3D Photo Converter

Add the photo and press Generate 3D. A depth model reads the picture 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. Then choose what to make from it: a side-by-side pair you view by aiming your eyes, a red-cyan anaglyph for coloured glasses, an animated wiggle that needs neither, or the depth map itself. The model runs once, so after that every control updates the preview as you drag it and you can try all four without waiting.

Only for the anaglyph. The wiggle output needs nothing at all — it animates between the two viewpoints, and that movement gives your brain the same depth cue that leaning sideways does, which is why it still reads as depth for someone who sees with one eye only. The side-by-side pair also needs no hardware, but it does need free-viewing, which is a knack some people pick up in minutes and others never find comfortable. If you want to hand a 3D photo to somebody else with no instructions, use the wiggle.

Choose the Wiggle output after generating depth. A wigglegram is an animation that rocks between slightly different viewpoints of the same scene, and it originally required a film camera with several lenses set side by side. Here the extra viewpoints are synthesised from the depth map instead, so one ordinary photo is enough. Set the frame count — two gives the classic hard alternation, more gives smoother motion — pick a speed around 10 to 14 frames a second, and download an animated GIF. The preview animates live at whatever settings you choose, so you can judge it before encoding anything.

Dubois, in almost every case. The obvious approach is to give the red channel to one eye and green and blue to the other, but real filters leak, and a saturated red object then appears bright to one eye and black to the other — the brain refuses to fuse that, which is what makes cheap anaglyphs uncomfortable. Dubois instead solves for the pair of images whose appearance through actual filters lands closest to the original, so reds survive and ghosting largely disappears. If a picture still shimmers, lower the colour strength before trying another method; greyscale and monochrome remove the problem entirely at the cost of all colour.

No. The depth model runs through WebGPU or WebAssembly inside your browser tab, and every other step is arithmetic on a canvas. Your photo is read from disk into the tab's memory and the finished image is written back out as a download. The depth model itself is downloaded from the network the first time you use the tool, but the photo and generated result are never uploaded.

Three common causes, in order of likelihood. First, the glasses are on the other way round, or the side-by-side pair is being viewed with the wrong technique — both invert the depth, and an inverted picture still looks convincingly three-dimensional, so it is easy to miss. Use the swap control for anaglyphs, or switch between parallel and cross-eyed. Second, the depth model misread the scene, which the depth map view will show you immediately: reflections, glass, and printed pictures within the shot are the usual culprits, and no setting recovers from a wrong depth map. Third, the strength is simply too high, which reads as smearing around objects rather than as depth.

Yes — choose the Depth map output and download it. For another program, select Greyscale and PNG: that gives you an 8-bit map at the output size you picked, aligned pixel for pixel with the scaled photo, with bright meaning near by default and an invert switch for software that expects the opposite. JPEG, WebP, and the three colour ramps are for sharing or inspecting the map, not geometry. It is relative depth: it tells you what is in front of what within that one picture, not how far away anything is in metres. That is what parallax effects, displacement, distance-based masking, and depth-conditioned image generation need.

They are two ways of aiming your eyes, and each needs the halves in a different order. Parallel viewing relaxes the eyes 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. Most people find cross-eyed easier to learn. It avoids parallel viewing's hard divergence ceiling, but a moderate display size is still more comfortable. Parallel viewing is often more comfortable once it clicks, but it has a firm size limit because your eyes cannot rotate further apart than straight ahead.

Almost certainly because the image is drawn too large, and this catches nearly everyone with parallel viewing. Fusing a parallel pair requires your eyes not to rotate further apart than straight ahead, since human eyes cannot diverge, which caps the separation of matching points at about 63 mm — the distance between your pupils. Each half of a full-width pair on a large monitor is tens of centimetres across, so it cannot be fused at any resolution. The size control under the preview exists for this: shrink it until each half is around 63 mm, or press the fit button. Cross-eyed viewing avoids that divergence limit, though a moderate size is still more comfortable.

Ones with a clear foreground, a distinct middle, and a real background — a landscape with layered distance, a person against a room, a single object on a plain surface. Those give the model unambiguous cues and leave little to invent behind edges. What converts badly is anything flat or ambiguous: a photograph of a photograph, a wall of text, heavy motion blur, large reflections, glass, and shots where subject and background are the same brightness and texture. Very tight close-ups also have little depth structure to work with, so the effect is subtle whatever the strength.

Two separate things. Shifting pixels sideways to build a second viewpoint exposes areas that were hidden behind nearer objects, and nothing ever photographed them, so they have to be invented — each gap is patched from the background side of the edge and mirrored inward, but a patch is still a patch, and it grows with strength. Separately, the model works at a much lower resolution than your photo, so its depth edges arrive soft and slightly off the object, which puts a band of background at the wrong depth around heads and shoulders. The edge alignment setting is aimed squarely at the second one.

Yes, and a phone is the best screen for looking at the result: it is the one display where a parallel side-by-side pair is comfortably inside the width limit, so free-viewing is much easier there than on a desktop monitor. Generating depth is slower on a phone than on a desktop with WebGPU, but it is one photo rather than thousands of frames, so it stays a matter of seconds. Keep the output width modest on mobile — the memory estimate shown before you generate will tell you when the settings are asking for too much.