Image → 3D
Early version
Photo → Point Cloud
Turn a single photograph into a 3D point cloud, in your browser.
The problem
Students constantly need quick 3D from flat reference — a site photo, a precedent image, a sketch — but photogrammetry needs dozens of images and scanning gear no one owns. A single image is usually all you have.
What this does
Drop in one image. A monocular depth model runs entirely on your device (no upload, no server), estimates a depth value per pixel, and unprojects every pixel into a coloured point in 3D space that you can orbit. It's a fast way to read the spatial structure of a precedent or a site.
Try it
Loading tool…
What it does
This tool estimates a depth map from a single image using a neural network that runs locally in your browser (WebGPU / WASM — nothing is uploaded). Each pixel is then back-projected through a simple pinhole-camera model into a coloured 3D point, producing a point cloud you can orbit and inspect.
Why it's useful
- Read the spatial depth of a precedent image or site photo in seconds.
- Generate rough point clouds for collage, diagramming, or as a base to model over.
- No account, no upload, no GPU server — it runs on your laptop.
Honest limits (it's an early version)
- Monocular depth is relative, not metric — proportions are plausible, not survey-accurate.
- One image only sees one side; there's no information behind objects.
- Best on images with clear depth cues (interiors, streets, landscapes).
The real pipeline (for the roadmap)
For metric, multi-view accuracy you'd move to photogrammetry (RealityCapture, Metashape, COLMAP) or learned multi-view depth. This artifact is the instant, single-image end of that spectrum.
Built on
- Depth Anything V2 (depth model)Apache-2.0 (small model)
- Transformers.js (in-browser inference)Apache-2.0