stable-diffusion.cpp Sana support

https://github.com/NVlabs/Sana

nvidia just release their model that claims 100x faster than flux. If true, it will be nice to support it in sd.cpp.

Jan 14 '25 19:01 chinshou

I actually started working on it last month, but I think I accidentally lost all the progress I made with a bad git command😬. If I can recover it, I'll start working on it again. Edit: git fsck saved the day

Jan 14 '25 19:01 stduhpf

Any progress on this?

Mar 09 '25 00:03 SkutteOleg

Any progress on this?

Not really. I did do some of the work to implement the backbone diffusion model (no idea if my code even works, you can find it here: https://github.com/stduhpf/stable-diffusion.cpp/tree/sana-wip), but I lost motivation at some point.

It's still completely missing the 32x32 DC-VAE (that one is annoying because VAE is assumed to compress from 8x8 all over the code), and the text "encoder" (gemma2).

i might come back to it, but no promises

Mar 09 '25 01:03 stduhpf

Hey @stduhpf sorry to disturb if this is a waste of time, I asked an "AI agent" to try to implement Sana. It failed of course, but could this be at least remotely useful? https://github.com/SkutteOleg/stable-diffusion.cpp/tree/sana-ai-slop

Jun 05 '25 01:06 SkutteOleg