Intrinsic PAPR: Tackling Misattribution in 3D Intrinsic Decomposition via Proximity Attention Point Rendering

Intrinsic PAPR targets a subtle but important failure mode in point-based inverse rendering: primitives can converge to the right final image while learning the wrong appearance features internally. That matters for any pipeline trying to separate shading from albedo, because the decomposition can look plausible in aggregate while being unusable for editing, relighting, or downstream reconstruction.

The core change is Proximity Attention Point Rendering, which drops translucent volume-style accumulation and instead predicts appearance at ray-surface intersections. In practice, that means each point gets direct supervision rather than only being judged through the final composited color. The method is designed to reduce the misattribution that happens when multiple primitives along a ray share responsibility for the same pixel.

To deal with monocular ambiguity, the system adds a 2D albedo prior using conditional Implicit Maximum Likelihood Estimation, plus a space-carving loss to enforce multi-view consistency. The authors report gains on both synthetic and real-world datasets, with better novel-view synthesis and albedo estimation than point-based inverse rendering, NeRF-based intrinsic decomposition, and diffusion-based PBR approaches.

For graphics teams, the practical takeaway is that supervision strategy still matters as much as model capacity. If you are building scene capture, relighting, or asset extraction tools, this is another reminder that the rendering formulation can determine whether learned features are actually meaningful per surface point.