Video Frame Extractor: Scenes, Keyframes & On-Screen Text

Turn any video file into the still frames that matter, ready for AI vision, thumbnails and search.

Give this Actor links to video files, or upload them, and get back images: one frame per scene (every shot, every slide, every new view), a frame every N seconds, N frames evenly spaced, or frames at the exact times you list. Fades, blank frames and repeated pictures are skipped, and you are not charged for them. Each video also gets a contact sheet, and you can have the text on screen read with OCR in 27 languages.

✨ Why use it

• 🎬 One frame per scene, not one per second: cuts, fades through black, dissolves and crossfades are recognised, so you get each shot or slide once, taken where the picture is still, not mid-motion and not half-blended.
• 🧹 No duplicates, no black frames: near-identical frames (a static shot, a slide shown twice, a shot that comes back) and blank frames are left out. Skipped frames are free.
• 🖼️ Images ready for AI vision: JPEG, WebP or PNG, scaled to 1,280 px by default (the size vision models work best with), with a link to each image and its time code.
• 🔤 On-screen text (OCR): slides, titles, captions, labels and credits read into a text field, in 27 languages. Only words the OCR is confident about are kept.
• 🗂️ Contact sheet: every frame of a video in one image, with time codes. Free.
• 📱 Any video file: MP4, MOV, WebM, MKV, AVI, M4V, MPEG-TS and more, H.264, HEVC, VP9, AV1. Phone videos are turned upright.
• 🛡️ Robust: a broken file, a web page instead of a video or a dead link becomes a failed row with a clear message, is not charged, and does not stop the run.
• 💸 Pay per frame: $2 per 1,000 frames, no subscription.

🎞️ Which frames

Mode	What you get	Good for
One per scene (default)	The stillest frame of every shot, slide or new view. Sensitivity low / medium / high; frame from the start, middle or end of each scene.	AI video analysis, storyboards, slide decks from recordings, ad creatives
Every N seconds	A frame every N seconds (0.5 is two a second), duplicates skipped.	Timelines, monitoring, sampling for datasets
N frames evenly spaced	N frames spread across the video.	Thumbnails, previews, quick summaries
At exact times	Frames at the times you list (90, 1:30, 1:02:03.5).	Known moments, chapters, QA

A Start time and End time limit any mode to part of the video.

💡 Use cases

• AI vision and video understanding: send a few representative frames to a vision model instead of the whole video: cheaper, faster, and nothing important left out.
• Lectures, webinars and screen recordings: get each slide once (use Frame from each scene: End for slides that build up), with its text, to make notes, search or a PDF.
• Ads and social videos: break creatives into their shots for review, competitor research or tagging.
• Thumbnails and previews: pick the best stills for a video page or a catalogue.
• Content review and moderation: check what a video shows without watching it.
• Datasets: sample frames for labelling or training, without near-duplicates.

🚀 How to use it

1. Click Try for free.
2. Paste links to your video files in Video URLs, or upload files.
3. Pick Which frames. For slides or text on screen, turn on Read on-screen text (OCR).
4. Run it. Each frame is a row with its image link, time and (with OCR) text. Download as JSON, CSV or Excel, open the images from the Storage tab, or read everything through the Apify API, Make, Zapier or n8n.

📥 Input examples

One frame per scene (the default):

{ "videoUrls": ["https://images-assets.nasa.gov/video/A1Launch/A1Launch~large.mp4"] }

A recorded presentation: one frame per slide, the finished slide, with its text:

{ "videoUrls": ["https://example.com/webinar.mp4"], "framePosition": "end", "ocr": true, "ocrLanguages": ["eng", "deu"] }

A frame every 5 seconds of the first two minutes, as WebP:

{ "videoUrls": ["https://example.com/clip.mp4"], "mode": "interval", "intervalSeconds": 5, "endTime": "2:00", "imageFormat": "webp" }

Frames at exact times:

