nas3r

Research code for a self-supervised method that reconstructs 3D scenes from images without camera poses or prior 3D data. It accompanies a CVPR 2026 paper.

Share on XLicense: MIT

Overview

NAS3R is research code from a CVPR 2026 paper on reconstructing 3D scenes from ordinary images. It is a self-supervised, feed-forward framework that learns 3D geometry and camera parameters together, without ground-truth annotations or pretrained priors. The repository provides installation steps, checkpoints hosted on Hugging Face, and instructions for datasets, training and evaluation in Python with PyTorch.

Key features

  • Learns 3D geometry and camera parameters jointly
  • Needs no ground-truth labels or pretrained priors
  • Checkpoints for two-view and multi-view models on Hugging Face
  • Optional camera intrinsics input for some checkpoints
  • VGGT-based and MASt3R-style architecture variants

Best for

Researchers working on pose-free 3D reconstruction or novel view synthesis who want to reproduce or build on the paper. It is research code that expects a CUDA PyTorch setup.

Upstream
ranrhuang/NAS3R
Fork on GitHub
Guo-astro/nas3r
Upstream stars
121
Category
Finance, data and research
Language
Python
License
MIT
Forked
2026-05-03
Sync status
In syncLast synced 2026-09-29