Original title: Any3R
Any3R reconstructs scenes from images captured by any central projection camera using a single forward computation: standard pinhole, distorted lenses, fisheye lenses exceeding 180°, 360° panoramas, and mixed usage of these within the same scene. Each view is described by a light field (the direction of each pixel), and the model predicts depth along the rays, camera poses, and a fused point cloud for each view. The paper, code, and project page are publicly available.
Why it matters 3D reconstruction no longer requires unifying lens types or performing distortion correction, allowing panoramic and fisheye footage to be used directly.
Hugging Face likes 2 · Hugging Face trending score 2 · Hugging Face downloads 321
License: Non-commercial · ComfyUI: Not available · VRAM: ≈3 GB estimated
Reasons
- Hugging Face image-to-3d trending list
- Score 27 (fresh 5 / source 5 / heat 0 / comfy 0 / vram 13 / demo 0 / license 4)
Usage
- git clone the official repository and run pip install -e ., then load using Any3R.from_pretrained
Risks
- FAIR non-commercial research license, not for commercial use