stable-diffusion-webui
stable-diffusion-webui copied to clipboard
AUTOMATIC1111
[Bug]: Highresfix denoising strength behaviour has changed since a day or two ago and is causing blurryness.
Is there an existing issue for this?
- [X] I have searched the existing issues and checked the recent builds/commits
What happened?
Images below 0.7 Denoinsing strength generate blurry in high res fix.. especially at high resolutions. Before you could do 0.4 to apply slight noise for it to re-step through, but instead it now it seems to resolve noise worse and worse the lower you set it, and is also applying the incorrect noise to resolve at these levels.
It wasn't like this a the other day, I'm pretty sure image2image broke the other day too and after that was fixed, I noticed this issue. Somehow related?
Steps to reproduce the problem
- Enable high res fix on something close to like 1400x1000, euler a 28steps, any prompt.
- Try any denoising strengths below 0.7
- Observe and compare denoising strengths results ranging from 0-0.7 it just makes the sharpness and quality worse and worse the lower you go.
What should have happened?
A nice sharp image. This never used to happen, instead less things would change, now it's as if it doesn't resolve the the noise at all.
Commit where the problem happens
No clue.
What platforms do you use to access UI ?
Other/Cloud
What browsers do you use to access the UI ?
Google Chrome
Command Line Arguments
--share --gradio-debug --no-half-vae --xformers
Additional information, context and logs
Using novel AI model, have never changed my setup apart from keep up to date with the repo.
was ca5a9e79dc28eeaa3a161427a82e34703bf15765 the commit the broke the denoise? I'll test it later locally to see if I get worse results.
Not sure about highres, but I've been using img2img a lot and under the same settings (near 512, same prompts, 0.45 strength, 28 steps 15cfg) I'm getting way worse quality results. Like blurry lines, and faces look worse, hands used to be sometimes good, now they're messed up 90% of the time, you get the idea.
Even 0.3-0.35 was fine for good outputs for me, but it just looks so bad for some reason. I haven't been keeping up with changelogs so I don't know what happened, but it wasn't like this when I used it last time about 5 days ago.
Oh and for me text2img seems worse in quality as well.
Example when of denoise levels with highresfix. It shouldn't be applying noise like this, if at all at level 0.
1girl, smile Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, (((deformed))), [blurry], bad anatomy, disfigured, poorly drawn face, mutation, mutated, (extra_limb), (ugly), (poorly drawn hands), messy drawing, broken legs Steps: 28, Sampler: Euler a, CFG scale: 11, Seed: 1624647752, Size: 1024x1024, Model hash: 925997e9, Denoising strength: 0.7, Clip skip: 2, First pass size: 0x0
old be1596ce30b1ead6998da0c62003003dcce5eb2c
--medvram --xformers
new 99d728b5b18829c8a6b7b2d69c9b9327dd257896
--medvram --xformers
Control Images
default model, seed 5,6,7 euler a, 20 steps, prompt: "frog" cfg 7
Img2img
denoise 0.7, seed 1, prompt: "frog" euler a, 20 steps
old
new
So its not anything to do with colab or the novel ai model, img2img differences are in the normal 1.4 model as well.
Same issue using txt2img. It is very noticeable for denoising from 0.4 and below.
Steps: 20, Sampler: Euler, CFG scale: 7, Seed: 1336575059, Size: 1024x1024, Model hash: 81761151, Model: v1-5, Denoising strength: 0.0, First pass size: 0x0
As a comparison, I want to just add that i2i denoise seems to work fine.
I am unsure if that i2i is what the highres fix result SHOULD look like, but if so, the result is desireable.
Just wondering if this would be related to the setting/feature where less steps are taken the lower the denoising level
Has there been any further response to this? I am running into these issues as well using t2i -- finding that anything under 0.4 with latent upscaling results in unusable blurry images, which kind of eliminates the possibility of using latent upscaling. I've moved to Lanczos upscaling but it doesn't have the ability to add detail the way latent can...
Thought that was normal, but maybe not? Using nearest for upscale and being below 0.5 denoise definitely results in blurry or pixelated images.
lora:aurora-style-richy-v1:1, aurorastyle, official art, unity 8k wallpaper, ultra detailed, beautiful and aesthetic, masterpiece, best quality, (zentangle, mandala, tangle, entangle), (fractal art:1.3), 1girl, extremely detailed, dynamic angle, cowboyshot, the most beautiful form of chaos, elegant, a brutalist designed, vivid colours, romanticism, by james jean, roby dwi antono, ross tran, francis bacon, michal mraz, adrian ghenie, petra cortright, gerhard richter, takato yamamoto, ashley wood, atmospheric, ecstasy of musical notes, streaming musical notes visible, Negative prompt: (worst quality:2,low quality:2), (interlocked fingers,badly drawn hands and fingers, anatomically incorrect hands), Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 9, Seed: 1072721686, Size: 384x704, Model hash: f1270c1f02, Model: zavychromaxl_v30,
p1:Denoising strength: 0.6
p2:Denoising strength: 0.4
Hypertile VAE: True, Hires resize: 768x1368, Hires upscaler: Latent, Lora hashes: "aurora-style-richy-v1: d331bdff9591", Version: v1.7.0
p2:Denoising strength: 0.4
Hires upscaler: Latent
Latent upscaler demands denoisinig strength to be larger than 0.5
p2:Denoising strength: 0.4
Hires upscaler: Latent
Latent upscaler demands denoisinig strength to be larger than 0.5
This is really old. But I'll explain. At the time of posting high-res fix was quite different, and until recently before that post, you used to be able to use Denise strength values of below 0.5 without a problem, or weird artifacts, some behaviour or code was changed, fairly certain of it.
Will close this. If anyone happens to have info related to the code behind this issue and how it was in the past feel free to share and reopen.