45ck/content-machine/face-or-screen-reframe

Face Or Screen Reframe

Use When

• The source footage is either a talking-head style clip or a screen

recording.

• One generic portrait crop keeps failing because the subject type is

different.

• You need a concrete branching rule before building a 9:16 crop.

Core Rule

• Decide what kind of source you have first.
• Face video and screen video should not share the same crop logic.

Inputs

• source video
• target portrait dimensions
• optional safe-zone or caption-lane constraints

Outputs

• chosen source classification:

face-led or screen-led

• crop strategy notes
• stable portrait crop or crop-keyframe plan

Workflow

1. Sample the early part of the clip and check for a stable face.
2. If a face is found, choose one stable face-centered crop rather than

hyperactive per-frame tracking.

1. If no face is found, treat the clip as screen-led footage and follow

the moving area instead.

1. Smooth screen-led crop movement heavily and update it at a controlled

interval.

1. Clamp the crop to source bounds and preserve room for captions.

Face-Led Pattern

• detect faces in the opening sample window
• pick the strongest or largest face
• use a stable central crop with slight bias if needed
• prefer calm framing over twitchy precision

Screen-Led Pattern

• resize if needed so the crop has room to move
• estimate the area of strongest motion or cursor activity
• update the crop slowly, not every frame
• fall back to a readable centered crop if motion confidence is weak

Bad Pattern

• one portrait crop heuristic for every source type
• chasing every face box change in a talking-head clip
• following noisy motion every frame in a screen recording
• forgetting caption safe zones when choosing the crop window

Pair With

• Use as a concrete sub-strategy under

`reframe-vertical`.

• Use with `longform-to-shorts` when

extracting portrait clips from wider source video.

Aggregated From

• SamurAIGPT/AI-Youtube-Shorts-Generator Components/FaceCrop.py
• source-type-specific reframe patterns from clipper repos

Validation Checklist

• The source is explicitly classified before crop logic is chosen.
• Face-led crops stay calm and readable.
• Screen-led crops follow the area of attention without jitter.
• Crop bounds and caption-safe areas are preserved throughout.