Summary
The current ALab Gaussian splat was generated from 8-bit sRGB PNG images (both HDR and SDR versions) due to pipeline limitations at the time of generation. This falls short of a full ACEScg pipeline and limits the quality and colour accuracy of the splat. This issue tracks the work to regenerate the splat using a proper ACEScg 16-bit half float pipeline.
More documentation available here
Current Limitations
- Training images were 8-bit sRGB PNGs
- No full ACEScg colour pipeline was used during generation
Proposed Work
Goal
Produce a reference-quality ACEScg Gaussian splat of ALab alongside clear documentation so the community can reproduce the result with their own assets.
References
Maybe useful diagram
Doesn't need to be 32bit!! - 16bit half float is good enough! this was theorycrafting :)

Summary
The current ALab Gaussian splat was generated from 8-bit sRGB PNG images (both HDR and SDR versions) due to pipeline limitations at the time of generation. This falls short of a full ACEScg pipeline and limits the quality and colour accuracy of the splat. This issue tracks the work to regenerate the splat using a proper ACEScg 16-bit half float pipeline.
More documentation available here
Current Limitations
Proposed Work
Goal
Produce a reference-quality ACEScg Gaussian splat of ALab alongside clear documentation so the community can reproduce the result with their own assets.
References
Maybe useful diagram
Doesn't need to be 32bit!! - 16bit half float is good enough! this was theorycrafting :)