|
--- |
|
license: openrail++ |
|
tags: |
|
- stable-diffusion |
|
- text-to-image |
|
--- |
|
|
|
# Clean Diffusion 2.0 PoC Model Card |
|
|
|
> That's one small step for artists, one giant leap for engineers. |
|
|
|
 |
|
|
|
Clean Diffusion 2.0 PoC is Latent Diffusion Model made of public domain images. |
|
Clean Diffusion 2.0 PoC is for the proof of the concept: Stable Diffusion can be made of public domain images. |
|
Therefore, the model can only express the ocean. |
|
If you are Japanese, I recommend Clean Diffusion For Japanese (TBA) instead of Clean Diffusion (For Global). |
|
The model is more powerful than this global version. |
|
|
|
# Note |
|
> With great power comes great responsibility. |
|
|
|
If you **CANNOT UNDERSTAND THESE WORDS**, I recommend that **YOU SHOULD NOT USE ALL OF DIFFUSION MODELS** what have great powers. |
|
|
|
# Getting Started |
|
You would be able to use Clean Diffusion by the following code soon. |
|
|
|
```python |
|
from diffusers import StableDiffusionPipeline |
|
import torch |
|
|
|
model_id = "alfredplpl/clean-diffusion-2-0-poc" |
|
pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float32) |
|
pipe = pipe.to("cuda") |
|
|
|
prompt = "Cartoon, ocean." |
|
image = pipe(prompt).images[0] |
|
|
|
image.save("ocean.png") |
|
``` |
|
|
|
# Tuning |
|
Clean Diffusion is less powerful than Stable Diffusion. |
|
Therefore, I recommend to tune Clean Diffusion like Stable Diffusion |
|
because Clean Diffusion of the network architecture and Stable Diffusion of the network architecture are same. |
|
And I repeat the words before I explain the detail. |
|
|
|
> With great power comes great responsibility. |
|
|
|
Please consider the words before you tune Clean Diffusion. |
|
|
|
## Textual Inversion |
|
TBA on Colab. |
|
|
|
## Dreambooth on Stable Diffusion |
|
TBA on Colab. |
|
|
|
## Pure fine-tuning |
|
TBA |
|
|
|
# Transparency of Clean Diffusion |
|
I proof that clean diffusion is clean by following explanation. |
|
|
|
## Legal information |
|
TBA |
|
|
|
### Training |
|
Clean Diffusion is legal and ethical. |
|
|
|
Clean Diffusion is MADE IN JAPAN. |
|
Therefore, Clean Diffusion is subject to [Japanese copyright laws](https://en.wikipedia.org/wiki/Copyright_law_of_Japan). |
|
|
|
TBA |
|
|
|
### Generating |
|
|
|
TBA |
|
|
|
## Training Images |
|
TBA |
|
|
|
### List of works |
|
- [ArtBench](https://github.com/liaopeiyuan/artbench) (public domain is True) |
|
- Popeye the Sailor Meets Sindbad the Sailor |
|
|
|
### Tiny training images |
|
I would like to the all training raw images because these images are public domain. |
|
However, these images are huge (70GB+). |
|
Therefore, I have opened the tiny version like this. |
|
|
|
[Tiny Images](https://1drv.ms/u/s!ApxVlgxlqLRliLpSC58y5qyAlt52tQ?e=3Yfwbt) |
|
|
|
### Training Process of VAE |
|
|
|
TBA |
|
|
|
## Training text-image pairs |
|
TBA |
|
|
|
## Trainning code and config |
|
TBA |
|
|
|
# Acknowledgement |
|
> Standing on the shoulders of giants |
|
|
|
# Citations |
|
|
|
```bibtex |
|
@misc{rombach2021highresolution, |
|
title={High-Resolution Image Synthesis with Latent Diffusion Models}, |
|
author={Robin Rombach and Andreas Blattmann and Dominik Lorenz and Patrick Esser and Björn Ommer}, |
|
year={2021}, |
|
eprint={2112.10752}, |
|
archivePrefix={arXiv}, |
|
primaryClass={cs.CV} |
|
} |
|
``` |
|
```bibtex |
|
@article{liao2022artbench, |
|
title={The ArtBench Dataset: Benchmarking Generative Models with Artworks}, |
|
author={Liao, Peiyuan and Li, Xiuyu and Liu, Xihui and Keutzer, Kurt}, |
|
journal={arXiv preprint arXiv:2206.11404}, |
|
year={2022} |
|
} |
|
``` |
|
|
|
|
