moved contents of presentations repo

Kandinsky-3/.gitignore

@@ -0,0 +1,160 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/#use-with-ide
+.pdm.toml
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+#.idea/

Kandinsky-3/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

Kandinsky-3/README.md

@@ -0,0 +1,230 @@
+# Kandinsky-3: Text-to-image diffusion model
+
+![](assets/title.jpg)
+
+[Kandinsky 3.0 Post](https://habr.com/ru/companies/sberbank/articles/775590/) | [Kandinsky 3.1 Post](https://habr.com/ru/companies/sberbank/articles/805337/) | [Project Page](https://ai-forever.github.io/Kandinsky-3) | [Generate](https://fusionbrain.ai) | [Telegram-bot](https://t.me/kandinsky21_bot) | [Technical Report](https://arxiv.org/pdf/2312.03511.pdf) |  [HuggingFace](https://huggingface.co/kandinsky-community/kandinsky-3) |
+
+# Kandinsky 3.1:
+
+## Description:
+
+We present Kandinsky 3.1, the follow-up to the Kandinsky 3.0 model, a large-scale text-to-image generation model based on latent diffusion, continuing the series of text-to-image Kandinsky models and reflecting our progress to achieve higher quality and realism of image generation, which we have enhanced and enriched with a variety of useful features and modes to give users more opportunities to fully utilise the power of our new model.
+
+## Kandinsky Flash (Kandinsky 3.0 Refiner)
+
+<figure>
+  <img src="assets/butterly_effect.jpg">
+</figure>
+
+
+Diffusion models have problems with fast image generation. To address this problem, we trained a Kandinksy Flash model based on the [Adversarial Diffusion Distillation](https://arxiv.org/abs/2311.17042) approach with some modifications: we trained the model on latents, which reduced the memory overhead and removed distillation loss as it did not affect the training. Also, we applied Kandinsky Flash model to images generated from Kandinsky 3.0 to improve visual quality of generated images.
+
+### Architecture
+
+For training Kandinsky Flash we used the following architecture of discriminator. It is the half of Kandinsky 3.0 U-Net encoder with additional head predictions.
+
+<img src="assets/architecture.png">
+
+### How to use:
+Check our jupyter notebooks with examples in `./examples` folder
+
+```python
+from kandinsky3 import get_T2I_Flash_pipeline
+
+device_map = torch.device('cuda:0')
+dtype_map = {
+    'unet': torch.float32,
+    'text_encoder': torch.float16,
+    'movq': torch.float32,
+}
+
+t2i_pipe = get_T2I_Flash_pipeline(
+    device_map, dtype_map
+)
+
+res = t2i_pipe("A cute corgi lives in a house made out of sushi.")
+```
+### Kandinsky Inpainting
+
+Also, we released a newer version of inpainting model, which we additionally trained the model on the object detection dataset. This allowed to get more stable generation of objects. The new weights are available at [ai-forever/Kandinsky3.1](https://huggingface.co/ai-forever/Kandinsky3.1). Check the usage [example](https://github.com/ai-forever/Kandinsky-3?tab=readme-ov-file#2-inpainting).
+
+
+## Prompt beautification
+
+<figure>
+  <img src="assets/prompt_beautifcation.png">
+</figure>
+
+
+Prompt plays crucial role in text-to-image generation. So, in Kandinsky 3.1 we decided to use language model for making prompt better. We used Intel's [neural-chat-7b-v3-1](https://huggingface.co/Intel/neural-chat-7b-v3-1) with the following system prompt as the LLM:
+
+```
+### System: You are a prompt engineer. Your mission is to expand prompts written by user. You should provide the best prompt for text to image generation in English.
+### User:
+{prompt}
+### Assistant:
+{answer of the model}
+```
+
+## KandiSuperRes
+
+<figure>
+  <img src="assets/superres.png">
+</figure>
+
+To learn more about KandiSuperRes, please checkout: https://github.com/ai-forever/KandiSuperRes/
+
+## Kandinsky IP-Adapter & Kandinsky ControlNet
+
+<figure>
+  <img src="assets/ip-adapter.png">
+</figure>
+
+To allow using image as condition in Kandinsky model, we trained IP-Adapter and HED-based ControlNet model. For more details please check out: https://github.com/ai-forever/kandinsky3-diffusers
+
+# Kandinsky 3.0:
+
+## Description:
+
+Kandinsky 3.0 is an open-source text-to-image diffusion model built upon the Kandinsky2-x model family. In comparison to its predecessors, Kandinsky 3.0 incorporates more data and specifically related to Russian culture, which allows to generate pictures related to Russin culture. Furthermore, enhancements have been made to the text understanding and visual quality of the model, achieved by increasing the size of the text encoder and Diffusion U-Net models, respectively.
+
+For more information: details of training, example of generations check out our [post](). The english version will be released in a couple of days.
+
+## Architecture details:
+
+
+![](assets/kandinsky.jpg)
+
+
+Architecture consists of three parts:
+
++ Text encoder Flan-UL2 (encoder part) - 8.6B
++ Latent Diffusion U-Net - 3B
++ MoVQ encoder/decoder - 267M
+
+
+## Models
+
+We release our two models:
+
++ [Base](): Base text-to-image diffusion model. This model was trained over 2M steps on 400 A100
++ [Inpainting](): Inpainting version of the model. The model was initialized from final checkpoint of base model and trained 250k steps on 300 A100.
+
+## Installing
+
+To install repo first one need to create conda environment:
+
+```
+conda create -n kandinsky -y python=3.8;
+source activate kandinsky;
+pip install torch==1.10.1+cu111 torchvision==0.11.2+cu111 torchaudio==0.10.1 -f https://download.pytorch.org/whl/cu113/torch_stable.html;
+pip install -r requirements.txt;
+```
+The exact dependencies is got using `pip freeze` and can be found in `exact_requirements.txt`
+
+## How to use:
+
+Check our jupyter notebooks with examples in `./examples` folder
+
+### 1. text2image
+
+```python
+import sys
+sys.path.append('..')
+
+import torch
+from kandinsky3 import get_T2I_pipeline
+
+device_map = torch.device('cuda:0')
+dtype_map = {
+    'unet': torch.float32,
+    'text_encoder': torch.float16,
+    'movq': torch.float32,
+}
+
+t2i_pipe = get_T2I_pipeline(
+    device_map, dtype_map,
+)
+res = t2i_pipe("A cute corgi lives in a house made out of sushi.")
+
+res[0]
+```
+
+### 2. inpainting
+
+```python
+from kandinsky3 import get_inpainting_pipeline
+
+device_map = torch.device('cuda:0')
+dtype_map = {
+    'unet': torch.float16,
+    'text_encoder': torch.float16,
+    'movq': torch.float32,
+}
+
+pipe = get_inpainting_pipeline(
+    device_map, dtype_map,
+)
+
+image = ... # PIL Image
+mask = ... # Numpy array (HxW). Set 1 where image should be masked
+image = inp_pipe( "A cute corgi lives in a house made out of sushi.", image, mask)
+```
+
+## Examples of generations
+
+<hr>
+
+<table class="center">
+<tr>
+  <td><img src="assets/photo_8.jpg" raw=true></td>
+  <td><img src="assets/photo_15.jpg"></td>
+  <td><img src="assets/photo_16.jpg"></td>
+  <td><img src="assets/photo_17.jpg"></td>
+</tr>
+<tr>
+  <td width=25% align="center">"A beautiful landscape outdoors scene in the crochet knitting art style, drawing in style by Alfons Mucha"</td>
+  <td width=25% align="center">"gorgeous phoenix, cosmic, darkness, epic, cinematic, moonlight, stars, high - definition, texture,Oscar-Claude Monet"</td>
+  <td width=25% align="center">"a yellow house at the edge of the danish fjord, in the style of eiko ojala, ingrid baars, ad posters, mountainous vistas, george ault, realistic details, dark white and dark gray, 4k"</td>
+  <td width=25% align="center">"dragon fruit head, upper body, realistic, illustration by Joshua Hoffine Norman Rockwell, scary, creepy, biohacking, futurism, Zaha Hadid style"</td>
+</tr>
+<tr>
+  <td><img src="assets/photo_2.jpg" raw=true></td>
+  <td><img src="assets/photo_19.jpg"></td>
+  <td><img src="assets/photo_13.jpg"></td>
+  <td><img src="assets/photo_14.jpg"></td>
+</tr>
+<tr>
+  <td width=25% align="center">"Amazing playful nice cute strawberry character, dynamic poze, surreal fantazy garden background, gorgeous masterpice, award winning photo, soft natural lighting, 3d, Blender, Octane render, tilt - shift, deep field, colorful, I can't believe how beautiful this is, colorful, cute and sweet baby - loved photo"</td>
+  <td width=25% align="center">"beautiful fairy-tale desert, in the sky a wave of sand merges with the milky way, stars, cosmism, digital art, 8k"</td>
+  <td width=25% align="center">"Car, mustang, movie, person, poster, car cover, person, in the style of alessandro gottardo, gold and cyan, gerald harvey jones, reflections, highly detailed illustrations, industrial urban scenes""</td>
+  <td width=25% align="center">"cloud in blue sky, a red lip, collage art, shuji terayama, dreamy objects, surreal, criterion collection, showa era, intricate details, mirror"</td>
+</tr>
+
+</table>
+
+<hr>
+
+## Authors
+
++ Vladimir Arkhipkin: [Github](https://github.com/oriBetelgeuse)
++ Anastasia Maltseva [Github](https://github.com/NastyaMittseva)
++ Andrei Filatov [Github](https://github.com/anvilarth), 
++ Igor Pavlov: [Github](https://github.com/boomb0om)
++ Julia Agafonova 
++ Arseniy Shakhmatov: [Github](https://github.com/cene555), [Blog](https://t.me/gradientdip)
++ Andrey Kuznetsov: [Github](https://github.com/kuznetsoffandrey), [Blog](https://t.me/complete_ai)
++ Denis Dimitrov: [Github](https://github.com/denndimitrov), [Blog](https://t.me/dendi_math_ai)
+
+## Citation
+```
+@misc{arkhipkin2023kandinsky,
+      title={Kandinsky 3.0 Technical Report}, 
+      author={Vladimir Arkhipkin and Andrei Filatov and Viacheslav Vasilev and Anastasia Maltseva and Said Azizov and Igor Pavlov and Julia Agafonova and Andrey Kuznetsov and Denis Dimitrov},
+      year={2023},
+      eprint={2312.03511},
+      archivePrefix={arXiv},
+      primaryClass={cs.CV}
+}
+```
+

