SDXL Img2Img problems - any ideas? #12187
Replies: 4 comments 7 replies
-
In the latest commits you can also enable "Automaticlly revert VAE to 32-bit floats" in Settings > Stable Diffusion without Actually saw it in action for the first time today, image generation stopped for a bit but then continued and finished normally without black squares. |
Beta Was this translation helpful? Give feedback.
-
Still not working for me after OP suggestions. Getting this error now: *** Error completing request 0%| | 0/20 [00:01<?, ?it/s] Any ideas? Also I don't have SDXL vae, i'm not sure if it matters. In fact simply selecting SDXL VAE results in an error of its own: |
Beta Was this translation helpful? Give feedback.
-
Same problem same GPU, tried all suggestions above nothing works. |
Beta Was this translation helpful? Give feedback.
-
Hello I'm using (img2img alternative test) option in Img2img and when I click on generate I get this error : TypeError: expected Tensor as element 0 in argument 0, but got dict Is this related to your issue? |
Beta Was this translation helpful? Give feedback.
-
Problem fixed! (can't delete it, and might help others)
Original problem:
Using SDXL in A1111.
no problems in txt2img, but when I use img2img, I get:
"NansException: A tensor with all NaNs was produced in Unet. This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the "Upcast cross attention layer to float32" option in Settings > Stable Diffusion or using the --no-half commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check."
I get this for both xl_base_1.0.safetensors and xl_refiner_1.0.safetensors
img2img works fine for all other checkpoints.
Tried SD VAE on both automatic and sdxl_vae-safetensors
Running on Windows system with Nvidia 12GB GeForce RTX 3060
--disable-nan-check results in a black image
--no-half causes an error in the startup batch file
changing the setting for Upcase cross attention... makes no difference
changing the resolution of the input and output images makes no difference
adding --medvram and --no-half startup arguments resolved the issue.
Thought I'd tried that combination previously - obviously not!!!!! (doh!)
Beta Was this translation helpful? Give feedback.
All reactions