RigPAPR: Rig-Based Animation of Static Neural Point Clouds from a Fixed-Viewpoint Video

RigPAPR is trying to fix a representation problem, not just a deformation problem. The authors argue that Gaussian splats with per-primitive shape can break at joints under linear blend skinning because the canonical tiling no longer holds when bones rotate. Their alternative, PAPR, removes per-primitive shape and recomposes pixels from deformed point positions at render time, which should let the surface “re-form” more naturally during articulation.

For developers, the interesting part is the pipeline: the system auto-rigs a static neural point cloud from a monocular fixed-view driving video, then animates it with direct LBS. No mesh proxy, no pose-dependent correction, and no category template are required. That makes it relevant to anyone exploring digital doubles, avatar retargeting, or fast character reconstruction from limited capture.

The paper reports results on synthetic subjects, where RigPAPR matches the strongest baseline at the supervised view and beats mesh-based and Gaussian-splatting baselines at novel views by more than 3 dB PSNR. It also claims cleaner joint-boundary renderings on both synthetic and real subjects. The submission is dated 4 Jun 2026 and is listed under cs.CV / cs.GR.

Practically, this is another sign that neural scene representations are moving from “nice reconstructions” toward riggable assets. If the approach holds up outside the paper’s test cases, it could reduce the amount of manual cleanup artists and technical animators need when turning capture into reusable characters. The open question is still robustness: fixed-view monocular...