Library, need not be\n",
"included in conveying the object code work.\n",
"\n",
" A \"User Product\" is either (1) a \"consumer product\", which means any\n",
"tangible personal property which is normally used for personal, family,\n",
"or household purposes, or (2) anything designed or sold for incorporation\n",
"into a dwelling. In determining whether a product is a consumer product,\n",
"doubtful cases shall be resolved in favor of coverage. For a particular\n",
"product received by a particular user, \"normally used\" refers to a\n",
"typical or common use of that class of product, regardless of the status\n",
"of the particular user or of the way in which the particular user\n",
"actually uses, or expects or is expected to use, the product. A product\n",
"is a consumer product regardless of whether the product has substantial\n",
"commercial, industrial or non-consumer uses, unless such uses represent\n",
"the only significant mode of use of the product.\n",
"\n",
" \"Installation Information\" for a User Product means any methods,\n",
"procedures, authorization keys, or other information required to install\n",
"and execute modified versions of a covered work in that User Product from\n",
"a modified version of its Corresponding Source. The information must\n",
"suffice to ensure that the continued functioning of the modified object\n",
"code is in no case prevented or interfered with solely because\n",
"modification has been made.\n",
"\n",
" If you convey an object code work under this section in, or with, or\n",
"specifically for use in, a User Product, and the conveying occurs as\n",
"part of a transaction in which the right of possession and use of the\n",
"User Product is transferred to the recipient in perpetuity or for a\n",
"fixed term (regardless of how the transaction is characterized), the\n",
"Corresponding Source conveyed under this section must be accompanied\n",
"by the Installation Information. But this requirement does not apply\n",
"if neither you nor any third party retains the ability to install\n",
"modified object code on the User Product (for example, the work has\n",
"been installed in ROM).\n",
"\n",
" The requirement to provide Installation Information does not include a\n",
"requirement to continue to provide support service, warranty, or updates\n",
"for a work that has been modified or installed by the recipient, or for\n",
"the User Product in which it has been modified or installed. Access to a\n",
"network may be denied when the modification itself materially and\n",
"adversely affects the operation of the network or violates the rules and\n",
"protocols for communication across the network.\n",
"\n",
" Corresponding Source conveyed, and Installation Information provided,\n",
"in accord with this section must be in a format that is publicly\n",
"documented (and with an implementation available to the public in\n",
"source code form), and must require no special password or key for\n",
"unpacking, reading or copying.\n",
"\n",
" 7. Additional Terms.\n",
"\n",
" \"Additional permissions\" are terms that supplement the terms of this\n",
"License by making exceptions from one or more of its conditions.\n",
"Additional permissions that are applicable to the entire Program shall\n",
"be treated as though they were included in this License, to the extent\n",
"that they are valid under applicable law. If additional permissions\n",
"apply only to part of the Program, that part may be used separately\n",
"under those permissions, but the entire Program remains governed by\n",
"this License without regard to the additional permissions.\n",
"\n",
" When you convey a copy of a covered work, you may at your option\n",
"remove any additional permissions from that copy, or from any part of\n",
"it. (Additional permissions may be written to require their own\n",
"removal in certain cases when you modify the work.) You may place\n",
"additional permissions on material, added by you to a covered work,\n",
"for which you have or can give appropriate copyright permission.\n",
"\n",
" Notwithstanding any other provision of this License, for material you\n",
"add to a covered work, you may (if authorized by the copyright holders of\n",
"that material) supplement the terms of this License with terms:\n",
"\n",
" a) Disclaiming warranty or limiting liability differently from the\n",
" terms of sections 15 and 16 of this License; or\n",
"\n",
" b) Requiring preservation of specified reasonable legal notices or\n",
" author attributions in that material or in the Appropriate Legal\n",
" Notices displayed by works containing it; or\n",
"\n",
" c) Prohibiting misrepresentation of the origin of that material, or\n",
" requiring that modified versions of such material be marked in\n",
" reasonable ways as different from the original version; or\n",
"\n",
" d) Limiting the use for publicity purposes of names of licensors or\n",
" authors of the material; or\n",
"\n",
" e) Declining to grant rights under trademark law for use of some\n",
" trade names, trademarks, or service marks; or\n",
"\n",
" f) Requiring indemnification of licensors and authors of that\n",
" material by anyone who conveys the material (or modified versions of\n",
" it) with contractual assumptions of liability to the recipient, for\n",
" any liability that these contractual assumptions directly impose on\n",
" those licensors and authors.\n",
"\n",
" All other non-permissive additional terms are considered \"further\n",
"restrictions\" within the meaning of section 10. If the Program as you\n",
"received it, or any part of it, contains a notice stating that it is\n",
"governed by this License along with a term that is a further\n",
"restriction, you may remove that term. If a license document contains\n",
"a further restriction but permits relicensing or conveying under this\n",
"License, you may add to a covered work material governed by the terms\n",
"of that license document, provided that the further restriction does\n",
"not survive such relicensing or conveying.\n",
"\n",
" If you add terms to a covered work in accord with this section, you\n",
"must place, in the relevant source files, a statement of the\n",
"additional terms that apply to those files, or a notice indicating\n",
"where to find the applicable terms.\n",
"\n",
" Additional terms, permissive or non-permissive, may be stated in the\n",
"form of a separately written license, or stated as exceptions;\n",
"the above requirements apply either way.\n",
"\n",
" 8. Termination.\n",
"\n",
" You may not propagate or modify a covered work except as expressly\n",
"provided under this License. Any attempt otherwise to propagate or\n",
"modify it is void, and will automatically terminate your rights under\n",
"this License (including any patent licenses granted under the third\n",
"paragraph of section 11).\n",
"\n",
" However, if you cease all violation of this License, then your\n",
"license from a particular copyright holder is reinstated (a)\n",
"provisionally, unless and until the copyright holder explicitly and\n",
"finally terminates your license, and (b) permanently, if the copyright\n",
"holder fails to notify you of the violation by some reasonable means\n",
"prior to 60 days after the cessation.\n",
"\n",
" Moreover, your license from a particular copyright holder is\n",
"reinstated permanently if the copyright holder notifies you of the\n",
"violation by some reasonable means, this is the first time you have\n",
"received notice of violation of this License (for any work) from that\n",
"copyright holder, and you cure the violation prior to 30 days after\n",
"your receipt of the notice.\n",
"\n",
" Termination of your rights under this section does not terminate the\n",
"licenses of parties who have received copies or rights from you under\n",
"this License. If your rights have been terminated and not permanently\n",
"reinstated, you do not qualify to receive new licenses for the same\n",
"material under section 10.\n",
"\n",
" 9. Acceptance Not Required for Having Copies.\n",
"\n",
" You are not required to accept this License in order to receive or\n",
"run a copy of the Program. Ancillary propagation of a covered work\n",
"occurring solely as a consequence of using peer-to-peer transmission\n",
"to receive a copy likewise does not require acceptance. However,\n",
"nothing other than this License grants you permission to propagate or\n",
"modify any covered work. These actions infringe copyright if you do\n",
"not accept this License. Therefore, by modifying or propagating a\n",
"covered work, you indicate your acceptance of this License to do so.\n",
"\n",
" 10. Automatic Licensing of Downstream Recipients.\n",
"\n",
" Each time you convey a covered work, the recipient automatically\n",
"receives a license from the original licensors, to run, modify and\n",
"propagate that work, subject to this License. You are not responsible\n",
"for enforcing compliance by third parties with this License.\n",
"\n",
" An \"entity transaction\" is a transaction transferring control of an\n",
"organization, or substantially all assets of one, or subdividing an\n",
"organization, or merging organizations. If propagation of a covered\n",
"work results from an entity transaction, each party to that\n",
"transaction who receives a copy of the work also receives whatever\n",
"licenses to the work the party's predecessor in interest had or could\n",
"give under the previous paragraph, plus a right to possession of the\n",
"Corresponding Source of the work from the predecessor in interest, if\n",
"the predecessor has it or can get it with reasonable efforts.\n",
"\n",
" You may not impose any further restrictions on the exercise of the\n",
"rights granted or affirmed under this License. For example, you may\n",
"not impose a license fee, royalty, or other charge for exercise of\n",
"rights granted under this License, and you may not initiate litigation\n",
"(including a cross-claim or counterclaim in a lawsuit) alleging that\n",
"any patent claim is infringed by making, using, selling, offering for\n",
"sale, or importing the Program or any portion of it.\n",
"\n",
" 11. Patents.\n",
"\n",
" A \"contributor\" is a copyright holder who authorizes use under this\n",
"License of the Program or a work on which the Program is based. The\n",
"work thus licensed is called the contributor's \"contributor version\".\n",
"\n",
" A contributor's \"essential patent claims\" are all patent claims\n",
"owned or controlled by the contributor, whether already acquired or\n",
"hereafter acquired, that would be infringed by some manner, permitted\n",
"by this License, of making, using, or selling its contributor version,\n",
"but do not include claims that would be infringed only as a\n",
"consequence of further modification of the contributor version. For\n",
"purposes of this definition, \"control\" includes the right to grant\n",
"patent sublicenses in a manner consistent with the requirements of\n",
"this License.\n",
"\n",
" Each contributor grants you a non-exclusive, worldwide, royalty-free\n",
"patent license under the contributor's essential patent claims, to\n",
"make, use, sell, offer for sale, import and otherwise run, modify and\n",
"propagate the contents of its contributor version.\n",
"\n",
" In the following three paragraphs, a \"patent license\" is any express\n",
"agreement or commitment, however denominated, not to enforce a patent\n",
"(such as an express permission to practice a patent or covenant not to\n",
"sue for patent infringement). To \"grant\" such a patent license to a\n",
"party means to make such an agreement or commitment not to enforce a\n",
"patent against the party.\n",
"\n",
" If you convey a covered work, knowingly relying on a patent license,\n",
"and the Corresponding Source of the work is not available for anyone\n",
"to copy, free of charge and under the terms of this License, through a\n",
"publicly available network server or other readily accessible means,\n",
"then you must either (1) cause the Corresponding Source to be so\n",
"available, or (2) arrange to deprive yourself of the benefit of the\n",
"patent license for this particular work, or (3) arrange, in a manner\n",
"consistent with the requirements of this License, to extend the patent\n",
"license to downstream recipients. \"Knowingly relying\" means you have\n",
"actual knowledge that, but for the patent license, your conveying the\n",
"covered work in a country, or your recipient's use of the covered work\n",
"in a country, would infringe one or more identifiable patents in that\n",
"country that you have reason to believe are valid.\n",
"\n",
" If, pursuant to or in connection with a single transaction or\n",
"arrangement, you convey, or propagate by procuring conveyance of, a\n",
"covered work, and grant a patent license to some of the parties\n",
"receiving the covered work authorizing them to use, propagate, modify\n",
"or convey a specific copy of the covered work, then the patent license\n",
"you grant is automatically extended to all recipients of the covered\n",
"work and works based on it.\n",
"\n",
" A patent license is \"discriminatory\" if it does not include within\n",
"the scope of its coverage, prohibits the exercise of, or is\n",
"conditioned on the non-exercise of one or more of the rights that are\n",
"specifically granted under this License. You may not convey a covered\n",
"work if you are a party to an arrangement with a third party that is\n",
"in the business of distributing software, under which you make payment\n",
"to the third party based on the extent of your activity of conveying\n",
"the work, and under which the third party grants, to any of the\n",
"parties who would receive the covered work from you, a discriminatory\n",
"patent license (a) in connection with copies of the covered work\n",
"conveyed by you (or copies made from those copies), or (b) primarily\n",
"for and in connection with specific products or compilations that\n",
"contain the covered work, unless you entered into that arrangement,\n",
"or that patent license was granted, prior to 28 March 2007.\n",
"\n",
" Nothing in this License shall be construed as excluding or limiting\n",
"any implied license or other defenses to infringement that may\n",
"otherwise be available to you under applicable patent law.\n",
"\n",
" 12. No Surrender of Others' Freedom.\n",
"\n",
" If conditions are imposed on you (whether by court order, agreement or\n",
"otherwise) that contradict the conditions of this License, they do not\n",
"excuse you from the conditions of this License. If you cannot convey a\n",
"covered work so as to satisfy simultaneously your obligations under this\n",
"License and any other pertinent obligations, then as a consequence you may\n",
"not convey it at all. For example, if you agree to terms that obligate you\n",
"to collect a royalty for further conveying from those to whom you convey\n",
"the Program, the only way you could satisfy both those terms and this\n",
"License would be to refrain entirely from conveying the Program.\n",
"\n",
" 13. Use with the GNU Affero General Public License.\n",
"\n",
" Notwithstanding any other provision of this License, you have\n",
"permission to link or combine any covered work with a work licensed\n",
"under version 3 of the GNU Affero General Public License into a single\n",
"combined work, and to convey the resulting work. The terms of this\n",
"License will continue to apply to the part which is the covered work,\n",
"but the special requirements of the GNU Affero General Public License,\n",
"section 13, concerning interaction through a network will apply to the\n",
"combination as such.\n",
"\n",
" 14. Revised Versions of this License.\n",
"\n",
" The Free Software Foundation may publish revised and/or new versions of\n",
"the GNU General Public License from time to time. Such new versions will\n",
"be similar in spirit to the present version, but may differ in detail to\n",
"address new problems or concerns.\n",
"\n",
" Each version is given a distinguishing version number. If the\n",
"Program specifies that a certain numbered version of the GNU General\n",
"Public License \"or any later version\" applies to it, you have the\n",
"option of following the terms and conditions either of that numbered\n",
"version or of any later version published by the Free Software\n",
"Foundation. If the Program does not specify a version number of the\n",
"GNU General Public License, you may choose any version ever published\n",
"by the Free Software Foundation.\n",
"\n",
" If the Program specifies that a proxy can decide which future\n",
"versions of the GNU General Public License can be used, that proxy's\n",
"public statement of acceptance of a version permanently authorizes you\n",
"to choose that version for the Program.\n",
"\n",
" Later license versions may give you additional or different\n",
"permissions. However, no additional obligations are imposed on any\n",
"author or copyright holder as a result of your choosing to follow a\n",
"later version.\n",
"\n",
" 15. Disclaimer of Warranty.\n",
"\n",
" THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY\n",
"APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT\n",
"HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM \"AS IS\" WITHOUT WARRANTY\n",
"OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,\n",
"THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR\n",
"PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM\n",
"IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF\n",
"ALL NECESSARY SERVICING, REPAIR OR CORRECTION.\n",
"\n",
" 16. Limitation of Liability.\n",
"\n",
" IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING\n",
"WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS\n",
"THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY\n",
"GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE\n",
"USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF\n",
"DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD\n",
"PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),\n",
"EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF\n",
"SUCH DAMAGES.\n",
"\n",
" 17. Interpretation of Sections 15 and 16.\n",
"\n",
" If the disclaimer of warranty and limitation of liability provided\n",
"above cannot be given local legal effect according to their terms,\n",
"reviewing courts shall apply local law that most closely approximates\n",
"an absolute waiver of all civil liability in connection with the\n",
"Program, unless a warranty or assumption of liability accompanies a\n",
"copy of the Program in return for a fee.\n",
"\n",
" END OF TERMS AND CONDITIONS\n",
"\n",
" How to Apply These Terms to Your New Programs\n",
"\n",
" If you develop a new program, and you want it to be of the greatest\n",
"possible use to the public, the best way to achieve this is to make it\n",
"free software which everyone can redistribute and change under these terms.\n",
"\n",
" To do so, attach the following notices to the program. It is safest\n",
"to attach them to the start of each source file to most effectively\n",
"state the exclusion of warranty; and each file should have at least\n",
"the \"copyright\" line and a pointer to where the full notice is found.\n",
"\n",
" {one line to give the program's name and a brief idea of what it does.}\n",
" Copyright (C) {year} {name of author}\n",
"\n",
" This program is free software: you can redistribute it and/or modify\n",
" it under the terms of the GNU General Public License as published by\n",
" the Free Software Foundation, either version 3 of the License, or\n",
" (at your option) any later version.\n",
"\n",
" This program is distributed in the hope that it will be useful,\n",
" but WITHOUT ANY WARRANTY; without even the implied warranty of\n",
" MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the\n",
" GNU General Public License for more details.\n",
"\n",
" You should have received a copy of the GNU General Public License\n",
" along with this program. If not, see <http://www.gnu.org/licenses/>.\n",
"\n",
"Also add information on how to contact you by electronic and paper mail.\n",
"\n",
" If the program does terminal interaction, make it output a short\n",
"notice like this when it starts in an interactive mode:\n",
"\n",
" {project} Copyright (C) {year} {fullname}\n",
" This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.\n",
" This is free software, and you are welcome to redistribute it\n",
" under certain conditions; type `show c' for details.\n",
"\n",
"The hypothetical commands `show w' and `show c' should show the appropriate\n",
"parts of the General Public License. Of course, your program's commands\n",
"might be different; for a GUI interface, you would use an \"about box\".\n",
"\n",
" You should also get your employer (if you work as a programmer) or school,\n",
"if any, to sign a \"copyright disclaimer\" for the program, if necessary.\n",
"For more information on this, and how to apply and follow the GNU GPL, see\n",
"<http://www.gnu.org/licenses/>.\n",
"\n",
" The GNU General Public License does not permit incorporating your program\n",
"into proprietary programs. If your program is a subroutine library, you\n",
"may consider it more useful to permit linking proprietary applications with\n",
"the library. If this is what you want to do, use the GNU Lesser General\n",
"Public License instead of this License. But first, please read\n",
"<http://www.gnu.org/philosophy/why-not-lgpl.html>."
