Peter Wonka

Feeding fewer, curated frames into novel view synthesis models yields higher reconstruction fidelity than using the entire capture sequence—all without requiring expensive runtime rendering to pick them.

Professional furnishing knowledge can be directly learned and applied to generate constraint-aware furniture layouts, bypassing costly iterative agentic inference.

Pre-trained vision encoders are shockingly bad at distinguishing identity when background context is controlled, but a simple contrastive learning scheme can fix it.

Reconstructing complete, animatable 3D avatars from heavily occluded YouTube videos is now possible, thanks to a hallucination-as-supervision pipeline using diffusion models.

Achieve real-time egocentric motion capture with 19% better accuracy and half the jitter of prior art, thanks to a transformer architecture and self-supervised pretraining on millions of unlabeled frames.

Key contribution not extracted.

Finally, you can generate a fully animatable, 360-degree 3D head avatar from a single image, without per-instance optimization.