Black output image using 4x_NMKD-Superscale-SP_178000_G

[reijin90]

Information:

• Chainner version: chaiNNer 0.19.3
• OS: macOS 12.6.7 (M1)

Description
Upscaling images with 4x_NMKD-Superscale-SP_178000_G.pth or 4x_NMKD-SuperscaleV2_46k.pth result in a black output image.
Update1: I tested with both fp16 mode enabled and disabled, same result. I ran the chain in CPU mode and it properly produced an image. So it is probably acceleration related. I re-installed pytorch, but still black images outside of CPU mode. Is there an other upscaler system I should use instead of pytorch?

Logs
https://www.dropbox.com/scl/fi/ksm8lhyxvik0pmumjc7a1/chainner_logs.zip?rlkey=vj9fategc42z1uu36miul0j1w&dl=0

Project:
upscale.chn.zip

Models:
https://nmkd.de/shared/ESRGAN/Models/Realistic%2C%20Multipurpose/4x_NMKD-Superscale-SP_178000_G.pth
https://nmkd.de/shared/ESRGAN/Models/Realistic%2C%20Multipurpose/4x_NMKD-SuperscaleV2_46k.pth

[joeyballentine]

I don't see anything in the logs that would indicate an issue, but just to check: do you have fp16 mode enabled? If you do, try disabling it. If that doesn't do anything, we might need @stonerl's help since this is probably a macos issue

[reijin90]

I had it enabled, but only as a test to see if it works when enabled. I disabled it, restarted the tool and it is still producing black images.
Updated the main question.

[stonerl]

@joeyballentine found the issue. We need to upgrade to PyTorch 2, at least on Apple Silicon.

[stonerl]

@reijin90 I made a test build for Apple Silicon that uses PyTorch 2.0.1 and TorchVision 0.15.2. You can download it from here:

https://www.dropbox.com/scl/fo/262tme9vznofa0wjt2j1e/h?rlkey=vt8vlbilkz4ant18sbrj9s01b&dl=0

Upgrade the dependencies and run your chain again.

[joeyballentine]

Well we wanna switch to 2 anyway in general so I guess I'll prioritize that

[reijin90]

works perfectly @stonerl , thank you! :)