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "ChangelogTop"
},
"source": [
"#### Changelog"
]
},
{
"cell_type": "code",
"metadata": {
"cellView": "form",
"id": "Changelog"
},
"source": [
"#@title <- View Changelog\n",
"skip_for_run_all = True #@param {type: 'boolean'}\n",
"\n",
"if skip_for_run_all == False:\n",
" print(\n",
" '''\n",
" v1 Update: Oct 29th 2021 - Somnai\n",
"\n",
" QoL improvements added by Somnai (@somnai_dreams), including user friendly UI, settings+prompt saving and improved google drive folder organization.\n",
"\n",
" v1.1 Update: Nov 13th 2021 - Somnai\n",
"\n",
" Now includes sizing options, intermediate saves and fixed image prompts and perlin inits. unexposed batch option since it doesn't work\n",
"\n",
" v2 Update: Nov 22nd 2021 - Somnai\n",
"\n",
" Initial addition of Katherine Crowson's Secondary Model Method (https://colab.research.google.com/drive/1mpkrhOjoyzPeSWy2r7T8EYRaU7amYOOi#scrollTo=X5gODNAMEUCR)\n",
"\n",
" Noticed settings were saving with the wrong name so corrected it. Let me know if you preferred the old scheme.\n",
"\n",
" v3 Update: Dec 24th 2021 - Somnai\n",
"\n",
" Implemented Dango's advanced cutout method\n",
"\n",
" Added SLIP models, thanks to NeuralDivergent\n",
"\n",
" Fixed issue with NaNs resulting in black images, with massive help and testing from @Softology\n",
"\n",
" Perlin now changes properly within batches (not sure where this perlin_regen code came from originally, but thank you)\n",
"\n",
" v4 Update: Jan 2022 - Somnai\n",
"\n",
" Implemented Diffusion Zooming\n",
"\n",
" Added Chigozie keyframing\n",
"\n",
" Made a bunch of edits to processes\n",
" \n",
" v4.1 Update: Jan 14th 2022 - Somnai\n",
"\n",
" Added video input mode\n",
"\n",
" Added license that somehow went missing\n",
"\n",
" Added improved prompt keyframing, fixed image_prompts and multiple prompts\n",
"\n",
" Improved UI\n",
"\n",
" Significant under the hood cleanup and improvement\n",
"\n",
" Refined defaults for each mode\n",
"\n",
" Added latent-diffusion SuperRes for sharpening\n",
"\n",
" Added resume run mode\n",
"\n",
" v4.9 Update: Feb 5th 2022 - gandamu / Adam Letts\n",
"\n",
" Added 3D\n",
"\n",
" Added brightness corrections to prevent animation from steadily going dark over time\n",
"\n",
" v4.91 Update: Feb 19th 2022 - gandamu / Adam Letts\n",
"\n",
" Cleaned up 3D implementation and made associated args accessible via Colab UI elements\n",
"\n",
" v4.92 Update: Feb 20th 2022 - gandamu / Adam Letts\n",
"\n",
" Separated transform code\n",
"\n",
" v5.01 Update: Mar 10th 2022 - gandamu / Adam Letts\n",
"\n",
" IPython magic commands replaced by Python code\n",
"\n",
" v5.1 Update: Mar 30th 2022 - zippy / Chris Allen and gandamu / Adam Letts\n",
"\n",
" Integrated Turbo+Smooth features from Disco Diffusion Turbo -- just the implementation, without its defaults.\n",
"\n",
" Implemented resume of turbo animations in such a way that it's now possible to resume from different batch folders and batch numbers.\n",
"\n",
" 3D rotation parameter units are now degrees (rather than radians)\n",
"\n",
" Corrected name collision in sampling_mode (now diffusion_sampling_mode for plms/ddim, and sampling_mode for 3D transform sampling)\n",
"\n",
" Added video_init_seed_continuity option to make init video animations more continuous\n",
"\n",
" v5.1 Update: Apr 4th 2022 - MSFTserver aka HostsServer\n",
"\n",
" Removed pytorch3d from needing to be compiled with a lite version specifically made for Disco Diffusion\n",
"\n",
" Remove Super Resolution\n",
"\n",
" Remove SLIP Models\n",
"\n",
" Update for crossplatform support\n",
"\n",
" v5.2 Update: Apr 10th 2022 - nin_artificial / Tom Mason\n",
"\n",
" VR Mode\n",
"\n",
" v5.3 Update: Jun 10th 2022 - nshepperd, huemin, cut_pow / Dmitrii Tochilkin\n",
"\n",
" Horizontal and Vertical symmetry\n",
"\n",
" Addition of ViT-L/14@336px model (requires high VRAM)\n",
"\n",
" v5.3 Update: Jun 14th 2022 - devdef / Alex Spirin, Alex's Warp changes integrated into DD main by gandamu / Adam Letts\n",
"\n",
" Warp mode - for smooth/continuous video input results leveraging optical flow estimation and frame blending\n",
"\n",
" Custom models support\n",
" '''\n",
" )"
],
"outputs": [],
"execution_count": null
},
{
"cell_type": "markdown",
"metadata": {
"id": "TutorialTop"
},
"source": [
"# Tutorial"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "DiffusionSet"
},
"source": [
"**Diffusion settings (Defaults are heavily outdated)**\n",
"---\n",
"Disco Diffusion is complex, and continually evolving with new features. The most current documentation on on Disco Diffusion settings can be found in the unofficial guidebook:\n",
"\n",
"[Zippy's Disco Diffusion Cheatsheet](https://docs.google.com/document/d/1l8s7uS2dGqjztYSjPpzlmXLjl5PM3IGkRWI3IiCuK7g/edit)\n",
"\n",
"We also encourage users to join the [Disco Diffusion User Discord](https://discord.gg/XGZrFFCRfN) to learn from the active user community.\n",
"\n",
"This section below is outdated as of v2\n",
"\n",
"Setting | Description | Default\n",
"--- | --- | ---\n",
"**Your vision:**\n",
"`text_prompts` | A description of what you'd like the machine to generate. Think of it like writing the caption below your image on a website. | N/A\n",
"`image_prompts` | Think of these images more as a description of their contents. | N/A\n",
"**Image quality:**\n",
"`clip_guidance_scale` | Controls how much the image should look like the prompt. | 1000\n",
"`tv_scale` | Controls the smoothness of the final output. | 150\n",
"`range_scale` | Controls how far out of range RGB values are allowed to be. | 150\n",
"`sat_scale` | Controls how much saturation is allowed. From nshepperd's JAX notebook. | 0\n",
"`cutn` | Controls how many crops to take from the image. | 16\n",
"`cutn_batches` | Accumulate CLIP gradient from multiple batches of cuts. | 2\n",
"**Init settings:**\n",
"`init_image` | URL or local path | None\n",
"`init_scale` | This enhances the effect of the init image, a good value is 1000 | 0\n",
"`skip_steps` | Controls the starting point along the diffusion timesteps | 0\n",
"`perlin_init` | Option to start with random perlin noise | False\n",
"`perlin_mode` | ('gray', 'color') | 'mixed'\n",
"**Advanced:**\n",
"`skip_augs` | Controls whether to skip torchvision augmentations | False\n",
"`randomize_class` | Controls whether the imagenet class is randomly changed each iteration | True\n",
"`clip_denoised` | Determines whether CLIP discriminates a noisy or denoised image | False\n",
"`clamp_grad` | Experimental: Using adaptive clip grad in the cond_fn | True\n",
"`seed` | Choose a random seed and print it at end of run for reproduction | random_seed\n",
"`fuzzy_prompt` | Controls whether to add multiple noisy prompts to the prompt losses | False\n",
"`rand_mag` | Controls the magnitude of the random noise | 0.1\n",
"`eta` | DDIM hyperparameter | 0.5\n",
"`use_vertical_symmetry` | Enforce symmetry over x axis of the image on [`tr_st`*`steps` for `tr_st` in `transformation_steps`] steps of the diffusion process | False\n",
"`use_horizontal_symmetry` | Enforce symmetry over y axis of the image on [`tr_st`*`steps` for `tr_st` in `transformation_steps`] steps of the diffusion process | False\n",
"`transformation_steps` | Steps (expressed in percentages) in which the symmetry is enforced | [0.01]\n",
"`video_init_flow_warp` | Flow warp enabled | True\n",
"`video_init_flow_blend` | 0 - you get raw input, 1 - you get warped diffused previous frame | 0.999\n",
"`video_init_check_consistency` | check forward-backward flow consistency (uncheck unless there are too many warping artifacts) | True\n",
"\n",
"..\n",
"\n",
"**Model settings**\n",
"---\n",
"\n",
"Setting | Description | Default\n",
"--- | --- | ---\n",
"**Diffusion:**\n",
"`timestep_respacing` | Modify this value to decrease the number of timesteps. | ddim100\n",
"`diffusion_steps` || 1000\n",
"**Diffusion:**\n",
"`clip_models` | Models of CLIP to load. Typically the more, the better but they all come at a hefty VRAM cost. | ViT-B/32, ViT-B/16, RN50x4"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "SetupTop"
},
"source": [
"# 1. Set Up"
]
},
{
"cell_type": "code",
"metadata": {
"cellView": "form",
"id": "CheckGPU"
},
"source": [
"#@title 1.1 Check GPU Status\n",
"import subprocess\n",
"simple_nvidia_smi_display = False#@param {type:\"boolean\"}\n",
"if simple_nvidia_smi_display:\n",
" #!nvidia-smi\n",
" nvidiasmi_output = subprocess.run(['nvidia-smi', '-L'], stdout=subprocess.PIPE).stdout.decode('utf-8')\n",
" print(nvidiasmi_output)\n",
"else:\n",
" #!nvidia-smi -i 0 -e 0\n",
" nvidiasmi_output = subprocess.run(['nvidia-smi'], stdout=subprocess.PIPE).stdout.decode('utf-8')\n",
" print(nvidiasmi_output)\n",
" nvidiasmi_ecc_note = subprocess.run(['nvidia-smi', '-i', '0', '-e', '0'], stdout=subprocess.PIPE).stdout.decode('utf-8')\n",
" print(nvidiasmi_ecc_note)"
],
"outputs": [],
"execution_count": null
},
{
"cell_type": "code",
"metadata": {
"cellView": "form",
"id": "PrepFolders"
},
"source": [
"#@title 1.2 Prepare Folders\n",
"import subprocess, os, sys, ipykernel\n",
"\n",
"def gitclone(url, targetdir=None):\n",
" if targetdir:\n",
" res = subprocess.run(['git', 'clone', url, targetdir], stdout=subprocess.PIPE).stdout.decode('utf-8')\n",
" else:\n",
" res = subprocess.run(['git', 'clone', url], stdout=subprocess.PIPE).stdout.decode('utf-8')\n",
" print(res)\n",
"\n",
"def pipi(modulestr):\n",
" res = subprocess.run(['pip', 'install', modulestr], stdout=subprocess.PIPE).stdout.decode('utf-8')\n",
" print(res)\n",
"\n",
"def pipie(modulestr):\n",
" res = subprocess.run(['git', 'install', '-e', modulestr], stdout=subprocess.PIPE).stdout.decode('utf-8')\n",
" print(res)\n",
"\n",
"def wget(url, outputdir):\n",
" res = subprocess.run(['wget', url, '-P', f'{outputdir}'], stdout=subprocess.PIPE).stdout.decode('utf-8')\n",
" print(res)\n",
"\n",
"try:\n",
" from google.colab import drive\n",
" print(\"Google Colab detected. Using Google Drive.\")\n",
" is_colab = True\n",
" #@markdown If you connect your Google Drive, you can save the final image of each run on your drive.\n",
" google_drive = True #@param {type:\"boolean\"}\n",
" #@markdown Click here if you'd like to save the diffusion model checkpoint file to (and/or load from) your Google Drive:\n",
" save_models_to_google_drive = True #@param {type:\"boolean\"}\n",
"except:\n",
" is_colab = False\n",
" google_drive = False\n",
" save_models_to_google_drive = False\n",
" print(\"Google Colab not detected.\")\n",
"\n",
"if is_colab:\n",
" if google_drive is True:\n",
" drive.mount('/content/drive')\n",
" root_path = '/content/drive/MyDrive/AI/Disco_Diffusion'\n",
" else:\n",
" root_path = '/content'\n",
"else:\n",
" root_path = os.getcwd()\n",
"\n",
"import os\n",
"def createPath(filepath):\n",
" os.makedirs(filepath, exist_ok=True)\n",
"\n",
"initDirPath = f'{root_path}/init_images'\n",
"createPath(initDirPath)\n",
"outDirPath = f'{root_path}/images_out'\n",
"createPath(outDirPath)\n",
"\n",
"if is_colab:\n",
" if google_drive and not save_models_to_google_drive or not google_drive:\n",
" model_path = '/content/models'\n",
" createPath(model_path)\n",
" if google_drive and save_models_to_google_drive:\n",
" model_path = f'{root_path}/models'\n",
" createPath(model_path)\n",
"else:\n",
" model_path = f'{root_path}/models'\n",
" createPath(model_path)\n",
"\n",
"# libraries = f'{root_path}/libraries'\n",
"# createPath(libraries)"
],
"outputs": [],
"execution_count": null
},
{
"cell_type": "code",
"metadata": {
"cellView": "form",
"id": "InstallDeps"
},
"source": [
"#@title ### 1.3 Install, import dependencies and set up runtime devices\n",
"\n",
"import pathlib, shutil, os, sys\n",
"\n",
"# There are some reports that with a T4 or V100 on Colab, downgrading to a previous version of PyTorch may be necessary.\n",
"# .. but there are also reports that downgrading breaks them! If you're facing issues, you may want to try uncommenting and running this code.\n",
"# nvidiasmi_output = subprocess.run(['nvidia-smi'], stdout=subprocess.PIPE).stdout.decode('utf-8')\n",
"# cards_requiring_downgrade = [\"Tesla T4\", \"V100\"]\n",
"# if is_colab:\n",
"# if any(cardstr in nvidiasmi_output for cardstr in cards_requiring_downgrade):\n",
"# print(\"Downgrading pytorch. This can take a couple minutes ...\")\n",
"# downgrade_pytorch_result = subprocess.run(['pip', 'install', 'torch==1.10.2', 'torchvision==0.11.3', '-q'], stdout=subprocess.PIPE).stdout.decode('utf-8')\n",
"# print(\"pytorch downgraded.\")\n",
"\n",
"#@markdown Check this if you want to use CPU\n",
"useCPU = False #@param {type:\"boolean\"}\n",
"\n",
"if not is_colab:\n",
" # If running locally, there's a good chance your env will need this in order to not crash upon np.matmul() or similar operations.\n",
" os.environ['KMP_DUPLICATE_LIB_OK']='TRUE'\n",
"\n",
"PROJECT_DIR = os.path.abspath(os.getcwd())\n",
"USE_ADABINS = True\n",
"\n",
"if is_colab:\n",
" if not google_drive:\n",
" root_path = f'/content'\n",
" model_path = '/content/models' \n",
"else:\n",
" root_path = os.getcwd()\n",
" model_path = f'{root_path}/models'\n",
"\n",
"multipip_res = subprocess.run(['pip', 'install', 'lpips', 'datetime', 'timm', 'ftfy', 'einops', 'pytorch-lightning', 'omegaconf'], stdout=subprocess.PIPE).stdout.decode('utf-8')\n",
"print(multipip_res)\n",
"\n",
"if is_colab:\n",
" subprocess.run(['apt', 'install', 'imagemagick'], stdout=subprocess.PIPE).stdout.decode('utf-8')\n",
"\n",
"try:\n",
" from CLIP import clip\n",
"except:\n",
" if not os.path.exists(\"CLIP\"):\n",
" gitclone(\"https://github.com/openai/CLIP\")\n",
" sys.path.append(f'{PROJECT_DIR}/CLIP')\n",
"\n",
"try:\n",
" from guided_diffusion.script_util import create_model_and_diffusion\n",
"except:\n",
" if not os.path.exists(\"guided-diffusion\"):\n",
" gitclone(\"https://github.com/kostarion/guided-diffusion\")\n",
" sys.path.append(f'{PROJECT_DIR}/guided-diffusion')\n",
"\n",
"try:\n",
" from resize_right import resize\n",
"except:\n",
" if not os.path.exists(\"ResizeRight\"):\n",
" gitclone(\"https://github.com/assafshocher/ResizeRight.git\")\n",
" sys.path.append(f'{PROJECT_DIR}/ResizeRight')\n",
"\n",
"try:\n",
" import py3d_tools\n",
"except:\n",
" if not os.path.exists('pytorch3d-lite'):\n",
" gitclone(\"https://github.com/MSFTserver/pytorch3d-lite.git\")\n",
" sys.path.append(f'{PROJECT_DIR}/pytorch3d-lite')\n",
"\n",
"try:\n",
" from midas.dpt_depth import DPTDepthModel\n",
"except:\n",
" if not os.path.exists('MiDaS'):\n",
" gitclone(\"https://github.com/isl-org/MiDaS.git\")\n",
" if not os.path.exists('MiDaS/midas_utils.py'):\n",
" shutil.move('MiDaS/utils.py', 'MiDaS/midas_utils.py')\n",
" if not os.path.exists(f'{model_path}/dpt_large-midas-2f21e586.pt'):\n",
" wget(\"https://github.com/intel-isl/DPT/releases/download/1_0/dpt_large-midas-2f21e586.pt\", model_path)\n",
" sys.path.append(f'{PROJECT_DIR}/MiDaS')\n",
"\n",
"try:\n",
" sys.path.append(PROJECT_DIR)\n",
" import disco_xform_utils as dxf\n",
"except:\n",
" if not os.path.exists(\"disco-diffusion\"):\n",
" gitclone(\"https://github.com/alembics/disco-diffusion.git\")\n",
" if not os.path.exists('disco_xform_utils.py'):\n",
" shutil.move('disco-diffusion/disco_xform_utils.py', 'disco_xform_utils.py')\n",
" sys.path.append(PROJECT_DIR)\n",
"\n",
"import torch\n",
"from dataclasses import dataclass\n",
"from functools import partial\n",
"import cv2\n",
"import pandas as pd\n",
"import gc\n",
"import io\n",
"import math\n",
"import timm\n",
"from IPython import display\n",
"import lpips\n",
"from PIL import Image, ImageOps\n",
"import requests\n",
"from glob import glob\n",
"import json\n",
"from types import SimpleNamespace\n",
"from torch import nn\n",
"from torch.nn import functional as F\n",
"import torchvision.transforms as T\n",
"import torchvision.transforms.functional as TF\n",
"from tqdm.notebook import tqdm\n",
"from CLIP import clip\n",
"from resize_right import resize\n",
"from guided_diffusion.script_util import create_model_and_diffusion, model_and_diffusion_defaults\n",
"from datetime import datetime\n",
"import numpy as np\n",
"import matplotlib.pyplot as plt\n",
"import random\n",
"from ipywidgets import Output\n",
"import hashlib\n",
"from functools import partial\n",
"if is_colab:\n",
" os.chdir('/content')\n",
" from google.colab import files\n",
"else:\n",
" os.chdir(f'{PROJECT_DIR}')\n",
"from IPython.display import Image as ipyimg\n",
"from numpy import asarray\n",
"from einops import rearrange, repeat\n",
"import torch, torchvision\n",
"import time\n",
"from omegaconf import OmegaConf\n",
"import warnings\n",
"warnings.filterwarnings(\"ignore\", category=UserWarning)\n",
"\n",
"# AdaBins stuff\n",
"if USE_ADABINS:\n",
" try:\n",
" from infer import InferenceHelper\n",
" except:\n",
" if not os.path.exists(\"AdaBins\"):\n",
" gitclone(\"https://github.com/shariqfarooq123/AdaBins.git\")\n",
" if not os.path.exists(f'{PROJECT_DIR}/pretrained/AdaBins_nyu.pt'):\n",
" createPath(f'{PROJECT_DIR}/pretrained')\n",
" wget(\"https://cloudflare-ipfs.com/ipfs/Qmd2mMnDLWePKmgfS8m6ntAg4nhV5VkUyAydYBp8cWWeB7/AdaBins_nyu.pt\", f'{PROJECT_DIR}/pretrained')\n",
" sys.path.append(f'{PROJECT_DIR}/AdaBins')\n",
" from infer import InferenceHelper\n",
" MAX_ADABINS_AREA = 500000\n",
"\n",
"import torch\n",
"DEVICE = torch.device('cuda:0' if (torch.cuda.is_available() and not useCPU) else 'cpu')\n",
"print('Using device:', DEVICE)\n",
"device = DEVICE # At least one of the modules expects this name..\n",
"\n",
"if not useCPU:\n",
