AnyDoor

AnyDoor is a zero-shot object-level image customization repo built around a control-diffusion pipeline, DINOv2 conditioning, mask-aware preprocessing, local Gradio demo support, and mixed-dataset training utilities.

Use this root skill as a router, not as a full manual. It points future agents to the smallest workflow-specific sub-skill that owns the task.

Start here

Read the root references first when you need the whole repo shape:

• references/overview.md for the top-level repository map and workflow split.
• references/environment-and-installation.md for supported install paths, Python/CUDA expectations, and the safest preflight order.
• references/checkpoints-and-configs.md for config placeholders, checkpoint files, and the paths that must be patched before generation.
• references/troubleshooting.md for cross-cutting import, checkpoint, and path failures.
• references/repo-provenance.md and references/repo-routing-metadata.json when you need staleness or router-placement metadata.

Run scripts/check_anydoor_environment.py before any generation or training workflow. It checks the repo layout, CUDA readiness, optional xformers, and placeholder config paths without running a full model.

If placeholder config values are still present, use scripts/patch_anydoor_configs.py to replace them before routing into an inference or training branch.

Routing map

User intent	Route to	Read next
Fresh clone, install, import, checkpoint, or CUDA questions	sub-skills/setup-and-checkpoints/	references/environment-and-installation.md and references/checkpoints-and-configs.md
Single-image customization, VITON-HD inference, Gradio demo, or Cog prediction	sub-skills/inference-and-demo/	sub-skills/inference-and-demo/references/inference-workflows.md, sub-skills/inference-and-demo/references/mask-and-image-formats.md, sub-skills/inference-and-demo/references/demo-and-cog-deployment.md
Dataset layout, preprocessing, training, debug sampling, or checkpoint conversion	sub-skills/data-and-training/	sub-skills/data-and-training/references/dataset-formats.md, sub-skills/data-and-training/references/training-workflows.md, sub-skills/data-and-training/references/model-architecture.md

What this root skill owns

• Project-wide repo orientation and workflow routing.
• Shared preflight checks for imports, CUDA, placeholder configs, and checkpoint presence.
• Cross-cutting troubleshooting for install, import, backend, and path issues.
• A consistent entry point for future agents so they do not have to rediscover the repo layout.

What this root skill does not own

• Detailed inference preprocessing and output composition.
• Dataset family-specific path rules and training recipes.
• Model architecture internals beyond what is needed to explain AnyDoor configuration and path requirements.

Common signals and destinations

• "I cannot import the repo" or "xformers is missing" → read setup first.
• "Generate from a reference image and mask" → go to inference and demo.
• "How do I prepare UVO / VITON-HD / SAM / MVImageNet data?" → go to data and training.
• "How do I patch checkpoint placeholders?" → go to setup and checkpoints.
• "How do I launch the demo or Cog predictor?" → go to inference and demo.
• "How do I start training or convert weights?" → go to data and training.

Short policy reminders

• Any actual generation in this repo is CUDA-backed; CPU checks are useful for support workflows only.
• Do not assume the placeholder checkpoint values in the configs are usable.
• Do not assume bundled examples or datasets are present on the target machine.
• Keep runtime links inside this skill tree only.

If you need deeper detail

Each sub-skill has its own focused references and scripts. Read the nearest one instead of turning this router into a manual.