---

name: pix2pix-hd description: "Route pix2pixHD setup, training, inference, and feature-conditioned workflows." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

pix2pixHD

Use this root skill as the router for pix2pixHD setup, training, checkpointed inference, and instance-feature workflows. The repository is script-based rather than a packaged library, so the bundled helpers take an explicit repo root instead of relying on an installable wheel.

Start here

1. Read Repository provenance if you need to know whether this skill still matches the current checkout.
2. Read Workflows to choose the right sub-skill.
3. Run the shared smoke check:
   • python scripts/check_environment.py --repo-root <repo-root>
4. Then route to the sub-skill that matches the task.

What this skill covers

• Cityscapes-style dataset setup and option defaults
• Training recipes, checkpointing, and memory planning
• Checkpointed inference, HTML output, and optional export/runtime paths
• Instance-aware feature encoding, clustering, and feature-conditioned workflows

Route map

setup-and-data

Use when the task is about prerequisites, dataroot layout, TrainOptions / TestOptions basics, bundled sample data, or quick loader smoke checks.

Read its helper scripts first when you need to verify label/instance/image folders or test the legacy resize path.

training

Use when the task asks about train.py, 512p or 1024p recipes, checkpoint cadence, resume behavior, VRAM planning, FP16, or multi-GPU training.

Read its command-builder helper first when you need a canonical recipe without launching a long run.

inference

Use when the task asks about test.py, HTML result browsing, checkpoint preflight, --export_onnx, --engine, --onnx, or result-file locations.

Read its checkpoint checker first when the requested experiment name or epoch may be wrong.

instance-features

Use when the task needs encode_features.py, precompute_feature_maps.py, --instance_feat, --label_feat, --load_features, or the clustered feature cache.

Read its cache checker first when feature-conditioned training or inference depends on cached feature files.

Minimal runtime expectations

• torch and torchvision are required for all workflows.
• dominate is required for HTML result rendering.
• scikit-learn is required for feature clustering.
• CUDA is required for the published training, inference, and feature workflows; CPU-only usage is limited to setup and smoke checks.

Reference files

• API reference — verified module signatures and object roles.
• Data layout — paired folder conventions, checkpoint roots, and result roots.
• Troubleshooting — cross-cutting install, backend, checkpoint, and compatibility failures.

Minimal import check

Use the shared smoke helper from any checkout:

• python scripts/check_environment.py --repo-root <repo-root>

If you need a deeper per-workflow check, use the sub-skill helper scripts linked above. Keep all runtime links inside this generated skill tree.