The example script produces a black image.

[maray29]

I just tried again and the output is a black image. I used the beach example.

Console output:

100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:03<00:00,  5.66it/s]
100%|████████████████████████████████████████████████████████████████████████████████| 236/236 [01:37<00:00,  2.42it/s]
D:\Code\stable-diffusion\ComfyUI_windows_portable\ComfyUI\nodes.py:1140: RuntimeWarning: invalid value encountered in cast
  img = Image.fromarray(np.clip(i, 0, 255).astype(np.uint8))
Prompt executed in 112.21 seconds

[BlenderNeko]

I can not reproduce this on my system. what is your GPU, OS, and what model did you use?

[maray29]

RTX 3090 FE, Win 10 22H2, Model: Silicon29.safetensors, CC model: control_v11f1e_sd15_tile_fp16.safetensors.

Also tried with vanilla vl-5-pruned-emaonly.safetensors.

The problem must be from my end, because I've had problems with almost all SD implementations. I'd appreciate if you have any suggestions what I can try/test.

[iamrohitanshu]

@maray29 Have you tried applying a VAE separately?

[Lerc]

I'm getting this error, but since I'm using a tiled VAE decoder, I get different symptoms. The VAE decodes most tiles correctly but some come out as black. It doesn't always happen for me though, perhaps dependent on the seed. Interestingly there's a 128px wide strip of correctly decoded within the black. The tile size for the Sampler and the VAE decode was both 512px.

[Lerc]

Doing some experiments with multiple generations and adjusting the VAE tile size. It seems that the black region is always associated with the bottom right corner.

Is it possible that a NaN is propagated by reading beyond the image data in the corner of the image? A literal corner-case if you will.

[Lerc]

A few more tests and the problem does not appear to occur on "random strict" strategy