reframe

reframe — horizontal to vertical 9:16

Turns 16:9 footage into subject-tracked 1080x1920 vertical video. Runs in WSL2 Ubuntu (GPU: RTX 4050).

Engine 2 — verthor (primary, verified)

Scene-aware reframe: PySceneDetect → YOLOv11-seg + ByteTrack → MediaPipe face/pose + saliency → smoothed virtual-camera pan/zoom → ffmpeg encode. Emits 1080x1920 H.264 with audio intact.

Run (from a FILE — never pipe the script in):

wsl -- bash /mnt/c/Users/agentuser/source/agent-skills-toolchain/skills/reframe/scripts/reframe.sh IN.mp4 OUT.mp4 preset

• Run the script from a file, NOT ... | bash: verthor/mediapipe read stdin, so a piped script gets consumed and corrupts after the verthor line. If you see a \r parse error, normalize once: wsl -- bash -lc "sed -i 's/\r//g' <script>".
• Paths are WSL paths: Windows D:\x.mp4 → /mnt/d/x.mp4.
• <preset>: talking_head (default) | sports | pets | cars.
• Or invoke verthor directly: wsl -- bash -lc "source ~/.local/share/reframe/verthor/bin/activate && cd /mnt/d/tools/reframe/verthor && python -m verthor <IN> <OUT> --preset talking_head --saliency-model handcrafted --video-encoder libx264"

Safety / pinned choices (do not change):

• --saliency-model handcrafted ALWAYS. NEVER deepgazemr (loads remote model via torch.hub = RCE risk).
• --video-encoder libx264 (deterministic; NVENC auto-degrades on stock ffmpeg).
• verthor pinned 33ac38d; venv at ~/.local/share/reframe/verthor (WSL).

Engine 1 — claude-shorts (comparison, see references/engines.md)

Acceptance

Setup (one-time, already done on this machine)