How frames are captured
This tool does not re-encode anything and does not use a video processing engine. It loads your file into an ordinary HTML video element, seeks to a timestamp, waits for the frame to be presented, and draws that frame onto a canvas at the native pixel dimensions of the video. The canvas is then exported as an image file. Nothing is uploaded and no engine is downloaded, so the tool starts working immediately.
Because the browser itself is doing the decoding, captured frames are exact decoded frames at full resolution. A 1920 x 1080 source produces 1920 x 1080 thumbnails, and there is no resizing option, so scaling down to a smaller thumbnail size is a step you do afterwards.
The same mechanism sets the limits. Your browser has to be able to play the file, and browsers support a narrower set of codecs than a desktop media player does. MP4 with H.264, WebM and many MOV files load reliably; MKV and AVI files, and anything using a codec your browser does not ship, may refuse to load or may load with no video track.
Single frames, intervals, counts and timestamps
In single frame mode you scrub the playback slider to the moment you want and capture the current frame, which is the right approach when you are choosing a cover image and care about the exact expression or composition.
Multiple frame mode has two sub modes. Interval takes a frame every N seconds through the whole clip, with a minimum interval of 0.1 seconds, and adds a final frame near the end if the last interval left a significant gap. Count distributes a fixed number of frames evenly from the first frame to the last, between 2 and 200 of them, which is the mode to use when you want a contact sheet of a predictable size regardless of duration.
You can also type an exact timestamp in seconds and capture that single frame, which is useful when you know the moment you want from an edit timeline. Captured frames accumulate in a list, so you can mix approaches: run an interval pass, then add two hand-picked frames, then download everything together.
Choosing PNG, JPG or WebP
PNG is lossless, so a captured frame is stored exactly as decoded. It is the right choice when the frame will be edited further, or when the content is a screen recording, a slide or animation with flat color areas, where lossless compression is both smaller and cleaner than a lossy codec. The quality slider does not apply to PNG.
JPG is the default at quality 92, and it suits photographic video content. At 92 the artifacts are hard to see and the file is a fraction of the PNG size. Lowering the slider trades visible quality for size in the usual way, and it is worth remembering that a frame from compressed video already carries the encoding artifacts of the source, which a second lossy pass can amplify.
WebP at the same quality setting is typically smaller than JPG for equivalent visual quality, and unlike JPG it supports transparency, though video frames are opaque so that does not matter here. Use it when the destination is a web page and you control the browser support requirements.
A worked example: a contact sheet from a 10 minute video
Load a 10 minute 1080p recording. In count mode, set the number to 24 and extract. The tool distributes 24 timestamps evenly across the 600 second duration, roughly every 26 seconds, seeks to each one, and captures the frame. The progress indicator counts up as it goes and the run can be stopped part way.
At JPG quality 92, 24 frames of 1080p content typically total somewhere between 4 and 10 MB depending on scene complexity. Download All packages them into video-thumbnails.zip with each file named by its timecode, so the sheet stays in order when extracted.
Interval mode on the same video with a 30 second interval produces 20 frames plus a final frame near the end. Interval mode is the better choice when what matters is coverage per unit time, for example finding every scene change in a lecture recording. Count mode is better when what matters is producing a fixed grid.
Seek accuracy and other practical notes
Seeking in a compressed video is not sample accurate in the way an editing application is. Browsers seek to a decodable position and present the nearest frame, so a capture at 12.000 seconds may be a frame or two away from what a frame-accurate tool would give you. For thumbnails and contact sheets this is irrelevant; for frame-exact work such as comparing two encodes frame by frame, it is not the right instrument.
Extraction is sequential because each frame requires a seek and a decode, so a 200 frame job on a long video takes a while and depends on how fast your device decodes the format. Very high resolution sources use more memory per captured frame, since each one is held as a full size image blob until you clear the list.
There is no audio involvement at all, no GIF or animation output, no sprite sheet assembly and no resizing or cropping of the captured frames. What you get is a set of full resolution stills.