Wrong colors in Gaussian Splats when using linear workflow

[MotivaCG]

I've noticed a significant color difference in Gaussian Splats when switching from the "Classic" workflow to the "Modern" (Linear) workflow. While I understand that the Modern workflow uses sRGB for better accuracy in general, the Splats appear to have gamma changes but color issues compared to the original look.

Are these color changes expected behavior for Splats in Linear space?

Is this a known issue with the current Splat rendering implementation?

Is there a specific tweak or pipeline setting I should adjust to match the original colors?

Screenshots Comparison below: Left: Classic/Default workflow | Right: Linear/Modern workflow.

Left classic-default workflow, right linear-modern workflow.

Thank you in advance!

[slimbuck]

Hi @MotivaCG,

Are you training your scenes (and so wanting to render) in linear space?

Most GS trainers assume gamma space and so the GS colors and spherical harmonics are all stored in gamma space. In this case the gaussian model must be blended in gamma space, otherwise the result will look wrong. This means you also can't correctly composite these GS models with linear-space traditional rendering.

However, if you're training in linear space it is possible to make this work. I understand that training in linear space is easy for trainers to do, but currently will result in lower-quality output.

If you have any experience with this on the trainer side I'd love to hear it!

Thanks!

[MotivaCG]

Sure I'm one one the (minor) contributors of lichfeld so I could code that on my own, but my question was more related to any posibility to tweak/fake the blending with pre-made splats.

[slimbuck]

Ah ok so I suppose you're working in the PC editor and using camera frame?

In that case you'll have to:

• render gaussian splats in gamma space by patching the GS shader (the supersplat-viewer does this here).
• after the main render, if you want to then perform linear-space postprocessing you must do a fullscreen blit converting gamma->linear
• then apply post effects as usual

I'm not sure the details of the camera-frame side of things, @mvaligursky should be able to help.