ByteDance-Seed/Depth-Anything-3

Health v2 · maintenance only

Flags: no_releases

Adoption not part of the score

6213 stars · 689 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

Depth Anything 3 (DA3) is a transformer-based model that predicts spatially consistent depth and geometry from any number of visual inputs, with or without known camera poses. It provides Python inference code and pretrained weights, including a streaming mode for ultra-long video sequences on modest GPU memory.

Use cases

• estimate depth from a single image
• reconstruct 3D geometry from multiple views or video
• estimate camera poses from unposed images
• render novel views with 3D Gaussian splatting
• run depth inference on long videos with limited GPU memory
• build SLAM or visual-geometry pipelines on top of a pretrained model

When to choose

• you need state-of-the-art monocular or multi-view depth estimation
• you need camera pose estimation and geometry from arbitrary unposed views
• you want a plain-transformer model trained only on public datasets
• you need streaming inference for long video under 12GB GPU memory

When to avoid

• you need real-time depth on CPU or edge devices without a GPU
• you need metric depth with guaranteed absolute scale out of the box
• you need a lightweight model for production mobile deployment

Facets

library · maturity active

machine-learning deep-learning computer-vision image-processing gpu-computing computer-vision deep-learning machine-learning artificial-intelligence graphics python cross-platform depth-estimation monocular-depth multi-view-geometry camera-pose-estimation 3d-reconstruction transformer novel-view-synthesis slam 3dgs pytorch gpu linux

Member repositories

For agents

markdown · JSON · MCP: product_card(name="ByteDance-Seed/Depth-Anything-3")