" if torch.cuda.get_device_capability(DEVICE) == (8,0): ## A100 fix thanks to Emad\n",
" print('Disabling CUDNN for A100 gpu', file=sys.stderr)\n",
" torch.backends.cudnn.enabled = False"
],
"outputs": [],
"execution_count": null
},
{
"cell_type": "code",
"metadata": {
"cellView": "form",
"id": "DefMidasFns"
},
"source": [
"#@title ### 1.4 Define Midas functions\n",
"\n",
"from midas.dpt_depth import DPTDepthModel\n",
"from midas.midas_net import MidasNet\n",
"from midas.midas_net_custom import MidasNet_small\n",
"from midas.transforms import Resize, NormalizeImage, PrepareForNet\n",
"\n",
"# Initialize MiDaS depth model.\n",
"# It remains resident in VRAM and likely takes around 2GB VRAM.\n",
"# You could instead initialize it for each frame (and free it after each frame) to save VRAM.. but initializing it is slow.\n",
"default_models = {\n",
" \"midas_v21_small\": f\"{model_path}/midas_v21_small-70d6b9c8.pt\",\n",
" \"midas_v21\": f\"{model_path}/midas_v21-f6b98070.pt\",\n",
" \"dpt_large\": f\"{model_path}/dpt_large-midas-2f21e586.pt\",\n",
" \"dpt_hybrid\": f\"{model_path}/dpt_hybrid-midas-501f0c75.pt\",\n",
" \"dpt_hybrid_nyu\": f\"{model_path}/dpt_hybrid_nyu-2ce69ec7.pt\",}\n",
"\n",
"\n",
"def init_midas_depth_model(midas_model_type=\"dpt_large\", optimize=True):\n",
" midas_model = None\n",
" net_w = None\n",
" net_h = None\n",
" resize_mode = None\n",
" normalization = None\n",
"\n",
" print(f\"Initializing MiDaS '{midas_model_type}' depth model...\")\n",
" # load network\n",
" midas_model_path = default_models[midas_model_type]\n",
"\n",
" if midas_model_type == \"dpt_large\": # DPT-Large\n",
" midas_model = DPTDepthModel(\n",
" path=midas_model_path,\n",
" backbone=\"vitl16_384\",\n",
" non_negative=True,\n",
" )\n",
" net_w, net_h = 384, 384\n",
" resize_mode = \"minimal\"\n",
" normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])\n",
" elif midas_model_type == \"dpt_hybrid\": #DPT-Hybrid\n",
" midas_model = DPTDepthModel(\n",
" path=midas_model_path,\n",
" backbone=\"vitb_rn50_384\",\n",
" non_negative=True,\n",
" )\n",
" net_w, net_h = 384, 384\n",
" resize_mode=\"minimal\"\n",
" normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])\n",
" elif midas_model_type == \"dpt_hybrid_nyu\": #DPT-Hybrid-NYU\n",
" midas_model = DPTDepthModel(\n",
" path=midas_model_path,\n",
" backbone=\"vitb_rn50_384\",\n",
" non_negative=True,\n",
" )\n",
" net_w, net_h = 384, 384\n",
" resize_mode=\"minimal\"\n",
" normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])\n",
" elif midas_model_type == \"midas_v21\":\n",
" midas_model = MidasNet(midas_model_path, non_negative=True)\n",
" net_w, net_h = 384, 384\n",
" resize_mode=\"upper_bound\"\n",
" normalization = NormalizeImage(\n",
" mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]\n",
" )\n",
" elif midas_model_type == \"midas_v21_small\":\n",
" midas_model = MidasNet_small(midas_model_path, features=64, backbone=\"efficientnet_lite3\", exportable=True, non_negative=True, blocks={'expand': True})\n",
" net_w, net_h = 256, 256\n",
" resize_mode=\"upper_bound\"\n",
" normalization = NormalizeImage(\n",
" mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]\n",
" )\n",
" else:\n",
" print(f\"midas_model_type '{midas_model_type}' not implemented\")\n",
" assert False\n",
"\n",
" midas_transform = T.Compose(\n",
" [\n",
" Resize(\n",
" net_w,\n",
" net_h,\n",
" resize_target=None,\n",
" keep_aspect_ratio=True,\n",
" ensure_multiple_of=32,\n",
" resize_method=resize_mode,\n",
" image_interpolation_method=cv2.INTER_CUBIC,\n",
" ),\n",
" normalization,\n",
" PrepareForNet(),\n",
" ]\n",
" )\n",
"\n",
" midas_model.eval()\n",
" \n",
" if optimize==True:\n",
" if DEVICE == torch.device(\"cuda\"):\n",
" midas_model = midas_model.to(memory_format=torch.channels_last) \n",
" midas_model = midas_model.half()\n",
"\n",
" midas_model.to(DEVICE)\n",
"\n",
" print(f\"MiDaS '{midas_model_type}' depth model initialized.\")\n",
" return midas_model, midas_transform, net_w, net_h, resize_mode, normalization"
],
"outputs": [],
"execution_count": null
},
{
"cell_type": "code",
"metadata": {
"cellView": "form",
"id": "DefFns"
},
"source": [
"#@title 1.5 Define necessary functions\n",
"\n",
"# https://gist.github.com/adefossez/0646dbe9ed4005480a2407c62aac8869\n",
"\n",
"import py3d_tools as p3dT\n",
"import disco_xform_utils as dxf\n",
"\n",
"def interp(t):\n",
" return 3 * t**2 - 2 * t ** 3\n",
"\n",
"def perlin(width, height, scale=10, device=None):\n",
" gx, gy = torch.randn(2, width + 1, height + 1, 1, 1, device=device)\n",
" xs = torch.linspace(0, 1, scale + 1)[:-1, None].to(device)\n",
" ys = torch.linspace(0, 1, scale + 1)[None, :-1].to(device)\n",
" wx = 1 - interp(xs)\n",
" wy = 1 - interp(ys)\n",
" dots = 0\n",
" dots += wx * wy * (gx[:-1, :-1] * xs + gy[:-1, :-1] * ys)\n",
" dots += (1 - wx) * wy * (-gx[1:, :-1] * (1 - xs) + gy[1:, :-1] * ys)\n",
" dots += wx * (1 - wy) * (gx[:-1, 1:] * xs - gy[:-1, 1:] * (1 - ys))\n",
" dots += (1 - wx) * (1 - wy) * (-gx[1:, 1:] * (1 - xs) - gy[1:, 1:] * (1 - ys))\n",
" return dots.permute(0, 2, 1, 3).contiguous().view(width * scale, height * scale)\n",
"\n",
"def perlin_ms(octaves, width, height, grayscale, device=device):\n",
" out_array = [0.5] if grayscale else [0.5, 0.5, 0.5]\n",
" # out_array = [0.0] if grayscale else [0.0, 0.0, 0.0]\n",
" for i in range(1 if grayscale else 3):\n",
" scale = 2 ** len(octaves)\n",
" oct_width = width\n",
" oct_height = height\n",
" for oct in octaves:\n",
" p = perlin(oct_width, oct_height, scale, device)\n",
" out_array[i] += p * oct\n",
" scale //= 2\n",
" oct_width *= 2\n",
" oct_height *= 2\n",
" return torch.cat(out_array)\n",
"\n",
"def create_perlin_noise(octaves=[1, 1, 1, 1], width=2, height=2, grayscale=True):\n",
" out = perlin_ms(octaves, width, height, grayscale)\n",
" if grayscale:\n",
" out = TF.resize(size=(side_y, side_x), img=out.unsqueeze(0))\n",
" out = TF.to_pil_image(out.clamp(0, 1)).convert('RGB')\n",
" else:\n",
" out = out.reshape(-1, 3, out.shape[0]//3, out.shape[1])\n",
" out = TF.resize(size=(side_y, side_x), img=out)\n",
" out = TF.to_pil_image(out.clamp(0, 1).squeeze())\n",
"\n",
" out = ImageOps.autocontrast(out)\n",
" return out\n",
"\n",
"def regen_perlin():\n",
" if perlin_mode == 'color':\n",
" init = create_perlin_noise([1.5**-i*0.5 for i in range(12)], 1, 1, False)\n",
" init2 = create_perlin_noise([1.5**-i*0.5 for i in range(8)], 4, 4, False)\n",
" elif perlin_mode == 'gray':\n",
" init = create_perlin_noise([1.5**-i*0.5 for i in range(12)], 1, 1, True)\n",
" init2 = create_perlin_noise([1.5**-i*0.5 for i in range(8)], 4, 4, True)\n",
" else:\n",
" init = create_perlin_noise([1.5**-i*0.5 for i in range(12)], 1, 1, False)\n",
" init2 = create_perlin_noise([1.5**-i*0.5 for i in range(8)], 4, 4, True)\n",
"\n",
" init = TF.to_tensor(init).add(TF.to_tensor(init2)).div(2).to(device).unsqueeze(0).mul(2).sub(1)\n",
" del init2\n",
" return init.expand(batch_size, -1, -1, -1)\n",
"\n",
"def fetch(url_or_path):\n",
" if str(url_or_path).startswith('http://') or str(url_or_path).startswith('https://'):\n",
" r = requests.get(url_or_path)\n",
" r.raise_for_status()\n",
" fd = io.BytesIO()\n",
" fd.write(r.content)\n",
" fd.seek(0)\n",
" return fd\n",
" return open(url_or_path, 'rb')\n",
"\n",
"def read_image_workaround(path):\n",
" \"\"\"OpenCV reads images as BGR, Pillow saves them as RGB. Work around\n",
" this incompatibility to avoid colour inversions.\"\"\"\n",
" im_tmp = cv2.imread(path)\n",
" return cv2.cvtColor(im_tmp, cv2.COLOR_BGR2RGB)\n",
"\n",
"def parse_prompt(prompt):\n",
" if prompt.startswith('http://') or prompt.startswith('https://'):\n",
" vals = prompt.rsplit(':', 2)\n",
" vals = [vals[0] + ':' + vals[1], *vals[2:]]\n",
" else:\n",
" vals = prompt.rsplit(':', 1)\n",
" vals = vals + ['', '1'][len(vals):]\n",
" return vals[0], float(vals[1])\n",
"\n",
"def sinc(x):\n",
" return torch.where(x != 0, torch.sin(math.pi * x) / (math.pi * x), x.new_ones([]))\n",
"\n",
"def lanczos(x, a):\n",
" cond = torch.logical_and(-a < x, x < a)\n",
" out = torch.where(cond, sinc(x) * sinc(x/a), x.new_zeros([]))\n",
" return out / out.sum()\n",
"\n",
"def ramp(ratio, width):\n",
" n = math.ceil(width / ratio + 1)\n",
" out = torch.empty([n])\n",
" cur = 0\n",
" for i in range(out.shape[0]):\n",
" out[i] = cur\n",
" cur += ratio\n",
" return torch.cat([-out[1:].flip([0]), out])[1:-1]\n",
"\n",
"def resample(input, size, align_corners=True):\n",
" n, c, h, w = input.shape\n",
" dh, dw = size\n",
"\n",
" input = input.reshape([n * c, 1, h, w])\n",
"\n",
" if dh < h:\n",
" kernel_h = lanczos(ramp(dh / h, 2), 2).to(input.device, input.dtype)\n",
" pad_h = (kernel_h.shape[0] - 1) // 2\n",
" input = F.pad(input, (0, 0, pad_h, pad_h), 'reflect')\n",
" input = F.conv2d(input, kernel_h[None, None, :, None])\n",
"\n",
" if dw < w:\n",
" kernel_w = lanczos(ramp(dw / w, 2), 2).to(input.device, input.dtype)\n",
" pad_w = (kernel_w.shape[0] - 1) // 2\n",
" input = F.pad(input, (pad_w, pad_w, 0, 0), 'reflect')\n",
" input = F.conv2d(input, kernel_w[None, None, None, :])\n",
"\n",
" input = input.reshape([n, c, h, w])\n",
" return F.interpolate(input, size, mode='bicubic', align_corners=align_corners)\n",
"\n",
"class MakeCutouts(nn.Module):\n",
" def __init__(self, cut_size, cutn, skip_augs=False):\n",
" super().__init__()\n",
" self.cut_size = cut_size\n",
" self.cutn = cutn\n",
" self.skip_augs = skip_augs\n",
" self.augs = T.Compose([\n",
" T.RandomHorizontalFlip(p=0.5),\n",
" T.Lambda(lambda x: x + torch.randn_like(x) * 0.01),\n",
" T.RandomAffine(degrees=15, translate=(0.1, 0.1)),\n",
" T.Lambda(lambda x: x + torch.randn_like(x) * 0.01),\n",
" T.RandomPerspective(distortion_scale=0.4, p=0.7),\n",
" T.Lambda(lambda x: x + torch.randn_like(x) * 0.01),\n",
" T.RandomGrayscale(p=0.15),\n",
" T.Lambda(lambda x: x + torch.randn_like(x) * 0.01),\n",
" # T.ColorJitter(brightness=0.1, contrast=0.1, saturation=0.1, hue=0.1),\n",
" ])\n",
"\n",
" def forward(self, input):\n",
" input = T.Pad(input.shape[2]//4, fill=0)(input)\n",
" sideY, sideX = input.shape[2:4]\n",
" max_size = min(sideX, sideY)\n",
"\n",
" cutouts = []\n",
" for ch in range(self.cutn):\n",
" if ch > self.cutn - self.cutn//4:\n",
" cutout = input.clone()\n",
" else:\n",
" size = int(max_size * torch.zeros(1,).normal_(mean=.8, std=.3).clip(float(self.cut_size/max_size), 1.))\n",
" offsetx = torch.randint(0, abs(sideX - size + 1), ())\n",
" offsety = torch.randint(0, abs(sideY - size + 1), ())\n",
" cutout = input[:, :, offsety:offsety + size, offsetx:offsetx + size]\n",
"\n",
" if not self.skip_augs:\n",
" cutout = self.augs(cutout)\n",
" cutouts.append(resample(cutout, (self.cut_size, self.cut_size)))\n",
" del cutout\n",
"\n",
" cutouts = torch.cat(cutouts, dim=0)\n",
" return cutouts\n",
"\n",
"cutout_debug = False\n",
"padargs = {}\n",
"\n",
"class MakeCutoutsDango(nn.Module):\n",
" def __init__(self, cut_size,\n",
" Overview=4, \n",
" InnerCrop = 0, IC_Size_Pow=0.5, IC_Grey_P = 0.2\n",
" ):\n",
" super().__init__()\n",
" self.cut_size = cut_size\n",
" self.Overview = Overview\n",
" self.InnerCrop = InnerCrop\n",
" self.IC_Size_Pow = IC_Size_Pow\n",
" self.IC_Grey_P = IC_Grey_P\n",
" if args.animation_mode == 'None':\n",
" self.augs = T.Compose([\n",
" T.RandomHorizontalFlip(p=0.5),\n",
" T.Lambda(lambda x: x + torch.randn_like(x) * 0.01),\n",
" T.RandomAffine(degrees=10, translate=(0.05, 0.05), interpolation = T.InterpolationMode.BILINEAR),\n",
" T.Lambda(lambda x: x + torch.randn_like(x) * 0.01),\n",
" T.RandomGrayscale(p=0.1),\n",
" T.Lambda(lambda x: x + torch.randn_like(x) * 0.01),\n",
" T.ColorJitter(brightness=0.1, contrast=0.1, saturation=0.1, hue=0.1),\n",
" ])\n",
" elif args.animation_mode == 'Video Input':\n",
" self.augs = T.Compose([\n",
" T.RandomHorizontalFlip(p=0.5),\n",
" T.Lambda(lambda x: x + torch.randn_like(x) * 0.01),\n",
" T.RandomAffine(degrees=15, translate=(0.1, 0.1)),\n",
" T.Lambda(lambda x: x + torch.randn_like(x) * 0.01),\n",
" T.RandomPerspective(distortion_scale=0.4, p=0.7),\n",
" T.Lambda(lambda x: x + torch.randn_like(x) * 0.01),\n",
" T.RandomGrayscale(p=0.15),\n",
" T.Lambda(lambda x: x + torch.randn_like(x) * 0.01),\n",
" # T.ColorJitter(brightness=0.1, contrast=0.1, saturation=0.1, hue=0.1),\n",
" ])\n",
" elif args.animation_mode == '2D' or args.animation_mode == '3D':\n",
" self.augs = T.Compose([\n",
" T.RandomHorizontalFlip(p=0.4),\n",
" T.Lambda(lambda x: x + torch.randn_like(x) * 0.01),\n",
" T.RandomAffine(degrees=10, translate=(0.05, 0.05), interpolation = T.InterpolationMode.BILINEAR),\n",
" T.Lambda(lambda x: x + torch.randn_like(x) * 0.01),\n",
" T.RandomGrayscale(p=0.1),\n",
" T.Lambda(lambda x: x + torch.randn_like(x) * 0.01),\n",
" T.ColorJitter(brightness=0.1, contrast=0.1, saturation=0.1, hue=0.3),\n",
" ])\n",
" \n",
"\n",
" def forward(self, input):\n",
" cutouts = []\n",
" gray = T.Grayscale(3)\n",
" sideY, sideX = input.shape[2:4]\n",
" max_size = min(sideX, sideY)\n",
" min_size = min(sideX, sideY, self.cut_size)\n",
" l_size = max(sideX, sideY)\n",
" output_shape = [1,3,self.cut_size,self.cut_size] \n",
" output_shape_2 = [1,3,self.cut_size+2,self.cut_size+2]\n",
" pad_input = F.pad(input,((sideY-max_size)//2,(sideY-max_size)//2,(sideX-max_size)//2,(sideX-max_size)//2), **padargs)\n",
" cutout = resize(pad_input, out_shape=output_shape)\n",
"\n",
" if self.Overview>0:\n",
" if self.Overview<=4:\n",
" if self.Overview>=1:\n",
" cutouts.append(cutout)\n",
" if self.Overview>=2:\n",
" cutouts.append(gray(cutout))\n",
" if self.Overview>=3:\n",
" cutouts.append(TF.hflip(cutout))\n",
" if self.Overview==4:\n",
" cutouts.append(gray(TF.hflip(cutout)))\n",
" else:\n",
" cutout = resize(pad_input, out_shape=output_shape)\n",
" for _ in range(self.Overview):\n",
" cutouts.append(cutout)\n",
"\n",
" if cutout_debug:\n",
" if is_colab:\n",
" TF.to_pil_image(cutouts[0].clamp(0, 1).squeeze(0)).save(\"/content/cutout_overview0.jpg\",quality=99)\n",
" else:\n",
" TF.to_pil_image(cutouts[0].clamp(0, 1).squeeze(0)).save(\"cutout_overview0.jpg\",quality=99)\n",
"\n",
" \n",
" if self.InnerCrop >0:\n",
" for i in range(self.InnerCrop):\n",
" size = int(torch.rand([])**self.IC_Size_Pow * (max_size - min_size) + min_size)\n",
" offsetx = torch.randint(0, sideX - size + 1, ())\n",
" offsety = torch.randint(0, sideY - size + 1, ())\n",
" cutout = input[:, :, offsety:offsety + size, offsetx:offsetx + size]\n",
" if i <= int(self.IC_Grey_P * self.InnerCrop):\n",
" cutout = gray(cutout)\n",
" cutout = resize(cutout, out_shape=output_shape)\n",
" cutouts.append(cutout)\n",
" if cutout_debug:\n",
" if is_colab:\n",
" TF.to_pil_image(cutouts[-1].clamp(0, 1).squeeze(0)).save(\"/content/cutout_InnerCrop.jpg\",quality=99)\n",
" else:\n",
" TF.to_pil_image(cutouts[-1].clamp(0, 1).squeeze(0)).save(\"cutout_InnerCrop.jpg\",quality=99)\n",
" cutouts = torch.cat(cutouts)\n",
" if skip_augs is not True: cutouts=self.augs(cutouts)\n",
" return cutouts\n",
"\n",
"def spherical_dist_loss(x, y):\n",
" x = F.normalize(x, dim=-1)\n",
" y = F.normalize(y, dim=-1)\n",
" return (x - y).norm(dim=-1).div(2).arcsin().pow(2).mul(2) \n",
"\n",
"def tv_loss(input):\n",
" \"\"\"L2 total variation loss, as in Mahendran et al.\"\"\"\n",
" input = F.pad(input, (0, 1, 0, 1), 'replicate')\n",
" x_diff = input[..., :-1, 1:] - input[..., :-1, :-1]\n",
" y_diff = input[..., 1:, :-1] - input[..., :-1, :-1]\n",
" return (x_diff**2 + y_diff**2).mean([1, 2, 3])\n",
"\n",
"\n",
"def range_loss(input):\n",
" return (input - input.clamp(-1, 1)).pow(2).mean([1, 2, 3])\n",
"\n",
"stop_on_next_loop = False # Make sure GPU memory doesn't get corrupted from cancelling the run mid-way through, allow a full frame to complete\n",
"TRANSLATION_SCALE = 1.0/200.0\n",
"\n",
"def do_3d_step(img_filepath, frame_num, midas_model, midas_transform):\n",
" if args.key_frames:\n",
" translation_x = args.translation_x_series[frame_num]\n",
" translation_y = args.translation_y_series[frame_num]\n",
" translation_z = args.translation_z_series[frame_num]\n",
" rotation_3d_x = args.rotation_3d_x_series[frame_num]\n",
" rotation_3d_y = args.rotation_3d_y_series[frame_num]\n",
" rotation_3d_z = args.rotation_3d_z_series[frame_num]\n",
" print(\n",
" f'translation_x: {translation_x}',\n",
" f'translation_y: {translation_y}',\n",
" f'translation_z: {translation_z}',\n",
" f'rotation_3d_x: {rotation_3d_x}',\n",
" f'rotation_3d_y: {rotation_3d_y}',\n",
" f'rotation_3d_z: {rotation_3d_z}',\n",
" )\n",
"\n",
" translate_xyz = [-translation_x*TRANSLATION_SCALE, translation_y*TRANSLATION_SCALE, -translation_z*TRANSLATION_SCALE]\n",
" rotate_xyz_degrees = [rotation_3d_x, rotation_3d_y, rotation_3d_z]\n",
" print('translation:',translate_xyz)\n",
" print('rotation:',rotate_xyz_degrees)\n",
" rotate_xyz = [math.radians(rotate_xyz_degrees[0]), math.radians(rotate_xyz_degrees[1]), math.radians(rotate_xyz_degrees[2])]\n",
" rot_mat = p3dT.euler_angles_to_matrix(torch.tensor(rotate_xyz, device=device), \"XYZ\").unsqueeze(0)\n",
" print(\"rot_mat: \" + str(rot_mat))\n",
" next_step_pil = dxf.transform_image_3d(img_filepath, midas_model, midas_transform, DEVICE,\n",
" rot_mat, translate_xyz, args.near_plane, args.far_plane,\n",
" args.fov, padding_mode=args.padding_mode,\n",
" sampling_mode=args.sampling_mode, midas_weight=args.midas_weight)\n",
" return next_step_pil\n",
"\n",
"def symmetry_transformation_fn(x):\n",
" if args.use_horizontal_symmetry:\n",
" [n, c, h, w] = x.size()\n",
" x = torch.concat((x[:, :, :, :w//2], torch.flip(x[:, :, :, :w//2], [-1])), -1)\n",
