Devjyoti Das

3D Lab · Computer vision in your browser

Turn a photo into a 3D model

Upload an image and a neural network estimates how far away every pixel is, then lifts the photo into 3D. Spin it around, tweak the shape, and download it. It all runs on your device: your image is never uploaded.

Private: nothing leaves your browser Depth Anything V2 Export GLB · OBJ · PLY · STL
Choose an image to begin. The model (about 25–50 MB) downloads the first time only.
Drop an image here or click to browse · JPG, PNG, WebP

How it works

  1. Estimate depth

    A neural network (Depth Anything V2, Small) looks at the photo and predicts, for every pixel, how near or far it is. People can do this too: we use shading, perspective and familiar object sizes.

  2. Lift to 3D

    Each pixel becomes a point on a grid, pushed out toward you by its predicted depth. Neighbouring points are joined into triangles, and the photo is draped over them as a texture.

  3. Clean up

    Smoothing removes noise. Triangles that stretch across a big depth jump are cut, and far pixels can be removed to isolate the object. A back plane and walls make it solid.

  4. Export

    Download the mesh as GLB, OBJ, PLY or STL for Blender, game engines, or a slicer for 3D printing.

How the distance in metres is estimated

  • A second network, Depth Anything V2 Metric (two versions: close range and far range), predicts the distance in metres from the camera to every pixel. Each of your two points is turned into a 3D position using the camera's field of view, and the straight-line distance between them is reported.
  • The field of view comes from the photo's EXIF data when it has it (most phone photos do). Otherwise a typical phone lens (26 mm) is assumed, so set the focal length if you know it. Cropped or edited photos break this.
  • Treat it as a rough estimate. It can easily be off by 20% or more, especially for scenes unlike its training data, glass, mirrors or distant objects. The error is often similar across one photo, so calibrating helps a lot: measure something whose real length you know (a door, a person, a ruler), enter it, and all other distances are rescaled to match.
  • Do not use it where accuracy matters for safety, building or legal purposes.

What this can and can't do

  • It reconstructs the visible surface from a single photo, so it is a 3D relief rather than a full object. The hidden back is guessed as flat.
  • The 3D relief is relative: "Depth strength" is an artistic control, not a measurement. Distances in metres come from a separate metric depth model and are rough estimates (see below).
  • It works best with one clear subject, good lighting and a plain background. Glass, mirrors and very dark or blurry areas confuse it.
  • Full all-around reconstruction from one image (for example TripoSR) needs a GPU server. That is a possible upgrade; see my 3D-Genesis work on parametric reconstruction.

Powered by Depth Anything V2 Small and the Metric variants (Apache-2.0) running through Transformers.js, with rendering by three.js. Only the model weights are downloaded; your photo stays on your device.