How AI video upscaling works

From compressed video to individual frames

A video file contains compressed pictures, timing information and, often, a separate soundtrack. Selecting a file gives this page permission to read that file locally. A worker inspects its tracks, then WebCodecs decodes the pictures. The worker processes frames in sequence instead of playing and recording the screen, so a slow GPU does not intentionally skip frames.

FrameLift runs the small, medium or large Anime4K CNN implemented by WebSR. Each network produces a 2× image. For a 1080p target from a 720p source, that intermediate image is resized down to the requested dimensions. This is a single inference pass, not a chain of repeated AI generations.

What the preview tells you

The comparison uses a frame around one-third of the way into your actual clip. Model compilation happens first; three subsequent renders provide a rough inference-time estimate for this device. Decoding, encoding and saving take additional time. The estimate is neither a benchmark nor a promise. Inspect faces, outlines, small lettering and gradients before starting the full clip.

Timing, audio and the final file

Mediabunny carries the frame timestamps into the output and copies compatible compressed audio packets. No audio re-encoding is requested. MP4 output uses H.264 video with AAC or MP3 audio; WebM uses VP9 video with Opus or Vorbis audio. If no supported output can retain every audio track, the job is blocked. Preview the finished video in motion: individual frames can look sharper while temporal flicker becomes more noticeable.