Kandinsky-3/exact_requirements.txt

@@ -0,0 +1,372 @@
+absl-py==1.0.0
+accelerate==0.20.3
+adal==1.2.7
+addict==2.4.0
+aioboto3==11.0.1
+aiobotocore==2.4.2
+aiofiles==23.2.1
+aiohttp==3.8.1
+aiohttp-cors==0.7.0
+aioitertools==0.11.0
+aioredis==1.3.1
+aiosignal==1.2.0
+albumentations==1.3.1
+alembic==1.4.1
+alt-profanity-check==1.0.1
+altair==5.0.1
+antlr4-python3-runtime==4.9.3
+anyio==3.5.0
+apex==0.1
+appdirs==1.4.4
+argon2-cffi==21.3.0
+argon2-cffi-bindings==21.2.0
+asgiref==3.4.1
+astunparse==1.6.3
+async-timeout==4.0.2
+asyncpool==1.0
+asynctest==0.13.0
+attrs==21.4.0
+audioread==2.1.9
+autopage==0.4.0
+avro==1.11.0
+awscli==1.22.38
+azure-common==1.1.27
+azure-storage-blob==2.1.0
+azure-storage-common==2.1.0
+Babel==2.9.1
+backcall==0.2.0
+basicsr==1.4.2
+beautifulsoup4 @ file:///home/conda/feedstock_root/build_artifacts/beautifulsoup4_1631087867185/work
+bezier==2021.2.12
+bitarray==2.8.3
+bitmath==1.3.3.1
+bleach==4.1.0
+blessings==1.7
+blis==0.7.11
+bokeh==2.4.2
+boto3==1.24.59
+botocore==1.27.59
+braceexpand==0.1.7
+brotlipy @ file:///home/conda/feedstock_root/build_artifacts/brotlipy_1636012184244/work
+cached-property==1.5.2
+cachetools==4.2.4
+catalogue==2.0.10
+certifi==2021.10.8
+cffi==1.15.0
+chardet @ file:///home/conda/feedstock_root/build_artifacts/chardet_1635814832679/work
+charset-normalizer==2.0.10
+click==8.0.3
+cliff==3.10.0
+clip @ git+https://github.com/openai/CLIP.git@a1d071733d7111c9c014f024669f959182114e33
+cloudevents==1.2.0
+cloudpathlib==0.16.0
+cloudpickle==2.0.0
+cmaes==0.8.2
+cmake==3.27.7
+cmd2==2.3.3
+colorama==0.4.3
+colorful==0.5.4
+colorlog==6.6.0
+conda==4.11.0
+conda-build==3.21.7
+conda-package-handling @ file:///home/conda/feedstock_root/build_artifacts/conda-package-handling_1636021712360/work
+confection==0.1.3
+configparser==5.2.0
+cryptography @ file:///tmp/build/80754af9/cryptography_1639414570729/work
+cycler==0.11.0
+cymem==2.0.8
+Cython==0.29.26
+dask==2022.1.0
+databricks-cli==0.16.2
+datasets==2.13.2
+DAWG-Python==0.7.2
+debugpy==1.5.1
+decorator==5.1.1
+deep-translator==1.11.4
+defusedxml==0.7.1
+Deprecated==1.2.13
+deprecation==2.1.0
+diffusers==0.21.4
+dill==0.3.6
+distributed==2022.1.0
+docker==5.0.3
+docker-pycreds==0.4.0
+docopt==0.6.2
+docutils==0.15.2
+einops==0.6.1
+entrypoints==0.3
+fairscale==0.4.6
+fairseq==0.12.2
+fastapi==0.72.0
+fastBPE==0.1.0
+fasttext==0.9.2
+fasttext-langdetect==1.0.5
+ffmpy==0.3.1
+filelock @ file:///home/conda/feedstock_root/build_artifacts/filelock_1641470428964/work
+Flask==2.0.2
+fonttools==4.28.5
+fpie==0.2.4
+frozenlist==1.3.0
+fsspec==2023.1.0
+ftfy==6.1.1
+future==0.18.2
+gitdb==4.0.9
+GitPython==3.1.26
+glob2==0.7
+google-api-core==2.4.0
+google-auth==1.35.0
+google-auth-oauthlib==0.4.6
+google-cloud-core==2.2.2
+google-cloud-language==2.3.1
+google-cloud-storage==2.0.0
+google-crc32c==1.3.0
+google-resumable-media==2.1.0
+googleapis-common-protos==1.54.0
+gorilla==0.4.0
+gpustat==0.6.0
+GPUtil==1.4.0
+gradio==3.34.0
+gradio_client==0.2.6
+grpcio==1.43.0
+grpcio-status==1.43.0
+gunicorn==20.1.0
+h11==0.14.0
+h5py==3.6.0
+HeapDict==1.0.1
+hiredis==2.0.0
+horovod==0.28.1
+httpcore==0.17.3
+httpx==0.24.1
+huggingface-hub==0.16.4
+hydra-core==1.3.2
+idna==3.3
+image-reward @ file:///home/jovyan/afilatov/Diffusion/imagen/notebooks/image_assessment/ImageReward
+imageio==2.31.2
+importlib-metadata==4.10.1
+importlib-resources==5.4.0
+inflect==5.3.0
+intervaltree==3.1.0
+ipykernel==6.7.0
+ipymarkup==0.9.0
+ipyplot==1.1.1
+ipython==7.31.1
+ipython-genutils==0.2.0
+ipywidgets==7.6.5
+itsdangerous==2.0.1
+jedi==0.18.1
+Jinja2 @ file:///home/conda/feedstock_root/build_artifacts/jinja2_1636510082894/work
+jmespath==0.10.0
+joblib==1.1.0
+json5==0.9.6
+jsonschema==4.4.0
+jupyter-archive==3.2.1
+jupyter-client==7.1.2
+jupyter-core==4.9.1
+jupyter-server==1.13.4
+jupyter-server-proxy==1.3.2
+jupyter-tensorboard @ file:///tmp/mlspace/packages/jupyter_tensorboard-0.2.2a0-py2.py3-none-any.whl
+jupyterlab==3.3.0a2
+jupyterlab-nvdashboard==0.6.0
+jupyterlab-pygments==0.1.2
+jupyterlab-server==2.10.3
+jupyterlab-tensorboard @ git+https://github.com/rhangelxs/jupyterlab_tensorboard.git@8dc7b1d5f24ece0e76e61b4dbbf36c58b84cbddd
+jupyterlab-widgets==1.0.2
+kfserving==0.6.1
+kiwisolver==1.3.2
+kubernetes==21.7.0
+langcodes==3.3.0
+langid==1.1.6
+libarchive-c @ file:///home/conda/feedstock_root/build_artifacts/python-libarchive-c_1643045751069/work
+libmambapy @ file:///home/conda/feedstock_root/build_artifacts/mamba-split_1643117251182/work/libmambapy
+librosa==0.8.1
+linkify-it-py==2.0.2
+llvmlite==0.38.0
+lmdb==1.4.1
+locket==0.2.1
+Mako==1.1.6
+mamba @ file:///home/conda/feedstock_root/build_artifacts/mamba-split_1643117251182/work/mamba
+Markdown==3.3.6
+markdown-it-py==2.2.0
+MarkupSafe @ file:///home/conda/feedstock_root/build_artifacts/markupsafe_1635833550185/work
+matplotlib==3.5.1
+matplotlib-inline==0.1.3
+mdit-py-plugins==0.3.3
+mdurl==0.1.2
+minio==6.0.2
+mistune==0.8.4
+mlflow @ file:///tmp/mlspace/packages/mlflow-1.7.2-py3-none-any.whl
+mmcv-full==1.4.3
+modin==0.12.1
+mpi4py==3.1.3
+msgpack==1.0.3
+multidict==5.2.0
+multiprocess==0.70.14
+murmurhash==1.0.10
+natasha==1.6.0
+navec==0.10.0
+nbclassic==0.3.5
+nbclient==0.5.10
+nbconvert==6.4.1
+nbformat==5.1.3
+nest-asyncio==1.5.4
+networkx==2.6.3
+nltk==3.6.7
+notebook @ file:///tmp/mlspace/packages/notebook-6.1.4-py3-none-any.whl
+npm==0.1.1
+numba==0.55.0
+numpy==1.21.5
+nvidia-ml-py3==7.352.0
+oauthlib==3.1.1
+omegaconf==2.3.0
+opencensus==0.8.0
+opencensus-context==0.1.2
+opencv-python==4.5.5.62
+opencv-python-headless==4.8.1.78
+optional-django==0.1.0
+optuna==2.10.0
+orjson==3.9.7
+packaging==21.3
+pandas==1.3.5
+pandocfilters==1.5.0
+parso==0.8.3
+partd==1.2.0
+pbr==5.8.0
+pexpect==4.8.0
+pickleshare==0.7.5
+Pillow==9.0.0
+pkginfo @ file:///home/conda/feedstock_root/build_artifacts/pkginfo_1638813452194/work
+pooch==1.5.2
+portalocker==2.3.2
+preshed==3.0.9
+prettytable==3.0.0
+prometheus-client==0.13.0
+prometheus-flask-exporter==0.18.7
+prompt-toolkit==3.0.26
+proto-plus==1.19.8
+protobuf==3.19.3
+psutil @ file:///home/conda/feedstock_root/build_artifacts/psutil_1640887121529/work
+ptyprocess==0.7.0
+py-spy==0.3.11
+pyarrow==12.0.1
+pyasn1==0.4.8
+pyasn1-modules==0.2.8
+pybind11==2.11.1
+pycosat @ file:///home/conda/feedstock_root/build_artifacts/pycosat_1636020357254/work
+pycparser==2.21
+pydantic==1.9.0
+pyDeprecate==0.3.2
+pydub==0.25.1
+pyee==8.2.2
+Pygments==2.16.1
+PyJWT==2.3.0
+pymorphy2==0.9.1
+pymorphy2-dicts-ru==2.4.417127.4579844
+pynvml==11.4.1
+pyOpenSSL @ file:///home/conda/feedstock_root/build_artifacts/pyopenssl_1633192417276/work
+pyparsing==3.0.7
+pyperclip==1.8.2
+pyppeteer==0.2.6
+pyrsistent==0.18.1
+PySocks @ file:///home/conda/feedstock_root/build_artifacts/pysocks_1635862409558/work
+python-dateutil==2.8.2
+python-editor==1.0.4
+python-multipart==0.0.6
+python-speech-features==0.6
+pytorch-lightning==1.7.5
+pytz @ file:///home/conda/feedstock_root/build_artifacts/pytz_1633452062248/work
+PyWavelets==1.3.0
+PyYAML @ file:///home/conda/feedstock_root/build_artifacts/pyyaml_1636139801027/work
+pyzmq==22.3.0
+qudida==0.0.4
+querystring-parser==1.2.4
+ray==1.6.0
+razdel==0.5.0
+redis==4.1.1
+regex==2022.1.18
+requests==2.27.1
+requests-oauthlib==1.3.0
+resampy==0.2.2
+rich==13.6.0
+rsa==4.8
+ruamel-yaml-conda @ file:///home/conda/feedstock_root/build_artifacts/ruamel_yaml_1636009153751/work
+s3fs==2023.1.0
+s3transfer==0.6.2
+sacrebleu==2.0.0
+sacremoses==0.0.53
+safetensors==0.4.0
+scikit-image==0.19.3
+scikit-learn==1.0.1
+scipy==1.7.3
+semantic-version==2.10.0
+Send2Trash==1.8.0
+sentence-transformers==2.2.2
+sentencepiece==0.1.96
+sentry-sdk==1.34.0
+setproctitle==1.3.3
+shortuuid==1.0.11
+simpervisor==0.4
+simplejson==3.17.6
+six==1.16.0
+slovnet==0.6.0
+smart-open==6.4.0
+smmap==5.0.0
+sniffio==1.2.0
+sortedcontainers==2.4.0
+SoundFile==0.10.3.post1
+soupsieve @ file:///home/conda/feedstock_root/build_artifacts/soupsieve_1638550740809/work
+sox==1.4.1
+spacy==3.7.2
+spacy-legacy==3.0.12
+spacy-loggers==1.0.5
+SQLAlchemy==1.3.13
+sqlparse==0.4.2
+srsly==2.4.8
+starlette==0.17.1
+stevedore==3.5.0
+table-logger==0.3.6
+tabulate==0.8.9
+taichi==1.6.0
+tblib==1.7.0
+tensorboard==2.11.2
+tensorboard-data-server==0.6.1
+tensorboard-plugin-wit==1.8.1
+termcolor==2.3.0
+terminado==0.13.1
+testpath==0.5.0
+thinc==8.2.1
+threadpoolctl==3.0.0
+tifffile==2021.11.2
+timm==0.9.7
+tokenizers==0.13.3
+toolz==0.11.2
+torch==1.10.1+cu111
+torchaudio==0.10.1+rocm4.1
+torchmetrics==0.11.4
+torchvision==0.11.2+cu111
+tornado==6.1
+tqdm @ file:///home/conda/feedstock_root/build_artifacts/tqdm_1632160078689/work
+traitlets==5.1.1
+transformers==4.30.2
+typer==0.4.0
+typing_extensions==4.0.1
+uc-micro-py==1.0.2
+urllib3==1.26.18
+uvicorn==0.17.0
+wandb==0.16.0
+wasabi==1.1.2
+wcwidth==0.2.5
+weasel==0.3.4
+webdataset==0.2.74
+webencodings==0.5.1
+websocket-client==1.2.3
+websockets==11.0.3
+Werkzeug==2.0.2
+widgetsnbextension==3.5.2
+wrapt==1.13.3
+xgboost==1.5.2
+xxhash==3.4.1
+yapf==0.32.0
+yargy==0.16.0
+yarl==1.7.2
+zict==2.0.0
+zipp==3.7.0

