[Bug]: Worse results with latest version

[Pisha99]

Checklist

• [X] The issue exists after disabling all extensions
• [ ] The issue exists on a clean installation of webui
• [ ] The issue is caused by an extension, but I believe it is caused by a bug in the webui
• [x] The issue exists in the current version of the webui
• [X] The issue has not been reported before recently
• [ ] The issue has been reported before but has not been fixed yet

What happened?

I use DPM++ 3M SDE Karras, after this new update, where sampler and schedule type are now separated, my results are much worse, they look like what my generations from a year ago looked like. It seems like it's kinda ignoring my negative prompts now. I get watermarks in all pictures, hands are horrible now. THIS didn't happen yesterday before I upgraded to this new version

Steps to reproduce the problem

DPM++ 3M SDE Karras 30 Steps CFG 5 Same Model Same Resolution

What should have happened?

Something related to this scheduler type. Negative prompt is weird now too

What browsers do you use to access the UI ?

Google Chrome, Other

Sysinfo

sysinfo-2024-04-14-22-16.json

Console logs

Already up to date.
venv "F:\DeepFaceLab\stable-diffusion-webui-SDXL\venv\Scripts\Python.exe"
Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug  1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]
Version: v1.9.0
Commit hash: adadb4e3c7382bf3e4f7519126cd6c70f4f8557b
CUDA 11.8
Launching Web UI with arguments: --opt-split-attention --xformers --no-half-vae --medvram --disable-safe-unpickle
==============================================================================
You are running torch 2.0.1+cu118.
The program is tested to work with torch 2.1.2.
To reinstall the desired version, run with commandline flag --reinstall-torch.
Beware that this will cause a lot of large files to be downloaded, as well as
there are reports of issues with training tab on the latest version.

Use --skip-version-check commandline argument to disable this check.
==============================================================================
=================================================================================
You are running xformers 0.0.20.
The program is tested to work with xformers 0.0.23.post1.
To reinstall the desired version, run with commandline flag --reinstall-xformers.

Use --skip-version-check commandline argument to disable this check.
=================================================================================
ControlNet preprocessor location: F:\DeepFaceLab\stable-diffusion-webui-SDXL\extensions\__sd-webui-controlnet\annotator\downloads
2024-04-14 23:24:54,441 - ControlNet - INFO - ControlNet v1.1.443
2024-04-14 23:24:54,694 - ControlNet - INFO - ControlNet v1.1.443
[-] ADetailer initialized. version: 24.4.1, num models: 12
23:24:56 - ReActor - STATUS - Running v0.7.0-b7 on Device: CUDA
23:24:56 - ReActor - STATUS - Running v0.7.0-b7 on Device: CUDA
[-] ADetailer initialized. version: 24.4.1, num models: 12
[AddNet] Updating model hashes...
0it [00:00, ?it/s]
[AddNet] Updating model hashes...
0it [00:00, ?it/s]
[Vec. CC] Style Sheet Loaded...
Loading weights [a6698c0f30] from F:\DeepFaceLab\stable-diffusion-webui-SDXL\models\Stable-diffusion\STOCK-NSFW.safetensors
Creating model from config: F:\DeepFaceLab\stable-diffusion-webui-SDXL\repositories\generative-models\configs\inference\sd_xl_base.yaml
Applying attention optimization: xformers... done.
Model loaded in 7.3s (load weights from disk: 0.6s, create model: 0.5s, apply weights to model: 5.5s, calculate empty prompt: 0.6s).
2024-04-14 23:25:06,073 - ControlNet - INFO - ControlNet UI callback registered.
Running on local URL:  http://127.0.0.1:7860

To create a public link, set `share=True` in `launch()`.
Startup time: 42.3s (prepare environment: 18.1s, import torch: 5.4s, import gradio: 1.6s, setup paths: 1.2s, initialize shared: 0.2s, other imports: 1.3s, list SD models: 0.1s, load scripts: 5.9s, create ui: 8.2s, gradio launch: 0.3s).
INFO:sd_dynamic_prompts.dynamic_prompting:Prompt matrix will create 5 images in a total of 5 batches.

