【Confusing】Potential Difference between code implementation and paper description

[lyf1212]

Thank you for your insightful work! However, some discrepancy between code implementation and paper description confuses me a lot.

In ./guided_diffusion/diffusion_llie.py#L269, it seems that you mix-up frequency value (which are complex numbers) directly, but in the paper Eq.(11), you mix-up only amplitude of low-/normal-light images.

It will be nice of you to clarify such important issue.

[lvxiaoqian]

Thank you for your interest! In fact, we only used the combined amplitude（L272）, which is equivalent to Eq.(11) in the paper.

[lyf1212]

Thank you for your interest! In fact, we only used the combined amplitude（L272）, which is equivalent to Eq.(11) in the paper.

In my opinion, L.272 combines the amplitude of mix-up frequency component and phase of low-light image. In other words, in your paper, you mix-up amplitude of low-light image and x_0_t first, then combines with phase of low-light image. This two are different process.

To be more clear, let us denote the low-light image in frequency by $a+bi$, x_0_t (which is maintained by DDIM prediction in your paper) by $c+di$. The implementation in your code is:
first, mix-up frequency is constructed, noted by $a+c * \gamma + (b + d * \gamma)i$, then compute its amplitude, noted by $\sqrt{(a+c*\gamma)^2 + (b+d*\gamma)^2}$. But in your paper, from my perspective, amplitude is combined first, noted as $\sqrt{a^2+b^2}+\gamma * \sqrt{c^2+d^2}$, then back-project frequency to image domain.

I'm just confusing about this. Could you provide more information about that? Thank you so much!