Kandinsky-3/kandinsky3/__init__.py

@@ -0,0 +1,267 @@
+import os
+from typing import Optional, Union
+
+import torch
+from huggingface_hub import hf_hub_download, snapshot_download
+
+from kandinsky3.model.unet import UNet
+from kandinsky3.movq import MoVQ
+from kandinsky3.condition_encoders import T5TextConditionEncoder
+from kandinsky3.condition_processors import T5TextConditionProcessor
+from kandinsky3.model.diffusion import BaseDiffusion, get_named_beta_schedule
+
+from .t2i_pipeline import Kandinsky3T2IPipeline
+from .inpainting_pipeline import Kandinsky3InpaintingPipeline
+
+
+def get_T2I_unet(
+        device: Union[str, torch.device],
+        weights_path: Optional[str] = None,
+        dtype: Union[str, torch.dtype] = torch.float32,
+) -> (UNet, Optional[torch.Tensor], Optional[dict]):
+    unet = UNet(
+        model_channels=384,
+        num_channels=4,
+        init_channels=192,
+        time_embed_dim=1536,
+        context_dim=4096,
+        groups=32,
+        head_dim=64,
+        expansion_ratio=4,
+        compression_ratio=2,
+        dim_mult=(1, 2, 4, 8),
+        num_blocks=(3, 3, 3, 3),
+        add_cross_attention=(False, True, True, True),
+        add_self_attention=(False, True, True, True),
+    )
+
+    null_embedding = None
+    if weights_path:
+        state_dict = torch.load(weights_path, map_location=torch.device('cpu'))
+        null_embedding = state_dict['null_embedding']
+        unet.load_state_dict(state_dict['unet'])
+
+    unet.to(device=device, dtype=dtype).eval()
+    return unet, null_embedding
+
+
+def get_T5encoder(
+        device: Union[str, torch.device],
+        weights_path: str,
+        projection_name: str,
+        dtype: Union[str, torch.dtype] = torch.float32,
+        low_cpu_mem_usage: bool = True,
+        load_in_8bit: bool = False,
+        load_in_4bit: bool = False,
+) -> (T5TextConditionProcessor, T5TextConditionEncoder):
+    tokens_length = 128
+    context_dim = 4096
+    processor = T5TextConditionProcessor(tokens_length, weights_path)
+    condition_encoder = T5TextConditionEncoder(
+        weights_path, context_dim, low_cpu_mem_usage=low_cpu_mem_usage, device=device,
+        dtype=dtype, load_in_8bit=load_in_8bit, load_in_4bit=load_in_4bit
+    )
+
+    if weights_path:
+        projections_weights_path = os.path.join(weights_path, projection_name)
+        state_dict = torch.load(projections_weights_path, map_location=torch.device('cpu'))
+        condition_encoder.projection.load_state_dict(state_dict)
+
+    condition_encoder.projection.to(device=device, dtype=dtype).eval()
+    return processor, condition_encoder
+
+
+def get_movq(
+        device: Union[str, torch.device],
+        weights_path: Optional[str] = None,
+        dtype: Union[str, torch.dtype] = torch.float32,
+) -> MoVQ:
+    generator_config = {
+        'double_z': False,
+        'z_channels': 4,
+        'resolution': 256,
+        'in_channels': 3,
+        'out_ch': 3,
+        'ch': 256,
+        'ch_mult': [1, 2, 2, 4],
+        'num_res_blocks': 2,
+        'attn_resolutions': [32],
+        'dropout': 0.0
+    }
+    movq = MoVQ(generator_config)
+
+    if weights_path:
+        state_dict = torch.load(weights_path, map_location=torch.device('cpu'))
+        movq.load_state_dict(state_dict)
+
+    movq.to(device=device, dtype=dtype).eval()
+    return movq
+
+
+def get_inpainting_unet(
+        device: Union[str, torch.device],
+        weights_path: Optional[str] = None,
+        dtype: Union[str, torch.dtype] = torch.float32,
+) -> (UNet, Optional[torch.Tensor], Optional[dict]):
+    unet = UNet(
+        model_channels=384,
+        num_channels=9,
+        init_channels=192,
+        time_embed_dim=1536,
+        context_dim=4096,
+        groups=32,
+        head_dim=64,
+        expansion_ratio=4,
+        compression_ratio=2,
+        dim_mult=(1, 2, 4, 8),
+        num_blocks=(3, 3, 3, 3),
+        add_cross_attention=(False, True, True, True),
+        add_self_attention=(False, True, True, True),
+    )
+
+    null_embedding = None
+    if weights_path:
+        state_dict = torch.load(weights_path, map_location=torch.device('cpu'))
+        null_embedding = state_dict['null_embedding']
+        unet.load_state_dict(state_dict['unet'])
+
+    unet.to(device=device, dtype=dtype).eval()
+    return unet, null_embedding
+
+
+def get_T2I_pipeline(
+        device_map: Union[str, torch.device, dict],
+        dtype_map: Union[str, torch.dtype, dict] = torch.float32,
+        low_cpu_mem_usage: bool = True,
+        load_in_8bit: bool = False,
+        load_in_4bit: bool = False,
+        cache_dir: str = '/tmp/kandinsky3/',
+        unet_path: str = None,
+        text_encoder_path: str = None,
+        movq_path: str = None,
+) -> Kandinsky3T2IPipeline:
+    # assert ((unet_path is not None) or (text_encoder_path is not None) or (movq_path is not None))
+    if not isinstance(device_map, dict):
+        device_map = {
+            'unet': device_map, 'text_encoder': device_map, 'movq': device_map
+        }
+    if not isinstance(dtype_map, dict):
+        dtype_map = {
+            'unet': dtype_map, 'text_encoder': dtype_map, 'movq': dtype_map
+        }
+
+    if unet_path is None:
+        unet_path = hf_hub_download(
+            repo_id="ai-forever/Kandinsky3.1", filename='weights/kandinsky3.pt', cache_dir=cache_dir
+        )
+    if text_encoder_path is None:
+        text_encoder_path = snapshot_download(
+            repo_id="ai-forever/Kandinsky3.1", allow_patterns="weights/flan_ul2_encoder/*", cache_dir=cache_dir
+        )
+        text_encoder_path = os.path.join(text_encoder_path, 'weights/flan_ul2_encoder')
+    if movq_path is None:
+        movq_path = hf_hub_download(
+            repo_id="ai-forever/Kandinsky3.1", filename='weights/movq.pt', cache_dir=cache_dir
+        )
+
+    unet, null_embedding = get_T2I_unet(device_map['unet'], unet_path, dtype=dtype_map['unet'])
+    processor, condition_encoder = get_T5encoder(
+        device_map['text_encoder'], text_encoder_path, 'projection.pt', dtype=dtype_map['text_encoder'],
+        low_cpu_mem_usage=low_cpu_mem_usage, load_in_8bit=load_in_8bit, load_in_4bit=load_in_4bit
+    )
+    movq = get_movq(device_map['movq'], movq_path, dtype=dtype_map['movq'])
+    return Kandinsky3T2IPipeline(
+        device_map, dtype_map, unet, null_embedding, processor, condition_encoder, movq, False
+    )
+
+
+def get_T2I_Flash_pipeline(
+        device_map: Union[str, torch.device, dict],
+        dtype_map: Union[str, torch.dtype, dict] = torch.float32,
+        low_cpu_mem_usage: bool = True,
+        load_in_8bit: bool = False,
+        load_in_4bit: bool = False,
+        cache_dir: str = '/tmp/kandinsky3/',
+        unet_path: str = None,
+        text_encoder_path: str = None,
+        movq_path: str = None,
+) -> Kandinsky3T2IPipeline:
+    # assert ((unet_path is not None) or (text_encoder_path is not None) or (movq_path is not None))
+    if not isinstance(device_map, dict):
+        device_map = {
+            'unet': device_map, 'text_encoder': device_map, 'movq': device_map
+        }
+    if not isinstance(dtype_map, dict):
+        dtype_map = {
+            'unet': dtype_map, 'text_encoder': dtype_map, 'movq': dtype_map
+        }
+
+    if unet_path is None:
+        unet_path = hf_hub_download(
+            repo_id="ai-forever/Kandinsky3.1", filename='weights/kandinsky3_flash.pt', cache_dir=cache_dir
+        )
+    if text_encoder_path is None:
+        text_encoder_path = snapshot_download(
+            repo_id="ai-forever/Kandinsky3.1", allow_patterns="weights/flan_ul2_encoder/*", cache_dir=cache_dir
+        )
+        text_encoder_path = os.path.join(text_encoder_path, 'weights/flan_ul2_encoder')
+    if movq_path is None:
+        movq_path = hf_hub_download(
+            repo_id="ai-forever/Kandinsky3.1", filename='weights/movq.pt', cache_dir=cache_dir
+        )
+
+    unet, null_embedding = get_T2I_unet(device_map['unet'], unet_path, dtype=dtype_map['unet'])
+    processor, condition_encoder = get_T5encoder(
+        device_map['text_encoder'], text_encoder_path, 'projection_flash.pt', dtype=dtype_map['text_encoder'],
+        low_cpu_mem_usage=low_cpu_mem_usage, load_in_8bit=load_in_8bit, load_in_4bit=load_in_4bit
+    )
+    movq = get_movq(device_map['movq'], movq_path, dtype=dtype_map['movq'])
+    return Kandinsky3T2IPipeline(
+        device_map, dtype_map, unet, null_embedding, processor, condition_encoder, movq, True
+    )
+
+
+def get_inpainting_pipeline(
+        device_map: Union[str, torch.device, dict],
+        dtype_map: Union[str, torch.dtype, dict] = torch.float32,
+        low_cpu_mem_usage: bool = True,
+        load_in_8bit: bool = False,
+        load_in_4bit: bool = False,
+        cache_dir: str = '/tmp/kandinsky3/',
+        unet_path: str = None,
+        text_encoder_path: str = None,
+        movq_path: str = None,
+) -> Kandinsky3InpaintingPipeline:
+    # assert ((unet_path is not None) or (text_encoder_path is not None) or (movq_path is not None))
+    if not isinstance(device_map, dict):
+        device_map = {
+            'unet': device_map, 'text_encoder': device_map, 'movq': device_map
+        }
+    if not isinstance(dtype_map, dict):
+        dtype_map = {
+            'unet': dtype_map, 'text_encoder': dtype_map, 'movq': dtype_map
+        }
+
+    if unet_path is None:
+        unet_path = hf_hub_download(
+            repo_id="ai-forever/Kandinsky3.1", filename='weights/kandinsky3_inpainting.pt', cache_dir=cache_dir
+        )
+    if text_encoder_path is None:
+        text_encoder_path = snapshot_download(
+            repo_id="ai-forever/Kandinsky3.1", allow_patterns="weights/flan_ul2_encoder/*", cache_dir=cache_dir
+        )
+        text_encoder_path = os.path.join(text_encoder_path, 'weights/flan_ul2_encoder')
+    if movq_path is None:
+        movq_path = hf_hub_download(
+            repo_id="ai-forever/Kandinsky3.1", filename='weights/movq.pt', cache_dir=cache_dir
+        )
+
+    unet, null_embedding = get_inpainting_unet(device_map['unet'], unet_path, dtype=dtype_map['unet'])
+    processor, condition_encoder = get_T5encoder(
+        device_map['text_encoder'], text_encoder_path, 'projection_inpainting.pt', dtype=dtype_map['text_encoder'],
+        low_cpu_mem_usage=low_cpu_mem_usage, load_in_8bit=load_in_8bit, load_in_4bit=load_in_4bit
+    )
+    movq = get_movq(device_map['movq'], movq_path, dtype=dtype_map['movq'])
+    return Kandinsky3InpaintingPipeline(
+        device_map, dtype_map, unet, null_embedding, processor, condition_encoder, movq
+    )

Kandinsky-3/kandinsky3/condition_encoders.py

@@ -0,0 +1,40 @@
+import torch
+from torch import nn
+from transformers import T5EncoderModel
+from typing import Optional, Union
+
+
+class T5TextConditionEncoder(nn.Module):
+
+    def __init__(
+            self, model_path, context_dim,
+            low_cpu_mem_usage: bool = True, device: Optional[str] = None,
+            dtype: Union[str, torch.dtype] = torch.float32, load_in_4bit: bool = False, load_in_8bit: bool = False
+    ):
+        super().__init__()
+        self.encoder = T5EncoderModel.from_pretrained(
+            model_path, low_cpu_mem_usage=low_cpu_mem_usage, device_map=device,
+            torch_dtype=dtype, load_in_8bit=load_in_8bit, load_in_4bit=load_in_4bit,
+        ).encoder
+        self.projection = nn.Sequential(
+            nn.Linear(self.encoder.config.d_model, context_dim, bias=False),
+            nn.LayerNorm(context_dim)
+        )
+
+    def forward(self, model_input):
+        embeddings = self.encoder(**model_input).last_hidden_state
+        context = self.projection(embeddings)
+        if 'attention_mask' in model_input:
+            context_mask = model_input['attention_mask']
+            context[context_mask == 0] = torch.zeros_like(context[context_mask == 0])
+            max_seq_length = context_mask.sum(-1).max() + 1
+            context = context[:, :max_seq_length]
+            context_mask = context_mask[:, :max_seq_length]
+        else:
+            context_mask = torch.ones(*embeddings.shape[:-1], dtype=torch.long, device=embeddings.device)
+        return context, context_mask
+
+
+def get_condition_encoder(conf):
+    return T5TextConditionEncoder(**conf)
+

Kandinsky-3/kandinsky3/condition_processors.py

@@ -0,0 +1,34 @@
+import torch
+from transformers import T5Tokenizer
+
+
+class T5TextConditionProcessor:
+
+    def __init__(self, tokens_length, processor_path):
+        self.tokens_length = tokens_length
+        self.processor = T5Tokenizer.from_pretrained(processor_path)
+
+    def encode(self, text=None, negative_text=None):
+        encoded = self.processor(text, max_length=self.tokens_length, truncation=True)
+        pad_length = self.tokens_length - len(encoded['input_ids'])
+        input_ids = encoded['input_ids'] + [self.processor.pad_token_id] * pad_length
+        attention_mask = encoded['attention_mask'] + [0] * pad_length
+        condition_model_input = {
+            'input_ids': torch.tensor(input_ids, dtype=torch.long),
+            'attention_mask': torch.tensor(attention_mask, dtype=torch.long)
+        }
+
+        if negative_text is not None:
+            negative_encoded = self.processor(negative_text, max_length=self.tokens_length, truncation=True)
+            negative_input_ids = negative_encoded['input_ids'][:len(encoded['input_ids'])]
+            negative_input_ids[-1] = self.processor.eos_token_id
+            negative_pad_length = self.tokens_length - len(negative_input_ids)
+            negative_input_ids = negative_input_ids + [self.processor.pad_token_id] * negative_pad_length
+            negative_attention_mask = encoded['attention_mask'] + [0] * pad_length
+            negative_condition_model_input = {
+                'input_ids': torch.tensor(negative_input_ids, dtype=torch.long),
+                'attention_mask': torch.tensor(negative_attention_mask, dtype=torch.long)
+            }
+        else:
+            negative_condition_model_input = None
+        return condition_model_input, negative_condition_model_input

Kandinsky-3/kandinsky3/inpainting_pipeline.py

@@ -0,0 +1,168 @@
+from typing import Union, List
+import PIL
+import numpy as np
+
+import torch
+import torchvision.transforms as T
+from einops import repeat
+
+from kandinsky3.model.unet import UNet
+from kandinsky3.movq import MoVQ
+from kandinsky3.condition_encoders import T5TextConditionEncoder
+from kandinsky3.condition_processors import T5TextConditionProcessor
+from kandinsky3.model.diffusion import BaseDiffusion, get_named_beta_schedule
+from kandinsky3.utils import resize_image_for_diffusion, resize_mask_for_diffusion
+
+
+class Kandinsky3InpaintingPipeline:
+
+    def __init__(
+            self,
+            device_map: Union[str, torch.device, dict],
+            dtype_map: Union[str, torch.dtype, dict],
+            unet: UNet,
+            null_embedding: torch.Tensor,
+            t5_processor: T5TextConditionProcessor,
+            t5_encoder: T5TextConditionEncoder,
+            movq: MoVQ,
+    ):
+        self.device_map = device_map
+        self.dtype_map = dtype_map
+        self.to_pil = T.ToPILImage()
+        self.to_tensor = T.ToTensor()
+
+        self.unet = unet
+        self.null_embedding = null_embedding
+        self.t5_processor = t5_processor
+        self.t5_encoder = t5_encoder
+        self.movq = movq
+
+    def shared_step(self, batch: dict) -> dict:
+        image = batch['image']
+        condition_model_input = batch['text']
+        negative_condition_model_input = batch['negative_text']
+
+        bs = image.shape[0]
+
+        masked_latent = None
+        mask = batch['mask']
+
+        if 'masked_image' in batch:
+            masked_latent = batch['masked_image']
+        elif self.unet.in_layer.in_channels == 9:
+            masked_latent = image.masked_fill((1 - mask).bool(), 0)
+        else:
+            raise ValueError()
+
+        with torch.cuda.amp.autocast(dtype=self.dtype_map['movq']):
+            masked_latent = self.movq.encode(masked_latent)
+        mask = torch.nn.functional.interpolate(mask, size=(masked_latent.shape[2], masked_latent.shape[3]))
+
+        with torch.cuda.amp.autocast(dtype=self.dtype_map['text_encoder']):
+            context, context_mask = self.t5_encoder(condition_model_input)
+
+        if negative_condition_model_input is not None:
+            negative_context, negative_context_mask = self.t5_encoder(negative_condition_model_input)
+        else:
+            negative_context, negative_context_mask = None, None
+
+        return {
+            'context': context,
+            'context_mask': context_mask,
+            'negative_context': negative_context,
+            'negative_context_mask': negative_context_mask,
+            'image': image,
+            'masked_latent': masked_latent,
+            'mask': mask
+        }
+
+    def prepare_batch(
+            self,
+            text: str,
+            negative_text: str,
+            image: PIL.Image.Image,
+            mask: np.ndarray,
+    ) -> dict:
+        condition_model_input, negative_condition_model_input = self.t5_processor.encode(
+            text=text, negative_text=negative_text
+        )
+        batch = {
+            'image': self.to_tensor(resize_image_for_diffusion(image.convert("RGB"))) * 2 - 1,
+            'mask': 1 - self.to_tensor(resize_mask_for_diffusion(mask)),
+            'text': condition_model_input,
+            'negative_text': negative_condition_model_input
+        }
+        batch['mask'] = batch['mask'].type(self.dtype_map['movq'])
+
+        batch['image'] = batch['image'].unsqueeze(0).to(self.device_map['movq'])
+        batch['text']['input_ids'] = batch['text']['input_ids'].unsqueeze(0).to(self.device_map['text_encoder'])
+        batch['text']['attention_mask'] = batch['text']['attention_mask'].unsqueeze(0).to(
+            self.device_map['text_encoder'])
+        batch['mask'] = batch['mask'].unsqueeze(0).to(self.device_map['movq'])
+
+        if negative_condition_model_input is not None:
+            batch['negative_text']['input_ids'] = batch['negative_text']['input_ids'].to(
+                self.device_map['text_encoder'])
+            batch['negative_text']['attention_mask'] = batch['negative_text']['attention_mask'].to(
+                self.device_map['text_encoder'])
+
+        return batch
+
+    def __call__(
+            self,
+            text: str,
+            image: PIL.Image.Image,
+            mask: np.ndarray,
+            negative_text: str = None,
+            images_num: int = 1,
+            bs: int = 1,
+            steps: int = 50,
+            guidance_weight_text: float = 4,
+            eta=1.0
+    ) -> List[PIL.Image.Image]:
+
+        with torch.no_grad():
+            batch = self.prepare_batch(text, negative_text, image, mask)
+            processed = self.shared_step(batch)
+            betas = get_named_beta_schedule('cosine', 1000)
+            base_diffusion = BaseDiffusion(betas, percentile=0.95)
+            times = list(range(999, 0, -1000 // steps))
+
+            pil_images = []
+            k, m = images_num // bs, images_num % bs
+            for minibatch in [bs] * k + [m]:
+                if minibatch == 0:
+                    continue
+
+                bs_context = repeat(processed['context'], '1 n d -> b n d', b=minibatch)
+                bs_context_mask = repeat(processed['context_mask'], '1 n -> b n', b=minibatch)
+
+                if processed['negative_context'] is not None:
+                    bs_negative_context = repeat(processed['negative_context'], '1 n d -> b n d', b=minibatch)
+                    bs_negative_context_mask = repeat(processed['negative_context_mask'], '1 n -> b n', b=minibatch)
+                else:
+                    bs_negative_context, bs_negative_context_mask = None, None
+
+                mask = processed['mask'].repeat_interleave(minibatch, dim=0)
+                masked_latent = processed['masked_latent'].repeat_interleave(minibatch, dim=0)
+
+                minibatch = masked_latent.shape[0]
+
+                with torch.cuda.amp.autocast(dtype=self.dtype_map['unet']):
+                    with torch.no_grad():
+                        images = base_diffusion.p_sample_loop(
+                            self.unet, (minibatch, 4, masked_latent.shape[2], masked_latent.shape[3]), times,
+                            self.device_map['unet'],
+                            bs_context, bs_context_mask, self.null_embedding, guidance_weight_text, eta,
+                            negative_context=bs_negative_context, negative_context_mask=bs_negative_context_mask,
+                            mask=mask, masked_latent=masked_latent, gan=False
+                        )
+
+                with torch.cuda.amp.autocast(dtype=self.dtype_map['movq']):
+                    images = torch.cat([self.movq.decode(image) for image in images.chunk(2)])
+                    images = torch.clip((images + 1.) / 2., 0., 1.).cpu()
+
+                for images_chunk in images.chunk(1):
+                    pil_images += [self.to_pil(image) for image in images_chunk]
+
+        return pil_images

Kandinsky-3/kandinsky3/model/diffusion.py

@@ -0,0 +1,200 @@
+import math
+
+import torch
+from einops import rearrange
+from tqdm import tqdm
+
+from .utils import get_tensor_items
+
+
+def get_named_beta_schedule(schedule_name, timesteps):
+    if schedule_name == "linear":
+        scale = 1000 / timesteps
+        beta_start = scale * 0.0001
+        beta_end = scale * 0.02
+        return torch.linspace(
+            beta_start, beta_end, timesteps, dtype=torch.float32
+        )
+    elif schedule_name == "cosine":
+        alpha_bar = lambda t: math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+        betas = []
+        for i in range(timesteps):
+            t1 = i / timesteps
+            t2 = (i + 1) / timesteps
+            betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), 0.999))
+        return torch.tensor(betas, dtype=torch.float32)
+
+
+class BaseDiffusion:
+
+    def __init__(self, betas, percentile=None, gen_noise=torch.randn_like):
+        self.betas = betas
+        self.num_timesteps = betas.shape[0]
+
+        alphas = 1. - betas
+        self.alphas_cumprod = torch.cumprod(alphas, dim=0)
+        self.alphas_cumprod_prev = torch.cat([torch.ones(1, dtype=betas.dtype), self.alphas_cumprod[:-1]])
+
+        # calculate q(x_t | x_{t-1})
+        self.sqrt_alphas_cumprod = torch.sqrt(self.alphas_cumprod)
+        self.sqrt_one_minus_alphas_cumprod = torch.sqrt(1. - self.alphas_cumprod)
+
+        # calculate q(x_{t-1} | x_t, x_0)
+        self.posterior_mean_coef_1 = torch.sqrt(self.alphas_cumprod_prev) * betas / (1. - self.alphas_cumprod)
+        self.posterior_mean_coef_2 = torch.sqrt(alphas) * (1. - self.alphas_cumprod_prev) / (1. - self.alphas_cumprod)
+        self.posterior_variance = betas * (1. - self.alphas_cumprod_prev) / (1. - self.alphas_cumprod)
+        self.posterior_log_variance = torch.log(
+            torch.cat([self.posterior_variance[1].unsqueeze(0), self.posterior_variance[1:]])
+        )
+
+        self.percentile = percentile
+        self.time_scale = 1000 // self.num_timesteps
+        self.gen_noise = gen_noise
+        self.jump_length = 3
+
+    def process_x_start(self, x_start):
+        bs, ndims = x_start.shape[0], len(x_start.shape[1:])
+        if self.percentile is not None:
+            quantile = torch.quantile(
+                rearrange(x_start, 'b ... -> b (...)').abs(),
+                self.percentile,
+                dim=-1
+            )
+            quantile = torch.clip(quantile, min=1.)
+            quantile = quantile.reshape(bs, *((1,) * ndims))
+            return torch.clip(x_start, -quantile, quantile) / quantile
+        else:
+            return torch.clip(x_start, -1., 1.)
+
+    def get_x_start(self, x, t, noise):
+        sqrt_one_minus_alphas_cumprod = get_tensor_items(self.sqrt_one_minus_alphas_cumprod, t, noise.shape)
+        sqrt_alphas_cumprod = get_tensor_items(self.sqrt_alphas_cumprod, t, noise.shape)
+        pred_x_start = (x - sqrt_one_minus_alphas_cumprod * noise) / sqrt_alphas_cumprod
+        return pred_x_start
+
+    def get_noise(self, x, t, x_start):
+        sqrt_one_minus_alphas_cumprod = get_tensor_items(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape)
+        sqrt_alphas_cumprod = get_tensor_items(self.sqrt_alphas_cumprod, t, x_start.shape)
+        pred_noise = (x - sqrt_alphas_cumprod * x_start) / sqrt_one_minus_alphas_cumprod
+        return pred_noise
+
+    def q_sample(self, x_start, t, noise=None):
+        if noise is None:
+            noise = self.gen_noise(x_start)
+        sqrt_alphas_cumprod = get_tensor_items(self.sqrt_alphas_cumprod, t, x_start.shape)
+        sqrt_one_minus_alphas_cumprod = get_tensor_items(self.sqrt_one_minus_alphas_cumprod, t, noise.shape)
+        x_t = sqrt_alphas_cumprod * x_start + sqrt_one_minus_alphas_cumprod * noise
+        return x_t
+
+    def q_posterior_mean_variance(self, x_start, x_t, t):
+        posterior_mean_coef_1 = get_tensor_items(self.posterior_mean_coef_1, t, x_start.shape)
+        posterior_mean_coef_2 = get_tensor_items(self.posterior_mean_coef_2, t, x_t.shape)
+        posterior_mean = posterior_mean_coef_1 * x_start + posterior_mean_coef_2 * x_t
+
+        posterior_variance = get_tensor_items(self.posterior_variance, t, x_start.shape)
+        posterior_log_variance = get_tensor_items(self.posterior_log_variance, t, x_start.shape)
+        return posterior_mean, posterior_variance, posterior_log_variance
+
+    def q_posterior_variance(self, t, prev_t, shape, eta=1., ):
+        alphas_cumprod = get_tensor_items(self.alphas_cumprod, t, shape)
+        prev_alphas_cumprod = get_tensor_items(self.alphas_cumprod, prev_t, shape)
+
+        posterior_variance = torch.sqrt(
+            eta * (1. - alphas_cumprod / prev_alphas_cumprod) * (1. - prev_alphas_cumprod) / (1. - alphas_cumprod)
+        )
+        return posterior_variance
+
+    def text_guidance(
+            self, model, x, t, context, context_mask, null_embedding, guidance_weight_text,
+            uncondition_context=None, uncondition_context_mask=None, mask=None, masked_latent=None
+    ):
+        large_x = x.repeat(2, 1, 1, 1)
+        large_t = t.repeat(2).to(x.dtype)
+
+        if uncondition_context is None:
+            uncondition_context = torch.zeros_like(context)
+            uncondition_context_mask = torch.zeros_like(context_mask)
+            uncondition_context[:, 0] = null_embedding
+            uncondition_context_mask[:, 0] = 1
+        large_context = torch.cat([context, uncondition_context])
+        large_context_mask = torch.cat([context_mask, uncondition_context_mask])
+
+        if mask is not None:
+            mask = mask.repeat(2, 1, 1, 1)
+        if masked_latent is not None:
+            masked_latent = masked_latent.repeat(2, 1, 1, 1)
+
+        if model.in_layer.in_channels == 9:
+            large_x = torch.cat([large_x, mask, masked_latent], dim=1)
+
+        pred_large_noise = model(large_x, large_t * self.time_scale, large_context, large_context_mask.bool())
+        pred_noise, uncond_pred_noise = torch.chunk(pred_large_noise, 2)
+        pred_noise = (guidance_weight_text + 1.) * pred_noise - guidance_weight_text * uncond_pred_noise
+        return pred_noise
+
+    def p_mean_variance(
+            self, model, x, t, prev_t, context, context_mask, null_embedding, guidance_weight_text, eta=1.,
+            negative_context=None, negative_context_mask=None, mask=None, masked_latent=None
+    ):
+
+        pred_noise = self.text_guidance(
+            model, x, t, context, context_mask, null_embedding, guidance_weight_text,
+            negative_context, negative_context_mask, mask, masked_latent
+        )
+
+        pred_x_start = self.get_x_start(x, t, pred_noise)
+        pred_x_start = self.process_x_start(pred_x_start)
+        pred_noise = self.get_noise(x, t, pred_x_start)
+        pred_var = self.q_posterior_variance(t, prev_t, x.shape, eta)
+
+        prev_alphas_cumprod = get_tensor_items(self.alphas_cumprod, prev_t, x.shape)
+        pred_mean = torch.sqrt(prev_alphas_cumprod) * pred_x_start
+        pred_mean += torch.sqrt(1. - prev_alphas_cumprod - pred_var ** 2) * pred_noise
+        return pred_mean, pred_var
+
+    @torch.no_grad()
+    def p_sample(
+            self, model, x, t, prev_t, context, context_mask, null_embedding, guidance_weight_text, eta=1.,
+            negative_context=None, negative_context_mask=None, mask=None, masked_latent=None
+    ):
+        bs = x.shape[0]
+        ndims = len(x.shape[1:])
+        pred_mean, pred_var = self.p_mean_variance(
+            model, x, t, prev_t, context, context_mask, null_embedding, guidance_weight_text, eta,
+            negative_context=negative_context, negative_context_mask=negative_context_mask,
+            mask=mask, masked_latent=masked_latent
+        )
+        noise = torch.randn_like(x)
+        mask = (prev_t != 0).reshape(bs, *((1,) * ndims))
+        sample = pred_mean + mask * pred_var * noise
+        return sample
+
+    @torch.no_grad()
+    def p_sample_loop(
+            self, model, shape, times, device, context, context_mask, null_embedding, guidance_weight_text, eta=1.,
+            negative_context=None, negative_context_mask=None, mask=None, masked_latent=None, gan=False,
+    ):
+        img = torch.randn(*shape, device=device)
+        times = times + [0, ]
+        times = list(zip(times[:-1], times[1:]))
+
+        for time, prev_time in tqdm(times):
+            time = torch.tensor([time] * shape[0], device=device)
+            if gan:
+                x_t = self.q_sample(img, time)
+                pred_noise = model(x_t, time.type(x_t.dtype), context, context_mask.bool())
+                img = self.get_x_start(x_t, time, pred_noise)
+            else:
+                prev_time = torch.tensor([prev_time] * shape[0], device=device)
+                img = self.p_sample(
+                    model, img, time, prev_time, context, context_mask, null_embedding, guidance_weight_text, eta,
+                    negative_context=negative_context, negative_context_mask=negative_context_mask,
+                    mask=mask, masked_latent=masked_latent
+                )
+        return img
+
+
+def get_diffusion(conf):
+    betas = get_named_beta_schedule(**conf.schedule_params)
+    base_diffusion = BaseDiffusion(betas, **conf.diffusion_params)
+    return base_diffusion

