vladmandic/automatic

[Issue]: Error: upcast=True failed to validate // Generating SDXL

Closed this issue · 9 comments

meditat2001 commented

Issue Description

First thanks and congratulations a lot for this incredible sd.next!!!!
Clean install with the latest SDnext (from yesterday) dev-master. Generating with trained model, 1k x 1k, no extensions, lora etc used. No upscaler, hires, etc.....
Openvino, cache is enabled.

Error:
ERROR Decode: sample=(1024, 1024, 3) invalid=3145728 mean=0.0 dtype=float32 vae=torch.float32
upcast=True failed to validate

This error comes in every version since sd.next updated=2024-07-24.

Version Platform Description

Win 11, latest Opera, Intel Iris Xe, 16GB Ram,

Relevant log output

Torch parameters: backend=openvino device=GPU config=Auto dtype=torch.float32 vae=torch.float32 unet=torch.float32 context=no_grad nohalf=True nohalfvae=True upscast=False deterministic=False test-fp16=False test-bf16=False optimization="Scaled-Dot-Product"
2024-10-23 19:23:40,299 | sd | INFO | shared | Device: device=Intel(R) Iris(R) Xe Graphics (iGPU) openvino=2024.3.0

Backend

Diffusers

UI

Standard

Branch

Master

Model

StableDiffusion XL

Acknowledgements

  • I have read the above and searched for existing issues
  • I confirm that this is classified correctly and its not an extension issue
meditat2001 commented

Yes the error is gone with the VAE! But now the image has only slightly(good as nothing), to do with the prompt. I used 2 SDXL models with built-in-VAE. I can try the SDXL 1.0 base model?
Thanks!

vladmandic commented

vae has no impact on image composition, it only does final pixel decode.

meditat2001 commented

ok. I am just saying, the error is coming if I use a trained sdxl model with a VAE built in. The error didn’t appear in the version 2024-07-24 with the same model.

And if I use the specific sdxl VAE, it works, but the output is so big different to the output of sd.next latest version (both same parameters)
output: version 2024-07-24. vs latest:

00312-realvisxl5_steps_0596000_float16-In watercolor beautiful scenary nature small house on a small river a few people Art zen style

download

meditat2001 commented

Maybe it is the openvino version that changed?
SD.NeXt version 2024-07-24 uses openvino=2023.3.0
vs latest version: openvino==2024.3.0

Disty0 commented

Might be. 2024.4.0 has this issue on Intel ARC.
Try manually downgrading to 2024.2.0 or upgrading to 2024.4.0

  • ❤️1
meditat2001 commented

Great Tip!!!
I guess, I can't do the version change pre-install?

I copied in venv\lib\site-packages\ the folders
openvino
openvino_telemetry
openvino_telemetry-2024.1.0.dist-info
openvino-2024.2.0.dist-info
and deleted the current version.

With 2024.2.0 it works! With 2024.4.0 it generates only a black image.
And also no additional VAE, but with the VAE included in the model.

Thanks a lot! Great Work!!!

Disty0 commented

exporting OPENVINO_PACKAGE="openvino==2024.2.0" environment variable will override the SD.Next's default version.

  • 👍1
meditat2001 commented

ah ok, thanks! ok, for me this thread can be closed now...