Feed-forward 3D Gaussian Splatting (3DGS) methods can reconstruct a scene from sparse multi-view images in a single pass, but existing scene-coding pipelines typically generate the full Gaussian representation first and then compress geometry and appearance attributes separately. That generate-then-compress approach often relies on non-differentiable operations, forcing attribute-specific codec designs and leaving cross-attribute redundancy unexploited.

The proposed framework, MFGS, takes a different route: it extracts multi-view features from the input images and encodes those features into a compact bitstream. End-to-end rate-distortion training learns features that jointly represent geometry and appearance, and the receiver predicts Gaussian attributes from the decoded features. An importance-guided Gaussian selection mechanism lets the receiver adjust how many Gaussians are used for rendering from the same bitstream, accommodating different rendering budgets.

On the RE10K, ACID, and DL3DV datasets, the authors report 99.76–99.86% bitrate savings over the generate-then-compress pipeline as measured by BD-rate, along with a 52–55% reduction in total GMACs and an approximately 12 times end-to-end speedup. The paper is a single-source preprint; no independent comparison is included in the source.