Kandinsky-3/kandinsky3/model/nn.py

@@ -0,0 +1,84 @@
+import math
+
+import torch
+from torch import nn, einsum
+from einops import rearrange, repeat
+
+from .utils import exist
+
+
+class Identity(nn.Module):
+    def __init__(self, *args, **kwargs):
+        super().__init__()
+
+    @staticmethod
+    def forward(x, *args, **kwargs):
+        return x
+
+
+class SinusoidalPosEmb(nn.Module):
+
+    def __init__(self, dim):
+        super().__init__()
+        self.dim = dim
+
+    def forward(self, x):
+        half_dim = self.dim // 2
+        emb = math.log(10000) / (half_dim - 1)
+        emb = torch.exp(torch.arange(half_dim, device=x.device, dtype=x.dtype) * -emb)
+        emb = rearrange(x, 'i -> i 1') * rearrange(emb, 'j -> 1 j')
+        return torch.cat((emb.sin(), emb.cos()), dim=-1)
+
+
+class ConditionalGroupNorm(nn.Module):
+
+    def __init__(self, groups, normalized_shape, context_dim):
+        super().__init__()
+        self.norm = nn.GroupNorm(groups, normalized_shape, affine=False)
+        self.context_mlp = nn.Sequential(
+            nn.SiLU(),
+            nn.Linear(context_dim, 2 * normalized_shape)
+        )
+        self.context_mlp[1].weight.data.zero_()
+        self.context_mlp[1].bias.data.zero_()
+
+    def forward(self, x, context):
+        context = self.context_mlp(context)
+        ndims = ' 1' * len(x.shape[2:])
+        context = rearrange(context, f'b c -> b c{ndims}')
+
+        scale, shift = context.chunk(2, dim=1)
+        x = self.norm(x) * (scale + 1.) + shift
+        return x
+
+
+class Attention(nn.Module):
+
+    def __init__(self, in_channels, out_channels, context_dim, head_dim=64):
+        super().__init__()
+        assert out_channels % head_dim == 0
+        self.num_heads = out_channels // head_dim
+        self.scale = head_dim ** -0.5
+
+        self.to_query = nn.Linear(in_channels, out_channels, bias=False)
+        self.to_key = nn.Linear(context_dim, out_channels, bias=False)
+        self.to_value = nn.Linear(context_dim, out_channels, bias=False)
+
+        self.output_layer = nn.Linear(out_channels, out_channels, bias=False)
+
+    def forward(self, x, context, context_mask=None):
+        query = rearrange(self.to_query(x), 'b n (h d) -> b h n d', h=self.num_heads)
+        key = rearrange(self.to_key(context), 'b n (h d) -> b h n d', h=self.num_heads)
+        value = rearrange(self.to_value(context), 'b n (h d) -> b h n d', h=self.num_heads)
+
+        attention_matrix = einsum('b h i d, b h j d -> b h i j', query, key) * self.scale
+        if exist(context_mask):
+            max_neg_value = -torch.finfo(attention_matrix.dtype).max
+            context_mask = rearrange(context_mask, 'b j -> b 1 1 j')
+            attention_matrix = attention_matrix.masked_fill(~context_mask, max_neg_value)
+        attention_matrix = attention_matrix.softmax(dim=-1)
+
+        out = einsum('b h i j, b h j d -> b h i d', attention_matrix, value)
+        out = rearrange(out, 'b h n d -> b n (h d)')
+        out = self.output_layer(out)
+        return out

