Depth Anything 3

Recovering the Visual Space from Any Views

Project Page Paper Awesome Optimized Fork Original

What This Demo Does

Upload images or videos → Get Metric Point Clouds, Cameras and Novel Views → Explore in 3D

Tip: Landscape-oriented images or videos are preferred for best 3D recovering.

Upload Video

Sampling FPS (Frames Per Second)

0.1 60

Upload Images

Preview

Please upload a video or images, then click Reconstruct.

3D Model

Model

da3-base: fast | da3-large: balanced | giant: best quality

Image Processing Method

low_res for much more images

Infer 3D Gaussian Splatting

Enable novel view rendering from 3DGS ( requires extra processing time)

Reference View Strategy

Strategy for selecting reference view from multiple inputs

Visualization Options: (Click Reconstruct to update)

Show Camera

Filter Black Background

Filter White Background

Filter Percentage

Confidence Threshold (%): Higher values filter more points.

0 100

Max Points (K points)

Maximum number of points to export to GLB (in thousands)

1000 100000

Example Scenes

SOH 2 images

robot_unitree 🎬 video

Research Credits & Acknowledgments

Original Research

Depth Anything 3

Previous Versions

Depth-Anything

•

Depth-Anything-V2

HF demo adapted from Map Anything

Depth Anything 3

What This Demo Does

Click points on the image to compute distance.

Visualization Options: (Click Reconstruct to update)

Example Scenes

Research Credits & Acknowledgments

Original Research

Previous Versions