On-device AI · WebGPU

Turn photos into 3D. Depth-aware. Export-ready.

Free photo-to-3D generator. AI estimates depth entirely on-device, rebuilds your picture as an interactive relief or point cloud, and exports GLB, OBJ or USDZ — no uploads, no sign-up.

Detecting GPU…
100% private
Free & open
Why Depth 3D

From flat photo to orbitable 3D. Privately.

No accounts, no API keys, no cloud renders. The depth model comes to your browser.

100% on-device & private

Depth estimation runs in your browser via WebGPU with automatic WASM fallback. Your photos never leave your device.

Real depth AI

Depth Anything V2 Small — a state-of-the-art monocular depth model (~50 MB, Apache-2.0) — maps every pixel to true scene distance.

Two 3D representations

Relief: a textured depth mesh under studio lighting. Points: a vertex-colored point cloud that reveals structure and depth layers.

Export everywhere

GLB with embedded texture, OBJ geometry, USDZ for AR Quick Look on iPhone/iPad, plus the grayscale depth map as PNG.

How it works

Three steps. Zero uploads.

  1. 01

    Add a photo

    Drag & drop, paste, or browse. Portraits, products, food, architecture — anything with layered depth works best.

  2. 02

    AI estimates depth

    Depth Anything V2 runs locally via WebGPU / WASM in seconds. The first run caches the model (~27–50 MB).

  3. 03

    Orbit, tune & export

    Drag to orbit, scroll to zoom. Adjust depth strength, smoothing and detail, then download GLB, OBJ or USDZ.

Complete guide

About Depth 3D

A free browser tool that turns a single photo into an interactive 3D object you can orbit, tune, and export. It is built for people who want a picture to feel more alive — a product shot for a store listing, a portrait, a plate of food, a building facade — or who need a starting 3D file for a design project without paying for cloud photo-to-3D services or waiting in their render queues.

Two representations come out of the same depth estimate: a textured relief mesh under studio-style lighting, and a vertex-colored point cloud that lays the scene's depth layers bare.

How it works

Depth Anything V2 Small, a state-of-the-art monocular depth network (Apache-2.0 licensed), runs in your browser through Transformers.js — on WebGPU where available, WebAssembly everywhere else. It maps every pixel to a relative scene distance, which the tool then turns into geometry: a subdivided plane displaced along the depth field and skinned with your original photo, or a point cloud sampled from the same field. Exports use the standard three.js exporters: GLB with the texture embedded, OBJ geometry, and USDZ that opens directly in AR Quick Look on iPhone and iPad.

Honest 2.5D — what it is and isn't

A single photograph contains no information about surfaces the camera never saw, so the tool reconstructs front-facing geometry only — in the spirit of the classic Facebook 3D photos rather than a full generative image-to-3D model like TripoSR or Meshy, which run on data-center GPUs and have no browser-capable quantized weights today. The back of an object stays flat. What you get instead is speed and privacy: depth in a couple of seconds, entirely on your own hardware, with standard 3D files as the output.

Getting better reconstructions

Photos with clear depth separation work best: a subject well in front of its background, layered foreground and scenery, a building seen at a slight angle. Depth strength controls how far the geometry protrudes toward the camera; smoothing softens the depth field before displacement, which helps on noisy scans; mesh detail trades vertex count for fidelity, while point size adapts the cloud. The grayscale depth map can be downloaded separately for use in other pipelines.

Limits & requirements

The model downloads once (~50 MB on WebGPU, ~27 MB on WebAssembly) and is cached by the browser; each photo afterwards processes in a few seconds. Chrome, Edge, and other Chromium browsers get the WebGPU path; Safari and Firefox run on WASM with identical output. Very flat subjects — document scans, perfectly frontal walls — have little depth to reveal. Mesh exports carry your photo as an embedded texture; OBJ is geometry only.

Privacy

Photos are decoded, depth-estimated, and rebuilt into 3D entirely inside your browser's memory. Nothing is uploaded, queued, or stored on a server, and there is no watermark or usage limit. Once the model is cached, the tool keeps working offline.

Support

Questions, answered.

An interactive 3D preview plus four downloads: GLB (geometry + texture, works in Blender, three.js, glTF viewers and most game engines), OBJ (geometry only), USDZ (opens in AR Quick Look on iPhone/iPad), and the grayscale depth map as PNG.