Peking University's NeuOctree predicts ray-surface intersections for direct ray-traced rendering
ssh4net · x · 2026-10-30
Researchers at Peking University (Jun-Hao Wang, Changan Lu, Baoquan Chen, Peng-Shuai Wang) introduce NeuOctree, a neural explicit 3D shape representation.
- Existing neural 3D representations handle surface modeling, rendering, or compression in isolation; none jointly provide explicit surface queries for ray tracing and a structured, predictable output space
- NeuOctree directly predicts whether and where a ray intersects a surface, along with normal and albedo
- It stores shared features at sparse octree voxel corners, interpolates them into a continuous local field, and refines queries near the predicted surface with a lightweight decoder using recurrent intersection-guided sampling
- Enables accurate, efficient ray-traced rendering; at 823 KiB compression it outperforms QNDF, NGF, VQAD, and N-BVH qualitatively
Paper is out; the code link is not live yet.
More from Multimodal
- Full AI Short Film Workflow: Character/Props/Location Sheets + Seedance Reference Mode — socialwithaayan · 2026-10-30
- 100% AI-made short film reimagines ancient Egypt with BMX bikes, full 5-step workflow shared — socialwithaayan · 2026-10-30
- 90 open-source apps in two weeks: the full #OpenSourceHiggsfield community catalogue — Visual_Tiger_2400 · 2026-10-30
- OpenSuperComputer open-sources a Higgsfield-style automated content generation pipeline — iAM_A_NiceGuy · 2026-10-30
- Open SuperComputer: open-source chat-driven workspace for ComfyUI media workflows — iAM_A_NiceGuy · 2026-10-30
- Reddit users push back on calling open models "Open-Source Higgsfield": you're doing their marketing — curious-scribbler · 2026-10-30