ByteDance-Seed/Depth-Anything-3

Depth Anything 3

What it solves

Depth Anything 3 (DA3) addresses the challenge of recovering spatially consistent 3D geometry from arbitrary visual inputs. It simplifies the process of depth and pose estimation, allowing users to predict depth maps and camera positions from single images (monocular) or multiple images (multi-view) without requiring complex multi-task learning architectures.

How it works

DA3 utilizes a single plain transformer backbone (such as a vanilla DINO encoder) and a unified depth-ray representation. This approach removes the need for architectural specialization. The model can be configured to perform different tasks based on the input configuration, including monocular and multi-view depth estimation, camera pose estimation, and 3D Gaussian estimation for novel view synthesis.

Who it’s for

This project is designed for researchers and developers working in computer vision, 3D reconstruction, robotics, and creative pipelines (such as Blender or VR/WebXR viewers) who need high-fidelity depth and geometry recovery from images or video.

Highlights

  • Unified Model Zoo: Offers a range of models including the Main Series (for general geometry tasks), Metric Series (for real-world scale depth), and Monocular Series (for relative depth).
  • Versatile Capabilities: Supports monocular depth estimation, multi-view depth estimation, pose-conditioned depth, camera pose estimation, and 3D Gaussian estimation.
  • DA3-Streaming: Enables ultra-long video sequence inference with less than 12GB of GPU memory using sliding-window streaming.
  • Comprehensive Tooling: Includes an interactive Gradio web UI, a flexible CLI for batch processing, and multiple export formats like .glb, .ply, and .npz.

Related

  • Project
    microsoft/MoGe

    MoGe is a model for recovering accurate 3D geometry, including metric depth and point maps, from single open-domain images.

  • Project
    PozzettiAndrea/ComfyUI-DepthAnythingV3

    Custom ComfyUI nodes that integrate Depth Anything V3 for high-quality depth estimation, 3D point cloud reconstruction, and consistent video depth mapping.

  • Project
    facebookresearch/VLM3

    VLM³ is a framework that enables standard Vision Language Models to perform 3D vision tasks like depth estimation and camera pose estimation using only text-based outputs and data scaling.

  • Project
    localai-org/depth-anything.cpp

    A dependency-free C++ port of Depth Anything 3 and V2 for fast, monocular metric depth and camera pose inference on CPU and GPU.

  • Project
    facebookresearch/map-anything

    An open-source research framework for universal metric 3D reconstruction that uses a single transformer model to handle multiple 3D tasks from various input combinations.