stable-diffusion-webui
stable-diffusion-webui copied to clipboard
AUTOMATIC1111
[Feature Request]: Old high-res fix mode
Is there an existing issue for this?
- [X] I have searched the existing issues and checked the recent builds/commits
What would your feature do ?
add a option to make use of the old high-res fix.
Proposed workflow
- Go to settings
- Press "old high-res fix"
- use old high-res fix in the txt2img tab
Additional information
No response
or you could add the old selection menu, for the higher fix + the new upscaler thingie(with the first pass size and so on)
Not opposed to this, but what exactly are you trying to do? Old hires. fix made it very easy to commit errors, and it had only a few upsides - like repeating canvases from selecting a high width & height.
There seem to be some differences in how the scaler works, maybe even the first generation pass too, which also are influencing things.
Yes this is like forcing to use latent upscaler,and even when using other upscalers,it takes a lot of time to produce results and most of the time UI gets stuck when you press cancel and only way to get it back to working is to reset the UI and you lose your prompt,settings etc. Old behavior was much better and giving faster results. (at least defaults needs to be changed)
The latent scaler is how it used to work too - the problem is that there seems to be some other things changed as well. There's a bunch of issues and some of them are being fixed up, like the hanging problem - I had that earlier, but it seems to be fixed. Try out the new updates, you might like them.
The latent scaler is how it used to work too - the problem is that there seems to be some other things changed as well. There's a bunch of issues and some of them are being fixed up, like the hanging problem - I had that earlier, but it seems to be fixed. Try out the new updates, you might like them.
I still have the hanging problem(it is annoying getting stuck when you try to cancel) and using the latest commit. I am not sure what has been changed with the high res fix or is it just the defaults was different? So we are having some problem...
Old images aren't the same with this new highres when using the same settings, that's already a reason to get the option. With my tests the gens look considerably different and seem to take longer.
Also anyone knows which version I should revert my installation to get the old hires?
One major benefit to the old high res fix was I could use it to generate a 768x768, then upscale to a much larger resolution AND a different aspect ratio while preserving the subject.
One issue with generating images larger than 768x768 at a 2:3 or 3:2 aspect ratio for instance is long necks, multiple heads or stacked torsos etc... Now most of the images I generate have those odd artifacts that just get upscaled into the final image.
Repeating myself again, but: If you are having long generations or multiples, check that you are setting the width * height as if it's the firstpass value from the old hires. fix. Because that's what it is now. Use the multiplier to get your final canvas value.
Repeating myself again, but: If you are having long generations or multiples, check that you are setting the width * height as if it's the firstpass value from the old hires. fix. Because that's what it is now. Use the multiplier to get your final canvas value.
Either you're wrong or PNG info/Image Browser isn't working properly with the update, since sending the images straight from them results in different images in the newest commit.
first pass res is the main res now.. u set upscaler multiply to set the final resolution when using new highres fix. it fixes the hassle to calculate aspect ratio or crop ins.. but still there is a lot of bugs now .. it can not find the upscaler sometimes.. u have to set img2img upscaler from settings then... progress bar is broken
Also anyone knows which version I should revert my installation to get the old hires?
This commit introduces the 'hires fix rework' changes: https://github.com/AUTOMATIC1111/stable-diffusion-webui/commit/ef27a18b6b7cb1a8eebdc9b2e88d25baf2c2414d
Here is a discussion about a defects in that commit: https://github.com/AUTOMATIC1111/stable-diffusion-webui/commit/ef27a18b6b7cb1a8eebdc9b2e88d25baf2c2414d#comments
if you want to revert to the last commit before the 'highres fxi rework', this is it: https://github.com/AUTOMATIC1111/stable-diffusion-webui/commit/fd4461d44c7256d56889f5b5ed9fb660a859172f
also, I too find the image coherence and speed to be drastically impaired by the new hires fix rework, when it does manage not to hang my a1111 instance.
Repeating myself again, but: If you are having long generations or multiples, check that you are setting the width * height as if it's the firstpass value from the old hires. fix. Because that's what it is now. Use the multiplier to get your final canvas value.
Either you're wrong or PNG info/Image Browser isn't working properly with the update, since sending the images straight from them results in different images in the newest commit.
I didn't say you're guaranteed to get identical results. There seem to be some other changes to basic image generation and the scaler too, but that's not anything to do with changing how you select the first pass and canvas values.
@spinferno: I'm actually getting better images sometimes. Slowness could be bugs, could also be setting the width * height too high as that's first pass values now.
@AUTOMATIC1111 , please change it to what it looked like before, I don't wanna do math every time I wanna change my res
Old hires. fix made it very easy to commit errors
Old hires. fix was incredibly easy to use without trouble. You set the resolution you wanted, ticked the box, and it just worked. The floating buildings, double-torsos and heads, and so on mostly went away. With the new version, it seems like it's just upscaling an incoherent image instead of making it more coherent.
the old style of the ui(instead of this new "scale by thing"), not how the scaling works
the old style of the ui(instead of this new "scale by thing"), not how the scaling works
https://github.com/AUTOMATIC1111/stable-diffusion-webui/commit/ef27a18b6b7cb1a8eebdc9b2e88d25baf2c2414d#commitcomment-95019921 check this. Maybe this would help.. just type the 1st pass value from high res fix in main resolution section.. select the upscaler you like. Set upscale by how many times. For upscale by 2 I would recommend .45 denoising strength.
like this image would generate this quality image
https://github.com/AUTOMATIC1111/stable-diffusion-webui/commit/81490780949fffed77493b4bd741e96ec737fe27
Repeating myself again, but: If you are having long generations or multiples, check that you are setting the width * height as if it's the firstpass value from the old hires. fix. Because that's what it is now. Use the multiplier to get your final canvas value.
Either you're wrong or PNG info/Image Browser isn't working properly with the update, since sending the images straight from them results in different images in the newest commit.
I didn't say you're guaranteed to get identical results. There seem to be some other changes to basic image generation and the scaler too, but that's not anything to do with changing how you select the first pass and canvas values.
@spinferno: I'm actually getting better images sometimes. Slowness could be bugs, could also be setting the width * height too high as that's first pass values now.
hey I really appreciate the effort you are putting in to help us in this issue!! if I wanted to replicate the old hires fix, what upscaler model do i select? my current results are super grainy