Kandinsky-3/kandinsky3/model/unet.py

@@ -0,0 +1,516 @@
+import torch
+from torch import nn, einsum
+from einops import rearrange
+
+from .nn import Identity, Attention, SinusoidalPosEmb, ConditionalGroupNorm
+from .utils import exist, set_default_item, set_default_layer
+import torch.nn.functional as F
+
+
+class Block(nn.Module):
+
+    def __init__(self, in_channels, out_channels, time_embed_dim, kernel_size=3, norm_groups=32, up_resolution=None):
+        super().__init__()
+        self.group_norm = ConditionalGroupNorm(norm_groups, in_channels, time_embed_dim)
+        self.activation = nn.SiLU()
+        self.up_sample = set_default_layer(
+            exist(up_resolution) and up_resolution,
+            nn.ConvTranspose2d, (in_channels, in_channels), {'kernel_size': 2, 'stride': 2}
+        )
+        padding = set_default_item(kernel_size == 1, 0, 1)
+        self.projection = nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, padding=padding)
+        self.down_sample = set_default_layer(
+            exist(up_resolution) and not up_resolution,
+            nn.Conv2d, (out_channels, out_channels), {'kernel_size': 2, 'stride': 2}
+        )
+
+    def forward(self, x, time_embed):
+        x = self.group_norm(x, time_embed)
+        x = self.activation(x)
+        x = self.up_sample(x)
+        x = self.projection(x)
+        x = self.down_sample(x)
+        return x
+
+
+class ResNetBlock(nn.Module):
+
+    def __init__(
+            self, in_channels, out_channels, time_embed_dim, norm_groups=32, compression_ratio=2, up_resolutions=4*[None]
+    ):
+        super().__init__()
+        kernel_sizes = [1, 3, 3, 1]
+        hidden_channel = max(in_channels, out_channels) // compression_ratio
+        hidden_channels = [(in_channels, hidden_channel)] + [(hidden_channel, hidden_channel)] * 2 + [(hidden_channel, out_channels)]
+        self.resnet_blocks = nn.ModuleList([
+            Block(in_channel, out_channel, time_embed_dim, kernel_size, norm_groups, up_resolution)
+            for (in_channel, out_channel), kernel_size, up_resolution in zip(hidden_channels, kernel_sizes, up_resolutions)
+        ])
+
+        self.shortcut_up_sample = set_default_layer(
+            True in up_resolutions,
+            nn.ConvTranspose2d, (in_channels, in_channels), {'kernel_size': 2, 'stride': 2}
+        )
+        self.shortcut_projection = set_default_layer(
+            in_channels != out_channels,
+            nn.Conv2d, (in_channels, out_channels), {'kernel_size': 1}
+        )
+        self.shortcut_down_sample = set_default_layer(
+            False in up_resolutions,
+            nn.Conv2d, (out_channels, out_channels), {'kernel_size': 2, 'stride': 2}
+        )
+
+    def forward(self, x, time_embed):
+        out = x
+        for resnet_block in self.resnet_blocks:
+            out = resnet_block(out, time_embed)
+
+        x = self.shortcut_up_sample(x)
+        x = self.shortcut_projection(x)
+        x = self.shortcut_down_sample(x)
+        x = x + out
+        return x
+
+
+class AttentionPolling(nn.Module):
+
+    def __init__(self, num_channels, context_dim, head_dim=64):
+        super().__init__()
+        self.attention = Attention(context_dim, num_channels, context_dim, head_dim)
+
+    def forward(self, x, context, context_mask=None):
+        context = self.attention(context.mean(dim=1, keepdim=True), context, context_mask)
+        return x + context.squeeze(1)
+
+
+class AttentionBlock(nn.Module):
+
+    def __init__(self, num_channels, time_embed_dim, context_dim=None, norm_groups=32, head_dim=64, expansion_ratio=4):
+        super().__init__()
+        self.in_norm = ConditionalGroupNorm(norm_groups, num_channels, time_embed_dim)
+        self.attention = Attention(num_channels, num_channels, context_dim or num_channels, head_dim)
+
+        hidden_channels = expansion_ratio * num_channels
+        self.out_norm = ConditionalGroupNorm(norm_groups, num_channels, time_embed_dim)
+        self.feed_forward = nn.Sequential(
+            nn.Conv2d(num_channels, hidden_channels, kernel_size=1, bias=False),
+            nn.SiLU(),
+            nn.Conv2d(hidden_channels, num_channels, kernel_size=1, bias=False),
+        )
+
+    def forward(self, x, time_embed, context=None, context_mask=None):
+        height, width = x.shape[-2:]
+        out = self.in_norm(x, time_embed)
+        out = rearrange(out, 'b c h w -> b (h w) c', h=height, w=width)
+        context = set_default_item(exist(context), context, out)
+        out = self.attention(out, context, context_mask)
+        out = rearrange(out, 'b (h w) c -> b c h w', h=height, w=width)
+        x = x + out
+
+        out = self.out_norm(x, time_embed)
+        out = self.feed_forward(out)
+        x = x + out
+        return x
+
+
+class DownSampleBlock(nn.Module):
+
+    def __init__(
+            self, in_channels, out_channels, time_embed_dim, context_dim=None,
+            num_blocks=3, groups=32, head_dim=64, expansion_ratio=4, compression_ratio=2,
+            down_sample=True, self_attention=True
+    ):
+        super().__init__()
+        self.self_attention_block = set_default_layer(
+            self_attention,
+            AttentionBlock,
+            (in_channels, time_embed_dim, None, groups, head_dim, expansion_ratio),
+            layer_2=Identity
+        )
+
+        up_resolutions = [[None] * 4] * (num_blocks - 1) + [[None, None, set_default_item(down_sample, False), None]]
+        hidden_channels = [(in_channels, out_channels)] + [(out_channels, out_channels)] * (num_blocks - 1)
+        self.resnet_attn_blocks = nn.ModuleList([
+            nn.ModuleList([
+                ResNetBlock(in_channel, out_channel, time_embed_dim, groups, compression_ratio),
+                set_default_layer(
+                    exist(context_dim),
+                    AttentionBlock,
+                    (out_channel, time_embed_dim, context_dim, groups, head_dim, expansion_ratio),
+                    layer_2=Identity
+                ),
+                ResNetBlock(out_channel, out_channel, time_embed_dim, groups, compression_ratio, up_resolution),
+            ]) for (in_channel, out_channel), up_resolution in zip(hidden_channels, up_resolutions)
+        ])
+
+    def forward(self, x, time_embed, context=None, context_mask=None, control_net_residual=None):
+        x = self.self_attention_block(x, time_embed)
+        for in_resnet_block, attention, out_resnet_block in self.resnet_attn_blocks:
+            x = in_resnet_block(x, time_embed)
+            x = attention(x, time_embed, context, context_mask)
+            x = out_resnet_block(x, time_embed)
+        return x
+
+
+class UpSampleBlock(nn.Module):
+
+    def __init__(
+            self, in_channels, cat_dim, out_channels, time_embed_dim, context_dim=None,
+            num_blocks=3, groups=32, head_dim=64, expansion_ratio=4, compression_ratio=2,
+            up_sample=True, self_attention=True
+    ):
+        super().__init__()
+        up_resolutions = [[None, set_default_item(up_sample, True), None, None]] + [[None] * 4] * (num_blocks - 1)
+        hidden_channels = [(in_channels + cat_dim, in_channels)] + [(in_channels, in_channels)] * (num_blocks - 2) + [(in_channels, out_channels)]
+        self.resnet_attn_blocks = nn.ModuleList([
+            nn.ModuleList([
+                ResNetBlock(in_channel, in_channel, time_embed_dim, groups, compression_ratio, up_resolution),
+                set_default_layer(
+                    exist(context_dim),
+                    AttentionBlock,
+                    (in_channel, time_embed_dim, context_dim, groups, head_dim, expansion_ratio),
+                    layer_2=Identity
+                ),
+                ResNetBlock(in_channel, out_channel, time_embed_dim, groups, compression_ratio),
+            ]) for (in_channel, out_channel), up_resolution in zip(hidden_channels, up_resolutions)
+        ])
+
+        self.self_attention_block = set_default_layer(
+            self_attention,
+            AttentionBlock,
+            (out_channels, time_embed_dim, None, groups, head_dim, expansion_ratio),
+            layer_2=Identity
+        )
+
+    def forward(self, x, time_embed, context=None, context_mask=None):
+        for in_resnet_block, attention, out_resnet_block in self.resnet_attn_blocks:
+            x = in_resnet_block(x, time_embed)
+            x = attention(x, time_embed, context, context_mask)
+            x = out_resnet_block(x, time_embed)
+        x = self.self_attention_block(x, time_embed)
+        return x
+
+class ControlNetModel(nn.Module):
+    def __init__(self,
+                 model_channels,
+                 init_channels=None,
+                 num_channels=3,
+                 out_channels=4,
+                 time_embed_dim=None,
+                 context_dim=None,
+                 groups=32,
+                 head_dim=64,
+                 expansion_ratio=4,
+                 compression_ratio=2,
+                 dim_mult=(1, 2, 4, 8),
+                 num_blocks=(3, 3, 3, 3),
+                 add_cross_attention=(False, True, True, True),
+                 add_self_attention=(False, True, True, True)
+                 ):
+        super().__init__()
+        init_channels = init_channels or model_channels
+        self.to_time_embed = nn.Sequential(
+            SinusoidalPosEmb(init_channels),
+            nn.Linear(init_channels, time_embed_dim),
+            nn.SiLU(),
+            nn.Linear(time_embed_dim, time_embed_dim)
+        )
+        self.feature_pooling = AttentionPolling(time_embed_dim, context_dim, head_dim)
+
+        self.in_layer = nn.Conv2d(num_channels, init_channels, kernel_size=3, padding=1)
+
+        hidden_dims = [init_channels, *map(lambda mult: model_channels * mult, dim_mult)]
+        in_out_dims = list(zip(hidden_dims[:-1], hidden_dims[1:]))
+        text_dims = [set_default_item(is_exist, context_dim) for is_exist in add_cross_attention]
+        layer_params = [num_blocks, text_dims, add_self_attention]
+        rev_layer_params = map(reversed, layer_params)
+
+        cat_dims = []
+        self.num_levels = len(in_out_dims)
+        self.down_samples = nn.ModuleList([])
+        for level, ((in_dim, out_dim), res_block_num, text_dim, self_attention) in enumerate(zip(in_out_dims, *layer_params)):
+            down_sample = level != (self.num_levels - 1)
+            cat_dims.append(set_default_item(level != (self.num_levels - 1), out_dim, 0))
+            self.down_samples.append(
+                DownSampleBlock(
+                    in_dim, out_dim, time_embed_dim, text_dim, res_block_num, groups, head_dim, expansion_ratio,
+                    compression_ratio, down_sample, self_attention
+                )
+            )
+
+    def forward(self, x, time, context=None, context_mask=None):
+        time_embed = self.to_time_embed(time)
+        if exist(context):
+            time_embed = self.feature_pooling(time_embed, context, context_mask)
+
+        hidden_states = []
+        x = self.in_layer(x)
+        for level, down_sample in enumerate(self.down_samples):
+            x = down_sample(x, time_embed, context, context_mask)
+            if level != self.num_levels - 1:
+                hidden_states.append(x)
+        return hidden_states
+
+class UNet(nn.Module):
+
+    def __init__(self,
+                 model_channels,
+                 init_channels=None,
+                 num_channels=3,
+                 out_channels=4,
+                 time_embed_dim=None,
+                 context_dim=None,
+                 groups=32,
+                 head_dim=64,
+                 expansion_ratio=4,
+                 compression_ratio=2,
+                 dim_mult=(1, 2, 4, 8),
+                 num_blocks=(3, 3, 3, 3),
+                 add_cross_attention=(False, True, True, True),
+                 add_self_attention=(False, True, True, True),
+                 *args,
+                 **kwargs,
+                 ):
+        super().__init__()
+        init_channels = init_channels or model_channels
+        self.to_time_embed = nn.Sequential(
+            SinusoidalPosEmb(init_channels),
+            nn.Linear(init_channels, time_embed_dim),
+            nn.SiLU(),
+            nn.Linear(time_embed_dim, time_embed_dim)
+        )
+        self.feature_pooling = AttentionPolling(time_embed_dim, context_dim, head_dim)
+
+        self.in_layer = nn.Conv2d(num_channels, init_channels, kernel_size=3, padding=1)
+
+        hidden_dims = [init_channels, *map(lambda mult: model_channels * mult, dim_mult)]
+        in_out_dims = list(zip(hidden_dims[:-1], hidden_dims[1:]))
+        text_dims = [set_default_item(is_exist, context_dim) for is_exist in add_cross_attention]
+        layer_params = [num_blocks, text_dims, add_self_attention]
+        rev_layer_params = map(reversed, layer_params)
+
+        cat_dims = []
+        self.num_levels = len(in_out_dims)
+        self.down_samples = nn.ModuleList([])
+        for level, ((in_dim, out_dim), res_block_num, text_dim, self_attention) in enumerate(zip(in_out_dims, *layer_params)):
+            down_sample = level != (self.num_levels - 1)
+            cat_dims.append(set_default_item(level != (self.num_levels - 1), out_dim, 0))
+            self.down_samples.append(
+                DownSampleBlock(
+                    in_dim, out_dim, time_embed_dim, text_dim, res_block_num, groups, head_dim, expansion_ratio,
+                    compression_ratio, down_sample, self_attention
+                )
+            )
+
+        self.up_samples = nn.ModuleList([])
+        for level, ((out_dim, in_dim), res_block_num, text_dim, self_attention) in enumerate(zip(reversed(in_out_dims), *rev_layer_params)):
+            up_sample = level != 0
+            self.up_samples.append(
+                UpSampleBlock(
+                    in_dim, cat_dims.pop(), out_dim, time_embed_dim, text_dim, res_block_num, groups, head_dim,
+                    expansion_ratio, compression_ratio, up_sample, self_attention
+                )
+            )
+
+        self.out_layer = nn.Sequential(
+            nn.GroupNorm(groups, init_channels),
+            nn.SiLU(),
+            nn.Conv2d(init_channels, out_channels, kernel_size=3, padding=1)
+        )
+
+        self.control_net = None
+
+    def forward(self, x, time, context=None, context_mask=None, control_net_residual=None):
+        time_embed = self.to_time_embed(time)
+        if exist(context):
+            time_embed = self.feature_pooling(time_embed, context, context_mask)
+
+        hidden_states = []
+        x = self.in_layer(x)
+        for level, down_sample in enumerate(self.down_samples):
+            x = down_sample(x, time_embed, context, context_mask, control_net_residual)
+            if level != self.num_levels - 1:
+                hidden_states.append(x)
+        for level, up_sample in enumerate(self.up_samples):
+            if level != 0:
+                x = torch.cat([x, hidden_states.pop()], dim=1)
+            x = up_sample(x, time_embed, context, context_mask)
+        x = self.out_layer(x)
+        return x
+
+
+class ControlNetModel(nn.Module):
+    def __init__(self,
+                 model_channels,
+                 init_channels=None,
+                 num_channels=3,
+                 out_channels=4,
+                 time_embed_dim=None,
+                 context_dim=None,
+                 groups=32,
+                 head_dim=64,
+                 expansion_ratio=4,
+                 compression_ratio=2,
+                 dim_mult=(1, 2, 4, 8),
+                 num_blocks=(3, 3, 3, 3),
+                 add_cross_attention=(False, True, True, True),
+                 add_self_attention=(False, True, True, True),
+                *args,
+                 **kwargs,
+                 ):
+        super().__init__()
+        init_channels = init_channels or model_channels
+        self.to_time_embed = nn.Sequential(
+            SinusoidalPosEmb(init_channels),
+            nn.Linear(init_channels, time_embed_dim),
+            nn.SiLU(),
+            nn.Linear(time_embed_dim, time_embed_dim)
+        )
+        self.feature_pooling = AttentionPolling(time_embed_dim, context_dim, head_dim)
+
+        self.in_layer = nn.Conv2d(num_channels, init_channels, kernel_size=3, padding=1)
+
+        hidden_dims = [init_channels, *map(lambda mult: model_channels * mult, dim_mult)]
+        in_out_dims = list(zip(hidden_dims[:-1], hidden_dims[1:]))
+        text_dims = [set_default_item(is_exist, context_dim) for is_exist in add_cross_attention]
+        layer_params = [num_blocks, text_dims, add_self_attention]
+        rev_layer_params = map(reversed, layer_params)
+
+        cat_dims = []
+        self.num_levels = len(in_out_dims)
+        self.down_samples = nn.ModuleList([])
+        for level, ((in_dim, out_dim), res_block_num, text_dim, self_attention) in enumerate(zip(in_out_dims, *layer_params)):
+            down_sample = level != (self.num_levels - 1)
+            cat_dims.append(set_default_item(level != (self.num_levels - 1), out_dim, 0))
+            self.down_samples.append(
+                DownSampleBlock(
+                    in_dim, out_dim, time_embed_dim, text_dim, res_block_num, groups, head_dim, expansion_ratio,
+                    compression_ratio, down_sample, self_attention
+                )
+            )
+
+    def forward(self, x, time, context=None, context_mask=None):
+        time_embed = self.to_time_embed(time)
+        if exist(context):
+            time_embed = self.feature_pooling(time_embed, context, context_mask)
+
+        hidden_states = []
+        x = self.in_layer(x)
+        for level, down_sample in enumerate(self.down_samples):
+            x = down_sample(x, time_embed, context, context_mask)
+            if level != self.num_levels - 1:
+                hidden_states.append(x)
+        return hidden_states
+
+class ControlUNet(nn.Module):
+
+    def __init__(self,
+                 model_channels,
+                 init_channels=None,
+                 num_channels=3,
+                 out_channels=4,
+                 time_embed_dim=None,
+                 context_dim=None,
+                 groups=32,
+                 head_dim=64,
+                 expansion_ratio=4,
+                 compression_ratio=2,
+                 dim_mult=(1, 2, 4, 8),
+                 num_blocks=(3, 3, 3, 3),
+                 add_cross_attention=(False, True, True, True),
+                 add_self_attention=(False, True, True, True),
+                 control_net_channels=5,
+                 *args,
+                 **kwargs,
+                 ):
+        super().__init__()
+        init_channels = init_channels or model_channels
+        self.to_time_embed = nn.Sequential(
+            SinusoidalPosEmb(init_channels),
+            nn.Linear(init_channels, time_embed_dim),
+            nn.SiLU(),
+            nn.Linear(time_embed_dim, time_embed_dim)
+        )
+        self.feature_pooling = AttentionPolling(time_embed_dim, context_dim, head_dim)
+
+        self.in_layer = nn.Conv2d(num_channels, init_channels, kernel_size=3, padding=1)
+
+        hidden_dims = [init_channels, *map(lambda mult: model_channels * mult, dim_mult)]
+        in_out_dims = list(zip(hidden_dims[:-1], hidden_dims[1:]))
+        text_dims = [set_default_item(is_exist, context_dim) for is_exist in add_cross_attention]
+        layer_params = [num_blocks, text_dims, add_self_attention]
+        rev_layer_params = map(reversed, layer_params)
+
+        cat_dims = []
+        self.num_levels = len(in_out_dims)
+        self.down_samples = nn.ModuleList([])
+        for level, ((in_dim, out_dim), res_block_num, text_dim, self_attention) in enumerate(zip(in_out_dims, *layer_params)):
+            down_sample = level != (self.num_levels - 1)
+            cat_dims.append(set_default_item(level != (self.num_levels - 1), out_dim, 0))
+            self.down_samples.append(
+                DownSampleBlock(
+                    in_dim, out_dim, time_embed_dim, text_dim, res_block_num, groups, head_dim, expansion_ratio,
+                    compression_ratio, down_sample, self_attention
+                )
+            )
+
+        self.up_samples = nn.ModuleList([])
+        for level, ((out_dim, in_dim), res_block_num, text_dim, self_attention) in enumerate(zip(reversed(in_out_dims), *rev_layer_params)):
+            up_sample = level != 0
+            self.up_samples.append(
+                UpSampleBlock(
+                    in_dim, cat_dims.pop(), out_dim, time_embed_dim, text_dim, res_block_num, groups, head_dim,
+                    expansion_ratio, compression_ratio, up_sample, self_attention
+                )
+            )
+
+        self.out_layer = nn.Sequential(
+            nn.GroupNorm(groups, init_channels),
+            nn.SiLU(),
+            nn.Conv2d(init_channels, out_channels, kernel_size=3, padding=1)
+        )
+
+        self.control_net = ControlNetModel(model_channels,
+                                            init_channels,
+                                            control_net_channels,
+                                            out_channels,
+                                            time_embed_dim,
+                                            context_dim,
+                                            groups,
+                                            head_dim,
+                                            expansion_ratio,
+                                            compression_ratio,
+                                            dim_mult,
+                                            num_blocks,
+                                            add_cross_attention,
+                                            add_self_attention)
+
+    def forward(self, x, time, context=None, context_mask=None, control_net_data=None):
+        time_embed = self.to_time_embed(time)
+        if exist(context):
+            time_embed = self.feature_pooling(time_embed, context, context_mask)
+
+        control_net_hiddens =  self.control_net(control_net_data, time, context, context_mask)
+        hidden_states = []
+        x = self.in_layer(x)
+        for level, down_sample in enumerate(self.down_samples):
+            x = down_sample(x, time_embed, context, context_mask)
+            if level != self.num_levels - 1:
+                x += control_net_hiddens.pop(0)
+                hidden_states.append(x)
+        for level, up_sample in enumerate(self.up_samples):
+            if level != 0:
+                x = torch.cat([x, hidden_states.pop()], dim=1)
+            x = up_sample(x, time_embed, context, context_mask)
+        x = self.out_layer(x)
+        return x
+
+
+def get_control_unet(conf):
+    unet = ControlUNet(**conf)
+    return unet
+
+
+def get_unet(conf):
+    unet = UNet(**conf)
+    return unet

Kandinsky-3/kandinsky3/model/utils.py

@@ -0,0 +1,62 @@
+from torch.nn import Identity
+from einops import rearrange
+
+
+def exist(item):
+    return item is not None
+
+
+def set_default_item(condition, item_1, item_2=None):
+    if condition:
+        return item_1
+    else:
+        return item_2
+
+
+def set_default_layer(condition, layer_1, args_1=[], kwargs_1={}, layer_2=Identity, args_2=[], kwargs_2={}):
+    if condition:
+        return layer_1(*args_1, **kwargs_1)
+    else:
+        return layer_2(*args_2, **kwargs_2)
+
+
+def get_tensor_items(x, pos, broadcast_shape):
+    device = pos.device
+    bs = pos.shape[0]
+    ndims = len(broadcast_shape[1:])
+    x = x.cpu()[pos.cpu()]
+    return x.reshape(bs, *((1,) * ndims)).to(device)
+
+
+def local_patching(x, height, width, group_size):
+    if group_size > 0:
+        x = rearrange(
+            x, 'b c (h g1) (w g2) -> b (h w) (g1 g2) c',
+            h=height//group_size, w=width//group_size, g1=group_size, g2=group_size
+        )
+    else:
+        x = rearrange(x, 'b c h w -> b (h w) c', h=height, w=width)
+    return x
+
+
+def local_merge(x, height, width, group_size):
+    if group_size > 0:
+        x = rearrange(
+            x, 'b (h w) (g1 g2) c -> b c (h g1) (w g2)',
+            h=height//group_size, w=width//group_size, g1=group_size, g2=group_size
+        )
+    else:
+        x = rearrange(x, 'b (h w) c -> b c h w', h=height, w=width)
+    return x
+
+
+def global_patching(x, height, width, group_size):
+    x = local_patching(x, height, width, height//group_size)
+    x = x.transpose(-2, -3)
+    return x
+
+
+def global_merge(x, height, width, group_size):
+    x = x.transpose(-2, -3)
+    x = local_merge(x, height, width, height//group_size)
+    return x

