Depth Anything 3
Scene reconstruction · Images or a video sequence
About Depth Anything 3
Depth Anything 3 estimates scene depth and visual geometry from images. Its Space accepts multiple images or a video, with a sampling control for choosing video frames.
The interface provides depth-map views, a 3D reconstruction preview, and an optional Gaussian-splatting path. Use these outputs to examine the spatial structure of an observed scene. This is a scene-understanding and reconstruction workflow, rather than an object generator driven by a text description.
How to use Depth Anything 3
- Upload images or a video and review the selected input frames.
- Run reconstruction with the available processing and Gaussian options.
- Inspect individual depth maps and the 3D view; compare them with the original frames.
What to know before you start
- The Gaussian novel-view rendering controls require at least two input views.
- Predicted geometry can contain errors around reflections, thin objects, sky, and occluded surfaces.
The provider controls availability, sign-in, processing, and downloads. Save any result before reloading or closing the demo.
Can I use a video instead of individual images?
Yes. The Space includes video upload and a frame-sampling control for selecting images from the clip.
