3DGS Vendor Evaluation

In August 2023, a paper called '3D Gaussian Splatting for Real-Time Radiance Field Rendering' landed at SIGGRAPH. Within 6 months: 2,000+ citations, GitHub implementations in every major engine, and venture capital pouring into 3D content startups. The technology is genuinely transformative—where NeRF requires 10+ seconds per frame, 3DGS renders at 100+ FPS in real-time. Game studios, AR platforms, and film VFX pipelines are racing to adopt it.

The problem nobody solved: enterprise content pipelines need standardized, auditable quality from vendors. What does a 'good' 3DGS asset even mean? No standards existed. Appen needed to source 3DGS content at scale for AI training datasets, which required evaluating vendors systematically—not just 'looks nice' but measurable, reproducible quality criteria.

We designed the evaluation framework from first principles. Geometry: does the HP mesh have a correct world-space pivot point? Is the LP version provably retopologized (not just decimated)? Texture: are all 4 PBR maps present (BaseColor, Normal, Metallic, AO) with correct naming conventions and no duplicate sets? File management: zero spaces in filenames, no double extensions (.json.txt), proper directory hierarchy, metadata.json with real values (not TODO placeholders)? Preview images: 8+ renders at defined camera angles with standardized naming?

What we discovered was diagnostic: the vendors who failed weren't technically incapable of producing beautiful 3DGS assets. They were operationally immature—no internal QA, no naming conventions, no understanding of downstream pipeline requirements. The evaluation framework didn't just score vendors; it revealed which ones could be trained to meet enterprise standards versus which ones needed to be cut. The multi-round CSV tracker made every revision cycle auditable: vendor X failed geometry in round 1, fixed it in round 2, introduced a new naming error, fixed that in round 3—full history, zero ambiguity.