Optimization Discussion #732
Replies: 4 comments 5 replies
-
CompVis/stable-diffusion#177 is very small but still quite effective. |
Beta Was this translation helpful? Give feedback.
-
I think someone added this but it got lost in the merge |
Beta Was this translation helpful? Give feedback.
-
I love the upgrades you put on this SD! I wanted to know if anyone could increase the optimization to allow 12gb users to use 1024 x 1024 or 1280 x 768 like Disco Diffusion. 1280 X 768 or 1280 x 1024 would be incredible!! It adds time but this is still better than Disco Diffusion. So it would be an awesome upgrade to add ASAP. |
Beta Was this translation helpful? Give feedback.
-
I'm collecting a list of useful optimizations at https://links.sd |
Beta Was this translation helpful? Give feedback.
-
I've seen some interesting modifications people have made, some just to attention.py - cleaning up after stuff is used.
A good discussion ongoing here:
https://gist.github.com/mchaker/e0cf180db6849fa03696293b3fc8f485
Could any/all of these be adapted for this repo? For the attention mods that slow things down, I think that should perhaps require a checkmark or config option, or automatic memory threshold determination - that is, if you don't have a 24GB VRAM card and instead have 11-12 or less, the slow but stable stepping option should kick in for images greater than 768 or a bit more pixels on either dimension....
I think we may already somewhat do this?
CompVis/stable-diffusion#142
Then, more interesting - Doggettx memory optimization: allows rendering 1024x1024 images and even up to 4096x1024:
CompVis/stable-diffusion@main...Doggettx:stable-diffusion:main
Another currently incompatible (or similar vein?) change:
basujindal/stable-diffusion#122
Beta Was this translation helpful? Give feedback.
All reactions