This is wonderful! I'm an English speaker, but can I help?

#4
by MatthewWaller - opened

I was able to get Stable Diffusion running on phones even before Apple did, but I never released for concern over copyright. https://www.cephalopod.studio/blog/adventures-in-ai-ethics-part-2-stable-diffusion-v2-and-the-curse-of-scale

Would you be able to share the resources on how to get training up and running? Like how you procured your data and got started?

As for the training, we used a diffusers based script.
The source code here also allows for full scratch training of U-Net by replacing from_pretrained with from_config.
https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image.py#L378-L380

And the creation of the dataset is very hard work. For each source, we implemented the appropriate downloader and captioning scripts.
Publishing this as a dataset is also a daunting task, and we do not plan to do so at this time.

I would also love to help by contributing to this model . I think this is exactly what we need right now !

Not sure I have the skills to help, but as a human artist whose copyrighted work has likely ended up in a dataset without my consent, I think this a great initiative. Good luck!

Sign up or log in to comment