Video to Slides - Extract Unique Slides to PDF + Text

Video to slides in one API call: give this Actor a lecture, webinar, conference talk or screen recording file and it returns every unique slide as an image, with its start and end timestamp, the slide text (OCR) and one combined PDF of all slides. A slide shown for five minutes becomes one slide. Bullet-by-bullet builds are merged and you get the most complete version. Slides the speaker goes back to are not repeated. The camera shots of the speaker are left out.

No API key, no LLM, no browser extension, no manual screenshots. It runs on CPU on Apify, so you can call it from n8n, Make, Zapier, Python, JavaScript or an AI agent (MCP).

What you get

For every video:

• One dataset row per slide: slideIndex, startSeconds / endSeconds, timecode / endTimecode, durationSeconds, imageUrl (full-size JPEG/PNG/WEBP), text (OCR, reading order), ocrConfidence, width, height, buildSteps (how many build steps were merged into it), alsoShownAt (later times the same slide came back).
• One summary row (type: "video"): slideCount, pdfUrl (all slides, one page each), contactSheetUrl (one overview image), allText (every slide's text with --- slide N (hh:mm:ss) --- markers, ready to paste into ChatGPT, Claude or a RAG index), duration, resolution, billedMinutes, note, error.

Field names match Video Frame Extractor where they overlap (videoUrl, fileName, timestampSeconds, timecode, width, height), so you can chain both Actors in one workflow.

Use cases

• Lecture video to PDF notes: turn a recorded class or MOOC lesson into a slide handout you can annotate.
• Webinar and meeting recordings: get the deck that was presented, even if nobody sent it.
• Conference talks: index the slides of every talk with their timestamps.
• Screen recordings and tutorials: one image per screen, with the text on it.
• Video RAG / LLM input: send allText and the slide images to a vision model instead of hundreds of near-identical frames.

How to extract slides from a video

1. Paste direct video file links into Video file URLs (MP4, MOV, WEBM, MKV, AVI, M4V, MPEG-TS...). Your own storage, S3/R2/GCS, a CDN, an Apify key-value store, or Dropbox / Google Drive links shared as "anyone with the link" all work.
2. Pick the Slide language for OCR (English/Chinese/Japanese, accented Latin languages, Cyrillic, Korean, Arabic, Devanagari, Greek, Thai).
3. Optional: set Sensitivity, Minimum slide duration, or a Slide area (crop) when the slides share the frame with a large speaker camera or burned-in subtitles.
4. Click Start, then open the Slides tab of the dataset, or the pdfUrl in the Videos tab.

Example input:

{
  "videoUrls": ["https://example.com/recordings/lecture-03.mp4"],
  "sensitivity": "medium",
  "ocr": true,
  "ocrLanguage": "en",
  "outputPdf": true
}

Example slide row (shortened):

{
  "type": "slide",
  "fileName": "lecture-03.mp4",
  "slideIndex": 5,
  "startSeconds": 146.25,
  "endSeconds": 213.75,
  "timecode": "00:02:26.250",
  "endTimecode": "00:03:33.750",
  "durationSeconds": 67.5,
  "timestampSeconds": 213.25,
  "imageUrl": "https://api.apify.com/v2/key-value-stores/.../records/...-slide-005.jpg",
  "width": 1920,
  "height": 1080,
  "text": "Uncitedness (publications):\n12% - 82% (Larivière & Gingras 2009)\nUncitedness (data):\n85% (Peters et al. 2016)",
  "ocrConfidence": 0.9749,
  "buildSteps": 8,
  "alsoShownAt": []
}

How slide detection works

The video is decoded once at low resolution (1-2 pictures per second). A slide is a picture that stays still for at least the minimum slide duration. Regions that keep moving all the time - a small webcam picture-in-picture, a playing video, a ticking clock - are learned from the whole video and ignored. When a later picture only adds content on empty background (a bullet appears, a chart builds), it is merged into the same slide and the final state is kept. Slides identical to an earlier one are dropped and recorded in alsoShownAt. Only then is each slide grabbed at full resolution, OCR'd and written to the PDF.

Because camera footage never stays still, a talk filmed on stage returns only the moments where a slide fills the screen - the speaker shots are not billed as slides.

Pricing

Pay per event, no subscription, no start fee beyond Apify's tiny default:

Event	Price	Store discount
Slide (one unique slide: image + text + PDF page)	$0.004 ($4 per 1,000 slides)	Silver $0.0036, Gold $0.0032
Video minute (each started minute scanned, per video)	$0.001	Silver $0.0009, Gold $0.0008

Worked examples:

• 60-minute lecture with 40 slides: 60 x $0.001 + 40 x $0.004 = $0.06 + $0.16 = $0.22.
• 10-minute screen recording with 15 screens: $0.01 + $0.06 = $0.07.
• 20-minute talk filmed on stage with 6 full-screen slides: $0.02 + $0.024 = $0.044.

Free: failed downloads, refused links, non-video files, videos where no slide was found, blank slides, duplicates, the PDF, the contact sheet and the OCR text. Set Max cost per run in the run options and the Actor stops cleanly at that amount: it scans only the minutes that fit, delivers the slides that fit, and marks later videos as skipped (free).

Video to slides vs. doing it yourself

A frame every N seconds gives you hundreds of near-duplicates and half-built slides. ffmpeg's scene filter misses bullet builds and fires on every camera cut. Browser extensions and desktop apps need a person to press the button. This Actor dedupes, keeps the complete build state, skips the speaker shots and returns structured JSON plus a PDF - from a schedule, a webhook or an API call.

Limits and tips

• Direct video files only. Links to YouTube, Vimeo, TikTok, Instagram, Facebook, X, LinkedIn, Twitch, Douyin or Bilibili pages are refused (free) because their terms forbid downloading.
• Default limits: 240 minutes and 2,000 MB per video, 300 slides per video (all adjustable).
• Burned-in subtitles are ignored when the rest of the slide stays the same. If a large speaker camera sits next to the slides, use Slide area (crop), e.g. {"x": 0, "y": 0, "width": 75, "height": 100}.
• Screen recordings: every screen that stays still for the minimum duration counts (scrolling is skipped). Use Sensitivity: low to merge small highlights and pop-ups.
• A slide shown again at a different zoom or crop may appear twice. Photos shown full screen for several seconds count as slides.
• OCR is fast CPU OCR (PP-OCR): very small text, handwriting and fancy fonts can come out wrong. Pick the right Slide language - the default model does not read Cyrillic.

FAQ

How do I convert a video to slides for free?

Apify's free plan includes monthly credit, which covers hundreds of slides with this Actor. You only pay per slide and per scanned minute.

Can I extract slides from a YouTube lecture?

Not from a YouTube page link. Use a video file you own or are allowed to process (for example your own Zoom, Teams or OBS recording uploaded to cloud storage).

Does it work with Zoom, Teams and Google Meet recordings?

Yes - upload the MP4 file and link it. Small webcam thumbnails are ignored automatically; for a large side-by-side camera use the crop option.

Can I get the slides as PowerPoint?

The Actor returns a PDF and one image per slide. Most tools (PowerPoint, Google Slides, Keynote) import the images or the PDF directly.

Related Actors by Kanto Labs

• Video Frame Extractor - keyframes, scene changes, evenly spaced frames or the best thumbnails from any video file.
• Video to LLM Context - scenes, one keyframe per scene and the spoken transcript aligned to each scene, for AI agents.
• Video Scene Detection & Shot List - every cut with timecodes, CSV/EDL and optional clips, for edited videos.

Support

Found a video where slides are missed or duplicated? Open an issue on this Actor's Issues tab with the video link and the run ID - we answer within a day.

The built-in sample video is a 3-minute cut of the webinar "Academic SEO" by Peter Kraker (Open Knowledge Maps), licensed CC BY-SA 4.0, from Wikimedia Commons.