Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Vikram/sd3.5m skiplayercfg #11

Merged
merged 10 commits into from
Oct 29, 2024
Merged

Vikram/sd3.5m skiplayercfg #11

merged 10 commits into from
Oct 29, 2024

Conversation

voletiv
Copy link
Collaborator

@voletiv voletiv commented Oct 29, 2024

Adds SkipLayerCFGDenoiser for SD3.5-Medium

Vikram Voleti and others added 10 commits October 25, 2024 21:01
Co-authored-by: Vikram Voleti <vikram@ip-26-0-153-234.us-west-2.compute.internal>
Co-authored-by: Vikram Voleti <vikram@ip-26-0-153-234.us-west-2.compute.internal>
@voletiv voletiv self-assigned this Oct 29, 2024
@voletiv voletiv merged commit 4e484e0 into main Oct 29, 2024
@voletiv voletiv deleted the vikram/sd3.5m_skiplayercfg branch October 29, 2024 14:23
@CodeExplode
Copy link

CodeExplode commented Oct 29, 2024

Heya, just wondering what layers it is recommended to use this on, for which noise ranges?

edit: Never mind, for anybody else wondering it seems to be:

layers: 7,8,9
scale: 3
start_percent: default: 0.01
end_percent: 0.15

@stduhpf
Copy link

stduhpf commented Oct 30, 2024

Can someone help me understand what is the point of those changes? How does this affect output, speed and all?

@Dango233
Copy link

SLG is for enhancing the guidance vector of a "certain part" of the model - in this case, we regard it as a way to enhance structural information like hands and anatomy,

It requires an additional forward pass to work so will slowdown the generation by around 33%

@CodeExplode
Copy link

Dango233 do you know if this idea is unique to this model? Or has it been tried on 3.5 large, SDXL, SD1.5, etc?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

4 participants