AUTOMATIC1111
[Bug]: High-res fix is breaking 768 model gens. (Making themm garbled/blurry)
Is there an existing issue for this?
- [X] I have searched the existing issues and checked the recent builds/commits
What happened?
I have noticed for the last 2-3 pulls adding "high res fix" breaks 2.0 768 model output. Even at low strengths. What happens is images come out garbled, blurry, with very odd backgrounds; often creating strange outputs. Turn it off and everything is fine; but it was originally working.
Steps to reproduce the problem
- Text2img
- 768 model
- select Euler a or two new upscalers
- Generate
- See issues.
- Turn off high res fix
- See proper Gen.
What should have happened?
High res fix should be essentially creating high quality slightly variant output from the first seed, but not garbled blurry messes.
Commit where the problem happens
Latest
What platforms do you use to access UI ?
Windows
What browsers do you use to access the UI ?
Mozilla Firefox, Google Chrome
Command Line Arguments
--xformers cuda116
Additional information, context and logs
None
Did you try changing the base resolution to 768x768? It's probably defaulting to a base res of 512x512, which looks about as nasty as you'd expect forcing a 768x768 model to output 512x512 to look.
yeah, its probably not designed for the 768 model, it likely that the 768 model itself produces bad generations at lower resolution. Can you post some output, or perhaps show a better slider adjustment for the 768 model?
If highres fix starts at 256x256 for base model (any lower and you get a mess) I'd expect the 768 to be way worse at 256x256, so we'd have to test a good minimum size for 768 model
I just tried using highres.fix on the latest version (As of this second), and it still works fine with the firstpass res set to 768x768, and the outres set to 1024x1024. "A guy trying to diagnose a bug on a mainframe, Kodachrome photo from 1974"
Leaving the firstpass res at 0x0 though, things look extremely oversaturated and a bit chaotic, almost like it's running at 512x512...
The best minimum resolution for the 768 model is 768x768, anything lower increases the saturation and chaoticness of the output. Even 704x704 is a touch too saturated, and the highres.fix results end up being too blurry.
Additionally, putting on the face-fixing system causes a similar breakage.
Yes codeformer especially breaks v2 completely.
Sent from Mobile.
On Mon, Dec 5, 2022, 9:51 PM azureprophet @.***> wrote:
Additionally, putting on the face-fixing system causes a similar breakage.
— Reply to this email directly, view it on GitHub https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/5430#issuecomment-1338658527, or unsubscribe https://github.com/notifications/unsubscribe-auth/AE2TIUH3HR5K35HT75BZMPTWL2S3ZANCNFSM6AAAAAASTS76SM . You are receiving this because you authored the thread.Message ID: @.***>
Still happening even with 768 as the first pass... High res fix OFF, picture is SHARP and coherent. High res fix on, it's BLURRY and less coherent.
