Terrible generations

[VenomousValentine]

Has anyone noticed their generations looking absolutely terrible since the latest update? Bad anatomy, jank faces, just lots of inconsistency.

[pw405]

Yes... and I have been tinkering with command line arguments trying to figure out what is happening.

I've even had results like this, but when using --precision autocast in the batch file. HOWEVER, I determined this was the result of using textual inversions "Bad Dream" and "Unrealistic Dream" (from Lykon, creator of popular Dreamshapers model). Without the textual inversions, it would generate successfully. (Using a 7900XTX).

It isn't completely broken, but something just seems "off" lately. I haven't got to test much.

This is generally safe to use for me:
--medvram --opt-sdp-attention --no-half --precision full

Supposed to be Aragorn from Lord of the Rings :

[EygenCat]

Thank you so much, I couldn't understand what was wrong with Textual Inversion. When using Textual Inversion, everything was breaking, but without Textual Inversion, everything was working fine. And after I added the command line arguments
--medvram --opt-sdp-attention --no-half --precision full,
everything became great.

[EygenCat]

But my Hires.fix function stopped working. I removed a few arguments, and now my arguments look like this:
set COMMANDLINE_ARGS=--opt-sub-quad-attention --disable-nan-check --precision autocast --autolaunch --listen --no-half --precision full
it seems like it's generating images perfectly fine.

[pw405]

Oh good! Note, the --opt-sub-quad-attention can be set in the Web UI options in the optimization section if you don't want to have it in the command line. I started selecting sdp attention there instead of command line.

[EygenCat]

Unfortunately, I couldn't make Hires.fix work properly. It only works intermittently and most of the time it throws a
RuntimeError: Could not allocate tensor with 1207959552 bytes. There is not enough GPU video memory available! error.
I have rx 6700 xt

Apparently I will be sitting on the old version

[pw405]

See, I've never used HiRes fix. Is this any different than simply upscaling using the Extras section?

Another method I use for Upscaling is the Control Net Tile Resample along with "Ultimate SD Upscale" script. It re-samples an image using ~15 512x512 images. Excellent results, and supposed to be easier on VRAM. May be worth checking out!

[EygenCat]

"Hires.fix" is needed to generate images in a higher resolution without distorting the composition. Essentially, it upscales the image using one of the provided methods and then performs img2img processing.

[VenomousValentine]

None of any of this helped. I even tried reinstalling, and my generations are still ugly as sin despite using the same prompts I used to.

[fishonainternet]

I'm currently in big regret after updating my Stable Diffusion to the newest version (1.4.1) haha, is there a way to revert back to old version?

[pw405]

What GPU are you using? This seems really strange that the new version created so many issues.

[VenomousValentine]

6600 xt

[Siralob]

I had same problem. Everything generated was like above post. I had to reset my webui because I re-install my venv after I upgraded latest to try to solve this issue. I thought it's broken for me. I also tried switched to spd, other optimization... and webui said no to me.

I reset back from latest to Commit HASH:[265d626] and work again.
But, I still got plane image or black image while generating with Doggtexx randomly. So, I have no idea. Also, some error code like mat1 and mat2... something like that.(sorry I forgot to backup them)

Which commit hash you used before this upgrade? Maybe, you could find your commit hash you used before on here.

Before the reinstall or git reset --hard [CommitHASH] just put your desire HASH no need [ ].
I revert back to old version, then re-install my venv, too.

Before you reset, these files will be replaced to original.

webui-user.bat
config.json
ui-config.json
params.txt
styles.csv
user.css

my webui-user.bat settings

set PYTORCH_CUDA_ALLOC_CONF=garbage_collection_threshold:0.6,max_split_size_mb:128
set COMMANDLINE_ARGS= --medvram --precision full --no-half --opt-split-attention --disable-nan-check --api --autolaunch
set SAFETENSORS_FAST_GPU=1

My question, Does anyone know which commit hash is the most suitable and stable for 8GB card like RX400 and RX500?(e.g. RX580 8GB) Also, RDNA2 card can be different for stable version.
I have no luck with hires.fix nor above 768x768 resolution for OOM (sometimes works with 512x912, 512x760).

[EygenCat]

Version 1 5 1 works great!!!

[VenomousValentine]

I tried the latest update. its still terrible. Still haven't been able to install previous versions that work.

[pw405]

What command line arguments are you using?

[VenomousValentine]

I've tried a bunch. they don't change anything.