" print(\"horizontal symmetry applied\")\n",
" if args.use_vertical_symmetry:\n",
" [n, c, h, w] = x.size()\n",
" x = torch.concat((x[:, :, :h//2, :], torch.flip(x[:, :, :h//2, :], [-2])), -2)\n",
" print(\"vertical symmetry applied\")\n",
" return x\n",
"\n",
"def do_run():\n",
" seed = args.seed\n",
" print(range(args.start_frame, args.max_frames))\n",
"\n",
" if (args.animation_mode == \"3D\") and (args.midas_weight > 0.0):\n",
" midas_model, midas_transform, midas_net_w, midas_net_h, midas_resize_mode, midas_normalization = init_midas_depth_model(args.midas_depth_model)\n",
" for frame_num in range(args.start_frame, args.max_frames):\n",
" if stop_on_next_loop:\n",
" break\n",
" \n",
" display.clear_output(wait=True)\n",
"\n",
" # Print Frame progress if animation mode is on\n",
" if args.animation_mode != \"None\":\n",
" batchBar = tqdm(range(args.max_frames), desc =\"Frames\")\n",
" batchBar.n = frame_num\n",
" batchBar.refresh()\n",
"\n",
" \n",
" # Inits if not video frames\n",
" if args.animation_mode != \"Video Input\":\n",
" if args.init_image in ['','none', 'None', 'NONE']:\n",
" init_image = None\n",
" else:\n",
" init_image = args.init_image\n",
" init_scale = args.init_scale\n",
" skip_steps = args.skip_steps\n",
"\n",
" if args.animation_mode == \"2D\":\n",
" if args.key_frames:\n",
" angle = args.angle_series[frame_num]\n",
" zoom = args.zoom_series[frame_num]\n",
" translation_x = args.translation_x_series[frame_num]\n",
" translation_y = args.translation_y_series[frame_num]\n",
" print(\n",
" f'angle: {angle}',\n",
" f'zoom: {zoom}',\n",
" f'translation_x: {translation_x}',\n",
" f'translation_y: {translation_y}',\n",
" )\n",
" \n",
" if frame_num > 0:\n",
" seed += 1\n",
" if resume_run and frame_num == start_frame:\n",
" img_0 = cv2.imread(batchFolder+f\"/{batch_name}({batchNum})_{start_frame-1:04}.png\")\n",
" else:\n",
" img_0 = cv2.imread('prevFrame.png')\n",
" center = (1*img_0.shape[1]//2, 1*img_0.shape[0]//2)\n",
" trans_mat = np.float32(\n",
" [[1, 0, translation_x],\n",
" [0, 1, translation_y]]\n",
" )\n",
" rot_mat = cv2.getRotationMatrix2D( center, angle, zoom )\n",
" trans_mat = np.vstack([trans_mat, [0,0,1]])\n",
" rot_mat = np.vstack([rot_mat, [0,0,1]])\n",
" transformation_matrix = np.matmul(rot_mat, trans_mat)\n",
" img_0 = cv2.warpPerspective(\n",
" img_0,\n",
" transformation_matrix,\n",
" (img_0.shape[1], img_0.shape[0]),\n",
" borderMode=cv2.BORDER_WRAP\n",
" )\n",
"\n",
" cv2.imwrite('prevFrameScaled.png', img_0)\n",
" init_image = 'prevFrameScaled.png'\n",
" init_scale = args.frames_scale\n",
" skip_steps = args.calc_frames_skip_steps\n",
"\n",
" if args.animation_mode == \"3D\":\n",
" if frame_num > 0:\n",
" seed += 1 \n",
" if resume_run and frame_num == start_frame:\n",
" img_filepath = batchFolder+f\"/{batch_name}({batchNum})_{start_frame-1:04}.png\"\n",
" if turbo_mode and frame_num > turbo_preroll:\n",
" shutil.copyfile(img_filepath, 'oldFrameScaled.png')\n",
" else:\n",
" img_filepath = '/content/prevFrame.png' if is_colab else 'prevFrame.png'\n",
"\n",
" next_step_pil = do_3d_step(img_filepath, frame_num, midas_model, midas_transform)\n",
" next_step_pil.save('prevFrameScaled.png')\n",
"\n",
" ### Turbo mode - skip some diffusions, use 3d morph for clarity and to save time\n",
" if turbo_mode:\n",
" if frame_num == turbo_preroll: #start tracking oldframe\n",
" next_step_pil.save('oldFrameScaled.png')#stash for later blending \n",
" elif frame_num > turbo_preroll:\n",
" #set up 2 warped image sequences, old & new, to blend toward new diff image\n",
" old_frame = do_3d_step('oldFrameScaled.png', frame_num, midas_model, midas_transform)\n",
" old_frame.save('oldFrameScaled.png')\n",
" if frame_num % int(turbo_steps) != 0: \n",
" print('turbo skip this frame: skipping clip diffusion steps')\n",
" filename = f'{args.batch_name}({args.batchNum})_{frame_num:04}.png'\n",
" blend_factor = ((frame_num % int(turbo_steps))+1)/int(turbo_steps)\n",
" print('turbo skip this frame: skipping clip diffusion steps and saving blended frame')\n",
" newWarpedImg = cv2.imread('prevFrameScaled.png')#this is already updated..\n",
" oldWarpedImg = cv2.imread('oldFrameScaled.png')\n",
" blendedImage = cv2.addWeighted(newWarpedImg, blend_factor, oldWarpedImg,1-blend_factor, 0.0)\n",
" cv2.imwrite(f'{batchFolder}/{filename}',blendedImage)\n",
" next_step_pil.save(f'{img_filepath}') # save it also as prev_frame to feed next iteration\n",
" if vr_mode:\n",
" generate_eye_views(TRANSLATION_SCALE,batchFolder,filename,frame_num,midas_model, midas_transform)\n",
" continue\n",
" else:\n",
" #if not a skip frame, will run diffusion and need to blend.\n",
" oldWarpedImg = cv2.imread('prevFrameScaled.png')\n",
" cv2.imwrite(f'oldFrameScaled.png',oldWarpedImg)#swap in for blending later \n",
" print('clip/diff this frame - generate clip diff image')\n",
"\n",
" init_image = 'prevFrameScaled.png'\n",
" init_scale = args.frames_scale\n",
" skip_steps = args.calc_frames_skip_steps\n",
"\n",
" if args.animation_mode == \"Video Input\":\n",
" init_scale = args.video_init_frames_scale\n",
" skip_steps = args.calc_frames_skip_steps\n",
" if not video_init_seed_continuity:\n",
" seed += 1\n",
" if video_init_flow_warp:\n",
" if frame_num == 0: \n",
" skip_steps = args.video_init_skip_steps\n",
" init_image = f'{videoFramesFolder}/{frame_num+1:04}.jpg'\n",
" if frame_num > 0: \n",
" prev = PIL.Image.open(batchFolder+f\"/{batch_name}({batchNum})_{frame_num-1:04}.png\")\n",
" \n",
" frame1_path = f'{videoFramesFolder}/{frame_num:04}.jpg'\n",
" frame2 = PIL.Image.open(f'{videoFramesFolder}/{frame_num+1:04}.jpg')\n",
" flo_path = f\"/{flo_folder}/{frame1_path.split('/')[-1]}.npy\"\n",
" \n",
" init_image = 'warped.png'\n",
" print(video_init_flow_blend)\n",
" weights_path = None\n",
" if video_init_check_consistency:\n",
" weights_path = f\"/{flo_folder}/{frame1_path.split('/')[-1]}-21.txt\" \n",
" \n",
" warp(prev, frame2, flo_path, blend=video_init_flow_blend, weights_path=weights_path).save(init_image)\n",
" \n",
" else:\n",
" init_image = f'{videoFramesFolder}/{frame_num+1:04}.jpg'\n",
"\n",
"\n",
" loss_values = []\n",
" \n",
" if seed is not None:\n",
" np.random.seed(seed)\n",
" random.seed(seed)\n",
" torch.manual_seed(seed)\n",
" torch.cuda.manual_seed_all(seed)\n",
" torch.backends.cudnn.deterministic = True\n",
" \n",
" target_embeds, weights = [], []\n",
" \n",
" if args.prompts_series is not None and frame_num >= len(args.prompts_series):\n",
" frame_prompt = args.prompts_series[-1]\n",
" elif args.prompts_series is not None:\n",
" frame_prompt = args.prompts_series[frame_num]\n",
" else:\n",
" frame_prompt = []\n",
" \n",
" print(args.image_prompts_series)\n",
" if args.image_prompts_series is not None and frame_num >= len(args.image_prompts_series):\n",
" image_prompt = args.image_prompts_series[-1]\n",
" elif args.image_prompts_series is not None:\n",
" image_prompt = args.image_prompts_series[frame_num]\n",
" else:\n",
" image_prompt = []\n",
"\n",
" print(f'Frame {frame_num} Prompt: {frame_prompt}')\n",
"\n",
" model_stats = []\n",
" for clip_model in clip_models:\n",
" cutn = 16\n",
" model_stat = {\"clip_model\":None,\"target_embeds\":[],\"make_cutouts\":None,\"weights\":[]}\n",
" model_stat[\"clip_model\"] = clip_model\n",
" \n",
" for prompt in frame_prompt:\n",
" txt, weight = parse_prompt(prompt)\n",
" txt = clip_model.encode_text(clip.tokenize(prompt).to(device)).float()\n",
" \n",
" if args.fuzzy_prompt:\n",
" for i in range(25):\n",
" model_stat[\"target_embeds\"].append((txt + torch.randn(txt.shape).cuda() * args.rand_mag).clamp(0,1))\n",
" model_stat[\"weights\"].append(weight)\n",
" else:\n",
" model_stat[\"target_embeds\"].append(txt)\n",
" model_stat[\"weights\"].append(weight)\n",
" \n",
" if image_prompt:\n",
" model_stat[\"make_cutouts\"] = MakeCutouts(clip_model.visual.input_resolution, cutn, skip_augs=skip_augs) \n",
" for prompt in image_prompt:\n",
" path, weight = parse_prompt(prompt)\n",
" img = Image.open(fetch(path)).convert('RGB')\n",
" img = TF.resize(img, min(side_x, side_y, *img.size), T.InterpolationMode.LANCZOS)\n",
" batch = model_stat[\"make_cutouts\"](TF.to_tensor(img).to(device).unsqueeze(0).mul(2).sub(1))\n",
" embed = clip_model.encode_image(normalize(batch)).float()\n",
" if fuzzy_prompt:\n",
" for i in range(25):\n",
" model_stat[\"target_embeds\"].append((embed + torch.randn(embed.shape).cuda() * rand_mag).clamp(0,1))\n",
" weights.extend([weight / cutn] * cutn)\n",
" else:\n",
" model_stat[\"target_embeds\"].append(embed)\n",
" model_stat[\"weights\"].extend([weight / cutn] * cutn)\n",
" \n",
" model_stat[\"target_embeds\"] = torch.cat(model_stat[\"target_embeds\"])\n",
" model_stat[\"weights\"] = torch.tensor(model_stat[\"weights\"], device=device)\n",
" if model_stat[\"weights\"].sum().abs() < 1e-3:\n",
" raise RuntimeError('The weights must not sum to 0.')\n",
" model_stat[\"weights\"] /= model_stat[\"weights\"].sum().abs()\n",
" model_stats.append(model_stat)\n",
" \n",
" init = None\n",
" if init_image is not None:\n",
" init = Image.open(fetch(init_image)).convert('RGB')\n",
" init = init.resize((args.side_x, args.side_y), Image.LANCZOS)\n",
" init = TF.to_tensor(init).to(device).unsqueeze(0).mul(2).sub(1)\n",
" \n",
" if args.perlin_init:\n",
" if args.perlin_mode == 'color':\n",
" init = create_perlin_noise([1.5**-i*0.5 for i in range(12)], 1, 1, False)\n",
" init2 = create_perlin_noise([1.5**-i*0.5 for i in range(8)], 4, 4, False)\n",
" elif args.perlin_mode == 'gray':\n",
" init = create_perlin_noise([1.5**-i*0.5 for i in range(12)], 1, 1, True)\n",
" init2 = create_perlin_noise([1.5**-i*0.5 for i in range(8)], 4, 4, True)\n",
" else:\n",
" init = create_perlin_noise([1.5**-i*0.5 for i in range(12)], 1, 1, False)\n",
" init2 = create_perlin_noise([1.5**-i*0.5 for i in range(8)], 4, 4, True)\n",
" # init = TF.to_tensor(init).add(TF.to_tensor(init2)).div(2).to(device)\n",
" init = TF.to_tensor(init).add(TF.to_tensor(init2)).div(2).to(device).unsqueeze(0).mul(2).sub(1)\n",
" del init2\n",
" \n",
" cur_t = None\n",
" \n",
" def cond_fn(x, t, y=None):\n",
" with torch.enable_grad():\n",
" x_is_NaN = False\n",
" x = x.detach().requires_grad_()\n",
" n = x.shape[0]\n",
" if use_secondary_model is True:\n",
" alpha = torch.tensor(diffusion.sqrt_alphas_cumprod[cur_t], device=device, dtype=torch.float32)\n",
" sigma = torch.tensor(diffusion.sqrt_one_minus_alphas_cumprod[cur_t], device=device, dtype=torch.float32)\n",
" cosine_t = alpha_sigma_to_t(alpha, sigma)\n",
" out = secondary_model(x, cosine_t[None].repeat([n])).pred\n",
" fac = diffusion.sqrt_one_minus_alphas_cumprod[cur_t]\n",
" x_in = out * fac + x * (1 - fac)\n",
" x_in_grad = torch.zeros_like(x_in)\n",
" else:\n",
" my_t = torch.ones([n], device=device, dtype=torch.long) * cur_t\n",
" out = diffusion.p_mean_variance(model, x, my_t, clip_denoised=False, model_kwargs={'y': y})\n",
" fac = diffusion.sqrt_one_minus_alphas_cumprod[cur_t]\n",
" x_in = out['pred_xstart'] * fac + x * (1 - fac)\n",
" x_in_grad = torch.zeros_like(x_in)\n",
" for model_stat in model_stats:\n",
" for i in range(args.cutn_batches):\n",
" t_int = int(t.item())+1 #errors on last step without +1, need to find source\n",
" #when using SLIP Base model the dimensions need to be hard coded to avoid AttributeError: 'VisionTransformer' object has no attribute 'input_resolution'\n",
" try:\n",
" input_resolution=model_stat[\"clip_model\"].visual.input_resolution\n",
" except:\n",
" input_resolution=224\n",
"\n",
" cuts = MakeCutoutsDango(input_resolution,\n",
" Overview= args.cut_overview[1000-t_int], \n",
" InnerCrop = args.cut_innercut[1000-t_int], IC_Size_Pow=args.cut_ic_pow, IC_Grey_P = args.cut_icgray_p[1000-t_int]\n",
" )\n",
" clip_in = normalize(cuts(x_in.add(1).div(2)))\n",
" image_embeds = model_stat[\"clip_model\"].encode_image(clip_in).float()\n",
" dists = spherical_dist_loss(image_embeds.unsqueeze(1), model_stat[\"target_embeds\"].unsqueeze(0))\n",
" dists = dists.view([args.cut_overview[1000-t_int]+args.cut_innercut[1000-t_int], n, -1])\n",
" losses = dists.mul(model_stat[\"weights\"]).sum(2).mean(0)\n",
" loss_values.append(losses.sum().item()) # log loss, probably shouldn't do per cutn_batch\n",
" x_in_grad += torch.autograd.grad(losses.sum() * clip_guidance_scale, x_in)[0] / cutn_batches\n",
" tv_losses = tv_loss(x_in)\n",
" if use_secondary_model is True:\n",
" range_losses = range_loss(out)\n",
" else:\n",
" range_losses = range_loss(out['pred_xstart'])\n",
" sat_losses = torch.abs(x_in - x_in.clamp(min=-1,max=1)).mean()\n",
" loss = tv_losses.sum() * tv_scale + range_losses.sum() * range_scale + sat_losses.sum() * sat_scale\n",
" if init is not None and init_scale:\n",
" init_losses = lpips_model(x_in, init)\n",
" loss = loss + init_losses.sum() * init_scale\n",
" x_in_grad += torch.autograd.grad(loss, x_in)[0]\n",
" if torch.isnan(x_in_grad).any()==False:\n",
" grad = -torch.autograd.grad(x_in, x, x_in_grad)[0]\n",
" else:\n",
" # print(\"NaN'd\")\n",
" x_is_NaN = True\n",
" grad = torch.zeros_like(x)\n",
" if args.clamp_grad and x_is_NaN == False:\n",
" magnitude = grad.square().mean().sqrt()\n",
" return grad * magnitude.clamp(max=args.clamp_max) / magnitude #min=-0.02, min=-clamp_max, \n",
" return grad\n",
" \n",
" if args.diffusion_sampling_mode == 'ddim':\n",
" sample_fn = diffusion.ddim_sample_loop_progressive\n",
" else:\n",
" sample_fn = diffusion.plms_sample_loop_progressive\n",
"\n",
"\n",
" image_display = Output()\n",
" for i in range(args.n_batches):\n",
" if args.animation_mode == 'None':\n",
" display.clear_output(wait=True)\n",
" batchBar = tqdm(range(args.n_batches), desc =\"Batches\")\n",
" batchBar.n = i\n",
" batchBar.refresh()\n",
" print('')\n",
" display.display(image_display)\n",
" gc.collect()\n",
" torch.cuda.empty_cache()\n",
" cur_t = diffusion.num_timesteps - skip_steps - 1\n",
" total_steps = cur_t\n",
"\n",
" if perlin_init:\n",
" init = regen_perlin()\n",
"\n",
" if args.diffusion_sampling_mode == 'ddim':\n",
" samples = sample_fn(\n",
" model,\n",
" (batch_size, 3, args.side_y, args.side_x),\n",
" clip_denoised=clip_denoised,\n",
" model_kwargs={},\n",
" cond_fn=cond_fn,\n",
" progress=True,\n",
" skip_timesteps=skip_steps,\n",
" init_image=init,\n",
" randomize_class=randomize_class,\n",
" eta=eta,\n",
" transformation_fn=symmetry_transformation_fn,\n",
" transformation_percent=args.transformation_percent\n",
" )\n",
" else:\n",
" samples = sample_fn(\n",
" model,\n",
" (batch_size, 3, args.side_y, args.side_x),\n",
" clip_denoised=clip_denoised,\n",
" model_kwargs={},\n",
" cond_fn=cond_fn,\n",
" progress=True,\n",
" skip_timesteps=skip_steps,\n",
" init_image=init,\n",
" randomize_class=randomize_class,\n",
" order=2,\n",
" )\n",
" \n",
" \n",
" # with run_display:\n",
" # display.clear_output(wait=True)\n",
" for j, sample in enumerate(samples): \n",
" cur_t -= 1\n",
" intermediateStep = False\n",
" if args.steps_per_checkpoint is not None:\n",
" if j % steps_per_checkpoint == 0 and j > 0:\n",
" intermediateStep = True\n",
" elif j in args.intermediate_saves:\n",
" intermediateStep = True\n",
" with image_display:\n",
" if j % args.display_rate == 0 or cur_t == -1 or intermediateStep == True:\n",
" for k, image in enumerate(sample['pred_xstart']):\n",
" # tqdm.write(f'Batch {i}, step {j}, output {k}:')\n",
" current_time = datetime.now().strftime('%y%m%d-%H%M%S_%f')\n",
" percent = math.ceil(j/total_steps*100)\n",
" if args.n_batches > 0:\n",
" #if intermediates are saved to the subfolder, don't append a step or percentage to the name\n",
" if cur_t == -1 and args.intermediates_in_subfolder is True:\n",
" save_num = f'{frame_num:04}' if animation_mode != \"None\" else i\n",
" filename = f'{args.batch_name}({args.batchNum})_{save_num}.png'\n",
" else:\n",
" #If we're working with percentages, append it\n",
" if args.steps_per_checkpoint is not None:\n",
" filename = f'{args.batch_name}({args.batchNum})_{i:04}-{percent:02}%.png'\n",
" # Or else, iIf we're working with specific steps, append those\n",
" else:\n",
" filename = f'{args.batch_name}({args.batchNum})_{i:04}-{j:03}.png'\n",
" image = TF.to_pil_image(image.add(1).div(2).clamp(0, 1))\n",
" if j % args.display_rate == 0 or cur_t == -1:\n",
" image.save('progress.png')\n",
" display.clear_output(wait=True)\n",
" display.display(display.Image('progress.png'))\n",
" if args.steps_per_checkpoint is not None:\n",
" if j % args.steps_per_checkpoint == 0 and j > 0:\n",
" if args.intermediates_in_subfolder is True:\n",
" image.save(f'{partialFolder}/{filename}')\n",
" else:\n",
" image.save(f'{batchFolder}/{filename}')\n",
" else:\n",
" if j in args.intermediate_saves:\n",
" if args.intermediates_in_subfolder is True:\n",
" image.save(f'{partialFolder}/{filename}')\n",
" else:\n",
" image.save(f'{batchFolder}/{filename}')\n",
" if cur_t == -1:\n",
" if frame_num == 0:\n",
" save_settings()\n",
" if args.animation_mode != \"None\":\n",
" image.save('prevFrame.png')\n",
" image.save(f'{batchFolder}/{filename}')\n",
" if args.animation_mode == \"3D\":\n",
" # If turbo, save a blended image\n",
" if turbo_mode and frame_num > 0:\n",
" # Mix new image with prevFrameScaled\n",
" blend_factor = (1)/int(turbo_steps)\n",
" newFrame = cv2.imread('prevFrame.png') # This is already updated..\n",
" prev_frame_warped = cv2.imread('prevFrameScaled.png')\n",
" blendedImage = cv2.addWeighted(newFrame, blend_factor, prev_frame_warped, (1-blend_factor), 0.0)\n",
" cv2.imwrite(f'{batchFolder}/{filename}',blendedImage)\n",
