A depth map is a grayscale copy of your footage where brightness means distance: near things are bright, far things are dark. It is the raw material behind parallax animation, fake camera moves, selective blur, relighting, and most of the control tricks people use to steer AI video models.
Until recently, getting one out of a video meant a paid plugin, a render queue, or uploading your footage to somebody else's server. None of that is necessary any more. The model that does this well runs in a browser tab.
What you need
A short video clip in MP4, WebM, or MOV. Chrome or Edge, because both support WebGPU — that single detail makes the process roughly 20 to 50 times faster than falling back to CPU. Firefox and Safari will work, but you will be waiting a while.
Nothing else. No account, no install, no upload.
The steps
Open the tool and drop your clip onto the page. On the first run it downloads the depth model — about 25 MB — and caches it, so the first clip is slower than every clip after it.
Pick a quality setting. Fast is usually indistinguishable from High detail once the footage is moving, and it finishes several times sooner. Use High detail only when you plan to hold on a still frame.
Let it process. The model runs per-frame, which is why short clips work best. A five-second clip is a pleasant wait; a five-minute clip is a commitment.
Download the result. You now have a grayscale depth video matched frame-for-frame to your original.
Why short clips really do matter
Per-frame means exactly that: a 10-second clip at 30fps is 300 separate model passes. Doubling the length doubles the work. If you need depth for a long sequence, cut it into the few seconds that actually need the effect and process only those.
This is also why the tool asks nothing of your bandwidth. Everything happens on your own machine, so a long clip costs you time rather than an upload and a queue position.
What to do with the output
The most common use is parallax: split the frame into layers by depth value and move them at different rates to fake a camera push on static footage.
Depth also drives realistic depth-of-field, because you can blur by distance rather than by a hand-drawn mask. It feeds relighting, where you need to know which surfaces face the light. And it works as a control input for AI video models that accept depth conditioning, letting you keep the geometry of a scene while changing everything else about it.
The honest limits
This is monocular depth estimation — the model infers distance from a single view rather than measuring it. It is remarkably good, and it is still a guess. Reflections, glass, and heavy motion blur confuse it. Flat repeating textures sometimes read as the wrong distance.
For animation and compositing that is fine, because you are after a believable gradient rather than a survey. For anything where the measurement itself matters, this is the wrong instrument.