OmniRay: single-image 3D reconstruction across pinhole to panorama cameras, SOTA
anand_bhattad · x · 2026-10-24
The OmniRay paper introduces camera-agnostic monocular 3D reconstruction: one model recovers up-to-scale camera-frame point clouds from a single image without intrinsics or camera-type input, spanning pinhole, GoPro, fisheye, catadioptric, and panorama geometries (GoPro and catadioptric held out from training).
- Motivation: monocular models assume pinhole; camera-aware models use a single fixed basis for per-pixel ray fields, struggling to capture both overall projection and finer variations
- Method: a Ray Module combining a monotone radial base, a spherical-harmonic residual, and a free-form residual, composed as successive rotations on the sphere
- Establishes a new state of the art in camera-agnostic monocular 3D reconstruction
Related event: OmniRay: Camera-Agnostic Monocular 3D Reconstruction from a Single Image(4 posts)→
More from Multimodal
- A full AI music video workflow: treat generation like a live shoot with B-roll planning — lmoroney · 2026-10-24
- AI short film series made entirely with local open-source models in ComfyUI — OneManHorrorBand · 2026-10-24
- Open-source NVAI app turns a single photo into a voiced video scene locally — No-Tap6022 · 2026-10-24
- Retro AI music video blends 1960s Oktoberfest vibes with Halloween — ComputerArtClub · 2026-10-24
- Struggling with fl2va image-to-video: quality decay and prompt detail misses — Disastrous-Yak- · 2026-10-24
- AI short film 'Day Dealer' nails 1970s film grain aesthetic with muscle car vibe — SightsFilms · 2026-10-24