highcwu / anime_biggan_toy

Generate Amazing Anime Pictures With BigGAN. Just Have Fun !!!

Anime BigGAN Toy

Generate Amazing Anime Pictures With BigGAN. Just Have Fun !!!

EN | 中文

Free V100 GPU For Model Test

My AI Studio Project

You can download the pretrained model of PaddlePaddle for local play from Anime BigGAN Pretrained Models As A Dataset or Google Drive which just contain the Generator Parameters.

File Samples.ipynb is a backup of this project's notebook.

My QQ Group: 1044867291

My Discord: https://discord.gg/YwWcAS47qb

Play in colab

Play it right now!

Play with TF!

Steps to convert model to TFHub, PyTorch and PaddlePaddle

• 1. Convert Shawwn's Model to TFHub version.
• 2. If you want to play the TFHub model, come here.
• 3. Convert TFHub generator model to PyTorch.
• 4. Convert TFHub discrimator model to PyTorch.
• 5. If you want to play the PyTorch model, come here.
• 6. Convert TFHub generator model to PaddlePaddle.
• 7. Convert TFHub discrimator model to PaddlePaddle.
• 8. If you want to play the PaddlePaddle model, come here.
• Wish you can have fun and find out more interesting ways to play it.

**Update 2020-10-25: Add truncnorm. Add a fast way to play the source tf pretrained model.

**Update 2022-05-22: Add colab for directly converting tf model to pytorch.

TODO

• Test code release.
• Make my AI Studio project public.
• Publish the colab notebook of model converting from TF to PaddlePaddle and PyTorch.
• Publish a gradio demo on HuggingFace Spaces.
• Update PaddlePaddle version to 2.3.

References

Models converted from Shawwn's training [1] and Gwern's release [2]. 2D character picture (HatsuneMiku) is licensed under CC BY-NC by piapro [3].

• [1] Shawwn, The model is trained based on his fork of google's 'compare_gan', https://github.com/shawwn/compare_gan/
• [2] Gwern, "A prototype anime BigGAN 256px trained on Danbooru2019+e621 for 600k iterations is now available for download", https://www.gwern.net/Faces#danbooru2019e621-256px-biggan
• [3] "For Creators", http://piapro.net/en_for_creators.html

anime_biggan_toy's Issues

What's the definition of different levels?

I noticed that the pic can only be changed by shape or color during interpolation, controlled by levels.
If it's possible, is there any definition of these levels (e.g. z11; z32; z42...)
Thank you so much!

example code in sample.ipynb:
std_gen_interpolate(8, seed=2333, out_path='data/std_inter1', levels="z11;z12")
std_gen_interpolate(8, seed=2333, out_path='data/std_inter2', levels="z21;z22")
std_gen_interpolate(8, seed=2333, out_path='data/std_inter3', levels="z31;z32")
std_gen_interpolate(8, seed=2333, out_path='data/std_inter4', levels="z41;z42")
std_gen_interpolate(8, seed=2333, out_path='data/std_inter5', levels="z51;z52")
std_gen_interpolate(8, seed=2333, out_path='data/std_inter6', levels="z61;z62")

Is there a way to generate by class?

Correct me if I'm wrong but the model was trained on 1000 classes, is there a way to generate images by class number?
Like I want to generate images for 500th class for example, is there a lookup what anime character for the 500th class?

I know this one has it: https://www.gwern.net/Faces#constructing-d1k but that was trained on a different model and dataset (Danbooru2018-1K Dataset) so it is safe to say the classes are still the same on this one (DANBOORU2019+E621) ?

How to change truncation value?

I've been eyeing the BigGAN anime model for a while but I don't have the know-how to get it work to generate images, so thanks so much for making it work! (if only BigGAN isn't computationally expensive and complex to train and make it work 😢)

Anyway so in BigGAN there is something called a truncation trick that is also in StyleGAN, where I could possibly change the truncation value in the code?

For reference I am using this: https://colab.research.google.com/github/HighCWu/anime_biggan_toy/blob/main/colab/Play_Anime_BigGAN.ipynb

Such a great project! Is it possible to run with anaconda on Windows?

项目超赞，已星！！本地运行需要的requirement都是啥呢，是不是要安装一个paddlepaddle 1.8.4环境之类的？
如果可以的话，我打算试试做一期视频介绍一下~