Framegrab

The video is read with File + ObjectURL and never leaves the device: no API, no database, no analytics. That makes privacy an architectural property rather than a promise: there is no server that could receive the file.

Frames are exported individually or as a ZIP, encoded in the browser.

Frames are drawn to a canvas sized to the video’s real dimensions, not the CSS player size, so every export is native resolution. Batch extraction seeks sequentially, holds one full-resolution blob at a time, commits results to the store in chunks, yields to the event loop, and can be cancelled mid-run.

Analysis runs off the main thread in a worker that only receives ImageBitmaps; the video file itself is never copied into worker memory, which keeps long and 4K videos from crashing the tab.

A two-phase finder surfaces the shots worth keeping without allocating a single full-resolution image. The scan probes up to 150 evenly spaced timestamps at thumbnail size and scores them in the worker on sharpness by Laplacian variance, exposure quality, and a 64-bit perceptual hash.

Selection normalises sharpness against the scan’s own maximum, blends in exposure, drops blurry frames using an adaptive floor from the scan median, removes near-duplicates by hash distance, and enforces a minimum time gap between picks. Only the winners are extracted, at native resolution, when they are saved.