It depends on artistic intent, right? Sometimes you want to set an unsettling or surreal mood. Not everything should look like D65. And it is not like you cannot use smoother illuminants with spectral rendering, too.
Though, I would be genuinely curious to hear other opinions and insights on this.

The state of the art report on Monte Carlo volume rendering by Jan Novák et al. [2018] gives an excellent introduction:
www.jannovak.info/publications...

Fair point. The image at the bottom is a teaser for subsequent parts that I added last minute. For anything beyond that, you will have to wait for part 3.

The spectrum is that of a high-pressure sodium vapor lamp, not incandescent. That context is provided in the blog post. Part 3 will have lots of similar comparisons with different illuminants. There is only so much you can expect from a 280 character summary 😉.

Thanks for letting me know, it should be working now.

Full source code for the renderer is on github:
github.com/MomentsInGra...
You can play around with the RGB to spectrum conversion on Shadertoy:
www.shadertoy.com/view/X3cBRl
Part 2 and 3 will follow in the next two weeks. Stay tuned!

'Tis the season 🎃.

BTW: If things go as planned, there will be something new on momentsingraphics.de for five weeks in a row. Stay tuned!

The underlying theory of unique minimum variance unbiased (UMVU) estimates can provide high-quality Monte Carlo estimates of all sorts of non-linear functions of random variables. The paper also proposes new estimates for distance sampling MIS weights. What will you use it for?

Mariia Soroka, the first author, did excellent work here. And it has been a pleasure to work with Steve Marschner. Since they are not on Bluesky, I am sharing this here.
Mariia also made a project webpage for the paper:
mariasoroka.github.io/papers/EdgeS...

My original motivation is in this paper:
momentsingraphics.de/VMV2023.html
Computing AABBs for SH glyphs is facilitated there by SH gradient and Hessian computation.
Another application could be inverse rendering with SH that benefits from faster gradient evaluation. Maybe Gaussian splatting?

Sec. 3.5 discusses what we do (and do not do) regarding RGB previews.

If we'd only want to get tristimulus values right, we'd be storing an RGB image. Our rationale is that we do not know exactly what downstream tasks the spectral image will be used for, so we avoid strong assumptions on which parts matter and try to preserve each spectrum in its entirety.