" else:\n",
" image.save(f'{batchFolder}/{filename}')\n",
"\n",
" if vr_mode:\n",
" generate_eye_views(TRANSLATION_SCALE, batchFolder, filename, frame_num, midas_model, midas_transform)\n",
"\n",
" # if frame_num != args.max_frames-1:\n",
" # display.clear_output()\n",
" \n",
" plt.plot(np.array(loss_values), 'r')\n",
"\n",
"def generate_eye_views(trans_scale,batchFolder,filename,frame_num,midas_model, midas_transform):\n",
" for i in range(2):\n",
" theta = vr_eye_angle * (math.pi/180)\n",
" ray_origin = math.cos(theta) * vr_ipd / 2 * (-1.0 if i==0 else 1.0)\n",
" ray_rotation = (theta if i==0 else -theta)\n",
" translate_xyz = [-(ray_origin)*trans_scale, 0,0]\n",
" rotate_xyz = [0, (ray_rotation), 0]\n",
" rot_mat = p3dT.euler_angles_to_matrix(torch.tensor(rotate_xyz, device=device), \"XYZ\").unsqueeze(0)\n",
" transformed_image = dxf.transform_image_3d(f'{batchFolder}/{filename}', midas_model, midas_transform, DEVICE,\n",
" rot_mat, translate_xyz, args.near_plane, args.far_plane,\n",
" args.fov, padding_mode=args.padding_mode,\n",
" sampling_mode=args.sampling_mode, midas_weight=args.midas_weight,spherical=True)\n",
" eye_file_path = batchFolder+f\"/frame_{frame_num:04}\" + ('_l' if i==0 else '_r')+'.png'\n",
" transformed_image.save(eye_file_path)\n",
"\n",
"def save_settings():\n",
" setting_list = {\n",
" 'text_prompts': text_prompts,\n",
" 'image_prompts': image_prompts,\n",
" 'clip_guidance_scale': clip_guidance_scale,\n",
" 'tv_scale': tv_scale,\n",
" 'range_scale': range_scale,\n",
" 'sat_scale': sat_scale,\n",
" # 'cutn': cutn,\n",
" 'cutn_batches': cutn_batches,\n",
" 'max_frames': max_frames,\n",
" 'interp_spline': interp_spline,\n",
" # 'rotation_per_frame': rotation_per_frame,\n",
" 'init_image': init_image,\n",
" 'init_scale': init_scale,\n",
" 'skip_steps': skip_steps,\n",
" # 'zoom_per_frame': zoom_per_frame,\n",
" 'frames_scale': frames_scale,\n",
" 'frames_skip_steps': frames_skip_steps,\n",
" 'perlin_init': perlin_init,\n",
" 'perlin_mode': perlin_mode,\n",
" 'skip_augs': skip_augs,\n",
" 'randomize_class': randomize_class,\n",
" 'clip_denoised': clip_denoised,\n",
" 'clamp_grad': clamp_grad,\n",
" 'clamp_max': clamp_max,\n",
" 'seed': seed,\n",
" 'fuzzy_prompt': fuzzy_prompt,\n",
" 'rand_mag': rand_mag,\n",
" 'eta': eta,\n",
" 'width': width_height[0],\n",
" 'height': width_height[1],\n",
" 'diffusion_model': diffusion_model,\n",
" 'use_secondary_model': use_secondary_model,\n",
" 'steps': steps,\n",
" 'diffusion_steps': diffusion_steps,\n",
" 'diffusion_sampling_mode': diffusion_sampling_mode,\n",
" 'ViTB32': ViTB32,\n",
" 'ViTB16': ViTB16,\n",
" 'ViTL14': ViTL14,\n",
" 'ViTL14_336px': ViTL14_336px,\n",
" 'RN101': RN101,\n",
" 'RN50': RN50,\n",
" 'RN50x4': RN50x4,\n",
" 'RN50x16': RN50x16,\n",
" 'RN50x64': RN50x64,\n",
" 'cut_overview': str(cut_overview),\n",
" 'cut_innercut': str(cut_innercut),\n",
" 'cut_ic_pow': cut_ic_pow,\n",
" 'cut_icgray_p': str(cut_icgray_p),\n",
" 'key_frames': key_frames,\n",
" 'max_frames': max_frames,\n",
" 'angle': angle,\n",
" 'zoom': zoom,\n",
" 'translation_x': translation_x,\n",
" 'translation_y': translation_y,\n",
" 'translation_z': translation_z,\n",
" 'rotation_3d_x': rotation_3d_x,\n",
" 'rotation_3d_y': rotation_3d_y,\n",
" 'rotation_3d_z': rotation_3d_z,\n",
" 'midas_depth_model': midas_depth_model,\n",
" 'midas_weight': midas_weight,\n",
" 'near_plane': near_plane,\n",
" 'far_plane': far_plane,\n",
" 'fov': fov,\n",
" 'padding_mode': padding_mode,\n",
" 'sampling_mode': sampling_mode,\n",
" 'video_init_path':video_init_path,\n",
" 'extract_nth_frame':extract_nth_frame,\n",
" 'video_init_seed_continuity': video_init_seed_continuity,\n",
" 'turbo_mode':turbo_mode,\n",
" 'turbo_steps':turbo_steps,\n",
" 'turbo_preroll':turbo_preroll,\n",
" 'use_horizontal_symmetry':use_horizontal_symmetry,\n",
" 'use_vertical_symmetry':use_vertical_symmetry,\n",
" 'transformation_percent':transformation_percent,\n",
" #video init settings\n",
" 'video_init_steps': video_init_steps,\n",
" 'video_init_clip_guidance_scale': video_init_clip_guidance_scale,\n",
" 'video_init_tv_scale': video_init_tv_scale,\n",
" 'video_init_range_scale': video_init_range_scale,\n",
" 'video_init_sat_scale': video_init_sat_scale,\n",
" 'video_init_cutn_batches': video_init_cutn_batches,\n",
" 'video_init_skip_steps': video_init_skip_steps,\n",
" 'video_init_frames_scale': video_init_frames_scale,\n",
" 'video_init_frames_skip_steps': video_init_frames_skip_steps,\n",
" #warp settings\n",
" 'video_init_flow_warp':video_init_flow_warp,\n",
" 'video_init_flow_blend':video_init_flow_blend,\n",
" 'video_init_check_consistency':video_init_check_consistency,\n",
" 'video_init_blend_mode':video_init_blend_mode\n",
" }\n",
" # print('Settings:', setting_list)\n",
" with open(f\"{batchFolder}/{batch_name}({batchNum})_settings.txt\", \"w+\") as f: #save settings\n",
" json.dump(setting_list, f, ensure_ascii=False, indent=4)"
],
"outputs": [],
"execution_count": null
},
{
"cell_type": "code",
"metadata": {
"cellView": "form",
"id": "DefSecModel"
},
"source": [
"#@title 1.6 Define the secondary diffusion model\n",
"\n",
"def append_dims(x, n):\n",
" return x[(Ellipsis, *(None,) * (n - x.ndim))]\n",
"\n",
"\n",
"def expand_to_planes(x, shape):\n",
" return append_dims(x, len(shape)).repeat([1, 1, *shape[2:]])\n",
"\n",
"\n",
"def alpha_sigma_to_t(alpha, sigma):\n",
" return torch.atan2(sigma, alpha) * 2 / math.pi\n",
"\n",
"\n",
"def t_to_alpha_sigma(t):\n",
" return torch.cos(t * math.pi / 2), torch.sin(t * math.pi / 2)\n",
"\n",
"\n",
"@dataclass\n",
"class DiffusionOutput:\n",
" v: torch.Tensor\n",
" pred: torch.Tensor\n",
" eps: torch.Tensor\n",
"\n",
"\n",
"class ConvBlock(nn.Sequential):\n",
" def __init__(self, c_in, c_out):\n",
" super().__init__(\n",
" nn.Conv2d(c_in, c_out, 3, padding=1),\n",
" nn.ReLU(inplace=True),\n",
" )\n",
"\n",
"\n",
"class SkipBlock(nn.Module):\n",
" def __init__(self, main, skip=None):\n",
" super().__init__()\n",
" self.main = nn.Sequential(*main)\n",
" self.skip = skip if skip else nn.Identity()\n",
"\n",
" def forward(self, input):\n",
" return torch.cat([self.main(input), self.skip(input)], dim=1)\n",
"\n",
"\n",
"class FourierFeatures(nn.Module):\n",
" def __init__(self, in_features, out_features, std=1.):\n",
" super().__init__()\n",
" assert out_features % 2 == 0\n",
" self.weight = nn.Parameter(torch.randn([out_features // 2, in_features]) * std)\n",
"\n",
" def forward(self, input):\n",
" f = 2 * math.pi * input @ self.weight.T\n",
" return torch.cat([f.cos(), f.sin()], dim=-1)\n",
"\n",
"\n",
"class SecondaryDiffusionImageNet(nn.Module):\n",
" def __init__(self):\n",
" super().__init__()\n",
" c = 64 # The base channel count\n",
"\n",
" self.timestep_embed = FourierFeatures(1, 16)\n",
"\n",
" self.net = nn.Sequential(\n",
" ConvBlock(3 + 16, c),\n",
" ConvBlock(c, c),\n",
" SkipBlock([\n",
" nn.AvgPool2d(2),\n",
" ConvBlock(c, c * 2),\n",
" ConvBlock(c * 2, c * 2),\n",
" SkipBlock([\n",
" nn.AvgPool2d(2),\n",
" ConvBlock(c * 2, c * 4),\n",
" ConvBlock(c * 4, c * 4),\n",
" SkipBlock([\n",
" nn.AvgPool2d(2),\n",
" ConvBlock(c * 4, c * 8),\n",
" ConvBlock(c * 8, c * 4),\n",
" nn.Upsample(scale_factor=2, mode='bilinear', align_corners=False),\n",
" ]),\n",
" ConvBlock(c * 8, c * 4),\n",
" ConvBlock(c * 4, c * 2),\n",
" nn.Upsample(scale_factor=2, mode='bilinear', align_corners=False),\n",
" ]),\n",
" ConvBlock(c * 4, c * 2),\n",
" ConvBlock(c * 2, c),\n",
" nn.Upsample(scale_factor=2, mode='bilinear', align_corners=False),\n",
" ]),\n",
" ConvBlock(c * 2, c),\n",
" nn.Conv2d(c, 3, 3, padding=1),\n",
" )\n",
"\n",
" def forward(self, input, t):\n",
" timestep_embed = expand_to_planes(self.timestep_embed(t[:, None]), input.shape)\n",
" v = self.net(torch.cat([input, timestep_embed], dim=1))\n",
" alphas, sigmas = map(partial(append_dims, n=v.ndim), t_to_alpha_sigma(t))\n",
" pred = input * alphas - v * sigmas\n",
" eps = input * sigmas + v * alphas\n",
" return DiffusionOutput(v, pred, eps)\n",
"\n",
"\n",
"class SecondaryDiffusionImageNet2(nn.Module):\n",
" def __init__(self):\n",
" super().__init__()\n",
" c = 64 # The base channel count\n",
" cs = [c, c * 2, c * 2, c * 4, c * 4, c * 8]\n",
"\n",
" self.timestep_embed = FourierFeatures(1, 16)\n",
" self.down = nn.AvgPool2d(2)\n",
" self.up = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=False)\n",
"\n",
" self.net = nn.Sequential(\n",
" ConvBlock(3 + 16, cs[0]),\n",
" ConvBlock(cs[0], cs[0]),\n",
" SkipBlock([\n",
" self.down,\n",
" ConvBlock(cs[0], cs[1]),\n",
" ConvBlock(cs[1], cs[1]),\n",
" SkipBlock([\n",
" self.down,\n",
" ConvBlock(cs[1], cs[2]),\n",
" ConvBlock(cs[2], cs[2]),\n",
" SkipBlock([\n",
" self.down,\n",
" ConvBlock(cs[2], cs[3]),\n",
" ConvBlock(cs[3], cs[3]),\n",
" SkipBlock([\n",
" self.down,\n",
" ConvBlock(cs[3], cs[4]),\n",
" ConvBlock(cs[4], cs[4]),\n",
" SkipBlock([\n",
" self.down,\n",
" ConvBlock(cs[4], cs[5]),\n",
" ConvBlock(cs[5], cs[5]),\n",
" ConvBlock(cs[5], cs[5]),\n",
" ConvBlock(cs[5], cs[4]),\n",
" self.up,\n",
" ]),\n",
" ConvBlock(cs[4] * 2, cs[4]),\n",
" ConvBlock(cs[4], cs[3]),\n",
" self.up,\n",
" ]),\n",
" ConvBlock(cs[3] * 2, cs[3]),\n",
" ConvBlock(cs[3], cs[2]),\n",
" self.up,\n",
" ]),\n",
" ConvBlock(cs[2] * 2, cs[2]),\n",
" ConvBlock(cs[2], cs[1]),\n",
" self.up,\n",
" ]),\n",
" ConvBlock(cs[1] * 2, cs[1]),\n",
" ConvBlock(cs[1], cs[0]),\n",
" self.up,\n",
" ]),\n",
" ConvBlock(cs[0] * 2, cs[0]),\n",
" nn.Conv2d(cs[0], 3, 3, padding=1),\n",
" )\n",
"\n",
" def forward(self, input, t):\n",
" timestep_embed = expand_to_planes(self.timestep_embed(t[:, None]), input.shape)\n",
" v = self.net(torch.cat([input, timestep_embed], dim=1))\n",
" alphas, sigmas = map(partial(append_dims, n=v.ndim), t_to_alpha_sigma(t))\n",
" pred = input * alphas - v * sigmas\n",
" eps = input * sigmas + v * alphas\n",
" return DiffusionOutput(v, pred, eps)"
],
"outputs": [],
"execution_count": null
},
{
"cell_type": "markdown",
"metadata": {
"id": "DiffClipSetTop"
},
"source": [
"# 2. Diffusion and CLIP model settings"
]
},
{
"cell_type": "code",
"metadata": {
"id": "ModelSettings"
},
"source": [
"#@markdown ####**Models Settings:**\n",
"diffusion_model = \"512x512_diffusion_uncond_finetune_008100\" #@param [\"256x256_diffusion_uncond\", \"512x512_diffusion_uncond_finetune_008100\", \"custom\"]\n",
"use_secondary_model = True #@param {type: 'boolean'}\n",
"diffusion_sampling_mode = 'ddim' #@param ['plms','ddim']\n",
"#@markdown #####**Custom model:**\n",
"custom_path = '/content/drive/MyDrive/deep_learning/ddpm/ema_0.9999_058000.pt'#@param {type: 'string'}\n",
"\n",
"#@markdown #####**CLIP settings:**\n",
"use_checkpoint = True #@param {type: 'boolean'}\n",
"ViTB32 = True #@param{type:\"boolean\"}\n",
"ViTB16 = True #@param{type:\"boolean\"}\n",
"ViTL14 = False #@param{type:\"boolean\"}\n",
"ViTL14_336px = False #@param{type:\"boolean\"}\n",
"RN101 = False #@param{type:\"boolean\"}\n",
"RN50 = True #@param{type:\"boolean\"}\n",
"RN50x4 = False #@param{type:\"boolean\"}\n",
"RN50x16 = False #@param{type:\"boolean\"}\n",
"RN50x64 = False #@param{type:\"boolean\"}\n",
"\n",
"#@markdown If you're having issues with model downloads, check this to compare SHA's:\n",
"check_model_SHA = False #@param{type:\"boolean\"}\n",
"\n",
"def download_models(diffusion_model,use_secondary_model,fallback=False):\n",
" model_256_downloaded = False\n",
" model_512_downloaded = False\n",
" model_secondary_downloaded = False\n",
"\n",
" model_256_SHA = '983e3de6f95c88c81b2ca7ebb2c217933be1973b1ff058776b970f901584613a'\n",
" model_512_SHA = '9c111ab89e214862b76e1fa6a1b3f1d329b1a88281885943d2cdbe357ad57648'\n",
" model_secondary_SHA = '983e3de6f95c88c81b2ca7ebb2c217933be1973b1ff058776b970f901584613a'\n",
"\n",
" model_256_link = 'https://openaipublic.blob.core.windows.net/diffusion/jul-2021/256x256_diffusion_uncond.pt'\n",
" model_512_link = 'https://the-eye.eu/public/AI/models/512x512_diffusion_unconditional_ImageNet/512x512_diffusion_uncond_finetune_008100.pt'\n",
" model_secondary_link = 'https://v-diffusion.s3.us-west-2.amazonaws.com/secondary_model_imagenet_2.pth'\n",
"\n",
" model_256_link_fb = 'https://www.dropbox.com/s/9tqnqo930mpnpcn/256x256_diffusion_uncond.pt'\n",
" model_512_link_fb = 'https://huggingface.co/lowlevelware/512x512_diffusion_unconditional_ImageNet/resolve/main/512x512_diffusion_uncond_finetune_008100.pt'\n",
" model_secondary_link_fb = 'https://the-eye.eu/public/AI/models/v-diffusion/secondary_model_imagenet_2.pth'\n",
"\n",
" model_256_path = f'{model_path}/256x256_diffusion_uncond.pt'\n",
" model_512_path = f'{model_path}/512x512_diffusion_uncond_finetune_008100.pt'\n",
" model_secondary_path = f'{model_path}/secondary_model_imagenet_2.pth'\n",
"\n",
" if fallback:\n",
" model_256_link = model_256_link_fb\n",
" model_512_link = model_512_link_fb\n",
" model_secondary_link = model_secondary_link_fb\n",
" # Download the diffusion model\n",
" if diffusion_model == '256x256_diffusion_uncond':\n",
" if os.path.exists(model_256_path) and check_model_SHA:\n",
" print('Checking 256 Diffusion File')\n",
" with open(model_256_path,\"rb\") as f:\n",
" bytes = f.read() \n",
" hash = hashlib.sha256(bytes).hexdigest();\n",
" if hash == model_256_SHA:\n",
" print('256 Model SHA matches')\n",
" model_256_downloaded = True\n",
" else:\n",
" print(\"256 Model SHA doesn't match, redownloading...\")\n",
" wget(model_256_link, model_path)\n",
" if os.path.exists(model_256_path):\n",
" model_256_downloaded = True\n",
" else:\n",
" print('First URL Failed using FallBack')\n",
" download_models(diffusion_model,use_secondary_model,True)\n",
" elif os.path.exists(model_256_path) and not check_model_SHA or model_256_downloaded == True:\n",
" print('256 Model already downloaded, check check_model_SHA if the file is corrupt')\n",
" else: \n",
" wget(model_256_link, model_path)\n",
" if os.path.exists(model_256_path):\n",
" model_256_downloaded = True\n",
" else:\n",
" print('First URL Failed using FallBack')\n",
" download_models(diffusion_model,True)\n",
" elif diffusion_model == '512x512_diffusion_uncond_finetune_008100':\n",
" if os.path.exists(model_512_path) and check_model_SHA:\n",
" print('Checking 512 Diffusion File')\n",
" with open(model_512_path,\"rb\") as f:\n",
" bytes = f.read() \n",
" hash = hashlib.sha256(bytes).hexdigest();\n",
" if hash == model_512_SHA:\n",
" print('512 Model SHA matches')\n",
" if os.path.exists(model_512_path):\n",
" model_512_downloaded = True\n",
" else:\n",
" print('First URL Failed using FallBack')\n",
" download_models(diffusion_model,use_secondary_model,True)\n",
" else: \n",
" print(\"512 Model SHA doesn't match, redownloading...\")\n",
" wget(model_512_link, model_path)\n",
" if os.path.exists(model_512_path):\n",
" model_512_downloaded = True\n",
" else:\n",
" print('First URL Failed using FallBack')\n",
" download_models(diffusion_model,use_secondary_model,True)\n",
" elif os.path.exists(model_512_path) and not check_model_SHA or model_512_downloaded:\n",
" print('512 Model already downloaded, check check_model_SHA if the file is corrupt')\n",
" else: \n",
" wget(model_512_link, model_path)\n",
" model_512_downloaded = True\n",
" # Download the secondary diffusion model v2\n",
" if use_secondary_model:\n",
" if os.path.exists(model_secondary_path) and check_model_SHA:\n",
" print('Checking Secondary Diffusion File')\n",
" with open(model_secondary_path,\"rb\") as f:\n",
" bytes = f.read() \n",
" hash = hashlib.sha256(bytes).hexdigest();\n",
" if hash == model_secondary_SHA:\n",
" print('Secondary Model SHA matches')\n",
" model_secondary_downloaded = True\n",
" else: \n",
" print(\"Secondary Model SHA doesn't match, redownloading...\")\n",
" wget(model_secondary_link, model_path)\n",
" if os.path.exists(model_secondary_path):\n",
" model_secondary_downloaded = True\n",
" else:\n",
" print('First URL Failed using FallBack')\n",
" download_models(diffusion_model,use_secondary_model,True)\n",
" elif os.path.exists(model_secondary_path) and not check_model_SHA or model_secondary_downloaded:\n",
" print('Secondary Model already downloaded, check check_model_SHA if the file is corrupt')\n",
" else: \n",
" wget(model_secondary_link, model_path)\n",
" if os.path.exists(model_secondary_path):\n",
" model_secondary_downloaded = True\n",
" else:\n",
" print('First URL Failed using FallBack')\n",
" download_models(diffusion_model,use_secondary_model,True)\n",
"\n",
"download_models(diffusion_model,use_secondary_model)\n",
"\n",
"model_config = model_and_diffusion_defaults()\n",
"if diffusion_model == '512x512_diffusion_uncond_finetune_008100':\n",
" model_config.update({\n",
" 'attention_resolutions': '32, 16, 8',\n",
" 'class_cond': False,\n",
" 'diffusion_steps': 1000, #No need to edit this, it is taken care of later.\n",
" 'rescale_timesteps': True,\n",
" 'timestep_respacing': 250, #No need to edit this, it is taken care of later.\n",
" 'image_size': 512,\n",
" 'learn_sigma': True,\n",
" 'noise_schedule': 'linear',\n",
" 'num_channels': 256,\n",
" 'num_head_channels': 64,\n",
" 'num_res_blocks': 2,\n",
" 'resblock_updown': True,\n",
" 'use_checkpoint': use_checkpoint,\n",
" 'use_fp16': not useCPU,\n",
" 'use_scale_shift_norm': True,\n",
" })\n",
"elif diffusion_model == '256x256_diffusion_uncond':\n",
" model_config.update({\n",
" 'attention_resolutions': '32, 16, 8',\n",
" 'class_cond': False,\n",
" 'diffusion_steps': 1000, #No need to edit this, it is taken care of later.\n",
" 'rescale_timesteps': True,\n",
" 'timestep_respacing': 250, #No need to edit this, it is taken care of later.\n",
