Skip to main content
Back

GoDe Streams 3D Scenes 70x Faster Than Single-Rate Codecs

A new framework reorganizes any trained 3D Gaussian Splatting model into a single scalable bitstream, offering eight quality levels without retraining.

  • GoDe turns any trained 3D Gaussian Splatting model into one progressive bitstream with eight quality levels.
  • Three stages: gradient-based hierarchy, one quantization-aware fine-tune, chunked entropy coding with shared Zstd dictionaries.
  • Decoding takes 0.5 to 1.3 seconds, one to two orders of magnitude faster than existing splat codecs.
  • Works across five backbones (Gaussians, triangles, neural anchors, octrees, Beta Splatting) without modification.
  • Rate-distortion matches single-rate state of the art while covering the full curve from one model.
  • Code, demo and paper available at gaussians-on-demand.github.io.

Researchers from the University of Turin, Japan’s National Institute of Informatics, and Télécom Paris have introduced GoDe, a framework that converts a trained 3D splatting model into a progressive stream. A client downloads the base layer first, then adds enhancement chunks as bandwidth, memory, or rendering capacity allows. Every prefix ending at a layer boundary contains a complete, renderable scene.

One model per bitrate adds up

3D Gaussian Splatting represents a scene with millions of translucent primitives. Each primitive stores properties such as position, scale, rotation, opacity, and view-dependent color. These models render quickly, but their large parameter counts make compression essential for storage and delivery.

Most splat codecs optimize one model for one target bitrate. Serving several quality levels therefore requires separate training runs, files, and release workflows. Existing progressive approaches often depend on a specific representation, such as Scaffold-GS. GoDe works at the primitive level, allowing the same conversion pipeline to support several kinds of splat model.

Inside the progressive stack

GoDe takes a pretrained scene through three stages. The conversion preserves the underlying renderer, although it still requires optimization against training views.

GoDe pipeline for ranking, fine-tuning, and progressively encoding scene primitives
GoDe ranks primitives, fine-tunes the shared parameters under simulated quantization, and writes the hierarchy as progressive chunks.
  1. Build a nested hierarchy.