DA3 Metric Large

DA3METRIC-LARGE is a monocular metric depth estimation model that produces depth maps with real-world scale from single images. Developed by the ByteDance Seed Team, it uses a plain transformer architecture with a unified depth‑ray representation and is trained exclusively on public academic datasets (0.35 B parameters, Apache 2.0 license).

Capabilities

The model outputs metric depth maps, relative depth maps, confidence maps, and camera intrinsics/extrinsics. It also provides sky segmentation. It supports inference on individual images or batches and can export results in GLB, PLY, NPZ, and other formats.

Performance

DA3METRIC-LARGE achieves state‑of‑the‑art results on monocular and multi‑view depth estimation, significantly outperforming Depth Anything V2 and the prior multi‑view model VGGT. On the Visual Geometry Benchmark, it surpasses VGGT by 35.7 % in camera pose accuracy and 23.6 % in geometric accuracy (project page). The model was accepted as an Oral at ICLR 2026.

Key Strengths

• Single plain transformer backbone without architectural specialization.
• Unified depth‑ray representation eliminates complex multi‑task learning.
• Trained only on publicly available academic datasets.
• Additional DA3‑Streaming feature enables ultra‑long video inference with less than 12 GB GPU memory via sliding‑window streaming.

Limitations

Performance may vary on domain‑specific images, depending on image quality, lighting, and scene complexity. No latency or throughput figures are publicly available.