" 'image_size': 256,\n",
" 'learn_sigma': True,\n",
" 'noise_schedule': 'linear',\n",
" 'num_channels': 256,\n",
" 'num_head_channels': 64,\n",
" 'num_res_blocks': 2,\n",
" 'resblock_updown': True,\n",
" 'use_checkpoint': use_checkpoint,\n",
" 'use_fp16': not useCPU,\n",
" 'use_scale_shift_norm': True,\n",
" })\n",
"\n",
"model_default = model_config['image_size']\n",
"\n",
"if use_secondary_model:\n",
" secondary_model = SecondaryDiffusionImageNet2()\n",
" secondary_model.load_state_dict(torch.load(f'{model_path}/secondary_model_imagenet_2.pth', map_location='cpu'))\n",
" secondary_model.eval().requires_grad_(False).to(device)\n",
"\n",
"clip_models = []\n",
"if ViTB32: clip_models.append(clip.load('ViT-B/32', jit=False)[0].eval().requires_grad_(False).to(device))\n",
"if ViTB16: clip_models.append(clip.load('ViT-B/16', jit=False)[0].eval().requires_grad_(False).to(device))\n",
"if ViTL14: clip_models.append(clip.load('ViT-L/14', jit=False)[0].eval().requires_grad_(False).to(device))\n",
"if ViTL14_336px: clip_models.append(clip.load('ViT-L/14@336px', jit=False)[0].eval().requires_grad_(False).to(device))\n",
"if RN50: clip_models.append(clip.load('RN50', jit=False)[0].eval().requires_grad_(False).to(device))\n",
"if RN50x4: clip_models.append(clip.load('RN50x4', jit=False)[0].eval().requires_grad_(False).to(device))\n",
"if RN50x16: clip_models.append(clip.load('RN50x16', jit=False)[0].eval().requires_grad_(False).to(device))\n",
"if RN50x64: clip_models.append(clip.load('RN50x64', jit=False)[0].eval().requires_grad_(False).to(device))\n",
"if RN101: clip_models.append(clip.load('RN101', jit=False)[0].eval().requires_grad_(False).to(device))\n",
"\n",
"normalize = T.Normalize(mean=[0.48145466, 0.4578275, 0.40821073], std=[0.26862954, 0.26130258, 0.27577711])\n",
"lpips_model = lpips.LPIPS(net='vgg').to(device)"
],
"outputs": [],
"execution_count": null
},
{
"cell_type": "markdown",
"metadata": {
"id": "CustModelTop"
},
"source": [
"# Custom model settings \n",
"Modify in accordance with your training settings and run the cell"
]
},
{
"cell_type": "code",
"metadata": {
"id": "CustModel"
},
"source": [
"#@markdown ####**Custom Model Settings:**\n",
"if diffusion_model == 'custom':\n",
" model_config.update({\n",
" 'attention_resolutions': '16',\n",
" 'class_cond': False,\n",
" 'diffusion_steps': 1000,\n",
" 'rescale_timesteps': True,\n",
" 'timestep_respacing': 'ddim100',\n",
" 'image_size': 256,\n",
" 'learn_sigma': True,\n",
" 'noise_schedule': 'linear',\n",
" 'num_channels': 128,\n",
" 'num_heads': 1,\n",
" 'num_res_blocks': 2,\n",
" 'use_checkpoint': use_checkpoint,\n",
" 'use_fp16': True,\n",
" 'use_scale_shift_norm': False,\n",
" })"
],
"outputs": [],
"execution_count": null
},
{
"cell_type": "markdown",
"metadata": {
"id": "SettingsTop"
},
"source": [
"# 3. Settings"
]
},
{
"cell_type": "code",
"metadata": {
"id": "BasicSettings"
},
"source": [
"#@markdown ####**Basic Settings:**\n",
"batch_name = 'TimeToDisco' #@param{type: 'string'}\n",
"steps = 250 #@param [25,50,100,150,250,500,1000]{type: 'raw', allow-input: true}\n",
"width_height = [1280, 768] #@param{type: 'raw'}\n",
"clip_guidance_scale = 5000 #@param{type: 'number'}\n",
"tv_scale = 0#@param{type: 'number'}\n",
"range_scale = 150#@param{type: 'number'}\n",
"sat_scale = 0#@param{type: 'number'}\n",
"cutn_batches = 4#@param{type: 'number'}\n",
"skip_augs = False#@param{type: 'boolean'}\n",
"\n",
"#@markdown ####**Video Init Basic Settings:**\n",
"video_init_steps = 100 #@param [25,50,100,150,250,500,1000]{type: 'raw', allow-input: true}\n",
"video_init_clip_guidance_scale = 1000 #@param{type: 'number'}\n",
"video_init_tv_scale = 0.1#@param{type: 'number'}\n",
"video_init_range_scale = 150#@param{type: 'number'}\n",
"video_init_sat_scale = 300#@param{type: 'number'}\n",
"video_init_cutn_batches = 4#@param{type: 'number'}\n",
"video_init_skip_steps = 50 #@param{type: 'integer'}\n",
"\n",
"#@markdown ---\n",
"\n",
"#@markdown ####**Init Image Settings:**\n",
"init_image = None #@param{type: 'string'}\n",
"init_scale = 1000 #@param{type: 'integer'}\n",
"skip_steps = 10 #@param{type: 'integer'}\n",
"#@markdown *Make sure you set skip_steps to ~50% of your steps if you want to use an init image.*\n",
"\n",
"#Get corrected sizes\n",
"side_x = (width_height[0]//64)*64;\n",
"side_y = (width_height[1]//64)*64;\n",
"if side_x != width_height[0] or side_y != width_height[1]:\n",
" print(f'Changing output size to {side_x}x{side_y}. Dimensions must by multiples of 64.')\n",
"\n",
"#Update Model Settings\n",
"timestep_respacing = f'ddim{steps}'\n",
"diffusion_steps = (1000//steps)*steps if steps < 1000 else steps\n",
"model_config.update({\n",
" 'timestep_respacing': timestep_respacing,\n",
" 'diffusion_steps': diffusion_steps,\n",
"})\n",
"\n",
"#Make folder for batch\n",
"batchFolder = f'{outDirPath}/{batch_name}'\n",
"createPath(batchFolder)"
],
"outputs": [],
"execution_count": null
},
{
"cell_type": "markdown",
"metadata": {
"id": "AnimSetTop"
},
"source": [
"### Animation Settings"
]
},
{
"cell_type": "code",
"metadata": {
"id": "AnimSettings"
},
"source": [
"#@markdown ####**Animation Mode:**\n",
"animation_mode = 'None' #@param ['None', '2D', '3D', 'Video Input'] {type:'string'}\n",
"#@markdown *For animation, you probably want to turn `cutn_batches` to 1 to make it quicker.*\n",
"\n",
"\n",
"#@markdown ---\n",
"\n",
"#@markdown ####**Video Input Settings:**\n",
"if is_colab:\n",
" video_init_path = \"/content/drive/MyDrive/AI/Disco_Diffusion/init.mp4\" #@param {type: 'string'}\n",
"else:\n",
" video_init_path = \"init.mp4\" #@param {type: 'string'}\n",
"extract_nth_frame = 2 #@param {type: 'number'}\n",
"persistent_frame_output_in_batch_folder = True #@param {type: 'boolean'}\n",
"video_init_seed_continuity = False #@param {type: 'boolean'}\n",
"#@markdown #####**Video Optical Flow Settings:**\n",
"video_init_flow_warp = True #@param {type: 'boolean'}\n",
"# Call optical flow from video frames and warp prev frame with flow\n",
"video_init_flow_blend = 0.999#@param {type: 'number'} #0 - take next frame, 1 - take prev warped frame\n",
"video_init_check_consistency = True #@param {type: 'boolean'}\n",
"video_init_blend_mode = \"optical flow\" #@param ['None', 'linear', 'optical flow']\n",
"# Call optical flow from video frames and warp prev frame with flow\n",
"if animation_mode == \"Video Input\":\n",
" if persistent_frame_output_in_batch_folder or (not is_colab): #suggested by Chris the Wizard#8082 at discord\n",
" videoFramesFolder = f'{batchFolder}/videoFrames'\n",
" else:\n",
" videoFramesFolder = f'/content/videoFrames'\n",
" createPath(videoFramesFolder)\n",
" print(f\"Exporting Video Frames (1 every {extract_nth_frame})...\")\n",
" try:\n",
" for f in pathlib.Path(f'{videoFramesFolder}').glob('*.jpg'):\n",
" f.unlink()\n",
" except:\n",
" print('')\n",
" vf = f'select=not(mod(n\\,{extract_nth_frame}))'\n",
" if os.path.exists(video_init_path):\n",
" subprocess.run(['ffmpeg', '-i', f'{video_init_path}', '-vf', f'{vf}', '-vsync', 'vfr', '-q:v', '2', '-loglevel', 'error', '-stats', f'{videoFramesFolder}/%04d.jpg'], stdout=subprocess.PIPE).stdout.decode('utf-8')\n",
" else: \n",
" print(f'\\nWARNING!\\n\\nVideo not found: {video_init_path}.\\nPlease check your video path.\\n')\n",
" #!ffmpeg -i {video_init_path} -vf {vf} -vsync vfr -q:v 2 -loglevel error -stats {videoFramesFolder}/%04d.jpg\n",
"\n",
"\n",
"#@markdown ---\n",
"\n",
"#@markdown ####**2D Animation Settings:**\n",
"#@markdown `zoom` is a multiplier of dimensions, 1 is no zoom.\n",
"#@markdown All rotations are provided in degrees.\n",
"\n",
"key_frames = True #@param {type:\"boolean\"}\n",
"max_frames = 10000#@param {type:\"number\"}\n",
"\n",
"if animation_mode == \"Video Input\":\n",
" max_frames = len(glob(f'{videoFramesFolder}/*.jpg'))\n",
"\n",
"interp_spline = 'Linear' #Do not change, currently will not look good. param ['Linear','Quadratic','Cubic']{type:\"string\"}\n",
"angle = \"0:(0)\"#@param {type:\"string\"}\n",
"zoom = \"0: (1), 10: (1.05)\"#@param {type:\"string\"}\n",
"translation_x = \"0: (0)\"#@param {type:\"string\"}\n",
"translation_y = \"0: (0)\"#@param {type:\"string\"}\n",
"translation_z = \"0: (10.0)\"#@param {type:\"string\"}\n",
"rotation_3d_x = \"0: (0)\"#@param {type:\"string\"}\n",
"rotation_3d_y = \"0: (0)\"#@param {type:\"string\"}\n",
"rotation_3d_z = \"0: (0)\"#@param {type:\"string\"}\n",
"midas_depth_model = \"dpt_large\"#@param {type:\"string\"}\n",
"midas_weight = 0.3#@param {type:\"number\"}\n",
"near_plane = 200#@param {type:\"number\"}\n",
"far_plane = 10000#@param {type:\"number\"}\n",
"fov = 40#@param {type:\"number\"}\n",
"padding_mode = 'border'#@param {type:\"string\"}\n",
"sampling_mode = 'bicubic'#@param {type:\"string\"}\n",
"\n",
"#======= TURBO MODE\n",
"#@markdown ---\n",
"#@markdown ####**Turbo Mode (3D anim only):**\n",
"#@markdown (Starts after frame 10,) skips diffusion steps and just uses depth map to warp images for skipped frames.\n",
"#@markdown Speeds up rendering by 2x-4x, and may improve image coherence between frames.\n",
"#@markdown For different settings tuned for Turbo Mode, refer to the original Disco-Turbo Github: https://github.com/zippy731/disco-diffusion-turbo\n",
"\n",
"turbo_mode = False #@param {type:\"boolean\"}\n",
"turbo_steps = \"3\" #@param [\"2\",\"3\",\"4\",\"5\",\"6\"] {type:\"string\"}\n",
"turbo_preroll = 10 # frames\n",
"\n",
"#insist turbo be used only w 3d anim.\n",
"if turbo_mode and animation_mode != '3D':\n",
" print('=====')\n",
" print('Turbo mode only available with 3D animations. Disabling Turbo.')\n",
" print('=====')\n",
" turbo_mode = False\n",
"\n",
"#@markdown ---\n",
"\n",
"#@markdown ####**Coherency Settings:**\n",
"#@markdown `frame_scale` tries to guide the new frame to looking like the old one. A good default is 1500.\n",
"frames_scale = 1500 #@param{type: 'integer'}\n",
"#@markdown `frame_skip_steps` will blur the previous frame - higher values will flicker less but struggle to add enough new detail to zoom into.\n",
"frames_skip_steps = '60%' #@param ['40%', '50%', '60%', '70%', '80%'] {type: 'string'}\n",
"\n",
"#@markdown ####**Video Init Coherency Settings:**\n",
"#@markdown `frame_scale` tries to guide the new frame to looking like the old one. A good default is 1500.\n",
"video_init_frames_scale = 15000 #@param{type: 'integer'}\n",
"#@markdown `frame_skip_steps` will blur the previous frame - higher values will flicker less but struggle to add enough new detail to zoom into.\n",
"video_init_frames_skip_steps = '70%' #@param ['40%', '50%', '60%', '70%', '80%'] {type: 'string'}\n",
"\n",
"#======= VR MODE\n",
"#@markdown ---\n",
"#@markdown ####**VR Mode (3D anim only):**\n",
"#@markdown Enables stereo rendering of left/right eye views (supporting Turbo) which use a different (fish-eye) camera projection matrix. \n",
"#@markdown Note the images you're prompting will work better if they have some inherent wide-angle aspect\n",
"#@markdown The generated images will need to be combined into left/right videos. These can then be stitched into the VR180 format.\n",
"#@markdown Google made the VR180 Creator tool but subsequently stopped supporting it. It's available for download in a few places including https://www.patrickgrunwald.de/vr180-creator-download\n",
"#@markdown The tool is not only good for stitching (videos and photos) but also for adding the correct metadata into existing videos, which is needed for services like YouTube to identify the format correctly.\n",
"#@markdown Watching YouTube VR videos isn't necessarily the easiest depending on your headset. For instance Oculus have a dedicated media studio and store which makes the files easier to access on a Quest https://creator.oculus.com/manage/mediastudio/\n",
"#@markdown \n",
"#@markdown The command to get ffmpeg to concat your frames for each eye is in the form: `ffmpeg -framerate 15 -i frame_%4d_l.png l.mp4` (repeat for r)\n",
"\n",
"vr_mode = False #@param {type:\"boolean\"}\n",
"#@markdown `vr_eye_angle` is the y-axis rotation of the eyes towards the center\n",
"vr_eye_angle = 0.5 #@param{type:\"number\"}\n",
"#@markdown interpupillary distance (between the eyes)\n",
"vr_ipd = 5.0 #@param{type:\"number\"}\n",
"\n",
"#insist VR be used only w 3d anim.\n",
"if vr_mode and animation_mode != '3D':\n",
" print('=====')\n",
" print('VR mode only available with 3D animations. Disabling VR.')\n",
" print('=====')\n",
" vr_mode = False\n",
"\n",
"\n",
"def parse_key_frames(string, prompt_parser=None):\n",
" \"\"\"Given a string representing frame numbers paired with parameter values at that frame,\n",
" return a dictionary with the frame numbers as keys and the parameter values as the values.\n",
"\n",
" Parameters\n",
" ----------\n",
" string: string\n",
" Frame numbers paired with parameter values at that frame number, in the format\n",
" 'framenumber1: (parametervalues1), framenumber2: (parametervalues2), ...'\n",
" prompt_parser: function or None, optional\n",
" If provided, prompt_parser will be applied to each string of parameter values.\n",
" \n",
" Returns\n",
" -------\n",
" dict\n",
" Frame numbers as keys, parameter values at that frame number as values\n",
"\n",
" Raises\n",
" ------\n",
" RuntimeError\n",
" If the input string does not match the expected format.\n",
" \n",
" Examples\n",
" --------\n",
" >>> parse_key_frames(\"10:(Apple: 1| Orange: 0), 20: (Apple: 0| Orange: 1| Peach: 1)\")\n",
" {10: 'Apple: 1| Orange: 0', 20: 'Apple: 0| Orange: 1| Peach: 1'}\n",
"\n",
" >>> parse_key_frames(\"10:(Apple: 1| Orange: 0), 20: (Apple: 0| Orange: 1| Peach: 1)\", prompt_parser=lambda x: x.lower()))\n",
" {10: 'apple: 1| orange: 0', 20: 'apple: 0| orange: 1| peach: 1'}\n",
" \"\"\"\n",
" import re\n",
" pattern = r'((?P<frame>[0-9]+):[\\s]*[\\(](?P<param>[\\S\\s]*?)[\\)])'\n",
" frames = dict()\n",
" for match_object in re.finditer(pattern, string):\n",
" frame = int(match_object.groupdict()['frame'])\n",
" param = match_object.groupdict()['param']\n",
" if prompt_parser:\n",
" frames[frame] = prompt_parser(param)\n",
" else:\n",
" frames[frame] = param\n",
"\n",
" if frames == {} and len(string) != 0:\n",
" raise RuntimeError('Key Frame string not correctly formatted')\n",
" return frames\n",
"\n",
"def get_inbetweens(key_frames, integer=False):\n",
" \"\"\"Given a dict with frame numbers as keys and a parameter value as values,\n",
" return a pandas Series containing the value of the parameter at every frame from 0 to max_frames.\n",
" Any values not provided in the input dict are calculated by linear interpolation between\n",
" the values of the previous and next provided frames. If there is no previous provided frame, then\n",
" the value is equal to the value of the next provided frame, or if there is no next provided frame,\n",
" then the value is equal to the value of the previous provided frame. If no frames are provided,\n",
" all frame values are NaN.\n",
"\n",
" Parameters\n",
" ----------\n",
" key_frames: dict\n",
" A dict with integer frame numbers as keys and numerical values of a particular parameter as values.\n",
" integer: Bool, optional\n",
" If True, the values of the output series are converted to integers.\n",
" Otherwise, the values are floats.\n",
" \n",
" Returns\n",
" -------\n",
" pd.Series\n",
" A Series with length max_frames representing the parameter values for each frame.\n",
" \n",
" Examples\n",
" --------\n",
" >>> max_frames = 5\n",
" >>> get_inbetweens({1: 5, 3: 6})\n",
" 0 5.0\n",
" 1 5.0\n",
" 2 5.5\n",
" 3 6.0\n",
" 4 6.0\n",
" dtype: float64\n",
"\n",
" >>> get_inbetweens({1: 5, 3: 6}, integer=True)\n",
" 0 5\n",
" 1 5\n",
" 2 5\n",
" 3 6\n",
" 4 6\n",
" dtype: int64\n",
" \"\"\"\n",
" key_frame_series = pd.Series([np.nan for a in range(max_frames)])\n",
"\n",
" for i, value in key_frames.items():\n",
" key_frame_series[i] = value\n",
" key_frame_series = key_frame_series.astype(float)\n",
" \n",
" interp_method = interp_spline\n",
"\n",
" if interp_method == 'Cubic' and len(key_frames.items()) <=3:\n",
" interp_method = 'Quadratic'\n",
" \n",
" if interp_method == 'Quadratic' and len(key_frames.items()) <= 2:\n",
" interp_method = 'Linear'\n",
" \n",
" \n",
" key_frame_series[0] = key_frame_series[key_frame_series.first_valid_index()]\n",
" key_frame_series[max_frames-1] = key_frame_series[key_frame_series.last_valid_index()]\n",
" # key_frame_series = key_frame_series.interpolate(method=intrp_method,order=1, limit_direction='both')\n",
" key_frame_series = key_frame_series.interpolate(method=interp_method.lower(),limit_direction='both')\n",
" if integer:\n",
" return key_frame_series.astype(int)\n",
" return key_frame_series\n",
"\n",
"def split_prompts(prompts):\n",
" prompt_series = pd.Series([np.nan for a in range(max_frames)])\n",
" for i, prompt in prompts.items():\n",
" prompt_series[i] = prompt\n",
" # prompt_series = prompt_series.astype(str)\n",
" prompt_series = prompt_series.ffill().bfill()\n",
" return prompt_series\n",
"\n",
"if key_frames:\n",
" try:\n",
" angle_series = get_inbetweens(parse_key_frames(angle))\n",
" except RuntimeError as e:\n",
" print(\n",
" \"WARNING: You have selected to use key frames, but you have not \"\n",
" \"formatted `angle` correctly for key frames.\\n\"\n",
" \"Attempting to interpret `angle` as \"\n",
" f'\"0: ({angle})\"\\n'\n",
" \"Please read the instructions to find out how to use key frames \"\n",
" \"correctly.\\n\"\n",
" )\n",
" angle = f\"0: ({angle})\"\n",
