Style Dataset Lab Handbook

Style Dataset Lab turns approved visual work into versioned, review-backed datasets, splits, export packages, and eval packs — then puts trained models to work in a full production workflow and feeds the best outputs back into the corpus. Teams define what they are making (visual canon), curate outputs against project-specific rubrics, bind accepted work to constitution rules, produce reproducible dataset packages, and run a closed production loop through brief compilation, generation, selection, and re-ingest.

The pipeline ships as an npm package (@mcptoolshop/style-dataset-lab) with the sdlab CLI, shared library modules, and domain-specific starter templates.

The pipeline

canon → generate → curate → bind → snapshot → split → export → eval
  |        |          |        |        |         |        |       |
rules   ComfyUI   judgment  rules   frozen    subject  package  verify
                                    selection isolation
         ↓
    brief → run → critique → batch → select → re-ingest
              ↑                                    |
              └────────────────────────────────────┘

Each stage writes structured JSON records to projects/<name>/records/. All commands accept --project <name> (defaults to star-freight). Records accumulate provenance, judgment, and canon binding over time. Nothing is lost — a curated record still carries its original generation provenance.

The four dataset artifacts

The dataset spine produces four artifacts. These are the product.

Artifact	What it is	Command
Snapshot	Frozen, fingerprinted selection of eligible records with explicit reason traces	sdlab snapshot create
Split	Subject-isolated, lane-balanced train/val/test partition (seeded PRNG, zero leakage)	sdlab split build
Export package	Self-contained dataset: manifest, metadata.jsonl, images, splits, dataset card, checksums	sdlab export build
Eval pack	Canon-aware test instruments: lane coverage, forbidden drift, anchor/gold, subject continuity	sdlab eval-pack build

sdlab defines and owns the dataset. Downstream format conversion (TRL, LLaVA, Parquet) is handled by repo-dataset.

The training triangle

Every record in the dataset carries three things:

1. Provenance — generated by ComfyUI with full history (checkpoint, LoRA, seed, sampler, cfg scale, resolution, timing).
2. Canon binding — which constitution rules this asset passes, fails, or partially meets, with rationale strings.
3. Quality judgment — approved, rejected, or borderline, with per-dimension scores and cited failure modes.

This triple — provenance, canon, judgment — is what makes the dataset useful beyond simple image labeling. A model trained on images alone learns to generate. A model trained on images with grounded judgment learns to evaluate.

Domain templates

The pipeline is domain-agnostic. Five starter templates ship with real production worldviews:

Domain	Focus	Example lanes
game-art	Game concept art with faction systems	character, environment, prop, ship, interior
character-design	Character production for film/animation/games	portrait, full_body, turnaround, expression_sheet
creature-design	Creature lineup and species development	concept, orthographic, detail_study, action, habitat
architecture	Architectural worldbuilding and pre-viz	exterior, interior, streetscape, ruin, landscape
vehicle-mech	Vehicle and mech design	exterior, cockpit, component, schematic, damage_variant

Each template includes constitution rules, lane definitions, scoring rubrics, and group vocabulary designed for that production context.

Project model

Each project lives in projects/<name>/ with 5 JSON config files alongside its data:

• project.json — identity, domain, generation defaults
• constitution.json — rules array with rationale templates
• lanes.json — subject lanes with regex detection patterns
• rubric.json — scoring dimensions, thresholds, failure-to-rule mappings
• terminology.json — group vocabulary with detection order

Use sdlab project doctor to validate any project’s config completeness.

Example: Star Freight

The repo includes projects/star-freight/ as a complete working example — a gritty sci-fi RPG with 5 factions, 1,182 curated records, 28 prompt waves, 7 lanes, and 24 constitution rules. Clone the repo to explore it.

Next steps

• Getting Started — install, scaffold a project, first generation
• Dataset Workflow — end-to-end: snapshot, split, export, eval
• Production Loop — brief, run, critique, batch, select, re-ingest
• Reference — all CLI commands with flags and arguments
• Architecture — pipeline flow, record schemas, config files
• Security — threat model and trust boundaries