HDR ACEScg Gaussian Splatting

I’ve noticed that .ply files trained with HDR ACEScg EXRs don’t get rendered correctly in VRay. The HDR ply files have been varified to be correct, i.e. when rendering directly from Postshot.
Of course I chose an OIIO workflow in 3dsMax, and selected ACEScg primaries for the 3DGS file.
I’m assuming that the VRay GS tool isn’t interpreting the high dynamic range - values above 1 - correctly?

Mind if you attach/send such a GS so we can troubleshoot?

I’ve sent you an ACEScg .ply file.
Thanks for looking into this!

Received. Will test things out and report back.

Great! This was a VERY limited scene, just 16 bracketed camera postions, ensuring very fast turnarounds for 3DGS training variations.
Once you’ve found a solution - I’m guessing the GS floating point values currently get clamped to an 0-1 gamma - I’m happy to test a nightly build on a “REAL” scene.

There’s no clamping, and you can see it by looking at the pixel values in the VFB. It looks to me as if the splats are not read as Raw in the first place, which leads to the high contrast (as if an inverse gamma is wrongly applied within the splat). Will ask the devs for opinions.

Thanks for investigating this!
Looking forward to the solution from the devs :slight_smile:

Are you sure this splat is ACEScg? When I open it in PostShot, it says it is in sRGB (SDR). Check the attached screenshot - the results seem pretty close to me.

Yes, it’s fully ACEScg linear. Ply files usually don’t have any meta-information about the color space.
Re-importing the ply not trained with a new version of Postshot is therefore wrongly interpreted as sRGB.

But if one manually sets it, like the option Vray provides, they should then be interpreted as linear ACEScg, because that’s what the underlying floating point RGB values in the ply are.
Postshot has now added a comment field to plys with the color space (I reported it on their forum some time back) but that only applies for Plys trained in Postshot with a new pre-release update.
I can send you furher material illustrating this if you wish. Anyhows, the ply values are correct ACEScg, the interpretation isn’t ;-);-):wink:

And thanks so much for looking into this!

And here’s the backup of my above input, using a REAL scene, with all its 32-bit ACEScg-linear adanvtages :slight_smile:
Plys either have correct ACEScg values or not. Postshot has now added a comment field into the ply header: re-importing an ACES ply then also gets displayed correctly in its App.
But manually selecting could be a better option IMHO, as long as we don’t have a better container-format with clear, compabitible meta-data, rather than .ply.

Well, in Vray we can manualy select the color space :slight_smile:




Well, if that’s so, it’s strange that the results match only if I set the RGB primaries of the GaussianGeom to sRGB. Regarding the rawness I mentioned earlier - that’s a confusion on my end - splats values are not Raw (but depend on what they’re trained on). Do you get similar results with the second .ply?

Not quite sure what I can add here, but please feel free to DM me!
I think everything’ s pretty clear, but maybe not…
Don’t want this to become too confusing for the expert audience here…

Try unticking the “sRGB-trained splats” option in Render Setup > Settings > Gaussian splat options. Also, since there is some differences in the sharpness of the splats, you can lower the Filtering value to 0.2-3. I just got aware of those options myself.

Btw, back to the initial topic: my 16-/32-bit linear Aces 3dgs plys are working fine - and as expected - in Arnold/ Postshot, also concerning gamma.
Feel free to reach out if you need such 3dgs scenes for getting Vray 3dgs on track.