" angle_series = get_inbetweens(parse_key_frames(angle))\n",
"\n",
" try:\n",
" zoom_series = get_inbetweens(parse_key_frames(zoom))\n",
" except RuntimeError as e:\n",
" print(\n",
" \"WARNING: You have selected to use key frames, but you have not \"\n",
" \"formatted `zoom` correctly for key frames.\\n\"\n",
" \"Attempting to interpret `zoom` as \"\n",
" f'\"0: ({zoom})\"\\n'\n",
" \"Please read the instructions to find out how to use key frames \"\n",
" \"correctly.\\n\"\n",
" )\n",
" zoom = f\"0: ({zoom})\"\n",
" zoom_series = get_inbetweens(parse_key_frames(zoom))\n",
"\n",
" try:\n",
" translation_x_series = get_inbetweens(parse_key_frames(translation_x))\n",
" except RuntimeError as e:\n",
" print(\n",
" \"WARNING: You have selected to use key frames, but you have not \"\n",
" \"formatted `translation_x` correctly for key frames.\\n\"\n",
" \"Attempting to interpret `translation_x` as \"\n",
" f'\"0: ({translation_x})\"\\n'\n",
" \"Please read the instructions to find out how to use key frames \"\n",
" \"correctly.\\n\"\n",
" )\n",
" translation_x = f\"0: ({translation_x})\"\n",
" translation_x_series = get_inbetweens(parse_key_frames(translation_x))\n",
"\n",
" try:\n",
" translation_y_series = get_inbetweens(parse_key_frames(translation_y))\n",
" except RuntimeError as e:\n",
" print(\n",
" \"WARNING: You have selected to use key frames, but you have not \"\n",
" \"formatted `translation_y` correctly for key frames.\\n\"\n",
" \"Attempting to interpret `translation_y` as \"\n",
" f'\"0: ({translation_y})\"\\n'\n",
" \"Please read the instructions to find out how to use key frames \"\n",
" \"correctly.\\n\"\n",
" )\n",
" translation_y = f\"0: ({translation_y})\"\n",
" translation_y_series = get_inbetweens(parse_key_frames(translation_y))\n",
"\n",
" try:\n",
" translation_z_series = get_inbetweens(parse_key_frames(translation_z))\n",
" except RuntimeError as e:\n",
" print(\n",
" \"WARNING: You have selected to use key frames, but you have not \"\n",
" \"formatted `translation_z` correctly for key frames.\\n\"\n",
" \"Attempting to interpret `translation_z` as \"\n",
" f'\"0: ({translation_z})\"\\n'\n",
" \"Please read the instructions to find out how to use key frames \"\n",
" \"correctly.\\n\"\n",
" )\n",
" translation_z = f\"0: ({translation_z})\"\n",
" translation_z_series = get_inbetweens(parse_key_frames(translation_z))\n",
"\n",
" try:\n",
" rotation_3d_x_series = get_inbetweens(parse_key_frames(rotation_3d_x))\n",
" except RuntimeError as e:\n",
" print(\n",
" \"WARNING: You have selected to use key frames, but you have not \"\n",
" \"formatted `rotation_3d_x` correctly for key frames.\\n\"\n",
" \"Attempting to interpret `rotation_3d_x` as \"\n",
" f'\"0: ({rotation_3d_x})\"\\n'\n",
" \"Please read the instructions to find out how to use key frames \"\n",
" \"correctly.\\n\"\n",
" )\n",
" rotation_3d_x = f\"0: ({rotation_3d_x})\"\n",
" rotation_3d_x_series = get_inbetweens(parse_key_frames(rotation_3d_x))\n",
"\n",
" try:\n",
" rotation_3d_y_series = get_inbetweens(parse_key_frames(rotation_3d_y))\n",
" except RuntimeError as e:\n",
" print(\n",
" \"WARNING: You have selected to use key frames, but you have not \"\n",
" \"formatted `rotation_3d_y` correctly for key frames.\\n\"\n",
" \"Attempting to interpret `rotation_3d_y` as \"\n",
" f'\"0: ({rotation_3d_y})\"\\n'\n",
" \"Please read the instructions to find out how to use key frames \"\n",
" \"correctly.\\n\"\n",
" )\n",
" rotation_3d_y = f\"0: ({rotation_3d_y})\"\n",
" rotation_3d_y_series = get_inbetweens(parse_key_frames(rotation_3d_y))\n",
"\n",
" try:\n",
" rotation_3d_z_series = get_inbetweens(parse_key_frames(rotation_3d_z))\n",
" except RuntimeError as e:\n",
" print(\n",
" \"WARNING: You have selected to use key frames, but you have not \"\n",
" \"formatted `rotation_3d_z` correctly for key frames.\\n\"\n",
" \"Attempting to interpret `rotation_3d_z` as \"\n",
" f'\"0: ({rotation_3d_z})\"\\n'\n",
" \"Please read the instructions to find out how to use key frames \"\n",
" \"correctly.\\n\"\n",
" )\n",
" rotation_3d_z = f\"0: ({rotation_3d_z})\"\n",
" rotation_3d_z_series = get_inbetweens(parse_key_frames(rotation_3d_z))\n",
"\n",
"else:\n",
" angle = float(angle)\n",
" zoom = float(zoom)\n",
" translation_x = float(translation_x)\n",
" translation_y = float(translation_y)\n",
" translation_z = float(translation_z)\n",
" rotation_3d_x = float(rotation_3d_x)\n",
" rotation_3d_y = float(rotation_3d_y)\n",
" rotation_3d_z = float(rotation_3d_z)"
],
"outputs": [],
"execution_count": null
},
{
"cell_type": "code",
"metadata": {
"id": "InstallRAFT"
},
"source": [
"#@title Install RAFT for Video input animation mode only\n",
"#@markdown Run once per session. Doesn't download again if model path exists.\n",
"#@markdown Use force download to reload raft models if needed\n",
"force_download = False #@param {type:'boolean'}\n",
"if animation_mode == 'Video Input':\n",
" os.chdir(f'{PROJECT_DIR}')\n",
" try:\n",
" from raft import RAFT\n",
" except:\n",
" if not os.path.exists(os.path.join(PROJECT_DIR, 'RAFT')):\n",
" gitclone('https://github.com/princeton-vl/RAFT', os.path.join(PROJECT_DIR, 'RAFT'))\n",
" sys.path.append(f'{PROJECT_DIR}/RAFT')\n",
"\n",
" if not os.path.exists(os.path.join(PROJECT_DIR, 'neural-style-tf')):\n",
" gitclone('https://github.com/cysmith/neural-style-tf', os.path.join(PROJECT_DIR, 'neural-style-tf'))\n",
" \n",
" os.chdir(PROJECT_DIR)\n",
" if (not (os.path.exists(f'/{root_path}/RAFT/models'))) or force_download:\n",
" sub_p_res = subprocess.run(['bash', './download_models.sh'], stdout=subprocess.PIPE).stdout.decode('utf-8')\n",
" print(sub_p_res)"
],
"outputs": [],
"execution_count": null
},
{
"cell_type": "code",
"metadata": {
"id": "FlowFns1"
},
"source": [
"#@title Define optical flow functions for Video input animation mode only\n",
"if animation_mode == 'Video Input':\n",
" in_path = videoFramesFolder\n",
" flo_folder = f'{in_path}/out_flo_fwd'\n",
" #the main idea comes from neural-style-tf frame warping with optical flow maps\n",
" #https://github.com/cysmith/neural-style-tf\n",
" path = f'{PROJECT_DIR}/RAFT/core'\n",
" import sys\n",
" sys.path.append(f'{PROJECT_DIR}/RAFT/core')\n",
" os.chdir(f'{PROJECT_DIR}/RAFT/core')\n",
" print(os.getcwd())\n",
" \n",
" print(\"Renaming RAFT core's utils.utils to raftutils.utils (to avoid a naming conflict with AdaBins)\")\n",
" if not os.path.exists(\"f'{PROJECT_DIR}/RAFT/core/raftutils'\"):\n",
" os.rename(f'{PROJECT_DIR}/RAFT/core/utils', f'{PROJECT_DIR}/RAFT/core/raftutils')\n",
" sub_p_res = subprocess.run(['sed', '-i', 's/from utils.utils/from raftutils.utils/g', f'{PROJECT_DIR}/RAFT/core/corr.py'], stdout=subprocess.PIPE).stdout.decode('utf-8')\n",
" sub_p_res = subprocess.run(['sed', '-i', 's/from utils.utils/from raftutils.utils/g', f'{PROJECT_DIR}/RAFT/core/raft.py'], stdout=subprocess.PIPE).stdout.decode('utf-8')\n",
"\n",
" from raftutils.utils import InputPadder\n",
" from raft import RAFT\n",
" from raftutils import flow_viz\n",
" import numpy as np\n",
" import argparse, PIL, cv2\n",
" from PIL import Image\n",
" from tqdm.notebook import tqdm\n",
" from glob import glob\n",
" import torch\n",
" \n",
" args2 = argparse.Namespace()\n",
" args2.small = False\n",
" args2.mixed_precision = True\n",
" \n",
" \n",
" TAG_CHAR = np.array([202021.25], np.float32)\n",
" \n",
" def writeFlow(filename,uv,v=None):\n",
" \"\"\" \n",
" https://github.com/NVIDIA/flownet2-pytorch/blob/master/utils/flow_utils.py\n",
" Copyright 2017 NVIDIA CORPORATION\n",
" \n",
" Licensed under the Apache License, Version 2.0 (the \"License\");\n",
" you may not use this file except in compliance with the License.\n",
" You may obtain a copy of the License at\n",
" \n",
" http://www.apache.org/licenses/LICENSE-2.0\n",
" \n",
" Unless required by applicable law or agreed to in writing, software\n",
" distributed under the License is distributed on an \"AS IS\" BASIS,\n",
" WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n",
" See the License for the specific language governing permissions and\n",
" limitations under the License.\n",
" \n",
" Write optical flow to file.\n",
" \n",
" If v is None, uv is assumed to contain both u and v channels,\n",
" stacked in depth.\n",
" Original code by Deqing Sun, adapted from Daniel Scharstein.\n",
" \"\"\"\n",
" nBands = 2\n",
" \n",
" if v is None:\n",
" assert(uv.ndim == 3)\n",
" assert(uv.shape[2] == 2)\n",
" u = uv[:,:,0]\n",
" v = uv[:,:,1]\n",
" else:\n",
" u = uv\n",
" \n",
" assert(u.shape == v.shape)\n",
" height,width = u.shape\n",
" f = open(filename,'wb')\n",
" # write the header\n",
" f.write(TAG_CHAR)\n",
" np.array(width).astype(np.int32).tofile(f)\n",
" np.array(height).astype(np.int32).tofile(f)\n",
" # arrange into matrix form\n",
" tmp = np.zeros((height, width*nBands))\n",
" tmp[:,np.arange(width)*2] = u\n",
" tmp[:,np.arange(width)*2 + 1] = v\n",
" tmp.astype(np.float32).tofile(f)\n",
" f.close()\n",
" \n",
" def load_cc(path):\n",
" rows = open(path).readlines()\n",
" w, h = list(map(int, rows[0].split(' ')))\n",
" nums = np.zeros((h, w)).astype('float32')\n",
" for i in range(1, len(rows)):\n",
" row = rows[i].rstrip().split(' ')\n",
" nums[i-1] = np.array(list(map(np.float32, row)))\n",
" nums[i-1] = list(map(lambda x: 0. if x < 255. else 1., nums[i-1]))\n",
" weights = np.repeat(nums[...,None],3, axis=2)\n",
" return weights\n",
" \n",
" def load_img(img, size):\n",
" img = Image.open(img).convert('RGB').resize(size)\n",
" return torch.from_numpy(np.array(img)).permute(2,0,1).float()[None,...].cuda()\n",
" \n",
" def get_flow(frame1, frame2, model, iters=20):\n",
" padder = InputPadder(frame1.shape)\n",
" frame1, frame2 = padder.pad(frame1, frame2)\n",
" _, flow12 = model(frame1, frame2, iters=iters, test_mode=True)\n",
" flow12 = flow12[0].permute(1, 2, 0).detach().cpu().numpy()\n",
" \n",
" return flow12\n",
" \n",
" def warp_flow(img, flow):\n",
" h, w = flow.shape[:2]\n",
" flow = flow.copy()\n",
" flow[:, :, 0] += np.arange(w)\n",
" flow[:, :, 1] += np.arange(h)[:, np.newaxis]\n",
" res = cv2.remap(img, flow, None, cv2.INTER_LINEAR)\n",
" return res\n",
" \n",
" def makeEven(_x):\n",
" return _x if (_x % 2 == 0) else _x+1\n",
" \n",
" def fit(img,maxsize=512):\n",
" maxdim = max(*img.size)\n",
" if maxdim>maxsize:\n",
" # if True:\n",
" ratio = maxsize/maxdim\n",
" x,y = img.size\n",
" size = (makeEven(int(x*ratio)),makeEven(int(y*ratio))) \n",
" img = img.resize(size)\n",
" return img\n",
" \n",
" def warp(frame1, frame2, flo_path, blend=0.5, weights_path=None):\n",
" flow21 = np.load(flo_path)\n",
" frame1pil = np.array(frame1.convert('RGB').resize((flow21.shape[1],flow21.shape[0])))\n",
" frame1_warped21 = warp_flow(frame1pil, flow21)\n",
" # frame2pil = frame1pil\n",
" frame2pil = np.array(frame2.convert('RGB').resize((flow21.shape[1],flow21.shape[0])))\n",
" \n",
" if weights_path: \n",
" forward_weights = load_cc(weights_path)\n",
" blended_w = frame2pil*(1-blend) + blend*(frame1_warped21*forward_weights+frame2pil*(1-forward_weights))\n",
" else:\n",
" blended_w = frame2pil*(1-blend) + frame1_warped21*(blend)\n",
" \n",
" return PIL.Image.fromarray(blended_w.astype('uint8'))\n",
" \n",
" in_path = videoFramesFolder\n",
" flo_folder = f'{in_path}/out_flo_fwd'\n",
" \n",
" temp_flo = in_path+'/temp_flo'\n",
" flo_fwd_folder = in_path+'/out_flo_fwd'\n",
" flo_bck_folder = in_path+'/out_flo_bck'\n",
" \n",
" os.chdir(root_path)"
],
"outputs": [],
"execution_count": null
},
{
"cell_type": "code",
"metadata": {
"id": "FlowFns2"
},
"source": [
"#@title Generate optical flow and consistency maps\n",
"#@markdown Run once per init video\n",
"\n",
"if animation_mode == \"Video Input\":\n",
" import gc\n",
"\n",
" force_flow_generation = False #@param {type:'boolean'}\n",
" in_path = videoFramesFolder\n",
" flo_folder = f'{in_path}/out_flo_fwd'\n",
"\n",
" if not video_init_flow_warp:\n",
" print('video_init_flow_warp not set, skipping')\n",
"\n",
" if (animation_mode == 'Video Input') and (video_init_flow_warp):\n",
" flows = glob(flo_folder+'/*.*')\n",
" if (len(flows)>0) and not force_flow_generation:\n",
" print(f'Skipping flow generation:\\nFound {len(flows)} existing flow files in current working folder: {flo_folder}.\\nIf you wish to generate new flow files, check force_flow_generation and run this cell again.')\n",
" \n",
" if (len(flows)==0) or force_flow_generation:\n",
" frames = sorted(glob(in_path+'/*.*'));\n",
" if len(frames)<2: \n",
" print(f'WARNING!\\nCannot create flow maps: Found {len(frames)} frames extracted from your video input.\\nPlease check your video path.')\n",
" if len(frames)>=2:\n",
" \n",
" raft_model = torch.nn.DataParallel(RAFT(args2))\n",
" raft_model.load_state_dict(torch.load(f'/{root_path}/RAFT/models/raft-things.pth'))\n",
" raft_model = raft_model.module.cuda().eval()\n",
" \n",
" for f in pathlib.Path(f'{flo_fwd_folder}').glob('*.*'):\n",
" f.unlink()\n",
" \n",
" temp_flo = in_path+'/temp_flo'\n",
" flo_fwd_folder = in_path+'/out_flo_fwd'\n",
" \n",
" createPath(flo_fwd_folder)\n",
" createPath(temp_flo)\n",
" \n",
" #(c) Manuel Ruder, Alexey Dosovitskiy, Thomas Brox 2016\n",
" #https://github.com/cysmith/neural-style-tf/tree/master/video_input/consistencyChecker\n",
" cchecker_path = f'/{PROJECT_DIR}/neural-style-tf/video_input/consistencyChecker/consistencyChecker'\n",
" \n",
" framecount = 0\n",
" for frame1, frame2 in tqdm(zip(frames[:-1], frames[1:]), total=len(frames)-1):\n",
" \n",
" out_flow21_fn = f\"{flo_fwd_folder}/{frame1.split('/')[-1]}\"\n",
" \n",
" frame1 = load_img(frame1, width_height)\n",
" frame2 = load_img(frame2, width_height)\n",
" \n",
" flow21 = get_flow(frame2, frame1, raft_model)\n",
" np.save(out_flow21_fn, flow21)\n",
" \n",
" if video_init_check_consistency:\n",
" flow12 = get_flow(frame1, frame2, raft_model)\n",
" gc.collect()\n",
" writeFlow(f'{temp_flo}/flow21.flo', flow21)\n",
" writeFlow(f'{temp_flo}/flow12.flo', flow12)\n",
" sub_p_res = subprocess.run([f'{cchecker_path}', f'{temp_flo}/flow21.flo', f'{temp_flo}/flow12.flo', f'{out_flow21_fn}-21.txt'], stdout=subprocess.PIPE).stdout.decode('utf-8')\n",
" print(f\"Checked consistency for framecount: {framecount} {sub_p_res}\")\n",
" framecount += 1\n",
" gc.collect()\n",
" \n",
" del raft_model \n",
" gc.collect()"
],
"outputs": [],
"execution_count": null
},
{
"cell_type": "markdown",
"metadata": {
"id": "ExtraSetTop"
},
"source": [
"### Extra Settings\n",
" Partial Saves, Advanced Settings, Cutn Scheduling"
]
},
{
"cell_type": "code",
"metadata": {
"id": "ExtraSettings"
},
"source": [
"#@markdown ####**Saving:**\n",
"\n",
"intermediate_saves = 0#@param{type: 'raw'}\n",
"intermediates_in_subfolder = True #@param{type: 'boolean'}\n",
"#@markdown Intermediate steps will save a copy at your specified intervals. You can either format it as a single integer or a list of specific steps \n",
"\n",
"#@markdown A value of `2` will save a copy at 33% and 66%. 0 will save none.\n",
"\n",
"#@markdown A value of `[5, 9, 34, 45]` will save at steps 5, 9, 34, and 45. (Make sure to include the brackets)\n",
"\n",
"\n",
"if type(intermediate_saves) is not list:\n",
" if intermediate_saves:\n",
" steps_per_checkpoint = math.floor((steps - skip_steps - 1) // (intermediate_saves+1))\n",
" steps_per_checkpoint = steps_per_checkpoint if steps_per_checkpoint > 0 else 1\n",
" print(f'Will save every {steps_per_checkpoint} steps')\n",
" else:\n",
" steps_per_checkpoint = steps+10\n",
"else:\n",
" steps_per_checkpoint = None\n",
"\n",
"if intermediate_saves and intermediates_in_subfolder is True:\n",
" partialFolder = f'{batchFolder}/partials'\n",
" createPath(partialFolder)\n",
"\n",
"#@markdown ---\n",
"\n",
"#@markdown ####**Advanced Settings:**\n",
"#@markdown *There are a few extra advanced settings available if you double click this cell.*\n",
"\n",
"#@markdown *Perlin init will replace your init, so uncheck if using one.*\n",
"\n",
"perlin_init = False #@param{type: 'boolean'}\n",
"perlin_mode = 'mixed' #@param ['mixed', 'color', 'gray']\n",
"set_seed = 'random_seed' #@param{type: 'string'}\n",
"eta = 0.8#@param{type: 'number'}\n",
"clamp_grad = True #@param{type: 'boolean'}\n",
"clamp_max = 0.05 #@param{type: 'number'}\n",
"\n",
"\n",
"### EXTRA ADVANCED SETTINGS:\n",
"randomize_class = True\n",
"clip_denoised = False\n",
"fuzzy_prompt = False\n",
"rand_mag = 0.05\n",
"\n",
"\n",
"#@markdown ---\n",
"\n",
"#@markdown ####**Cutn Scheduling:**\n",
"#@markdown Format: `[40]*400+[20]*600` = 40 cuts for the first 400 /1000 steps, then 20 for the last 600/1000\n",
"\n",
"#@markdown cut_overview and cut_innercut are cumulative for total cutn on any given step. Overview cuts see the entire image and are good for early structure, innercuts are your standard cutn.\n",
"\n",
"cut_overview = \"[12]*400+[4]*600\" #@param {type: 'string'} \n",
"cut_innercut =\"[4]*400+[12]*600\"#@param {type: 'string'} \n",
"cut_ic_pow = 1#@param {type: 'number'} \n",
"cut_icgray_p = \"[0.2]*400+[0]*600\"#@param {type: 'string'}\n",
"\n",
"#@markdown ---\n",
"\n",
"#@markdown ####**Transformation Settings:**\n",
"use_vertical_symmetry = False #@param {type:\"boolean\"}\n",
"use_horizontal_symmetry = False #@param {type:\"boolean\"}\n",
"transformation_percent = [0.09] #@param"
],
"outputs": [],
"execution_count": null
},
{
"cell_type": "markdown",
"metadata": {
"id": "PromptsTop"
},
"source": [
"### Prompts\n",
"`animation_mode: None` will only use the first set. `animation_mode: 2D / Video` will run through them per the set frames and hold on the last one."
