The results of the experiment

[ZackH2002]

Hello esr-nerf authors,thanks for the great work!
The experimental visualization I ran seems to have some issues.
The following images show the NV results from the PDRA phase of the lego_v dataset.

Below are the results of editing light colors on the lego_v dataset.

[ecrireme]

Hi, thank you for taking the time to read our work with interest.

From the results, it looks like the effect of the emissive source has not fully propagated through the scene, which usually indicates that some hyperparameter adjustments are needed.

Here are a few practical tips for tuning:
• If the emissive influence does not sufficiently reach other parts of the object, try training for more iterations.
• If the object appearance changes abruptly with unintended colors or intensities, reduce the learning rate so radiances evolve more gradually.
• If the reconstructed emissive source properties diverge from the intended edit, the LTS satisfaction can collapse too sharply. In that case, slowing down the rate of change (via lower learning rates or longer schedules) is necessary.

This work was conducted about two years ago, so I may not recall every detail precisely. However, I do remember that scenes with complex geometry (e.g., the Lego example) were especially challenging—achieving the desired color and intensity often required very careful tuning.

For more details on relighting issues and alternative approaches to radiance fine-tuning, please see Section 6 and Appendix 8.10 of the paper. As we also noted in README:

“Although we demonstrated the feasibility of controlling emissive sources by reconstructing them from LDR images, there are limitations in re-lighting methods. Re-lighting via radiance fine-tuning is slow and unstable. Adjusting learning rates and iterations is often necessary to achieve the best scene-editing results.”

Developing relighting methods that are more robust to hyperparameter choices would indeed be valuable, and remains an important direction for future research.

Sincerely,
Jinseo 😊

[ZackH2002]

Thank you for your response. The hyperparameters and iterations I'm currently using are both sourced from this repository, and I haven't made any modifications.

[ecrireme]

In the repository, we provided the general settings that we found to work well across a variety of cases. These configurations can serve as a good starting point, but achieving optimal results still requires hyperparameter tuning for each editing scenario. This is why we highlighted the re-lighting technique as a limitation in the paper.

(Updated comments for better readability 🙂)