-
Notifications
You must be signed in to change notification settings - Fork 28.3k
Feature Request: Restore "Scale Latent" option in Hires Fix #2613
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
Comments
I concur, loading previous prompts that used HRfix now has dissimilar results, possibly from not having latent? |
I agree, it was useful in a lot of cases. |
I agree, you removed now something that was very useful. I don't know how to use this new thing - the results look bad without scale latent. |
and then stay with this version without any new update ? |
Could we please get the option back instead? |
Everything looks so much worse with the new |
Agreed, I'm getting inferior results without "Scale Latent" |
|
Just to add another voice, I also noticed better results with |
It's that kind of je ne sais quoi that just makes a good image great. I can't seem to pin down exactly what it is, but the sharpness and overall color science are definitely lacking without |
@zenblender 1st one up top with |
I'm guessing the 2nd, I see finer details in the eye, coat and reflection. Obvious difference with the face too. The latest fix improved things for sure but still not getting the quality I got before with scale latent. |
@RomanescoAI maybe it's not obvious to everyone. The second one, with the more defined wrinkles in the face, and overall higher contrast, is with |
@zenblender funny I'm the one failing your test, but for sure, it might be case-dependent. I've generated over 5,000 great-looking images with |
It's of course subjective @RomanescoAI and I am not sure that I always prefer the results of |
If you're getting overly mushy/smooth details, one cause may be having a bad img2img upscaler selected in your settings. RealESRGAN is particularly egregious for this. If you select a more detail-preserving upscaler model, like SwinIR or one of the good custom esrgan models like Remacri, you may find the problem reduced somewhat. I too support the return of the scale latent option though, it was nice to have the option of skipping the upscaler model step and instead having the upscaling entirely performed by SD. |
I'm not using and have not used an upscaler, as these are all vanilla generations. Glad to hear more support in favor of the `Scale latent' return. |
You are in fact using an upscaler now, though you may not have realized it. It's an automatic part of the hires fix process with scale latent gone. |
Thanks for clarifying that. I still wish 'Scale Latent' was back, to the way it was working originally. |
I've also noticed fine details from the original getting lost and also a graininess that wasn't there with scale latent on (at least for the images I'm generating). |
which version was this removed so I can revert it for now? |
@zenblender |
#2523 (comment) you’ll find the commit here |
For “scale latent” references, all of these I’ve been doing for the past week have all been “scale latent” on - https://twitter.com/rainisto/status/1580285490925735936?s=46&t=IxJ61DsmTWgBlzwj5kMF9A It’s very good for this “realistic” look. Don’t really know how to replicate this look without it… |
Scale latent have been giving me results with much more details. Please bring it back :( |
Is it then possible to even make "scale latent" a part of the img-img upscaling process? Just curious, thanks for clarifying the issue though. |
Bumping this i'm getting way worse results than before pulling |
nice thanks voldy |
Thanks for bringing this back! It's now a checkbox in the |
Describe the bug
Please add scale latent back into highres.fix, it was producing decent result for me , example :
Screenshots


If applicable, add screenshots to help explain your problem.
just highres
scale latent
Desktop (please complete the following information):
git rev-parse HEAD
]The text was updated successfully, but these errors were encountered: