Colorize

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

bashCopy code
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

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

Common Commands

bashCopy code
# 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

Effect Parameters

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

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.