Depth Anything 3 Large

Depth Anything 3 DA3-LARGE is a multi-view depth estimation and camera pose estimation model that predicts spatially consistent geometry from arbitrary visual inputs, with or without known camera poses. It uses a single plain transformer backbone with a unified depth-ray representation, eliminating the need for architectural specialization or complex multi-task learning.

Capabilities

• Relative depth estimation from single or multiple images
• Camera pose estimation (extrinsics and intrinsics) from unordered image sets
• Pose conditioning for geometry-aware inference

Key strengths

Trained exclusively on public academic datasets, DA3-LARGE (0.35B parameters) significantly outperforms prior state-of-the-art models. According to its published paper (arXiv:2511.10647), it surpasses VGGT by 44.3% in camera pose accuracy and 25.1% in geometric accuracy. It also substantially improves over Depth Anything 2 for monocular depth estimation. The work has been accepted as an ICLR 2026 Oral presentation.

Architecture insights

The model demonstrates that a single plain transformer (e.g., vanilla DINO encoder) suffices as a backbone, and a singular depth-ray representation replaces the need for multi-task learning—both insights contribute to its efficiency and performance.

Visual results

Limitations

• Trained on academic datasets; performance may degrade on domain-specific images
• Output quality depends on image quality, lighting, and scene complexity

References

Project page · Paper