What can a paper animator not do?

It cannot read a video

There is no video pipeline. If you hand it an MP4, a MOV or a WebM, it will tell you the file is not an image and stop. The only moving format it touches is an animated GIF, and even then it reads the first frame and nothing else — the animation is discarded.

The workaround is to export a single frame from whatever you already edit in, then bring that still in. In CapCut, scrub to the frame you want and use the export-frame option. In Premiere or DaVinci, set an in-point and export a still. In QuickTime or VLC, take a snapshot at the playhead. You are only after one image, so any of those takes seconds.

It cannot export MP4 or GIF

This is the limit people hit second, right after video input. The animation export here is WebM, encoded in the browser by MediaRecorder using VP8. That is not a design preference — it is what the browser ships for free. Chrome, Edge, Firefox and Safari expose no built-in MP4 encoder and no built-in GIF encoder, so a page that does its own encoding in the browser has exactly one animation container available to it, and that is WebM.

If you need MP4 or GIF, convert afterwards. Any editor will import the WebM and re-export it — that is one extra step, and it is honest about where the encoding actually happens. A site that offers one-click MP4 from the browser is either running a server to do the encoding, or is quietly charging you for the render.

The still export is unaffected: PNG comes straight off the canvas, so the image path has no such limit.

It cannot recognise what is in your photo

There is no segmentation model and no object detection. The cutout works by flood-filling inward from the four edges of the image, spreading across pixels that resemble the border colour. That is a colour operation, not an understanding of the picture.

In practice this means the cutout is excellent when your subject sits on a reasonably even background — a wall, a studio sweep, a plain table, a sky — and it falls apart on a busy, textured or gradient background. It has no idea that the thing you care about is a person, a bottle or a dog; it only knows which pixels look like the edge colour.

The tool is honest about this. The tolerance slider maps 0 to 1 onto a colour distance of roughly 26 to 156 in RGB, and after each pass the status line reports the percentage of the frame it actually removed. If that number comes back near zero, the background is not even enough for a flood-fill and the slider is the wrong tool. If it comes back above about 92%, the fill has eaten your subject. Either way you get a number instead of a guess.

If you genuinely need one-click subject extraction on a hard photo, use a tool with a real segmentation model, cut the subject out there, and bring the result in as a PNG with transparency. This tool will happily animate a cut-out PNG.

It cannot add sound, upscale, or rescue a bad photo

There is no audio pipeline at all — no music, no sound effects, no waveform. If you want audio, add it in your editor after the export; the WebM carries video only.

There is no enhancement. No upscaling, no denoise, no sharpening, no face detection, no style transfer, no model of any kind. The image is fitted to a working canvas of up to 1600 pixels on the long side, and that is the end of the processing.

That ceiling is worth knowing before you start: a 4000-pixel photo is scaled down to fit the working canvas, so the output resolution is not your input resolution. The status line reads both back — something like Loaded 3840×2400 → working at 1600×1000 — so you can see exactly what happened rather than finding out at export.

Very small images go the other way. A thumbnail has no room for a tear to exist in, so anything under roughly 560 pixels on the short side is scaled up, capped at 64×. That keeps the tool from rendering a blank square, but it also means a 40-pixel image will look soft. Feed it the largest version you have.

What it can do instead

The list above is the whole of the negative space. Inside it, the tool does a few things properly:

• Read one still image in any format your browser can decode — PNG, JPEG, WebP or GIF.
• Generate a torn, deckle or shredded paper edge procedurally, at ~500/180/70-pixel wavelengths, rather than stamping a fixed template.
• Composite a paper sheet, a drop shadow, loose fibres and a raw paper core along the tear.
• Flood-fill an even background away from the borders, and report the percentage it removed.
• Export the still frame as PNG, and the animation as WebM — with the frame rate it actually reached printed next to the result.
• Run entirely on your own machine, with no account, no upload and no export limit.