Getting Started

Prerequisites

You need three things running before you start:

1. ComfyUI — a local Stable Diffusion backend. Install from github.com/comfyanonymous/ComfyUI. Style Dataset Lab drives it over HTTP on localhost:8188.

2. Node.js 20+ — all scripts are ESM JavaScript. No build step, no TypeScript compilation.

3. repo-dataset (optional, for export) — the @mcptoolshop/repo-dataset CLI reads your records and images to produce training data. Install from github.com/mcp-tool-shop-org/repo-dataset.

Model weights

The default generation setup uses an SDXL checkpoint and a style LoRA. Configure these in your project’s project.json under defaults.checkpoint and defaults.loras. Place model files in your ComfyUI models/checkpoints/ and models/loras/ directories.

Install

# Install globally for the sdlab CLI
npm install -g @mcptoolshop/style-dataset-lab

# Or clone the repo for the Star Freight example
git clone https://github.com/mcp-tool-shop-org/style-dataset-lab
cd style-dataset-lab

Scaffold a project

# Create a project from a domain template
sdlab init my-project --domain character-design

# See available domains
sdlab init

Available domains: game-art, character-design, creature-design, architecture, vehicle-mech, or omit --domain for a generic starter.

This creates projects/my-project/ with config files, canon templates, directory structure, and a starter prompt pack.

Validate the project

sdlab project doctor --project my-project

Doctor checks config files, lane patterns, rubric dimensions, constitution rules, and directory structure. Fix any failures before generating.

Customize your canon

Edit the config files in your project root:

• constitution.json — add or modify rules that define your visual law
• lanes.json — define subject categories (portrait, full_body, etc.)
• rubric.json — set scoring dimensions and approval thresholds
• terminology.json — define style groups/factions with detection patterns
• canon/constitution.md — human-readable style rules for the team

Start ComfyUI

cd /path/to/ComfyUI
python main.py --listen 127.0.0.1 --port 8188

Verify it is running:

curl http://127.0.0.1:8188/system_stats

Generate your first wave

Edit projects/my-project/inputs/prompts/example-wave.json with your subjects, style prefix, and generation defaults. Then:

# Preview what would be generated (no ComfyUI calls)
sdlab generate inputs/prompts/example-wave.json --project my-project --dry-run

# Generate for real
sdlab generate inputs/prompts/example-wave.json --project my-project

Each generated image lands in outputs/candidates/ with a matching record in records/. The record captures full provenance: checkpoint, LoRA, seed, steps, cfg, sampler, scheduler, resolution, and the exact prompt used.

Curate

Once you have candidates, curate them:

# List uncurated candidates
sdlab curate --list --project my-project

# Approve with per-dimension scores
sdlab curate my_subject_v1 approved "Clean proportions, good gesture" \
  --scores proportion_accuracy:0.9,gesture_clarity:0.85 --project my-project

# Reject with failure modes
sdlab curate my_subject_v2 rejected "Broken anatomy, stiff pose" \
  --failures broken_anatomy,stiff_pose --project my-project

Curation moves the image from outputs/candidates/ to outputs/approved/ (or rejected/, borderline/) and writes the judgment into the record.

Bind to canon

After curation, bind records to constitution rules:

# Bind all curated records
sdlab bind --project my-project

# Preview without writing
sdlab bind --project my-project --dry-run

# Print coverage stats
sdlab bind --project my-project --stats

Each record gets canon.assertions — an array of rule citations with pass/fail/partial verdicts and rationale strings.

Export training data

Use repo-dataset to produce training data from your curated, canon-bound records:

repo-dataset visual generate ./projects/my-project --format trl
repo-dataset visual inspect ./projects/my-project

Explore the Star Freight example

If you cloned the repo, the Star Freight project has 1,182 records ready to explore:

sdlab bind --stats --project star-freight
sdlab project doctor --project star-freight