My third and final blog post on spectral rendering is out. This one compares results of RGB and spectral rendering under different illuminants and draws conclusions.
https://momentsingraphics.de/SpectralRendering3Results.html
Remote
Christoph Peters
@MomentsInGraphics@mastodon.gamedev.place
mastodon 4.7.3Graphics researcher at TU Delft. Formerly Intel, KIT, NVIDIA, Uni Bonn. Known for moment shadow maps, MBOIT, blue noise, spectra, light sampling. Opinions are my own.
https://MomentsInGraphics.de
0 Followers
0 Following
6 Posts
Joined November 08, 2022
Replying to
@yiningkarlli@mastodon.gamedev.place Gaussian splats are a "learning-based scene representation." Calling it AI is just trying to link it to the hype.
It may also be partially motivated by where we came from: Networks that do the rendering got displaced by NeRF, which got largely displaced by Gaussian splatting. Every step forward has removed an AI component, to the point where there is no neural network left, but people still associate novel view synthesis with AI.
Also the papers get published at AI venues.
In 2026, I reprise my role as High-Performance Graphics papers chair, joined by Markus Kettunen.
The papers deadline is 8 weeks from now. Submit your performance-oriented graphics research and present it in Los Angeles, just before SIGGRAPH (July 17-19 2026)!
https://www.highperformancegraphics.org/2026/call-for-participation/
My second blog post on spectral rendering is out. This one explains how to use wavelength importance sampling to minimize color noise while limiting the overhead to 0.3 ms and how to make BRDFs spectral.
https://momentsingraphics.de/SpectralRendering2Rendering.html
Replying to
@Dimas@social.edu.nl Short answer: No.
The product of multiple spectra can simply be a more complicated function than one on its own. Either you accept strong approximation errors or you use so many coefficients that you are not gaining anything by moving away from Monte Carlo integration.
That is already true for products of reflectance spectra. And emission spectra need more coefficients to begin with, see Fig. 9 in my SIGGRAPH 2019 paper:
https://momentsingraphics.de/Siggraph2019.html
Replying to
@Dimas@social.edu.nl Is it? Color noise is really not that much of a problem and evaluating multiple wavelengths does not cost much.