Avatar Frames Extractor

Extract production-ready avatar frames from video reels, interviews, or talking-head clips.

Setup

bash ~/.cursor/skills/avatar-frames/scripts/setup.sh

Quick Reference

Default — ready frames only

python3 ~/.cursor/skills/avatar-frames/scripts/extract_clean_frames.py video.mp4 -o frames/

Output: frame_0001.png, frame_0002.png, … (1 face, sharp, no subtitles)

Explicitly include inpaint candidates

python3 ~/.cursor/skills/avatar-frames/scripts/extract_clean_frames.py video.mp4 \
  -o frames/ --with-subtitles

Debug rejections

python3 ~/.cursor/skills/avatar-frames/scripts/extract_clean_frames.py video.mp4 \
  -o frames/ --save-rejected

Output Rules

When clean frames exist and --with-subtitles is not set, inpaint candidates are counted in manifest.json but not written to disk.

Options

Output Structure

Typical (ready frames found):

frames/
├── frame_0001.png
├── frame_0002.png
└── manifest.json

No ready frames (auto-fallback) or --with-subtitles:

frames/
├── frame_0001.png              # may be empty
├── with_subtitles/
│   ├── subtitle_0001.png
│   └── subtitle_0002.png
└── manifest.json

Pipeline

Per interval window (best frame every 2s):

1. Single face — short + full-range BlazeFace; reject 0 or 2+ faces
2. Face sharpness — Laplacian on face bbox (not full frame)
3. Subtitle check — EasyOCR on top (40%), overlay (30–85%), bottom (60–100%) zones
4. Route — ready → root; subtitled → inpaint pool; else → rejected
5. Dedup — perceptual hash per output group
6. Write — root always; with_subtitles/ only per output rules above

Manifest

manifest.json keys:

• frames — saved ready frames
• with_subtitles — saved inpaint frames (empty if skipped)
• inpaint_saved, inpaint_reason (user_requested | no_clean_frames | null)
• frames_with_subtitles_skipped — candidates found but not saved
• rejected — metadata only (images only with --save-rejected)
• subtitle_style — burned-in caption style profiled from subtitled frames (null if none)

Subtitle style profile

When a video has burned-in captions, the extractor profiles their style from the subtitled frames and writes both a subtitle_style block in manifest.json and a standalone subtitle_style.json:

{ "y_frac": 0.62, "text_height_frac": 0.059, "fontsize_frac": 0.082,
  "lines": 2, "words_per_caption": 3, "casing": "lower",
  "color_rgb": [240,228,224], "color_hex": "#f0e4e0",
  "progression": "replace", "mean_word_overlap": 0.10,
  "emphasis": { "auto_detected": false, "convention": "…" },
  "samples": 21, "note": "…" }

• Measured reliably: vertical position (y_frac), text size (text_height_frac / suggested fontsize_frac), lines, words_per_caption, text color.
• Progression (approximate): progression is replace vs accumulate, derived from the word overlap between consecutive caption samples (mean_word_overlap). It tells downstream tools whether captions swap out per phrase (replace) or build up within a phrase (accumulate). Sparse sampling makes it a hint, not a measurement.
• Emphasis (convention, not detected): OCR can't read weight/italic, so emphasis.convention documents what the originals highlight and why — on these reels the breath-ending payoff words are set in bold-italic of the same serif. Downstream reproduces this by emphasizing each breath group's completion.
• Low confidence: casing — EasyOCR lowercases its output, so only upper (all-caps) is trustworthy; treat a reported lower as natural.
• Not detected: font family (serif vs sans), weight, italic emphasis.

Downstream (e.g. avatar-reel-composer/finish_reel.py --style-from) consumes this to place/size new captions and to mirror the originals' replace-per-phrase progression + payoff emphasis. Disable with --no-subtitle-style.

Tuning

Batch

for f in videos/*.mp4; do
  name=$(basename "$f" .mp4)
  python3 ~/.cursor/skills/avatar-frames/scripts/extract_clean_frames.py "$f" -o "frames/$name"
done

Only merge ready frame_*.png files unless the user asks for inpaint candidates.

Integration

• Inpaint: run on with_subtitles/ after auto-fallback or --with-subtitles
• Asset generation: use root frame_*.png as --ref in asset-generator or gpt-image-2

Notes

• Face sharpness uses the face bbox, not the full frame.
• Social-media captions sit in the center overlay zone (35–85% height).
• EasyOCR is the bottleneck (~1–2s per candidate frame).