Convert Depth Video
Use the bundled Python application to create grayscale relative-depth MP4 videos
locally. Resolve every relative path from this skill directory.
Launch
- Confirm that Python 3.10 through 3.14 is available.
- Run:
text
python scripts/launch.py
- Keep the process running and report the printed localhost URL to the user.
- Do not add
--server-name 0.0.0.0or enable Gradio sharing unless the user
explicitly requests network access and understands the exposure.
The launcher creates a versioned virtual environment in the operating system's
user cache directory. It installs dependencies only when the environment or
requirements change.
Useful options:
python scripts/launch.py --inbrowser
python scripts/launch.py --force-cpu
python scripts/launch.py --port 7861
python scripts/launch.py --output-dir "/path/to/completed-videos"
python scripts/launch.py --setup-only
python scripts/launch.py --reinstall
Pass all arguments except --setup-only and --reinstall through to the
bundled app.py.
Convert
Use these defaults unless the user specifies otherwise:
- Model: Small.
- Resolution: Original.
- Temporal smoothing: 0.65.
- Preserve audio: enabled.
- Invert grayscale: disabled.
Explain that the result is relative depth, not distance measured in meters.
Uploading a video must never transmit it to an external service. Model files
download from Hugging Face on first use and then run locally.
Model guidance:
- Use Small for CPU systems, long videos, previews, and commercial workflows.
- Use Base when quality and speed need a better balance.
- Use Large only when detail matters more than processing time and memory.
- Read
references/model-licenses.mdbefore advising about commercial use.
Troubleshoot
Read references/platform-setup.md when installation, GPU acceleration,
ffmpeg, or model loading fails.
Apply these first responses:
- CUDA out of memory: select Small, lower resolution, or use
--force-cpu. - MPS unsupported operation: restart with
--force-cpu. - Model download failure: check network, cache permissions, and free space.
- Audio missing: verify the source has audio and Preserve Audio is enabled.
- MOV decoding failure: transcode the source to H.264/AAC MP4 with ffmpeg.
- Port already used: restart with
--portand another local port.
Do not install CUDA toolkits speculatively. Use the current PyTorch installation
selector when a Windows NVIDIA wheel must be replaced.
Maintain
Treat scripts/app.py as the installed application's source of truth inside
the Skill. Keep scripts/requirements.txt synchronized with it.
After changes, run:
python scripts/launch.py --setup-only
python -m pytest -q
python /path/to/skill-creator/scripts/quick_validate.py /path/to/trmachel-convert-depth-video
