Reframe Social

Reframe for social

reframe converts to a target aspect ratio two ways. Pick by what the footage is:

# Vertical 9:16 talking-head (full-bleed crop, subject centered)
uv run video-agent reframe in.mp4 -o reel.mp4 --aspect 9:16

# Vertical with blurred-background fill (nothing cropped)
uv run video-agent reframe in.mp4 -o reel.mp4 --aspect 9:16 --mode pad --width 1080

# Square / portrait
uv run video-agent reframe in.mp4 -o sq.mp4 --aspect 1:1
uv run video-agent reframe in.mp4 -o pt.mp4 --aspect 4:5

--width sets the final output width (height follows the aspect); omit to keep native resolution. For 9:16 social, --width 1080 → 1080×1920, the standard upload size.

Keeping an off-center subject in frame (crop mode)

Crop defaults to center (--focus 0.5). If the speaker sits left/right of center, the default crop clips them. There's no face tracking (by design). Set the crop bias by hand:

1. Read where the subject is: uv run video-agent position-grid in.mp4 --at <t> --spacing 200 -o grid.png, open it, note the subject's center x.
2. focus = subject_center_x / frame_width (0 = hug left edge, 1 = hug right edge).
3. uv run video-agent reframe in.mp4 -o reel.mp4 --aspect 9:16 --focus 0.62
4. Verify: uv run video-agent frame reel.mp4 --at <t> -o check.png and look.

Following a subject that MOVES across the shot

One fixed --focus can't track motion. Cut the clip at the points where the subject moves, reframe each piece with its own focus, then concat:

# subject starts left, walks right around 4s
uv run video-agent trim in.mp4 --start 0.0 --end 4.0 -o a.mp4
uv run video-agent trim in.mp4 --start 4.0 --end 8.0 -o b.mp4
uv run video-agent reframe a.mp4 -o ar.mp4 --aspect 9:16 --focus 0.3
uv run video-agent reframe b.mp4 -o br.mp4 --aspect 9:16 --focus 0.7
uv run video-agent concat ar.mp4 br.mp4 -o reel.mp4

Use detect over the range to find exactly where the subject crosses, so the cut lands on the move (not mid-gesture).

Gotchas

• Crop loses the edges — always eyeball a frame from the output before shipping; a gesture or second person at the edge may be gone.
• Pad mode never upscales the content — it scales to fit and fills the rest with a blurred, zoomed copy. Output canvas can be large without --width; pass --width to pin the size.
• Caption after reframing, not before — so caption size/wrapping match the final frame. See the captions skill.
• Audio is re-encoded (aac); video uses h264_videotoolbox. After reframe, a further join must use filter_complex concat or the concat command (it re-encodes audio).