ARS-Avatar: Animatable and Relightable Surfel Avatars with Learnable Ambient Occlusion

arXiv cs.GR details ARS-Avatar, a new approach for building human avatars that can be animated and relit from multi-view captures taken under unknown illumination. The method uses surfels rather than a more conventional mesh-only pipeline, aiming to preserve fine appearance while keeping the avatar editable and pose-aware.

The key technical move is to pull deformation priors from a template mesh and use them as extra guidance beyond the driving pose. That helps the system estimate surfel attributes more faithfully, which matters when you want the same character to deform cleanly across motion without smearing details or breaking surface consistency.

For lighting, ARS-Avatar uses deferred shading to estimate BRDF materials, then adds a differentiable screen-space ambient occlusion model. The interesting bit for graphics developers is that occlusion radii are optimized with gradient-based finite differences on a body-part basis, giving the model a practical way to approximate light visibility while training alongside the avatar itself.

The paper says the method produces high-fidelity appearance reconstruction and physically based material estimates, while supporting realistic animation and relighting under novel poses and illuminations. For teams working on digital humans, performance capture, or avatar pipelines, this is another sign that the hard part is no longer just geometry reconstruction, but disentangling motion, shading, and visibility well enough to make relighting believable.