Robots Atlas>ROBOTS ATLAS
Depth Anything 3
AI Models›Depth Anything

Depth Anything 3

3 · Family: Depth Anything
Visual depth and geometry model by ByteDance Seed. From any views it predicts depth, camera poses and 3D geometry using a single plain transformer with a DINOv2 backbone.
✓ Active✓ Public access⚖ Open weightsVisionSpecialized AI📁 Depth Anything
Parameters
0,08–1,40 mld (SMALL → NESTED-GIANT-LARGE)
parameters
Release date
14 November 2025
🏢ByteDanceProducer
Access:DownloadHostedDeployment:💻 Local☁ Cloud

Overview

Depth Anything 3 (DA3) is a computer-vision model developed by the ByteDance Seed team, introduced in the paper "Depth Anything 3: Recovering the Visual Space from Any Views" (arXiv:2511.10647, November 2025; ICLR 2026 Oral). DA3 recovers the visual space from any number of images or video: it predicts depth maps, camera poses (extrinsics and intrinsics) and 3D geometry, and optionally a 3D Gaussian Splatting representation for novel view synthesis.

Its two core design choices are (1) a single plain transformer — a vanilla DINOv2 encoder is sufficient, with no architectural specialization — and (2) a single depth-ray prediction target that replaces complex multi-task learning. The model uses a DualDPT decoder and is trained with a teacher–student paradigm. It ships as open weights in variants ranging from Small (0.08B parameters) to Nested-Giant-Large (1.40B), with dedicated metric-depth (DA3Metric-Large) and monocular (DA3Mono-Large) variants.

Classification
VisionSpecialized AI
Family: Depth Anything
Applications
Depth perception & 3D geometry3D world generation
Access & deployment
DownloadHosted
LocalCloud
Weights: Open weights
Key parameters
🧩 Parameters: 0,08–1,40 mld (SMALL → NESTED-GIANT-LARGE)
📥 Input: image, video, structured data

Technical specification

Parameters
0,08–1,40 mld (SMALL → NESTED-GIANT-LARGE)
parameters
License
Apache-2.0 (SMALL/BASE/METRIC/MONO); CC BY-NC 4.0 (LARGE/GIANT/NESTED)
Hardware requirements
Streaming inference runs on GPUs with <12 GB VRAM.
Modalities
⬇ Input
imagevideostructured_data
⬆ Output
3d_representationstructured_dataimage

Capabilities and applications

Native model capabilities
Depth estimation
Predicting a per-pixel depth map (distance from the camera) from an image. Covers both relative and metric depth.
Category: vision
Camera pose estimation
Recovering camera extrinsics and intrinsics from one or multiple images.
Category: vision
3D reconstruction
Recovering 3D scene geometry (point clouds, meshes, any-view geometry) from images.
Category: vision
Novel view synthesis
Rendering a scene from unseen viewpoints, e.g. via 3D Gaussian Splatting estimation.
Category: vision
Application domains
Depth perception & 3D geometry3D world generation

Benchmark results

1 benchmark
Visual Geometry Benchmark (VGB)
AUC / F1-score / Chamfer Distance · Any-view geometry, camera-pose estimation and novel-view rendering
SOTA: +44,3% dokładności pozy kamery i +25,1% dokładności geometrii vs VGGT
📄 paper
Figures from the abstract and README. DA3 surpasses the prior SOTA VGGT and Depth Anything 2 on monocular depth.

Technical architecture

Core Architecture
VIViTTRTransformer
Model Form
FMFoundation Model

Sources and related pages