]
},
{
"cell_type": "code",
"metadata": {
"id": "Prompts"
},
"source": [
"text_prompts = {\n",
" 0: [\"A beautiful painting of a singular lighthouse, shining its light across a tumultuous sea of blood by greg rutkowski and thomas kinkade, Trending on artstation.\", \"yellow color scheme\"],\n",
" 100: [\"This set of prompts start at frame 100\",\"This prompt has weight five:5\"],\n",
"}\n",
"\n",
"image_prompts = {\n",
" # 0:['ImagePromptsWorkButArentVeryGood.png:2',],\n",
"}"
],
"outputs": [],
"execution_count": null
},
{
"cell_type": "markdown",
"metadata": {
"id": "DiffuseTop"
},
"source": [
"# 4. Diffuse!"
]
},
{
"cell_type": "code",
"metadata": {
"id": "DoTheRun"
},
"source": [
"#@title Do the Run!\n",
"#@markdown `n_batches` ignored with animation modes.\n",
"display_rate = 20 #@param{type: 'number'}\n",
"n_batches = 50 #@param{type: 'number'}\n",
"\n",
"#Update Model Settings\n",
"timestep_respacing = f'ddim{steps}'\n",
"diffusion_steps = (1000//steps)*steps if steps < 1000 else steps\n",
"model_config.update({\n",
" 'timestep_respacing': timestep_respacing,\n",
" 'diffusion_steps': diffusion_steps,\n",
"})\n",
"\n",
"batch_size = 1 \n",
"\n",
"def move_files(start_num, end_num, old_folder, new_folder):\n",
" for i in range(start_num, end_num):\n",
" old_file = old_folder + f'/{batch_name}({batchNum})_{i:04}.png'\n",
" new_file = new_folder + f'/{batch_name}({batchNum})_{i:04}.png'\n",
" os.rename(old_file, new_file)\n",
"\n",
"#@markdown ---\n",
"\n",
"\n",
"resume_run = False #@param{type: 'boolean'}\n",
"run_to_resume = 'latest' #@param{type: 'string'}\n",
"resume_from_frame = 'latest' #@param{type: 'string'}\n",
"retain_overwritten_frames = False #@param{type: 'boolean'}\n",
"if retain_overwritten_frames:\n",
" retainFolder = f'{batchFolder}/retained'\n",
" createPath(retainFolder)\n",
"\n",
"\n",
"skip_step_ratio = int(frames_skip_steps.rstrip(\"%\")) / 100\n",
"calc_frames_skip_steps = math.floor(steps * skip_step_ratio)\n",
"\n",
"if animation_mode == 'Video Input':\n",
" frames = sorted(glob(in_path+'/*.*'));\n",
" if len(frames)==0: \n",
" sys.exit(\"ERROR: 0 frames found.\\nPlease check your video input path and rerun the video settings cell.\")\n",
" flows = glob(flo_folder+'/*.*')\n",
" if (len(flows)==0) and video_init_flow_warp:\n",
" sys.exit(\"ERROR: 0 flow files found.\\nPlease rerun the flow generation cell.\")\n",
"\n",
"if steps <= calc_frames_skip_steps:\n",
" sys.exit(\"ERROR: You can't skip more steps than your total steps\")\n",
"\n",
"if resume_run:\n",
" if run_to_resume == 'latest':\n",
" try:\n",
" batchNum\n",
" except:\n",
" batchNum = len(glob(f\"{batchFolder}/{batch_name}(*)_settings.txt\"))-1\n",
" else:\n",
" batchNum = int(run_to_resume)\n",
" if resume_from_frame == 'latest':\n",
" start_frame = len(glob(batchFolder+f\"/{batch_name}({batchNum})_*.png\"))\n",
" if animation_mode != '3D' and turbo_mode == True and start_frame > turbo_preroll and start_frame % int(turbo_steps) != 0:\n",
" start_frame = start_frame - (start_frame % int(turbo_steps))\n",
" else:\n",
" start_frame = int(resume_from_frame)+1\n",
" if animation_mode != '3D' and turbo_mode == True and start_frame > turbo_preroll and start_frame % int(turbo_steps) != 0:\n",
" start_frame = start_frame - (start_frame % int(turbo_steps))\n",
" if retain_overwritten_frames is True:\n",
" existing_frames = len(glob(batchFolder+f\"/{batch_name}({batchNum})_*.png\"))\n",
" frames_to_save = existing_frames - start_frame\n",
" print(f'Moving {frames_to_save} frames to the Retained folder')\n",
" move_files(start_frame, existing_frames, batchFolder, retainFolder)\n",
"else:\n",
" start_frame = 0\n",
" batchNum = len(glob(batchFolder+\"/*.txt\"))\n",
" while os.path.isfile(f\"{batchFolder}/{batch_name}({batchNum})_settings.txt\") or os.path.isfile(f\"{batchFolder}/{batch_name}-{batchNum}_settings.txt\"):\n",
" batchNum += 1\n",
"\n",
"print(f'Starting Run: {batch_name}({batchNum}) at frame {start_frame}')\n",
"\n",
"if set_seed == 'random_seed':\n",
" random.seed()\n",
" seed = random.randint(0, 2**32)\n",
" # print(f'Using seed: {seed}')\n",
"else:\n",
" seed = int(set_seed)\n",
"\n",
"args = {\n",
" 'batchNum': batchNum,\n",
" 'prompts_series':split_prompts(text_prompts) if text_prompts else None,\n",
" 'image_prompts_series':split_prompts(image_prompts) if image_prompts else None,\n",
" 'seed': seed,\n",
" 'display_rate':display_rate,\n",
" 'n_batches':n_batches if animation_mode == 'None' else 1,\n",
" 'batch_size':batch_size,\n",
" 'batch_name': batch_name,\n",
" 'steps': steps,\n",
" 'diffusion_sampling_mode': diffusion_sampling_mode,\n",
" 'width_height': width_height,\n",
" 'clip_guidance_scale': clip_guidance_scale,\n",
" 'tv_scale': tv_scale,\n",
" 'range_scale': range_scale,\n",
" 'sat_scale': sat_scale,\n",
" 'cutn_batches': cutn_batches,\n",
" 'init_image': init_image,\n",
" 'init_scale': init_scale,\n",
" 'skip_steps': skip_steps,\n",
" 'side_x': side_x,\n",
" 'side_y': side_y,\n",
" 'timestep_respacing': timestep_respacing,\n",
" 'diffusion_steps': diffusion_steps,\n",
" 'animation_mode': animation_mode,\n",
" 'video_init_path': video_init_path,\n",
" 'extract_nth_frame': extract_nth_frame,\n",
" 'video_init_seed_continuity': video_init_seed_continuity,\n",
" 'key_frames': key_frames,\n",
" 'max_frames': max_frames if animation_mode != \"None\" else 1,\n",
" 'interp_spline': interp_spline,\n",
" 'start_frame': start_frame,\n",
" 'angle': angle,\n",
" 'zoom': zoom,\n",
" 'translation_x': translation_x,\n",
" 'translation_y': translation_y,\n",
" 'translation_z': translation_z,\n",
" 'rotation_3d_x': rotation_3d_x,\n",
" 'rotation_3d_y': rotation_3d_y,\n",
" 'rotation_3d_z': rotation_3d_z,\n",
" 'midas_depth_model': midas_depth_model,\n",
" 'midas_weight': midas_weight,\n",
" 'near_plane': near_plane,\n",
" 'far_plane': far_plane,\n",
" 'fov': fov,\n",
" 'padding_mode': padding_mode,\n",
" 'sampling_mode': sampling_mode,\n",
" 'angle_series':angle_series,\n",
" 'zoom_series':zoom_series,\n",
" 'translation_x_series':translation_x_series,\n",
" 'translation_y_series':translation_y_series,\n",
" 'translation_z_series':translation_z_series,\n",
" 'rotation_3d_x_series':rotation_3d_x_series,\n",
" 'rotation_3d_y_series':rotation_3d_y_series,\n",
" 'rotation_3d_z_series':rotation_3d_z_series,\n",
" 'frames_scale': frames_scale,\n",
" 'skip_step_ratio': skip_step_ratio,\n",
" 'calc_frames_skip_steps': calc_frames_skip_steps,\n",
" 'text_prompts': text_prompts,\n",
" 'image_prompts': image_prompts,\n",
" 'cut_overview': eval(cut_overview),\n",
" 'cut_innercut': eval(cut_innercut),\n",
" 'cut_ic_pow': cut_ic_pow,\n",
" 'cut_icgray_p': eval(cut_icgray_p),\n",
" 'intermediate_saves': intermediate_saves,\n",
" 'intermediates_in_subfolder': intermediates_in_subfolder,\n",
" 'steps_per_checkpoint': steps_per_checkpoint,\n",
" 'perlin_init': perlin_init,\n",
" 'perlin_mode': perlin_mode,\n",
" 'set_seed': set_seed,\n",
" 'eta': eta,\n",
" 'clamp_grad': clamp_grad,\n",
" 'clamp_max': clamp_max,\n",
" 'skip_augs': skip_augs,\n",
" 'randomize_class': randomize_class,\n",
" 'clip_denoised': clip_denoised,\n",
" 'fuzzy_prompt': fuzzy_prompt,\n",
" 'rand_mag': rand_mag,\n",
" 'turbo_mode':turbo_mode,\n",
" 'turbo_steps':turbo_steps,\n",
" 'turbo_preroll':turbo_preroll,\n",
" 'use_vertical_symmetry': use_vertical_symmetry,\n",
" 'use_horizontal_symmetry': use_horizontal_symmetry,\n",
" 'transformation_percent': transformation_percent,\n",
" #video init settings\n",
" 'video_init_steps': video_init_steps,\n",
" 'video_init_clip_guidance_scale': video_init_clip_guidance_scale,\n",
" 'video_init_tv_scale': video_init_tv_scale,\n",
" 'video_init_range_scale': video_init_range_scale,\n",
" 'video_init_sat_scale': video_init_sat_scale,\n",
" 'video_init_cutn_batches': video_init_cutn_batches,\n",
" 'video_init_skip_steps': video_init_skip_steps,\n",
" 'video_init_frames_scale': video_init_frames_scale,\n",
" 'video_init_frames_skip_steps': video_init_frames_skip_steps,\n",
" #warp settings\n",
" 'video_init_flow_warp':video_init_flow_warp,\n",
" 'video_init_flow_blend':video_init_flow_blend,\n",
" 'video_init_check_consistency':video_init_check_consistency,\n",
" 'video_init_blend_mode':video_init_blend_mode\n",
"}\n",
"\n",
"# This isn't great in terms of what will get saved to the settings.. but it should work.\n",
"args['steps'] = args['video_init_steps']\n",
"args['clip_guidance_scale'] = args['video_init_clip_guidance_scale']\n",
"args['tv_scale'] = args['video_init_tv_scale']\n",
"args['range_scale'] = args['video_init_range_scale']\n",
"args['sat_scale'] = args['video_init_sat_scale']\n",
"args['cutn_batches'] = args['video_init_cutn_batches']\n",
"args['skip_steps'] = args['video_init_skip_steps']\n",
"args['frames_scale'] = args['video_init_frames_scale']\n",
"args['frames_skip_steps'] = args['video_init_frames_skip_steps']\n",
"\n",
"args = SimpleNamespace(**args)\n",
"\n",
"print('Prepping model...')\n",
"model, diffusion = create_model_and_diffusion(**model_config)\n",
"if diffusion_model == 'custom':\n",
" model.load_state_dict(torch.load(custom_path, map_location='cpu'))\n",
"else:\n",
" model.load_state_dict(torch.load(f'{model_path}/{diffusion_model}.pt', map_location='cpu'))\n",
"model.requires_grad_(False).eval().to(device)\n",
"for name, param in model.named_parameters():\n",
" if 'qkv' in name or 'norm' in name or 'proj' in name:\n",
" param.requires_grad_()\n",
"if model_config['use_fp16']:\n",
" model.convert_to_fp16()\n",
"\n",
"gc.collect()\n",
"torch.cuda.empty_cache()\n",
"try:\n",
" do_run()\n",
"except KeyboardInterrupt:\n",
" pass\n",
"finally:\n",
" print('Seed used:', seed)\n",
" gc.collect()\n",
" torch.cuda.empty_cache()"
],
"outputs": [],
"execution_count": null
},
{
"cell_type": "markdown",
"metadata": {
"id": "CreateVidTop"
},
"source": [
"# 5. Create the video"
]
},
{
"cell_type": "code",
"metadata": {
"id": "CreateVid",
"cellView": "form"
},
"source": [
"import PIL\n",
"# @title ### **Create video**\n",
"#@markdown Video file will save in the same folder as your images.\n",
"from tqdm.notebook import trange\n",
"skip_video_for_run_all = False #@param {type: 'boolean'}\n",
"\n",
"if animation_mode == 'Video Input':\n",
" frames = sorted(glob(in_path+'/*.*'));\n",
" if len(frames)==0: \n",
" sys.exit(\"ERROR: 0 frames found.\\nPlease check your video input path and rerun the video settings cell.\")\n",
" flows = glob(flo_folder+'/*.*')\n",
" if (len(flows)==0) and video_init_flow_warp:\n",
" sys.exit(\"ERROR: 0 flow files found.\\nPlease rerun the flow generation cell.\")\n",
"\n",
"if (video_init_blend_mode == \"optical flow\") and (animation_mode != 'Video Input'):\n",
" print('Please enable Video Input mode and generate optical flow maps to use optical flow video_init_blend_mode')\n",
"blend = 0.5#@param {type: 'number'}\n",
"video_init_check_consistency = True #@param {type: 'boolean'}\n",
"if skip_video_for_run_all == True:\n",
" print('Skipping video creation, uncheck skip_video_for_run_all if you want to run it')\n",
"\n",
"else:\n",
" # import subprocess in case this cell is run without the above cells\n",
" import subprocess\n",
" from base64 import b64encode\n",
"\n",
" latest_run = batchNum\n",
"\n",
" folder = batch_name #@param\n",
" run = latest_run #@param\n",
" final_frame = 'final_frame'\n",
"\n",
"\n",
" init_frame = 1#@param {type:\"number\"} This is the frame where the video will start\n",
" last_frame = final_frame#@param {type:\"number\"} You can change i to the number of the last frame you want to generate. It will raise an error if that number of frames does not exist.\n",
" fps = 12#@param {type:\"number\"}\n",
" # view_video_in_cell = True #@param {type: 'boolean'}\n",
"\n",
" frames = []\n",
" # tqdm.write('Generating video...')\n",
"\n",
" if last_frame == 'final_frame':\n",
" last_frame = len(glob(batchFolder+f\"/{folder}({run})_*.png\"))\n",
" print(f'Total frames: {last_frame}')\n",
"\n",
" image_path = f\"{outDirPath}/{folder}/{folder}({run})_%04d.png\"\n",
" filepath = f\"{outDirPath}/{folder}/{folder}({run}).mp4\"\n",
"\n",
" if (video_init_blend_mode == 'optical flow') and (animation_mode == 'Video Input'):\n",
" image_path = f\"{outDirPath}/{folder}/flow/{folder}({run})_%04d.png\"\n",
" filepath = f\"{outDirPath}/{folder}/{folder}({run})_flow.mp4\"\n",
" if last_frame == 'final_frame':\n",
" last_frame = len(glob(batchFolder+f\"/flow/{folder}({run})_*.png\"))\n",
" flo_out = batchFolder+f\"/flow\"\n",
" sub_p_res = subprocess.run(['mkdir', '\"{flo_out}\"'], stdout=subprocess.PIPE).stdout.decode('utf-8')\n",
" print(sub_p_res)\n",
" frames_in = sorted(glob(batchFolder+f\"/{folder}({run})_*.png\"))\n",
" sub_p_res = subprocess.run(['cp', '\"{frames_in[0]}\"', '\"{flo_out}\"'], stdout=subprocess.PIPE).stdout.decode('utf-8')\n",
" print(sub_p_res)\n",
" for i in trange(init_frame, min(len(frames_in), last_frame)):\n",
" frame1_path = frames_in[i-1]\n",
" frame2_path = frames_in[i]\n",
" \n",
" frame1 = PIL.Image.open(frame1_path)\n",
" frame2 = PIL.Image.open(frame2_path)\n",
" frame1_stem = f\"{(int(frame1_path.split('/')[-1].split('_')[-1][:-4])+1):04}.jpg\"\n",
" flo_path = f\"/{flo_folder}/{frame1_stem}.npy\"\n",
" weights_path = None\n",
" if video_init_check_consistency:\n",
" weights_path = f\"/{flo_folder}/{frame1_stem}-21.txt\" \n",
" warp(frame1, frame2, flo_path, blend=blend, weights_path=weights_path).save(batchFolder+f\"/flow/{folder}({run})_{i:04}.png\")\n",
" if video_init_blend_mode == 'linear':\n",
" image_path = f\"{outDirPath}/{folder}/blend/{folder}({run})_%04d.png\"\n",
" filepath = f\"{outDirPath}/{folder}/{folder}({run})_blend.mp4\"\n",
" if last_frame == 'final_frame':\n",
" last_frame = len(glob(batchFolder+f\"/blend/{folder}({run})_*.png\"))\n",
" blend_out = batchFolder+f\"/blend\"\n",
" sub_p_res = subprocess.run(['mkdir', '\"{blend_out}\"'], stdout=subprocess.PIPE).stdout.decode('utf-8')\n",
" print(sub_p_res)\n",
" frames_in = glob(batchFolder+f\"/{folder}({run})_*.png\")\n",
" sub_p_res = subprocess.run(['cp', '\"{frames_in[0]}\"', '\"{blend_out}\"'], stdout=subprocess.PIPE).stdout.decode('utf-8')\n",
" print(sub_p_res)\n",
" for i in trange(1, len(frames_in)):\n",
" frame1_path = frames_in[i-1]\n",
" frame2_path = frames_in[i]\n",
" \n",
" frame1 = PIL.Image.open(frame1_path)\n",
" frame2 = PIL.Image.open(frame2_path)\n",
" \n",
" frame = PIL.Image.fromarray((np.array(frame1)*(1-blend) + np.array(frame2)*(blend)).astype('uint8')).save(batchFolder+f\"/blend/{folder}({run})_{i:04}.png\")\n",
"\n",
"\n",
" cmd = [\n",
" 'ffmpeg',\n",
" '-y',\n",
" '-vcodec',\n",
" 'png',\n",
" '-r',\n",
" str(fps),\n",
" '-start_number',\n",
" str(init_frame),\n",
" '-i',\n",
" image_path,\n",
" '-frames:v',\n",
" str(last_frame+1),\n",
" '-c:v',\n",
" 'libx264',\n",
" '-vf',\n",
" f'fps={fps}',\n",
" '-pix_fmt',\n",
" 'yuv420p',\n",
" '-crf',\n",
" '17',\n",
" '-preset',\n",
" 'veryslow',\n",
" filepath\n",
" ]\n",
"\n",
" process = subprocess.Popen(cmd, cwd=f'{batchFolder}', stdout=subprocess.PIPE, stderr=subprocess.PIPE)\n",
" stdout, stderr = process.communicate()\n",
" if process.returncode != 0:\n",
" print(stderr)\n",
" raise RuntimeError(stderr)\n",
" else:\n",
" print(\"The video is ready and saved to the images folder\")\n",
"\n",
" # if view_video_in_cell:\n",
" # mp4 = open(filepath,'rb').read()\n",
" # data_url = \"data:video/mp4;base64,\" + b64encode(mp4).decode()\n",
" # display.HTML(f'<video width=400 controls><source src=\"{data_url}\" type=\"video/mp4\"></video>')"
],
"outputs": [],
"execution_count": null
}
],
"metadata": {
"anaconda-cloud": {},
"accelerator": "GPU",
"colab": {
"collapsed_sections": [
"CreditsChTop",
"TutorialTop",
"CheckGPU",
"InstallDeps",
"DefMidasFns",
"DefFns",
"DefSecModel",
"DefSuperRes",
"AnimSetTop",
"ExtraSetTop",
"InstallRAFT",
"CustModel",
"FlowFns1",
"FlowFns2"
],
"machine_shape": "hm",
"name": "Disco Diffusion v5.3.999 [Short-lived inclusions]",
"private_outputs": true,
"provenance": [],
"include_colab_link": true
},
"kernelspec": {
"display_name": "Python 3",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.6.1"
}
},
"nbformat": 4,
"nbformat_minor": 4
}