alfredplpl's picture
Update README
41cfcc2
metadata
license: openrail++
tags:
  - stable-diffusion
  - text-to-image
widget:
  - text: Cartoon, ocean.
    example_title: Ocean (Cartoon)

Clean Diffusion 2.0 PoC Model Card

That's one small step for artists, one giant leap for engineers.

ocean

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.

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.

TBA

Generating

TBA

Training Images

TBA

List of works

  • 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

Training Process of VAE

TBA

Training text-image pairs

TBA

Trainning code and config

TBA

Acknowledgement

Standing on the shoulders of giants

Citations

@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}
}
@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}
}