---

name: "still-image-demo" description: "Run 3DDFA_V2 still-image alignment, rendering, pose, texture, and mesh export workflows." metadata: disco-role: operating disable-model-invocation: true license: MIT

Still-image demo

Use this sub-skill for demo.py workflows on a single image: sparse or dense landmarks, 3D renderings, depth, PNCC, UV texture, pose boxes, and PLY/OBJ mesh exports.

When to read

Read this sub-skill when the task asks to:

• Run 3DDFA_V2 on a still image.
• Generate 2d_sparse, 2d_dense, 3d, depth, pncc, uv_tex, pose, ply, or obj outputs.
• Use --onnx for a single-image demo.
• Diagnose No face detected, missing result files, headless plotting, or mode-specific failures.

Before running

1. Use ../setup-and-assets/ if native extensions or checkpoints are not known to be ready.
2. Check ../../references/model-assets.md when switching configs or missing weight files.
3. In headless environments, keep --show_flag false; the bundled wrapper also sets a headless plotting backend.

Main wrapper

The bundled wrapper preserves the original demo.py CLI. Put original demo arguments after --:

python <skill-root>/sub-skills/still-image-demo/scripts/run-still-image.py \
  --repo-root <checkout> -- \
  -f <image-path> -o 3d --show_flag false --onnx

Use direct repo commands only when you already applied the same compatibility and headless setup the wrapper provides.

Output selection

Read references/workflows.md for the full output-mode table. The highest-use choices are:

• 2d_sparse for 68-point landmark overlays.
• 2d_dense for dense landmark visualization.
• 3d for rendered dense mesh overlay.
• depth, pncc, and uv_tex for specialized per-pixel visual products.
• pose for yaw/pitch/roll pose-box visualization.
• ply and obj for mesh serialization.

Outputs default to examples/results/ and use the input basename plus the selected option.

Decision points

• Prefer --onnx for CPU latency and when the .onnx assets are already present or can be auto-converted.
• Use configs/mb05_120x120.yml when the task values speed over the default backbone.
• Use dense reconstruction for 3d, depth, pncc, uv_tex, ply, and obj; sparse mode is enough for 2d_sparse and pose.
• If the user needs multi-frame tracking, route to ../video-and-tracking/.
• If the user asks about latency numbers rather than a saved image, route to ../onnx-and-benchmarking/.

Troubleshooting

Read references/troubleshooting.md for still-image-specific failures and ../../references/troubleshooting.md for shared build/import failures.