LUTs with values outside [0,1] range

[Tchaikovic]

Hi there,

This is probably more a modelling question rather than an issue with the code, but it might be worth discussing here anyway.

I am training your model and found that trained LUTs can actually map normalized RGB to values outside the range [0,1]. Don't you think this makes little sense?

[HuiZeng]

Hi， I also noticed this issue. This is a commom issue that many learning based image restoration or enhancement models have. That major reason may be that the model tend to optimize the entire data distribution and the values at corner are not fitted well. Usually, the output image is simply clamped before writing. In most cases, this will not be a big issue for the final visual quality, so I did not put explicit constraint on the output range of 3D LUT.

[Tchaikovic]

Thanks for your answer! Have you thought about clamping the LUTs after training and then fine-tune the weight predictor network? In this way at least the base LUTs would be in the same range as commercial luts. It might also be that trying to optimise the LUT and the weight predictor at the same time causes the issue.

[HuiZeng]

Hi, this sounds reasonable. But I have not tried it yet. You may update your findings if you make some attempts. Thanks.