Additional information

I have updated WebUI version. Not other changes were made.

[promptmachine]

I agree!

See also: https://www.reddit.com/r/StableDiffusion/comments/1c3jcia/confused_where_your_fav_sampling_methods_are_now/

I'm hating 80%+ of my generations now. They look far too "cartoonish", which is maybe fine if you're into that sort of things.

[AG-w]

Check your sampler and scheduler combination first and see if results are same

and if you know you use DPM++ 3M SDE Karras why don't you just pair DPM++ 3M SDE sampler with karras scheduler?

[Pisha99]

Check your sampler and scheduler combination first and see if results are same

and if you know you use DPM++ 3M SDE Karras why don't you just pair DPM++ 3M SDE sampler with karras scheduler?

What even are these questions? I used to use DPM++ 3M SDE Karras, now I use DPM++ 3M SDE and then I choose Karras scheduler

BAD RESULTS

[maxmarkeev]

Same. Last commit did changed UI. It added to pop up window for models. Now the result is seems to much worser quality.

[Woisek]

Can't confirm. But a memory issue still exists since v1.8.0.

[maxmarkeev]

Model: DPM++ 2M Schedule: Karras 20 Steps Resolution : 512x512 Prompt: photy face of a girl Neg: picture

Something is terribly wrong. It's not even a worse of results, just 3 in a row.

[liusan19]

I'm in the same situation. Results are worse than previous versions.

I used to use "DPM++ 2M SDE Karras". After upgrading, the Sampling method is "DPM++ 2M SDE" and the Schedule type is "Karras".

Other settings are the same, but the quality has obviously decreased.

Once I tried reverting the version back to 1.8.0. but the quality did not return.

[Pisha99]

I'm in the same situation. Results are worse than previous versions.

I used to use "DPM++ 2M SDE Karras". After upgrading, the Sampling method is "DPM++ 2M SDE" and the Schedule type is "Karras".

Other settings are the same, but the quality has obviously decreased.

Once I tried reverting the version back to 1.8.0. but the quality did not return.

Okay I thought I was going crazy after doing all I can to try and solve this by myself, when I found out that reverting to 1.8.0 version, which was working fine, didn't solve this and results are still worse. What's going on?

[Pisha99]

@AUTOMATIC1111 please we need help. I'm totally unable to continue my work.

[Woisek]

Model: DPM++ 2M Schedule: Karras 20 Steps Resolution : 512x512 Prompt: photy face of a girl Neg: picture

Something is terribly wrong. It's not even a worse of results, just 3 in a row.

What checkpoint did you use? This has surely nothing to do with A1111 itself.

[maxmarkeev]

Model: DPM++ 2M Schedule: Karras 20 Steps Resolution : 512x512 Prompt: photy face of a girl Neg: picture Something is terribly wrong. It's not even a worse of results, just 3 in a row.

What checkpoint did you use? This has surely nothing to do with A1111 itself.

v1-5-pruned-emaonly.safetensors

But it seems like checkpoint isn't really matter in that case

Here's the result wih realisticVisionV51_v51VAE.safetensors

This one is much better, as it should be. But clearly there's an obvious downgrade. Take a closer look on the arms. Eyes somewhere looks a bit of too. Hadn't have that kind of artifacts in a long run.

[Pisha99]

I made a collage of just hands to show the difference I'm getting. Prompts are mixed but they repeat equally in each version. 1.8.0 1.9.0

I think it's an obvious downgrade. And this is just hands!! Some prompts now give me bad skin textures, and overall, I'm getting watermarks almost 100% of the time now. Somehow I feel like the weight of the prompts have changed, like this hands thing may be because negative prompt is somehow broken, that's why faces are worse too, and watermarks are appearing now. I don't know where's the issue, but there's clearly an issue

[maxmarkeev]

By the way it's all going on on clean setup

[promptmachine]

I'm curious, how many of us are using Loras? Thumbs up = 1 Lora Heart = More than 1 Lora Thumbs down = No Loras

(If I don't use any Loras, my results don't seem so bad ...)

[liusan19]

This is a comparison between 1.8.0 and 1.9.0. The model and settings are the same, but the textures are no longer realistic.

This model is still better, but some models are more cartoony.

Same condition after complete reinstallation.

[AG-w]

I import old pic info then regenerate it again, the result is same

[Theliel]

ummmm I don't see any problems either., exactly the same as @Ag-W said, same model, same settings/seed... same result with 1.8.0 and 1.9.0. Tested with different models, with or without negative prompt.

Furthermore, for those who have the problem, it also doesn't make sense that when you put the previous version back, the problem you have now persists... if there is a particular real problem with this version, when you go back it would work fine again.

[liusan19]

For some reason, it suddenly reverted to its pre-update quality.

Today I reinstalled and lowered the version, but the quality remained poor.

I gave up and generated it with 1.9.0, and a few hours ago it suddenly reverted back to high quality.

Might be unrelated, but I updated the extension shortly before that.

[37OMKAR]

WE NEED THE OLD VERSION BACK WHERE THE SAMPLING METHOD WAS ONE, PLEAAAAASSSSSSSSEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEE ![Uploading download.jpg…]()

[37OMKAR]

version, when you go back it wo

the new one is more 3d, old one is more realsistic see the hairs, the colours are poping, its not what required, the old one was best

[37OMKAR]

also now i olnly need euler, i have to mandatory keep some type, previous i can select one single sampler only now why make life more complicated, also all the tutorials and all the pdf and tranning stuff will be useless, try to make it more easy not compicated by adding some thing that will make life more complicted, you can give a chek mark in extension if some one wants they can get it , those who dont want they will just tick mark off

[Theliel]

Good day!!

As has already been said in this thread, using exactly the same settings, the result is exactly the same, verified on several occasions.

On the other hand, it seems absurd to me to say that adding flexibility to the graphical interface means complicating something. If you used DPM++ 2M Karras before, you can continue using exactly the same. The difference is that now you can use DPM++ 2M SGM if you want.

This is a big positive change.

[37OMKAR]

Good day!!

As has already been said in this thread, using exactly the same settings, the result is exactly the same, verified on several occasions.

On the other hand, it seems absurd to me to say that adding flexibility to the graphical interface means complicating something. If you used DPM++ 2M Karras before, you can continue using exactly the same. The difference is that now you can use DPM++ 2M SGM if you want.

This is a big positive change.

how you use only elure ??? can you guide

[Asherathe]

Good day!! As has already been said in this thread, using exactly the same settings, the result is exactly the same, verified on several occasions. On the other hand, it seems absurd to me to say that adding flexibility to the graphical interface means complicating something. If you used DPM++ 2M Karras before, you can continue using exactly the same. The difference is that now you can use DPM++ 2M SGM if you want. This is a big positive change.

how you use only elure ??? can you guide

Change your sampling method to Euler, and make your other selections the way you prefer them. Click the Settings tab, then click Defaults at the bottom of the left menu, then click Apply at the top of the page. This will save Euler as your default sampling method.

[vixelfox]

I was a holdout in 1.7.0, finally updated straight to 1.9.0 and immediately experienced broken/unusable generations on lots of checkpoints that previously worked well. Found this bug report by googling similar issues.

I randomly loaded one of my old generations into img2img and found an override setting "Downcast alphas_cumprod: True" applied, and generations were fine as long as that flag was set. Changed the setting permanently and all generations (text2img and img2img) are good again. Dunno if it's related to this bug, or if my install coincidentally started working again at the same time.

Otherwise maybe there's a cached file or issue with an installation dependency (python, venv etc) for some people to see the problem fix itself on its own, or stay broken between installations/reinstallations/rollbacks.