Study: 3DGS Geometry Is Luminance-Dominated — Chroma-Only Supervision Costs 3.9-5.5 dB
kwangmoo_yi · x · 2026-10-03
A new arXiv paper, "What Builds the Scene? Luminance Dominates Geometry Formation in 3D Gaussian Splatting" by Joshaghani & Cutchin, disentangles how luminance and chroma contribute to 3DGS geometry. The method trains models under different channel supervision, freezes non-appearance parameters (position, scale, rotation, opacity), and re-estimates appearance with the same solver before comparing held-out reconstruction.
Key findings across eleven benchmark scenes (four runs each):
- Geometry learned from luminance alone scores just 0.085 dB below RGB-trained geometry on held-out reconstruction.
- After deleting chroma from a trained model, a sufficiently expressive solver can re-fit appearance on the frozen geometry to original quality or slightly better.
- Higher-order spherical harmonics benefit luminance far more than chroma: +1.44 dB vs +0.19 dB PSNR, though hue still shifts with viewpoint on mirror-like surfaces.
- The gap is larger during geometry formation: chroma-only supervision yields geometry 3.9-5.5 dB worse than luminance-only, partly explained by densification differences.
Conclusion: geometry formation in standard 3DGS is strongly luminance-dominated but not luminance-exclusive; much of chromatic appearance is a post-hoc layer. The author who shared it notes this echoes the old computer-vision practice of inferring shape from grayscale images.
Related event: Study Finds Luminance Dominates Geometry Formation in 3D Gaussian Splatting(2 posts)→
More from Research
- ENCODE preprint: neural nets predict super-additive GATA motif epistasis (r=0.55) — anshulkundaje · 2026-11-19
- New TLAPS-Bench blog: formally proving correctness of real-world system specs with TLA+ — tianyin_xu · 2026-11-19
- Stanford's Kundaje: ChIP-seq resolution limits motif-level deep learning analysis — anshulkundaje · 2026-11-19
- Mahina 4.0 LoRA Adapters Claim Fewer Hallucinations, Devs Seek Stress-Tests — External-Fee-8920 · 2026-11-19
- The biggest advantage of a top lab isn't money — it's saving you years — madeofAjala · 2026-11-19
- LEV: Looped decision model lifts 2.6B accuracy 63.2% to 76.2% with adaptive compute — jiqizhixin · 2026-11-19