Is it possible to also share the code for Figure 1b (ideal denoiser outputs) in the paper

[cuopbiensaysong]

It is very helpful for understanding Appendix B.3 (Denoising score matching (Eq.2 and Eq.3)).

[mehdeh]

Hi @tien-th ,I've implemented the ideal denoiser (Equation 57) from the EDM paper. It's an unofficial implementation, but it should help with understanding Appendix B.3.

You can find the code here: https://github.com/mehdeh/ideal-denoiser

The main implementation is in ideal_denoiser.py. It includes:

• Equation 57 implementation with numerical stability (log-sum-exp trick)
• A CLI tool to generate visualizations similar to Figure 1b
• CIFAR-10 examples showing the denoiser behavior across different noise levels

Quick example:

from ideal_denoiser import ideal_denoiser
denoised = ideal_denoiser(x_noisy, sigma, x_train)

Hope this helps! Let me know if you have any questions.