Video to LLM Context – Scenes, Keyframes & Transcript for AI
Pricing
from $2.00 / 1,000 keyframes
Turn any video file into AI-ready context: scene cuts, one keyframe per scene, a Whisper transcript aligned to each scene, a keyframe grid and a ready-to-paste Markdown block for GPT, Claude or Gemini. MP4, MOV, WebM links, Dropbox, Drive or Apify storage. Failed files free.
Direct links to video files: MP4, MOV, WebM, MKV, AVI, M4V, TS and more (your CDN, S3/R2/GCS bucket, Dropbox or Google Drive file shared as 'Anyone with the link', OneDrive). YouTube, TikTok, Instagram, Facebook, X, Vimeo, Twitch and other platform page links are rejected (free error row): their terms of service forbid downloading. Only submit videos you have the right to process.
[ "https://upload.wikimedia.org/wikipedia/commons/2/2c/Abbott_and_Costello_Go_to_Mars_Official_Trailer_%281953%29.webm"]Optional. Chain after another Actor: every value of the field below (string or array of strings) is processed like a video URL.
Name of the dataset field that holds the video file URL, e.g. url, videoUrl or fileUrl.
Optional. Every record in this store is analyzed as a video file, e.g. files you uploaded with the Apify API or another Actor's output. JSON/text records become free error rows.
Optional key prefix filter for the key-value store (e.g. video-).
Stop after this many inputs (0 = no limit). Useful when reading a large dataset or store.
ffmpeg scene-change score (0.05–0.95) above which a new scene starts. Lower = more scenes (0.2 for subtle cuts), higher = only hard cuts (0.4–0.5).
Cuts closer than this to the previous one are merged, so flashes and fast montages do not produce dozens of near-identical keyframes.
Longer scenes (talking heads, screen recordings, lectures) are split evenly so the context still gets a keyframe at least this often. 0 = never split.
Upper limit on scenes/keyframes per video (one keyframe per scene, each charged). If more cuts are found, evenly spread cut points are kept.
Off (default): fast scan that compares the video's own keyframes - cut times can be off by a second or two on files with sparse keyframes. On: every frame is decoded for frame-exact cuts (several times slower on HD video, same price).
Keyframes are scaled down to this width (never up). 768 px is a good size for vision LLMs (GPT-4o, Claude, Gemini) and keeps image tokens low.
JPEG quality of the keyframes (30–100).
Add keyframeBase64 (JPEG, base64) to every scene so you can send images to a vision model without downloading the links. Makes rows much larger; dropped automatically if a row would exceed 8 MB.
Save one contact-sheet JPEG with up to 48 keyframes, labelled with scene number and timecode (gridImageUrl). One image = one vision-model call for the whole video. Free.
Transcribe the audio track with Whisper (on-CPU, no API key) and align the text to the scenes. Charged per started audio minute. Turn off for keyframes only.
base: good accuracy for clear speech in 99 languages. tiny: about 2× faster, fine for clean single-speaker English. Same price.
auto detects the language from the first 30 seconds of speech. Or force a Whisper code (en, es, de, fr, pt, it, nl, tr, ru, ja, zh, ar, hi…) for short or noisy clips.
Optional names and jargon that help Whisper spell them right, e.g. Apify, Kubernetes, Dr. Nguyen.
Add the raw Whisper segments (start, end, text) to each row. The per-scene transcript and the Markdown already contain the text.
Put a Markdown image link for every keyframe (and the grid) into llmContext. Turn off for a text-only context block.
Only the first N minutes of each video are analyzed (scenes and transcript, truncated: true). You pay only for what is analyzed.
Larger downloads are stopped (free error row). 1 hour of 720p MP4 is about 500–1500 MB.