AtlasLC: Fast Codec-Ready Compression of Object-Centric 3D Gaussian Splatting

AtlasLC is a new compression pipeline for object-centric 3D Gaussian Splatting that is built around deployment realities rather than just image-space quality. It runs source-free and training-free, so teams can compress released Gaussian assets without needing original images, camera poses, or per-asset optimization.

The core idea is to combine local-competition pruning with deterministic atlas packing, while using a lightweight single-pass sort-based conditional transport as the shared coordinate backbone. That removes the usual mapping/remapping bottleneck and keeps foreground support intact, which matters when the asset is a reusable object rather than a full captured scene.

For developers building XR asset libraries, the practical win is speed. AtlasLC cuts atlas-preparation time by up to 25x and end-to-end compression time by up to 5x across the evaluated assets, while balancing payload size, decode latency, runtime FPS, and 3D geometry against the baselines it was tested with.

The broader takeaway is that object-centric 3DGS compression may need different priorities than scene-scale capture pipelines. In this case, the goal is not just smaller files, but a codec-ready asset format that can be packaged, transmitted, decoded, and instantiated repeatedly without turning preparation into a pipeline bottleneck.