stable-diffusion-webui
stable-diffusion-webui copied to clipboard
AUTOMATIC1111
Just black out while image2image
I use the final pruned version of that hyperwork supported model(we know that),but always get black area while using mask of image2image. seem the result just blurred the black mask. Setting :origin,cfg 7-8,dnoisy0.5-0.7 I dont know whether is it because of the model or the version.
I'm having same issue. It works for 512x512, but above that, there is high chance of black image output. With with image creating progress enabled you can see that it creates image normally for the first 5-10 steps, and then image becomes completely black and keeps being so for the rest of creation.
Same issue. Got black blocks. Downgraded to the previous commit - works fine.
without check the full resolution,things not happen. otherwise the masking area just could not be large
seems a issue for lack of memory or something as it is not happened on none vae loaded models. 8vram gpu
I'm also seeing this issue with a 3090Ti/24GB. I was able to generate high resolution images with img2img consistently before but now they turn out black most of the time.
I think it may be an issue with having a VAE used for the current model.
Thats too bad and may be added to the list 1st bug.I dont know whether a111 is well these days.Hope could be fix.
It does seem that output becomes solid black if using vae and more often with higher resolutions. If I don't use vae it doesn't happen at all.
Similar behaviour here in a 3090Ti/24GB. Original image 512x728, img2img doubling the size using a model including VAE. With Euler A at 12 steps it generated the new image, but after increasing the step number I also got the black screen. Also for some reason, looks like decreasing the number of steps after getting that black screen (even to a quantity which was working before) doesn't help.
Switched the model to WD 1.3 (no VAE), steps increased to 28: output generated again without issues.
Confirm. No vae: OK vae : option --no-half-vae fixes black outputs for me!
Thank you for sharing! I hope Automatic1111 will include this fix into the new release.
Just add to webui-user.bat, its already there, like set COMMANDLINE_ARGS=--xformers --no-half-vae
And (for example), SD upscale your prompts )
Just add to webui-user.bat, its already there, like set COMMANDLINE_ARGS=--xformers --no-half-vae
And (for example), SD upscale your prompts )
Thanks. I add set COMMANDLINE_ARGS=--xformers --no-half-vae to webui_user.bat and run it by webui_user,not webui.bat. And having a trial.
sorry for necroing the thread, very similar problem with inpainting in my a1111, I have been on
set COMMANDLINE_ARGS= --medvram --opt-split-attention --precision full --no-half --xformers
but suddenly it gave me black out in the masked area until I added --no-half-vae, now sometimes it's just the masked area being blurred out from the original and not inpainted, other times it gives me out of memory error.
Adding to issue: Had that after one generation in img2img, just generating new images, i would get consistent black result. --no-half-vae seem to fix it for me, and no black boxes, huge thanks to you guys <3
I'm having same issue. It works for 512x512, but above that, there is high chance of black image output. With with image creating progress enabled you can see that it creates image normally for the first 5-10 steps, and then image becomes completely black and keeps being so for the rest of creation.
Same issue, why is that?
Adding to issue: Had that after one generation in img2img, just generating new images, i would get consistent black result. --no-half-vae seem to fix it for me, and no black boxes, huge thanks to you guys <3
What does no-half-vae even do? And why does it work?
I'm always on last commit and I use SD almost daily, but this happened to me just now... what could it be?
