Stereo 3D Converter

Add the video, choose whether you want a parallel or cross-eyed pair, set the size per eye, and press Convert. A depth model reads every frame to work out what is near and what is far, and a second viewpoint is built by shifting pixels sideways in proportion to their depth. The result is an ordinary MP4 containing two images side by side — one for each eye. Preview a single frame first: depth strength is a judgement call that depends on the shot, and checking it on a still takes seconds against minutes for a full run.

No, and that is the point. The output is a side-by-side pair you view by aiming your eyes rather than by filtering them. Cross your eyes slightly, or relax them as though looking past the screen, until the two halves overlap into a third image in the middle — that middle one has depth. It is a knack rather than a trick, and most people get it within a few minutes. The same file also loads directly into a VR headset, if you would rather not learn.

They are two ways of aiming your eyes, and each needs the halves in a different order. Parallel viewing relaxes the eyes outward as though looking through the screen, and wants the left eye's image on the left. Cross-eyed viewing points them inward, in front of the screen, so the halves are swapped. Viewing a file the wrong way inverts the depth: the foreground sinks and the background floats, which looks wrong without being obviously wrong, so pick the mode before converting.

Almost certainly because the image is too large, and this catches nearly everyone with parallel viewing. To fuse a parallel pair your eyes must not rotate further apart than straight ahead, since human eyes cannot diverge — which caps how far apart matching points can sit at about 63 mm, the distance between your pupils. Each half of a full-width pair on a 27-inch monitor is roughly 30 cm across, so it cannot be fused at all. Shrink the window, move further back, or watch on a phone. Cross-eyed viewing has no such limit and works at any size.

No. The video engine is FFmpeg compiled to WebAssembly and the depth model runs through WebGPU or WebAssembly, both inside your browser tab. Your file is read from disk into the tab's memory and the finished video is written back out as a download. The only things fetched from the network are the engine and the model themselves, once, and you can confirm nothing else moves by watching the network panel during a conversion.

It depends on the device, the output size, the frame rate, and the footage, so there is no single figure worth quoting. Depth has to be estimated on every frame and that dominates everything else, which means a machine with WebGPU is far quicker than one without — the CPU fallback is workable for a short clip and impractical for a long one. Rather than predict, the progress bar measures actual throughput and shows an estimate that becomes accurate within the first few seconds.

There is no fixed threshold, but there is a ceiling. The clip is converted in segments, each extracted, converted, encoded, and handed out of the engine's memory before the next starts, so the work itself does not grow with duration. What does is the finished file: it has to be assembled in memory once before it can be saved, because the engine cannot write anywhere else. How long that allows depends on the output size, the frame rate, how compressible the footage is, and how much memory the device has. The tool works out whether your settings fit before starting and says so rather than failing partway, and choosing a shorter range or a smaller size per eye is what buys more length.

Usually because the depth model misread the scene, and the depth map preview will show you exactly that. Reflections, glass, flat surfaces with strong texture, printed photographs within the shot, and heavy motion blur are all read unreliably, and no amount of strength adjustment recovers from a wrong depth map. Scenes with a clear foreground and a real background — landscapes, drone footage, a person against a room — convert far better than flat or cluttered ones.

Shifting pixels sideways to build the second eye exposes areas that were hidden behind nearer objects, and nothing ever filmed them, so they have to be invented. Each gap is patched from the background side of the edge, which is where the missing content actually belongs, but a patch is still a patch. It grows with depth strength, which is the main reason modest settings usually look better than strong ones. Thin objects and hair are the hardest cases.

No, and nothing can from a flat video. Those billboards work by rendering a scene with the perspective distortion needed for one specific curved screen and one specific viewing position — the illusion lives in the geometry of the screen and where you stand relative to it. That information does not exist anywhere in an ordinary video, so it cannot be recovered from one. This tool makes stereo pairs, which is a different effect with a different mechanism.

Yes. Portrait sources keep their orientation, and rotation recorded in the file is handled — a clip shot vertically is treated as vertical rather than being reported as a sideways landscape frame. The resulting pair is two portrait frames side by side, which is quite wide overall but works well for free-viewing precisely because each half stays narrow.

The first audio track is muxed onto the finished video in a single pass at the end rather than being split and rejoined with every segment, which is deliberate — cutting audio at every boundary is how drift creeps in. Where the output container will carry the stream as it is, it is copied rather than re-encoded; otherwise it becomes AAC. Only that first track survives: extra language tracks and subtitles are not carried over. Since the conversion changes no timings, the sound lines up exactly where it did, and a silent source is handled without complaint.