Intrinsic decomposition and editing of 3D Gaussian splats

This paper takes a familiar image-editing idea—splitting appearance into diffuse albedo and shading—and applies it to 3D Gaussian splatting. The authors argue that doing this inside a radiance field makes edits much more robust than repainting a single rendered view, because the change can be re-rendered with plausible lighting from arbitrary camera angles.

The main technical twist is representing the intrinsic decomposition as independent sets of Gaussian primitives, rather than forcing one shared representation to do everything. That gives each layer more freedom to fit what it actually represents, which matters when you’re trying to disentangle texture, lighting, and view-dependent effects from multi-view photos of a real scene.

They also describe an optimization pipeline guided by data-driven predictions to separate the scene into those intrinsic components. For developers, the interesting part is less the academic framing and more the workflow: you can modify the albedo of a planar surface in one image, then propagate that edit through the intrinsic radiance field so the result survives viewpoint changes instead of collapsing into a flat paint-over.

The paper is dated 30 Jun 2026 on arXiv (cs.GR), with a journal reference to Proc. ACM Comput. Graph. Interact. Tech. 9, 1, Article 10 (May 2026). If Gaussian splats are part of your capture or editing pipeline, this points toward a more practical authoring model for relighting-aware scene edits, though it also suggests extra complexity in decomposition, optimization, and tool support.