SplatPainter: Interactive Authoring of 3D Gaussians from 2D Edits via Test-Time Training

SplatPainter tackles a practical gap in 3D Gaussian Splatting: creation is fast, but interactive editing has been awkward. The paper proposes a state-aware feedforward model that takes user-provided 2D edits and updates the underlying Gaussian representation directly, rather than relying on slow iterative optimization or heavy diffusion-based pipelines.

The key idea is to make the editor aware of the current asset state through test-time training, so repeated edits can stay consistent instead of drifting. That matters for production workflows where you want to preserve an asset’s identity while still making local changes, paint-overs, or broader recolors. The authors say the same architecture handles local detail refinement, local paint-over, and consistent global recoloring.

For game teams, the interesting part is not just the novelty of the ML method, but the workflow implication: faster iteration on photoreal assets with less destructive editing. If this holds up outside the paper’s demos, it could reduce the gap between scan/asset capture and art-direction-driven cleanup, especially for technical artists and 3D artists working with splats or similar neural representations.

The paper is an arXiv submission in cs.GR/cs.CV, first submitted on 5 Dec 2025 and revised on 24 Jun 2026. It’s still research, not a shipping tool, but it points toward a more interactive authoring loop for neural scene representations. The main question for developers is whether the quality and stability are good enough to fit into real production pipelines, or whether it remains a niche preview...