Colorize skill

Colorize

Workflow

1. Collect the user's inputs: images, videos, or Live Photos. Use absolute paths.
2. Resolve the skill directory and install dependencies once (see Requirements).
3. Run scripts/colorize.py with the inputs and an output path.
4. Wait for VALIDATION=PASS and OUTPUT_VIDEO=<absolute path>.
5. Optionally extract a few frames with ffmpeg to check the look before replying.
6. Return the output path and the settings used.

Quick Start

skill_dir=""
for base in "${AGENTS_HOME:-$HOME/.agents}" "${CLAUDE_HOME:-$HOME/.claude}" "${CODEX_HOME:-$HOME/.codex}"; do
  if [ -d "$base/skills/colorize" ]; then
    skill_dir="$base/skills/colorize"
    break
  fi
done
[ -n "$skill_dir" ] || { echo "colorize skill not found under ~/.agents, ~/.claude, or ~/.codex"; exit 1; }

python3 -m pip install -r "$skill_dir/scripts/requirements.txt"
python3 "$skill_dir/scripts/colorize.py" "/abs/path/photo.jpg" -o "$(pwd)/out/photo_colorize.mp4"

If the agent already knows the directory containing this SKILL.md, use that as skill_dir.

Inputs

Input	What happens
Image (png jpg jpeg webp heic heif avif bmp tif)	Static plate; the reveal lasts --duration seconds, then --hold seconds of the settled frame.
Video (mp4 mov m4v webm mkv avi gif)	Each source frame keeps moving; the reveal runs over the first --duration seconds and the rest of the clip keeps the settled look. Audio is kept.
iPhone Live Photo	Pass the still (IMG_0001.HEIC or .JPG). A same-name .MOV beside it is used automatically, with its audio. Pass the .MOV directly for the same result.
Android motion photo	Pass the .jpg; the embedded MP4 (Google MP.jpg / MVIMG, Samsung) is extracted automatically.
Several inputs	Rendered in order and joined into one MP4. Each input gets its own reveal.

Use --still to ignore Live Photo or motion photo video and animate only the still.

Common Commands

# One photo, 2 s reveal + 0.5 s hold
python3 "$skill_dir/scripts/colorize.py" photo.jpg --hold 0.5 -o out/photo.mp4

# Live Photo (IMG_0001.MOV next to it is picked up)
python3 "$skill_dir/scripts/colorize.py" IMG_0001.HEIC -o out/live.mp4

# Video: slower 3 s reveal, keep the first 6 s
python3 "$skill_dir/scripts/colorize.py" clip.mov --duration 3 --max-length 6 -o out/clip.mp4

# Several images and videos as one vertical reel
python3 "$skill_dir/scripts/colorize.py" a.jpg b.heic c.mov --size 1080x1920 -o out/reel.mp4

# Stronger spin and a little more glow
python3 "$skill_dir/scripts/colorize.py" photo.jpg --set angle=0.09 --set bloom=0.7 -o out/strong.mp4

Options

Option	Default	Meaning
-o, --output	<first input>_colorize.mp4	Output MP4 path.
--duration	2	Reveal length in seconds.
--hold	0	Images only: seconds to hold the settled frame.
--fps	30	Output frame rate.
--size WxH	first input's aspect, long edge ≤ --max-edge	Force output size; other inputs are center-cropped to it.
--max-edge	1920	Long-edge cap when --size is not set. Never upscales.
--max-length	full clip	Videos: keep at most this many seconds.
--still	off	Ignore Live Photo / motion photo video.
--mute	off	Drop audio.
--set KEY=VALUE	—	Override one effect parameter (repeatable).
--params file.json	—	Override several effect parameters.
--crf	17	x264 quality; lower is larger and sharper.

Effect Parameters

The defaults were fitted to the Jianying effect stack on static photos. Useful knobs for --set:

Key	Default	Effect
angle	0.0503	Spin-blur arc in radians; larger is stronger, the sign sets direction. Decays to zero.
focus_scale	0.9	Final width of the sharp center disc; larger settles the edges sooner.
motion	0.70	Horizontal motion-blur strength; decays to zero.
blur_power	1.0	Blur decay curve; larger fades faster.
front_speed	0.63	Speed of the color sweep.
edge	0.24	Softness of the color/mono boundary.
bloom / rays	0.54 / 0.29	Soft glow and radial light-ray strength.
softness	2.3	Residual softening; 0 is sharpest.
saturation	0.98	Final saturation.
center_x, center_y	0.5	Spin center in UV (y up).

Behavior

• Renders on the GPU with OpenGL 3.3 through moderngl (falls back to EGL on headless Linux) and encodes H.264 yuv420p BT.709 MP4 with ffmpeg.
• Output follows the first input's aspect ratio and orientation, including rotation metadata on phone videos.
• HDR (HLG/PQ) videos are tone-mapped to SDR when ffmpeg has zscale.
• HEIC/AVIF stills need pillow-heif, macOS sips, or an ffmpeg with HEIF support.
• When any input has audio, every segment carries an AAC track (silence for images) so joined clips stay in sync.
• Videos shorter than --duration get a compressed reveal that still ends fully settled.
• Prints VALIDATION=PASS after checking the output's size, duration, and audio with ffprobe.

Requirements

• Python 3.9+ with numpy, moderngl, Pillow (scripts/requirements.txt); optional pillow-heif
• ffmpeg and ffprobe on PATH
• A GPU or driver with OpenGL 3.3 (macOS, Windows, desktop Linux, or Linux with EGL)

On macOS, if moderngl has to build from source: CXXFLAGS='-std=c++11' python3 -m pip install moderngl.