LoRA Pipeline

Orchestrates the full LoRA dataset-to-model pipeline. Each phase is self-contained and can be delegated to a sub-agent independently.

Pipeline Overview

Phase 1: 蒐集範例照片   → collect 3–6 reference face photos
Phase 2: 確認人臉正確   → user confirms refs; deepface cross-check
Phase 3: 蒐集 datasets  → scrape web sources guided by face features
Phase 4: 確認照片正確   → face verify + dedup + quality filter + crop
Phase 5: 開始 caption   → WD14 local tagging + trigger word
Phase 6: LoRA training  → RunPod Kohya training → retrieve outputs

Phase Index

To load a specific phase: read skills/lora-pipeline/phases/ — each file is independently readable.

Directory Structure

~/.openclaw/workspace/
└── datasets/
    ├── face_references/
    │   └── <lora_name>/          # Phase 1–2: Gold standard refs (3–6 photos)
    │       ├── ref_01.jpg
    │       └── ...
    ├── <lora_name>_raw/          # Phase 3: Raw scraped images (pre-verification)
    │   └── ...
    └── <lora_name>/              # Phase 4–5: Verified + captioned training set
        ├── image001.png
        ├── image001.txt
        └── ...

Privacy Rules (CRITICAL — All Phases)

• NO DATA INSPECTION: Do NOT cat, read, or analyze image file contents or .txt caption files.
• NO CLOUD UPLOAD: All face verification (DeepFace) must run locally. Never send images to cloud APIs.
• NO DATA LEAKAGE: Do not describe dataset details (person names, attributes) to the LLM unnecessarily.
• Treat datasets as opaque binary blobs except when running local scripts.

Quality Standards (SDXL)

• Resolution: 1024×1024 minimum after crop
• Format: Convert all to PNG before training
• No black borders: Run autocrop before final save
• Dataset diversity: ≥30% clothed/natural skin shots

Scripts

Sub-Agent Protocol

Each phase file contains:

1. Input Contract — what must already exist before this phase starts
2. Output Contract — what this phase produces
3. Completion Signal — how to report back (sessions_send + status file fallback)
4. Error Escalation — sub-agent reports to parent, never self-escalates model tier