Unable to reproduce the Multinex LOL-v1 training performance using the official configuration #7

Hi authors,

Thank you for releasing the code and pretrained models for Multinex.

I successfully evaluated the official pretrained checkpoint on the LOL-v1 test set and obtained the following results:

• PSNR: 23.1828 dB
• SSIM: 0.8426
• LPIPS: 0.1248

These results are consistent with the performance reported in the paper, which suggests that my LOL-v1 test dataset and evaluation pipeline are likely correct.

However, when I retrained the model from scratch using the released configuration, I was unable to reproduce the reported performance.

My training command was:

python -m basicsr.train --opt Options/Multinex_LOL-v1.yaml

The best checkpoint obtained during training is shown below: