Breaking

ByteDance Releases Depth Anything 3, a Depth Model for Arbitrary Views

July 5, 2026 at 20:03 EDT

Media Generation
Open Source
Research & Papers

ByteDance's Seed team has released Depth Anything 3 (DA3), a new model that predicts spatially consistent geometry from arbitrary visual inputs, including single images, multiple views, and video. The arXiv paper, project page, GitHub repository, and Hugging Face models and demos were published together around November 13, 2025, drawing attention among developers primarily for 3D reconstruction use cases.

ByteDance Seed · Open Source

Depth Anything 3: One Transformer, Consistent 3D from Any View

DA3 predicts spatially consistent 3D geometry from single images, multi-view images, or video — with or without camera poses — using a plain Transformer and a single depth-ray target. No complex architecture, no multi-task tricks.

SOTA

On the new Visual Geometry Benchmark

Single depth-ray target — no 3D-task modifications

Public

Trained on academic datasets only

Advantage over prior leader VGGT

Higher column = larger improvement

+40%

+24%

beats DA2

Camera pose
35.7–44.3%

Geometric accuracy
23.6–25.1%

Monocular depth
vs DA2

The minimal pipeline

Any input

single · multi-view · video

→

Plain Transformer

DINOv2 + cross-view attention

→

Consistent 3D

dual-DPT depth + ray

Rapid ports across environments

Developers praise how easily DA3 ports and integrates — sizes from Small to Base released open source.

Apple Core ML (iPhone / macOS) LiteRT GGUF / ggml (CPU) ComfyUI + Blender 3D Gaussian Splatting Qualcomm AI Hub

Applications span 3D reconstruction, robotics, AR/VR, 4D face reconstruction, and use as a downstream backbone.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.