{ "videoUrls": ["https://example.com/match.mp4"], "mode": "timestamps", "timestamps": ["0:45", "12:03", "1:02:30.5"] }

📤 Output

One row per frame:

{
  "videoUrl": "https://images-assets.nasa.gov/video/A1Launch/A1Launch~large.mp4",
  "fileName": "A1Launch~large.mp4",
  "status": "ok",
  "mode": "scenes",
  "frameIndex": 3,
  "framesInVideo": 14,
  "scenesFound": 14,
  "timestamp": 6.6,
  "timecode": "00:00:06.600",
  "sceneStart": 5.4,
  "sceneEnd": 8.0,
  "imageUrl": "https://api.apify.com/v2/key-value-stores/.../records/video-001-frame-0003.jpg",
  "imageFormat": "jpg",
  "width": 1280,
  "height": 720,
  "fileSizeBytes": 118234,
  "contactSheetUrl": "https://api.apify.com/v2/key-value-stores/.../records/video-001-contact-sheet.jpg",
  "videoDuration": 59.6,
  "videoWidth": 1920,
  "videoHeight": 1080,
  "videoFps": 30,
  "videoCodec": "h264",
  "extractedAt": "2026-10-21T21:48:56Z"
}

With OCR on, rows also have text (lines and paragraphs as on screen) and textConfidence (0-100). A video that could not be read gives one row with status: "failed" and the reason in message. The dataset has three views: Frames (with previews), On-screen text and Videos (one contact sheet per video).

💸 Pricing

Pay per event, no subscription:

Event	Price	Charged when
Frame	$0.002 ($2 per 1,000)	A frame is returned.
OCR frame	$0.002	A frame is read with OCR (only with OCR on).
Scene detection minute	$0.004	Each started minute of video analysed in One per scene mode, counted at Full HD (a 4K minute counts as 4).

Free: skipped duplicates and blank frames, failed videos, and the contact sheet. Higher Apify plans get 10-20% off.

Example: the one-minute NASA launch video above gives 14 frames in One per scene mode: 14 × $0.002 + 1 × $0.004 = $0.032. A 30-minute webinar with 40 slides, with OCR: 40 × ($0.002 + $0.002) + 30 × $0.004 = $0.28.

Set a maximum cost per run and the Actor stops before it: it never starts a video it cannot pay at least one frame for.

❓ FAQ

Can I use YouTube, TikTok, Instagram or Vimeo links? No. Those are web pages that play a video, not video files, and their terms do not allow downloading. Use a direct link to the file (a cloud storage link, a CDN link, your own server) or upload the file.

How big can a video be? Up to 1 GB by default; raise Max video size up to 8 GB (for files of several GB, give the run more memory). Long videos work too: for videos of an hour or more, raise Timeout per video and the run's own timeout. Live streams and streaming playlists (HLS, DASH) are not supported.

How are scenes detected? The video is sampled five times a second and each sample is compared with the one before and the one a second earlier, block by block. Sharp changes are cuts; settled-before, settled-after changes are fades and dissolves; black or single-colour stretches separate scenes; a slow change of exposure is ignored. From each scene, the stillest and clearest frame is returned.

How good is the OCR? It reads printed text well: slides, titles, captions, signs, credits, screen recordings. Handwriting, stylised logos, tiny text and text over busy pictures may be missed or partly read. Words with low confidence are dropped rather than guessed.

Where are the images stored, and for how long? In the run's key-value store, with a link in each row. They are kept for as long as your Apify plan keeps run storage; download them or copy them elsewhere if you need them longer.

Is my video kept? No. Videos are downloaded into the run's temporary disk, processed and deleted when the run ends. Only the frames and the dataset you asked for are saved, in your own Apify storage.

🧩 Notes

• The input example video is "Artemis I Launch Collage" by NASA (public domain).
• Frame extraction uses ffmpeg; OCR uses Tesseract.
• Questions or ideas? Open an issue on the Actor's Issues tab.