Kandinsky-3/kandinsky3/movq.py

@@ -0,0 +1,431 @@
+import math
+import torch
+import torch.nn as nn
+import numpy as np
+import torch.nn.functional as F
+
+from .utils import freeze
+
+
+def nonlinearity(x):
+    return x*torch.sigmoid(x)
+
+
+class SpatialNorm(nn.Module):
+    def __init__(
+        self, f_channels, zq_channels=None, norm_layer=nn.GroupNorm, freeze_norm_layer=False, add_conv=False, **norm_layer_params
+    ):
+        super().__init__()
+        self.norm_layer = norm_layer(num_channels=f_channels, **norm_layer_params)
+        if zq_channels is not None:
+            if freeze_norm_layer:
+                for p in self.norm_layer.parameters:
+                    p.requires_grad = False
+            self.add_conv = add_conv
+            if self.add_conv:
+                self.conv = nn.Conv2d(zq_channels, zq_channels, kernel_size=3, stride=1, padding=1)
+            self.conv_y = nn.Conv2d(zq_channels, f_channels, kernel_size=1, stride=1, padding=0)
+            self.conv_b = nn.Conv2d(zq_channels, f_channels, kernel_size=1, stride=1, padding=0)
+    def forward(self, f, zq=None):
+        norm_f = self.norm_layer(f)
+        if zq is not None:
+            f_size = f.shape[-2:]
+            zq = torch.nn.functional.interpolate(zq, size=f_size, mode="nearest")
+            if self.add_conv:
+                zq = self.conv(zq)
+            norm_f = norm_f * self.conv_y(zq) + self.conv_b(zq)
+        return norm_f
+
+
+def Normalize(in_channels, zq_ch=None, add_conv=None):
+    return SpatialNorm(
+            in_channels, zq_ch, norm_layer=nn.GroupNorm,
+            freeze_norm_layer=False, add_conv=add_conv, num_groups=32, eps=1e-6, affine=True
+        )
+
+
+class Upsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            self.conv = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
+        if self.with_conv:
+            x = self.conv(x)
+        return x
+
+
+class Downsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            self.conv = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=3,
+                                        stride=2,
+                                        padding=0)
+
+    def forward(self, x):
+        if self.with_conv:
+            pad = (0,1,0,1)
+            x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
+            x = self.conv(x)
+        else:
+            x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
+        return x
+
+
+class ResnetBlock(nn.Module):
+    def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False,
+                 dropout, temb_channels=512, zq_ch=None, add_conv=False):
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+
+        self.norm1 = Normalize(in_channels, zq_ch, add_conv=add_conv)
+        self.conv1 = torch.nn.Conv2d(in_channels,
+                                     out_channels,
+                                     kernel_size=3,
+                                     stride=1,
+                                     padding=1)
+        if temb_channels > 0:
+            self.temb_proj = torch.nn.Linear(temb_channels,
+                                             out_channels)
+        self.norm2 = Normalize(out_channels, zq_ch, add_conv=add_conv)
+        self.dropout = torch.nn.Dropout(dropout)
+        self.conv2 = torch.nn.Conv2d(out_channels,
+                                     out_channels,
+                                     kernel_size=3,
+                                     stride=1,
+                                     padding=1)
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                self.conv_shortcut = torch.nn.Conv2d(in_channels,
+                                                     out_channels,
+                                                     kernel_size=3,
+                                                     stride=1,
+                                                     padding=1)
+            else:
+                self.nin_shortcut = torch.nn.Conv2d(in_channels,
+                                                    out_channels,
+                                                    kernel_size=1,
+                                                    stride=1,
+                                                    padding=0)
+
+    def forward(self, x, temb, zq=None):
+        h = x
+        h = self.norm1(h, zq)
+        h = nonlinearity(h)
+        h = self.conv1(h)
+
+        if temb is not None:
+            h = h + self.temb_proj(nonlinearity(temb))[:,:,None,None]
+
+        h = self.norm2(h, zq)
+        h = nonlinearity(h)
+        h = self.dropout(h)
+        h = self.conv2(h)
+
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                x = self.conv_shortcut(x)
+            else:
+                x = self.nin_shortcut(x)
+
+        return x+h
+
+
+class AttnBlock(nn.Module):
+    def __init__(self, in_channels, zq_ch=None, add_conv=False):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels, zq_ch, add_conv=add_conv)
+        self.q = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.k = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.v = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.proj_out = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=1,
+                                        stride=1,
+                                        padding=0)
+
+
+    def forward(self, x, zq=None):
+        h_ = x
+        h_ = self.norm(h_, zq)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b,c,h,w = q.shape
+        q = q.reshape(b,c,h*w)
+        q = q.permute(0,2,1)   # b,hw,c
+        k = k.reshape(b,c,h*w) # b,c,hw
+        w_ = torch.bmm(q,k)     # b,hw,hw    w[b,i,j]=sum_c q[b,i,c]k[b,c,j]
+        w_ = w_ * (int(c)**(-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = v.reshape(b,c,h*w)
+        w_ = w_.permute(0,2,1)   # b,hw,hw (first hw of k, second of q)
+        h_ = torch.bmm(v,w_)     # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j]
+        h_ = h_.reshape(b,c,h,w)
+
+        h_ = self.proj_out(h_)
+
+        return x+h_
+    
+    
+class Encoder(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, z_channels, double_z=True, **ignore_kwargs):
+        super().__init__()
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+
+        # downsampling
+        self.conv_in = torch.nn.Conv2d(in_channels,
+                                       self.ch,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        curr_res = resolution
+        in_ch_mult = (1,)+tuple(ch_mult)
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch*in_ch_mult[i_level]
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(AttnBlock(block_in))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions-1:
+                down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = AttnBlock(block_in)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        2*z_channels if double_z else z_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+
+    def forward(self, x):
+        temb = None
+
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1], temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions-1:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class Decoder(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, z_channels, give_pre_end=False, zq_ch=None, add_conv=False, **ignorekwargs):
+        super().__init__()
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.give_pre_end = give_pre_end
+
+        # compute in_ch_mult, block_in and curr_res at lowest res
+        in_ch_mult = (1,)+tuple(ch_mult)
+        block_in = ch*ch_mult[self.num_resolutions-1]
+        curr_res = resolution // 2**(self.num_resolutions-1)
+        self.z_shape = (1,z_channels,curr_res,curr_res)
+
+        # z to block_in
+        self.conv_in = torch.nn.Conv2d(z_channels,
+                                       block_in,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout,
+                                       zq_ch=zq_ch,
+                                       add_conv=add_conv)
+        self.mid.attn_1 = AttnBlock(block_in, zq_ch, add_conv=add_conv)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout,
+                                       zq_ch=zq_ch,
+                                       add_conv=add_conv)
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks+1):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout,
+                                         zq_ch=zq_ch,
+                                         add_conv=add_conv))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(AttnBlock(block_in, zq_ch, add_conv=add_conv))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up) # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in, zq_ch, add_conv=add_conv)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        out_ch,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, z, zq):
+        #assert z.shape[1:] == self.z_shape[1:]
+        self.last_z_shape = z.shape
+
+        # timestep embedding
+        temb = None
+
+        # z to block_in
+        h = self.conv_in(z)
+
+        # middle
+        h = self.mid.block_1(h, temb, zq)
+        h = self.mid.attn_1(h, zq)
+        h = self.mid.block_2(h, temb, zq)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks+1):
+                h = self.up[i_level].block[i_block](h, temb, zq)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h, zq)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        if self.give_pre_end:
+            return h
+
+        h = self.norm_out(h, zq)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+    
+class MoVQ(nn.Module):
+    
+    def __init__(self, generator_params):
+        super().__init__()
+        z_channels = generator_params["z_channels"]
+        self.encoder = Encoder(**generator_params)
+        self.quant_conv = torch.nn.Conv2d(z_channels, z_channels, 1)
+        self.post_quant_conv = torch.nn.Conv2d(z_channels, z_channels, 1)
+        self.decoder = Decoder(zq_ch=z_channels, **generator_params)
+        
+    @torch.no_grad()
+    def encode(self, x):
+        h = self.encoder(x)
+        h = self.quant_conv(h)
+        return h
+
+    @torch.no_grad()
+    def decode(self, quant):
+        decoder_input = self.post_quant_conv(quant)
+        decoded = self.decoder(decoder_input, quant)
+        return decoded
+    
+    
+def get_vae(conf):
+    movq = MoVQ(conf.params)
+    if conf.checkpoint is not None:
+        movq_state_dict = torch.load(conf.checkpoint)
+        movq.load_state_dict(movq_state_dict)
+    movq = freeze(movq)
+    return movq

Kandinsky-3/kandinsky3/setup.py

@@ -0,0 +1,38 @@
+from setuptools import setup
+
+setup(
+    name="kandinsky3",
+    packages=[
+        "kandinsky3",
+        "kandinsky3/model"
+    ],
+    install_requires=[
+                    "timm",
+                    "torch==1.10.1+cu111",
+                    "torchvision==0.11.2+cu111",
+                    "torchaudio==0.10.1",
+                    "pytorch_lightning==1.7.5",
+                    "transformers",
+                    "accelerate",
+                    "diffusers",
+                    "setuptools==59.5.0",
+                    "omegaconf",
+                    "datasets",
+                    "einops",
+                    "webdataset",
+                    "fsspec",
+                    "s3fs",
+                    "hydra-core",
+                    "scikit-image",
+                    "matplotlib",
+                    "wandb",
+                    "albumentations",
+                    "bezier",
+                    "scipy",
+                    "Pillow",
+                    "tqdm",
+                    "huggingface_hub"
+  
+    ],
+    author="",
+)

Kandinsky-3/kandinsky3/t2i_pipeline.py

@@ -0,0 +1,106 @@
+from typing import Union, List
+import PIL
+
+import torch
+import torchvision.transforms as T
+from einops import repeat
+
+from kandinsky3.model.unet import UNet
+from kandinsky3.movq import MoVQ
+from kandinsky3.condition_encoders import T5TextConditionEncoder
+from kandinsky3.condition_processors import T5TextConditionProcessor
+from kandinsky3.model.diffusion import BaseDiffusion, get_named_beta_schedule
+
+
+class Kandinsky3T2IPipeline:
+
+    def __init__(
+            self,
+            device_map: Union[str, torch.device, dict],
+            dtype_map: Union[str, torch.dtype, dict],
+            unet: UNet,
+            null_embedding: torch.Tensor,
+            t5_processor: T5TextConditionProcessor,
+            t5_encoder: T5TextConditionEncoder,
+            movq: MoVQ,
+            gan: bool,
+    ):
+        self.device_map = device_map
+        self.dtype_map = dtype_map
+        self.to_pil = T.ToPILImage()
+
+        self.unet = unet
+        self.null_embedding = null_embedding
+        self.t5_processor = t5_processor
+        self.t5_encoder = t5_encoder
+        self.movq = movq
+
+        self.gan = gan
+
+    def __call__(
+            self,
+            text: str,
+            negative_text: str = None,
+            images_num: int = 1,
+            bs: int = 1,
+            width: int = 1024,
+            height: int = 1024,
+            guidance_scale: float = 3.0,
+            steps: int = 50,
+            eta: float = 1.0
+    ) -> List[PIL.Image.Image]:
+
+        betas = get_named_beta_schedule('cosine', 1000)
+        base_diffusion = BaseDiffusion(betas, 0.99)
+        times = list(range(999, 0, -1000 // steps))
+        if self.gan:
+            times = list(range(979, 0, -250))
+
+        condition_model_input, negative_condition_model_input = self.t5_processor.encode(text, negative_text)
+        for input_type in condition_model_input:
+            condition_model_input[input_type] = condition_model_input[input_type][None].to(
+                self.device_map['text_encoder']
+            )
+
+        if negative_condition_model_input is not None:
+            for input_type in negative_condition_model_input:
+                negative_condition_model_input[input_type] = negative_condition_model_input[input_type][None].to(
+                    self.device_map['text_encoder']
+                )
+
+        pil_images = []
+        with torch.no_grad():
+            with torch.cuda.amp.autocast(dtype=self.dtype_map['text_encoder']):
+                context, context_mask = self.t5_encoder(condition_model_input)
+                if negative_condition_model_input is not None:
+                    negative_context, negative_context_mask = self.t5_encoder(negative_condition_model_input)
+                else:
+                    negative_context, negative_context_mask = None, None
+
+            k, m = images_num // bs, images_num % bs
+            for minibatch in [bs] * k + [m]:
+                if minibatch == 0:
+                    continue
+                bs_context = repeat(context, '1 n d -> b n d', b=minibatch)
+                bs_context_mask = repeat(context_mask, '1 n -> b n', b=minibatch)
+                if negative_context is not None:
+                    bs_negative_context = repeat(negative_context, '1 n d -> b n d', b=minibatch)
+                    bs_negative_context_mask = repeat(negative_context_mask, '1 n -> b n', b=minibatch)
+                else:
+                    bs_negative_context, bs_negative_context_mask = None, None
+
+                with torch.cuda.amp.autocast(dtype=self.dtype_map['unet']):
+                    images = base_diffusion.p_sample_loop(
+                        self.unet, (minibatch, 4, height // 8, width // 8), times, self.device_map['unet'],
+                        bs_context, bs_context_mask, self.null_embedding, guidance_scale, eta,
+                        negative_context=bs_negative_context, negative_context_mask=bs_negative_context_mask,
+                        gan=self.gan
+                    )
+
+                with torch.cuda.amp.autocast(dtype=self.dtype_map['movq']):
+                    images = torch.cat([self.movq.decode(image) for image in images.chunk(2)])
+                    images = torch.clip((images + 1.) / 2., 0., 1.)
+                    for images_chunk in images.chunk(1):
+                        pil_images += [self.to_pil(image) for image in images_chunk]
+
+        return pil_images

