---

name: data-preparation description: "Prepare instance images, prior-preservation bundles, and concept metadata for Custom Diffusion training." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Data preparation

Use this sub-skill when you need to stage and validate training inputs before a Custom Diffusion run.

It covers:

• instance image directories for one or more concepts
• generated or real-prior class image layouts
• images.txt / caption.txt / urls.txt bundles
• concept-list JSON manifests
• offline validation of regularization inputs

It does not cover training launch, sample generation, delta math, or benchmark evaluation. Route those tasks to:

• ../training/SKILL.md
• ../inference/SKILL.md
• ../checkpoint-tools/SKILL.md
• ../benchmarking/SKILL.md

Start here

1. Read references/data-formats.md.
2. Run scripts/validate_regularization_layout.py.
3. Read references/workflows.md if you need the end-to-end prep flow.
4. Check references/troubleshooting.md when a manifest or layout fails.

Network boundary

The real-prior retrieval helper is reference-only because it queries the LAION KNN service and downloads images. Do not make data-preparation depend on that network path when an offline bundle is available.