Depth Anything 3 Large

depth-anything/DA3-LARGE is a multi-view depth estimation and camera pose estimation model that uses a unified depth-ray representation to recover spatially consistent geometry from arbitrary visual inputs. Developed by the ByteDance Seed Team, it is a single plain transformer (0.35B parameters) trained exclusively on public academic datasets, and has been accepted as an oral at ICLR 2026.

Capabilities

The model jointly estimates relative depth, camera poses (extrinsics and intrinsics), and confidence maps from two or more images. It supports pose conditioning and can export results in formats such as GLB, PLY, and Gaussian splatting assets. A streaming variant (DA3-Streaming) handles ultra-long video sequences with less than 12 GB GPU memory.

Performance

DA3 significantly outperforms prior state-of-the-art models on the key benchmark tasks:

• Monocular depth estimation: surpasses Depth Anything 2 (DA2).
• Multi-view depth and pose estimation: outperforms VGGT by an average of 44.3% in camera pose accuracy and 25.1% in geometric accuracy (arXiv: 2511.10647).

It also achieves strong results in generalizable novel view synthesis via a frozen backbone that predicts 3D Gaussian parameters, and when used in SLAM pipelines it reduces drift in large-scale scenes compared to COLMAP.

Limitations

The model is trained on academic datasets only; performance may degrade on domain-specific images, low-quality inputs, or extreme lighting conditions. Results depend on image quality and scene complexity.