Same Scene, Same Edit: Canonical Gauge Fixing for Stable 3D Gaussian Editing

Editing a 3D Gaussian Splatting (3DGS) scene should be invariant to its Gaussian parameterization. Yet co-located $K$-way splits preserving optical thickness $\tau =-\log (1-\alpha )$ pass a strict $60$ dB input gate but induce up to $82.6\%$ relative RMSE deviation in a state-of-the-art editor, increasing with $K$ and persisting on held-out views. We trace this instability to opacity-threshold culling, duplicated optimizer states, and row-dependent selection. Our training-free wrapper, GaugeFix3D, canonicalizes refinement lineages before running the unchanged editor, using exact merging for co-located siblings, $\tau$-weighted barycenters under jitter, and signature-based recovery without metadata. Across material, local appearance, environment, and object-level semantic edits, GaugeFix3D removes $78.4$–$88.0\%$ of representation-induced edit error in every exact-regime cell and $65.2$–$75.6\%$ under metadata-free approximate refinements. In our main configuration, edit quality improves from $43.41$ to $58.61$ dB PSNR ($+15.2$ dB), within $2.2\%$ of the editor’s identical-input repeat floor, while retaining $\ge 99.98\%$ of the author’s edit magnitude. After continued optimization, $97\%$ of lineages remain recoverable at practical tolerance. The same signature structure occurs in unaugmented checkpoints, and the canonical interface transfers to a second, architecturally distinct editor family.