Video Scene Detection, Keyframe Extractor: extract scene frames
Pricing
from $20.00 / 1,000 minute of video processeds
Extract frames from a video file, one per scene, with scene change detection: scene timestamps, JPEG keyframes (40 by default, 100 max) and an ffprobe summary as JSON. No ffmpeg to install, no extra API key. $0.01 a run plus $0.02 a video minute; failed videos are not charged.
Direct media file URLs (ending in .mp4 and similar), processed straight away. Social page links (tiktok.com, instagram.com, youtube.com) will NOT work: those platforms block video fetching from cloud servers, so chain a scraper and use the dataset field below. If you paste a page link anyway, the Actor tells you which scraper to run.
The recommended input. Run a TikTok, Instagram or YouTube scraper first, then pick its dataset here: this Actor reads the media URLs the scraper already collected and splits each one. Recognised fields include videoUrl, videoUrlNoWaterMark, mediaUrl, webVideoUrl and url.
Hard stop on how many videos this run processes, so a large input list cannot spend more than you expect. Cost is per minute of video, so this is a ceiling on how many videos are attempted, not on the total bill. Use the run's maximum charge setting for a hard spend cap.
How different two frames must be to count as a cut, from 0.10 to 0.90. LOWER finds more scenes, including soft transitions, camera moves and animated text. HIGHER finds only hard cuts. Start at the default; if a fast-cut video comes back with too few scenes, lower it by 0.05 at a time.
Cuts closer together than this merge into the scene before them. Overlays, flashes and kinetic text all trip a scene detector, so without a floor a 3-minute video can come back with 300 'scenes'. RAISE it for cleaner, chunkier scenes; set it to 0 to keep every detection.
Cap on how many stills are saved per video. Every scene boundary is always returned in the dataset; this only limits the images. When a video has more scenes than this, stills are sampled evenly across the whole video rather than stopping partway through.
Longest edge of each saved still, in pixels. Aspect ratio is preserved and the image is never upscaled. 1024 is a good size for feeding a vision model; drop to 512 for cheaper storage and faster runs.
