Geometry Is All You Predict: Appearance Is Redundant in Feed-Forward 3D Gaussian Splatting

A feed-forward 3D Gaussian Splatting model predicts, per Gaussian, a geometry block (position, scale, rotation) and an appearance block (opacity, colour). Ap- pearance is the majority of that output — 74% of the parameters — and we show it is largely redundant. At a 1000-step test-time refinement budget, replacing the predicted appearance with a draw from an uninformative prior costs −1.25 dB, whereas replacing the geometry to the same degree costs −9.03 dB: a nearly 8 dB asymmetry, matched in information destroyed, on every scene tested. The asymme- try holds on a second architecture with its authors’ checkpoint (MVSplat, 4.60 dB), so a feed-forward predictor is in effect a geometry prior with a largely redundant appendage. We turn this into a method: SplatAnything is a training-free procedure converting a pretrained feed-forward prediction into a compact, higher-fidelity reconstruction in seconds on one 16 GB GPU by keeping a saliency-selected sub- set of the predicted Gaussians, restoring coverage, and briefly refining. It lifts a standard backbone from 23.08 to 29.39 dB, beats InstantSplat under sparse inputs by 3.23 dB, matches the strongest per-scene optimizer with 6.7× fewer Gaussians, and improves every out-of-domain scene — adding no parameters and training nothing.