A headless wrapper around the Wan2GP engine that enables programmatic video generation and queue-based processing without the Gradio UI.

• Create venv: python3 -m venv venv_headless_wan2gp && source venv_headless_wan2gp/bin/activate
• Install deps:
  • python -m pip install --upgrade pip wheel setuptools
  • python -m pip install -r Wan2GP/requirements.txt
  • python -m pip install -r requirements.txt
• Run baseline tests (real generation):
  • python test_model_comparison.py --output-dir outputs/baseline_real_$(date +%Y%m%d_%H%M%S)

When no GPU is available, you can validate the integration path and end‑to‑end flow without heavy generation.

• Usage:
  • HEADLESS_WAN2GP_SMOKE=1 HEADLESS_WAN2GP_FORCE_CPU=1 python test_model_comparison.py --output-dir outputs/smoke_$(date +%Y%m%d_%H%M%S)
• Behavior:
  • Skips Wan2GP import and model loading
  • Produces placeholder outputs under Wan2GP/outputs/ and returns those paths
  • Leaves the upstream Wan2GP/ code untouched

• scripts/run_baseline_smoke.sh – Runs the smoke-mode baseline tests
• scripts/run_baseline_real.sh – Runs the real baseline tests (requires GPU)

• GitHub Actions workflow Smoke Test runs the smoke-mode baseline on push/PR to validate end-to-end integration without GPU.
• GPU Baseline (workflow_dispatch) runs the real baseline on a self-hosted GPU runner. You can optionally set frames and steps inputs to avoid OOM.

• AI_AGENT_MAINTENANCE_GUIDE.md – Complete agent procedure for updates and validation
• maintenance_analysis/ – Living documents during maintenance cycles:
  • wan2gp_diff_analysis.md, system_integration_analysis.md, update_checklist.md, milestone_progress.txt, attempt_log.txt

• If you encounter CUDA OOM during baseline, reduce video_length in test_model_comparison.py (e.g., 65 → 33) and retry.
• The wrapper prefers upstream‑first fixes to maintain clean Wan2GP updates.