Kandinsky-3/kandinsky3/utils.py

@@ -0,0 +1,71 @@
+from omegaconf import OmegaConf
+import numpy as np
+from scipy import ndimage
+import torch.nn as nn
+from skimage.transform import resize
+
+
+def load_conf(config_path):
+    conf = OmegaConf.load(config_path)
+    conf.data.tokens_length = conf.common.tokens_length
+    conf.data.processor_names = conf.model.encoders.model_names
+    conf.data.dataset.seed = conf.common.seed
+    conf.data.dataset.image_size = conf.common.image_size
+
+    conf.trainer.trainer_params.max_steps = conf.common.train_steps
+    conf.scheduler.params.total_steps = conf.common.train_steps
+    conf.logger.tensorboard.name = conf.common.experiment_name
+
+    conf.model.encoders.context_dim = conf.model.unet_params.context_dim
+    return conf
+
+
+def freeze(model):
+    for p in model.parameters():
+        p.requires_grad = False
+    return model
+
+def unfreeze(model):
+    for p in model.parameters():
+        p.requires_grad = True
+    return model
+
+def zero_module(module):
+    for p in module.parameters():
+        nn.init.zeros_(p)
+    return module
+
+def resize_mask_for_diffusion(mask):
+    reduce_factor = max(1, (mask.size / 1024**2)**0.5)
+    resized_mask = resize(
+        mask,
+        (
+            (round(mask.shape[0] / reduce_factor) // 64) * 64,
+            (round(mask.shape[1] / reduce_factor) // 64) * 64
+        ),
+        preserve_range=True,
+        anti_aliasing=False
+    )
+
+    return resized_mask
+
+def resize_image_for_diffusion(image):
+    reduce_factor = max(1, (image.size[0] * image.size[1] / 1024**2)**0.5)
+    image = image.resize((
+        (round(image.size[0] / reduce_factor) // 64) * 64, (round(image.size[1] / reduce_factor) // 64) * 64
+    ))
+
+    return image
+
+def prepare_mask(mask):
+    ker = np.array([[1, 1,  1, 1, 1],
+        [1, 5,  5, 5, 1],
+        [1, 5, 44, 5, 1],
+        [1, 5,  5, 5, 1],
+        [1, 1,  1, 1, 1]]) / 100
+    out = ndimage.convolve(mask, ker)
+    out = ndimage.convolve(out, ker)
+    out = ndimage.convolve(out, ker)
+
+    mask = (out > 0).astype(int)
+    return mask

Kandinsky-3/requirements.txt

@@ -0,0 +1,23 @@
+timm
+
+pytorch_lightning==1.7.5
+transformers
+accelerate
+diffusers
+setuptools==59.5.0
+omegaconf
+datasets
+einops
+webdataset
+fsspec
+s3fs
+hydra-core
+scikit-image
+matplotlib
+wandb
+albumentations
+bezier
+scipy
+Pillow
+tqdm
+huggingface_hub

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2024 Said Azizov
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -1,13 +1,137 @@
+
+# README
+
+## Overview
+
+This project generates a PowerPoint presentation based on user-provided descriptions. It leverages language models to generate text content and an image generation API to create images for the slides. The architecture is modular, allowing for easy extension and customization of the text and image generation components.
+
+## How to Use
+
+### Prerequisites
+
+- Python 3.10 or higher
+- Required Python packages (listed in `requirements.txt`)
+
+### Setup
+
+1. **Clone the repository**:
+
+   ```bash
+   git clone --recurse-submodules https://github.com/ai-forever/presentations.git
+   cd presentations
+   ```
+
+2. **Install dependencies**:
+
+   ```bash
+   pip install -r requirements.txt
+   ```
+
+3. **Create a .env file** in the root directory with GigaChat credentials:
+
+Here is the [documentation](https://developers.sber.ru/portal/products/gigachat-api) on how to get access token.
+
+   ```plaintext
+   AUTH_TOKEN=XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
+   COOKIE=XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
+   ```
+
+
+4. **Run the FastAPI server** for the image generation API:
+
+   ```bash
+   python src/kandinsky.py
+   ```
+
+### Running the Script
+
+To generate a presentation, use the following command:
+
+```bash
+python main.py -d "Description of the presentation"  -l 'en'
+```
+
+This will generate a presentation based on the provided description and save it in the `logs` directory with a timestamp.
+
+## Examples
+
+```bash
+python main.py -d "Сгенерируй презентацию про планеты солнечной системы" -l 'ru'
+```
+
+```bash
+python main.py -d "Generate presentation about planets of Solar system" -l 'en'
+```
+
+This command will create a presentation on the topic "Planets of the Solar System" using the configured text and image generation functions.
+
+## Architecture
+
+### Main Components
+
+1. **main.py**: The entry point of the application. It parses command-line arguments, initializes required components, and orchestrates the presentation generation process.
+
+2. **Font Class (src/font.py)**: Manages fonts used in the presentation. It can select a random font with basic and bold styles and provide paths to various font styles (basic, bold, italic, and italic bold).
+
+3. **Presentation Generation Functions (src/constructor.py)**: Functions that generate different types of slides in the presentation. They handle the layout, font settings, and placement of text and images.
+
+4. **Text Generation (src/gigachat.py)**: Contains the `giga_generate` function, which generates text based on a given prompt.
+
+5. **Image Generation (src/kandinsky.py)**: Includes the `api_k31_generate` function, which generates images based on a prompt using an external API. Additionally, it provides a FastAPI server for the image generation API.
+
+6. **Prompt Configuration (src/prompt_configs.py)**: Defines the structure of prompts used for generating titles, text, images, and backgrounds for slides.
+
+### How It Works
+
+1. **Initialization**:
+    - `main.py` parses command-line arguments to get the presentation description.
+    - It initializes the `Font` class with the directory containing font files and sets a random font.
+
+2. **Prompt Configuration**:
+    - The `ru_gigachat_config` defines the structure and content of prompts used for generating slide components (titles, text, images, backgrounds).
+
+3. **Text and Image Generation**:
+    - The `giga_generate` function generates text based on the provided description.
+    - The `api_k31_generate` function generates images based on prompts using the FastAPI server.
+
+4. **Slide Generation**:
+    - The `generate_presentation` function orchestrates the creation of slides by calling appropriate functions to generate text and images, and then formats them into slides.
+
+## Extending the Project
+
+### Adding New Font Styles
+
+To add new font styles, place the font files in the `fonts` directory and update the `Font` class if necessary to recognize the new styles.
+
+### Changing Text Generation
+
+To use a different text generation function, replace the `giga_generate` function from `src/gigachat.py` or add a new function and update the call in `main.py`.
+
+### Changing Image Generation
+
+To use a different image generation API, modify the `api_k31_generate` function in `src/kandinsky.py` or add a new function and update the call in `main.py`.
+
+## Acknowledgements
+
+This project leverages the `python-pptx` library for PowerPoint generation, PIL for image processing, and other Python libraries for various functionalities. The text and image generation models are based on external APIs and language models.
+
 ---
-title: Slides Generaror
-emoji: 🏆
-colorFrom: gray
-colorTo: pink
-sdk: gradio
-sdk_version: 4.44.0
-app_file: app.py
-pinned: false
-license: mit
----
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+Feel free to reach out with any questions or suggestions!
+
+## Authors
+
++ Said Azizov: [Github](https://github.com/stazizov), [Blog](https://t.me/said_azizau)
+
+## Citation
+
+```
+@misc{arkhipkin2023kandinsky,
+      title={Kandinsky 3.0 Technical Report}, 
+      author={Vladimir Arkhipkin and Andrei Filatov and Viacheslav Vasilev and Anastasia Maltseva and Said Azizov and Igor Pavlov and Julia Agafonova and Andrey Kuznetsov and Denis Dimitrov},
+      year={2023},
+      eprint={2312.03511},
+      archivePrefix={arXiv},
+      primaryClass={cs.CV}
+}
+```

app.py

@@ -1,7 +1,95 @@
 import gradio as gr
+import time
+from src.constructor import generate_presentation 
+from src.prompt_configs import en_gigachat_config, ru_gigachat_config
+from src.gigachat import giga_generate
+from src.kandinsky import api_k31_generate
+from src.font import Font
 
-def greet(name):
-    return "Hello " + name + "!!"
+logs_dir = "logs"
+fonts_dir = "fonts"
 
-demo = gr.Interface(fn=greet, inputs="text", outputs="text")
-demo.launch()
+def create_presentation(description: str, language: str):
+    # Select the appropriate prompt configuration based on the selected language
+    if language == "English":
+        prompt_config = en_gigachat_config
+    elif language == "Русский":
+        prompt_config = ru_gigachat_config
+    else: 
+        # set default to prevent interruptions in unexpected scenario
+        prompt_config = en_gigachat_config
+        
+    font = Font(fonts_dir)
+    font.set_random_font() 
+    
+    output_dir = f'{logs_dir}/{int(time.time())}'
+
+    generate_presentation(
+        llm_generate=giga_generate, 
+        generate_image=api_k31_generate,
+        prompt_config=prompt_config, 
+        description=description,
+        font=font,
+        output_dir=output_dir,
+    )
+
+    filename = f'{output_dir}/presentation.pptx'
+    
+    return filename
+
+# Updated examples to include language selection
+examples = [
+    ["Generate a presentation on economics, 7 slides", "English"],
+    ["Сгенерируйте презентацию по экономике, 7 слайдов", "Русский"],
+    ["Create a presentation on climate change, 6 slides", "English"],
+    ["Создайте презентацию об изменении климата, 6 слайдов", "Русский"],
+    ["Create a presentation on artificial intelligence, 8 slides", "English"],
+    ["Создайте презентацию об искусственном интеллекте, 8 слайдов", "Русский"],
+    ["Design a presentation on space exploration, 10 slides", "English"],
+    ["Разработайте презентацию о космических исследованиях, 10 слайдов", "Русский"],
+    ["Prepare a presentation on the future of renewable energy, 7 slides", "English"],
+    ["Подготовьте презентацию о будущем возобновляемой энергетики, 7 слайдов", "Русский"],
+    ["Develop a presentation on the history of art movements, 9 slides", "English"],
+    ["Разработайте презентацию о истории художественных движений, 9 слайдов", "Русский"],
+    ["Generate a presentation on the impact of social media, 6 slides", "English"],
+    ["Сгенерируйте презентацию о влиянии социальных сетей, 6 слайдов", "Русский"],
+    ["Create a presentation on sustainable urban planning, 8 slides", "English"],
+    ["Создайте презентацию о устойчивом градостроительстве, 8 слайдов", "Русский"],
+    ["Разработайте презентацию о новшествах в области медицинских технологий, 7 слайдов", "Русский"],
+    ["Design a presentation on innovations in healthcare technology, 7 slides", "English"],
+    ["Подготовьте презентацию о глобальных экономических тенденциях, 5 слайдов", "Русский"],
+    ["Prepare a presentation on global economic trends, 5 slides", "English"],
+    ["Разработайте презентацию о психологии потребительского поведения, 6 слайдов", "Русский"],
+    ["Develop a presentation on the psychology of consumer behavior, 6 slides", "English"],
+    ["Сгенерируйте презентацию о преимуществах осознанности и медитации, 7 слайдов", "Русский"],
+    ["Generate a presentation on the benefits of mindfulness and meditation, 7 slides", "English"],
+    ["Создайте презентацию о достижениях в области автономных транспортных средств, 8 слайдов", "Русский"],
+    ["Create a presentation on advancements in autonomous vehicles, 8 slides", "English"],
+    ["Разработайте презентацию о влиянии изменений климатической политики, 5 слайдов", "Русский"],
+    ["Design a presentation on the impact of climate policy changes, 5 slides", "English"],
+]
+
+iface = gr.Interface(
+    fn=create_presentation,
+    inputs=[
+        gr.Textbox(
+            label="Presentation Description", 
+            placeholder="Enter the description for the presentation..."
+        ),
+        gr.Dropdown(
+            label="Language",
+            choices=["English", "Russian"],
+            value="English"
+        )
+    ],
+    outputs=gr.File(
+        label="Download Presentation"
+    ),
+    title="Presentation Generator",
+    description="Generate a presentation based on the provided description and selected language. Click the button to download the presentation.",
+    css="footer {visibility: hidden}",
+    allow_flagging="never",  
+    examples=examples  
+)
+
+iface.launch()

fonts/README

@@ -0,0 +1,28 @@
+copr-some-nice-fonts
+====================
+
+This is a Fedora/CentOS repo for easy installing of some nice fonts. This
+includes the following fonts:
+
+- Arial
+- Comic Sans MS
+- Courier New
+- Georgia
+- Helvetica Neue
+- Helvetica
+- Lucida Sans Unicode
+- Tahoma
+- Times New Roman
+- Trebuchet MS
+- Verdana
+
+Using it
+--------
+
+sudo dnf copr enable adrienverge/some-nice-fonts
+sudo dnf install some-nice-fonts
+
+Building it
+-----------
+
+cp *.ttf ~/rpmbuild/SOURCES && rpmbuild -ba some-nice-fonts.spec