Stable Diffusion Explained
Series of videos going in-depth into how Stable Diffusion works
I’d like to create a series of videos covering the internal workings of Stable Diffusion.
I already have done the live version of the first part both online and in-person, and should be able to do a proper recording with some minimal refining. It covers the basic mechanism of how the VAE, CLIP and U-NET parts interact to extract appealing images for random noise.
After that, I will do a video on DreamBooth, as a refinement of an interactive live presentation I gave in
person, and move on to the goodies that are covered in Stable Diffusion 2.0.
Besides that, we have some crazy awesome developments like Control Net. I’m not sure if it would be feasible to include it under the current grant, but if the project is successful and community feedback is positive, maybe we can extend it to another time?
- Video on vanilla Stable Diffusion inner workings
- Video on Dreambooth training
- The inner workings of SD 2.0
Each video on the order of 30 minutes, will include diagrams, live code execution, and commentary.
Squad Lead: xl0
- Twitter handle: @xl0xl0xl0
- discord handle: Alexey Zaytsev#6786
I have already received a grant from Algovera in round 7
The project was a complete success, going well past the proposed deliverables:
It also resulted in Lovely Numpy, an even more popular off-shoot of Lovely Tensors than the proposed Lovely JAX: