FacuM/youtube-step-extractor

YouTube Step Extractor

Download a YouTube video, extract frames at regular intervals, and analyze them to identify a specific sequence of steps from the visual content.

Prerequisites

Requires yt-dlp and ffmpeg:

# Ubuntu/Debian
sudo apt-get install -y ffmpeg
pip install yt-dlp

# macOS
brew install ffmpeg yt-dlp

Workflow

Step 1: Download the YouTube video

yt-dlp -f "bestvideo[height<=1080]+bestaudio/best[height<=1080]" \
  -o "/tmp/yt_video.mp4" \
  --merge-output-format mp4 \
  "YOUTUBE_URL"

For faster download (lower quality is fine for frame analysis):

yt-dlp -f "bestvideo[height<=720]+bestaudio/best[height<=720]" \
  -o "/tmp/yt_video.mp4" \
  --merge-output-format mp4 \
  "YOUTUBE_URL"

Step 2: Extract frames

Use the bundled script:

{baseDir}/scripts/extract_frames.sh /tmp/yt_video.mp4 /tmp/yt_frames 1

Arguments:

• video_path (required): Path to the downloaded video
• output_dir (optional): Where to save frames. Default: ./frames_<video_name>
• fps (optional): Frames per second. Default: 1 (one frame per second)

For longer videos, reduce fps to avoid too many frames:

# 1 frame every 2 seconds for videos > 5 min
{baseDir}/scripts/extract_frames.sh /tmp/yt_video.mp4 /tmp/yt_frames 0.5

# 1 frame every 5 seconds for videos > 15 min
{baseDir}/scripts/extract_frames.sh /tmp/yt_video.mp4 /tmp/yt_frames 0.2

Or use the all-in-one script:

{baseDir}/scripts/download_and_extract.sh "YOUTUBE_URL" /tmp/yt_frames 1

Step 3: Analyze the frames

1. List the extracted frames: ls /tmp/yt_frames/
2. Read key frames using the image viewing tool
3. For comprehensive analysis, sample frames at regular intervals (e.g., every 5th frame)
4. Identify distinct steps by looking for:
   • Scene/screen transitions
   • UI changes (new dialogs, menus, pages)
   • Text overlays, titles, or captions
   • Actions being performed (clicks, typing, navigation)
   • Before/after states
5. Build a numbered step-by-step summary of the process shown

Step 4: (Optional) Extract transcript for context

Subtitles add context to what's visible in the frames:

yt-dlp --write-auto-sub --sub-lang en --skip-download --sub-format vtt \
  -o "/tmp/yt_transcript" "YOUTUBE_URL"

Clean to plain text:

sed -e '/^$/d' -e '/^[0-9]/d' -e '/-->/d' -e 's/<[^>]*>//g' \
  /tmp/yt_transcript.en.vtt | sort -u > /tmp/yt_transcript.txt

Tips

• Short videos (<2 min): Use fps=1, review all frames
• Medium videos (2-10 min): Use fps=0.5, sample every 3-5 frames
• Long videos (>10 min): Use fps=0.2, focus on scene changes
• Tutorials/screencasts: Higher fps (1-2) captures more UI transitions
• Presentations/talks: Lower fps (0.2-0.5) is sufficient
• Combine frame analysis with transcript for best results
• Look for: screen transitions, text changes, button clicks, new panels/dialogs

Output

The extracted frames are numbered sequentially: frame_001.jpg, frame_002.jpg, etc.

Each frame filename corresponds to its position in time:

• At fps=1: frame_001.jpg = ~1s, frame_060.jpg = ~60s
• At fps=0.5: frame_001.jpg = ~2s, frame_030.jpg = ~60s

Cleanup

rm -rf /tmp/yt_video.mp4 /tmp/yt_frames /tmp/yt_transcript*