4 ms·
It is from the Heidelberg giants, Patrick Esser and Robin Rombach They continued building off their latent diffusion direction (encode with vqgan-vae and then
by lucidrains 4y ago
It is from the Heidelberg giants, Patrick Esser and Robin Rombach
They continued building off their latent diffusion direction (encode with vqgan-vae and then diffusion in latent space)
All roads lead to rome
https://github.com/CompVis/stable-diffusion https://github.com/CompVis/stable-diffusion
https://arxiv.org/abs/2112.10752 https://arxiv.org/abs/2112.10752
- SleekEagle 4y agoInteresting, thanks for the link. It seems that CLIP encodings aren't as useful as frozen encoders from the textual domain, which is a little unintuitive imo. Can't keep up with all these advancements!
- lucidrains 4y agoIkr! Hang on, this ride is about to get crazy :)