Support XYZ scripts / split hires path from unet #14126
Merged
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Description
Supports XYZ options for Hypertile setups
Extends swap size option
Separates HR pass option from Unet option, so that hypertile can only work for hr pass
Hypertile has Swap size problem (as far as I know)
And as conclusion - I'd just let swap size to biggest one that we can use.
All images were generated with Blessed2 VAE, CuteYukiMix, with prompt
1girl
,easynegative, nsfw
Here is example with 512x768 without any hr-pass.
With 1024x1536, of course it won't work without hr pass, so maybe with kohya's hires fix:
Or default
With swap size 1, we can observe significant artifacts which caused by 'unknown nearby pixels', splitted by hypertile.
By allowing bigger swap sizes, we can usually avoid the artifacts with minimal performance degradation.
(There will be additional patch to pre-calculate list of divisors and cache it too)
Also, I observed that LCM scheduler + LCM Lora will improve performance and reduce artifact in some aspect too.
(1024x1024 with kohya's hires fix, 10 step test)
Screenshots/videos:
1024x1408 with kohya hires fix
512x704, 2x hires fix with Anime6B (denoise 0.5):
Some close-up, almost no artifact for 512x704 + 2x hires fix
TODO : fix some code snippets, cache divisorsChecklist: