https://snappedcut.com/blog/ai-upscaling

← All articles

What "AI upscaling" actually does

Snapped Cut · 2 min read

Every upscaler makes the same implicit promise: your 720p clip becomes a 4K clip, with detail. It is worth separating what is possible from what is being sold.

Enlarging versus inventing

Enlarging is arithmetic. A resampler decides what colour to put between the pixels you have; a good one plus a sharpening pass produces a clean, crisp enlargement. Nothing is added that was not implied by the source.

Neural super-resolution does something else: it hallucinates plausible detail from a model trained on millions of images. On a still photograph it can be remarkable. On video it has a second problem — the invented detail has to stay consistent from frame to frame, or the picture crawls.

Why not in a browser

A super-resolution model runs per frame, and at a useful size that is seconds per frame on a laptop CPU. Thirty seconds of video is nine hundred frames. Any tool offering it instantly in a browser is either doing it on a server — which means uploading your footage — or doing something much simpler than it implies.

Snapped Cut's upscaler resamples and sharpens, and the page says so. It is fast, it runs on your device, and it makes an enlargement look deliberate rather than soft. It does not invent detail, and neither does anything else that finishes in seconds.

When to enlarge at all

When something has to be delivered larger than it was shot: a 720p clip going into a 1080p timeline, an old export being reused. Doing it once, cleanly, on export beats letting every player do it badly on every playback.

When you are hoping to recover detail that was never recorded, no tool will help, and the ones that claim to are showing you a guess.

See also

0 comments

  • Be the first to comment.