Depth Studio
Drop a photo and a real AI depth model turns it into a 3D parallax animation you can move with your mouse. The model and the 3D rendering both run on your device.
🔒 100% private — your photo never leaves your browserDrop a photo here
or click to choose a file · JPG, PNG, WebP
How it works
Predict depth
A monocular depth model (Depth Anything V2) runs on your device and estimates how far every pixel sits from the camera.
Displace on the GPU
That depth map becomes a displacement map. Your photo is drawn on a fine mesh that pushes near pixels toward you and far pixels back.
Move the camera
As the camera drifts, near and far pixels shift by different amounts. That is parallax — the cue your eyes read as real 3D.
How is this running in my browser?
Two things that used to need a server now run on the open web. The depth model is a real neural network compiled to run with WebGPU (or WebAssembly as a fallback) through Transformers.js, so the prediction happens on your own GPU. The 3D scene is built with WebGL: the photo is a texture on a subdivided plane, and a vertex shader reads the depth map to move each point along the camera axis. Because both steps are local, your photo is never uploaded, there is no queue, and there is no per-image cost. The model file downloads once and the browser caches it, so every photo after the first is fast.
What it is good for
Turn a flat snapshot into a moving 3D photo for a post or a story, pull a clean depth map out of any image for design and compositing work, or just see how good on-device monocular depth has become. It works best on photos with a clear near subject and a more distant background, like portraits, streets, landscapes, and product shots. Very flat images, like a straight-on document scan, have little depth to find.
Is my photo uploaded anywhere?
No. The AI model and the 3D rendering both run inside your browser. Your image never touches a server, which is why it is private and there are no limits.
Why does the first photo take a moment?
The depth model (about 50 MB) downloads once on first use, then the browser caches it. After that, depth estimation takes a second or two and every later photo is quick.
Do I need a powerful computer?
A recent laptop or desktop with WebGPU runs it fastest. Without WebGPU it falls back to WebAssembly, which still works but is slower. Phones can run it but the model load is heavier on mobile data.
Why are there small stretched edges when I move a lot?
A single photo has no information about what is hidden behind a near object, so when the camera moves far enough those hidden areas stretch. Keeping the motion modest gives the cleanest look. This is a property of single-image 3D, not a bug.
Can I save the result?
Yes. Record the moving 3D view as a video clip to share, or download the raw depth map as a PNG to use elsewhere.
Does it work offline?
After the model has been downloaded once, the depth estimation works with no network. The page itself needs to load first.
I build tools like this every day.
Senior full-stack engineer, available for senior or contract work, fully remote. See the rest of the lab or get in touch.