@mcptoolshop/style-dataset-lab 1.1.0

package/CHANGELOG.md

@@ -0,0 +1,37 @@
+# Changelog
+
+All notable changes to this project will be documented in this file.
+
+## [1.1.0] - 2026-04-14
+
+- Monorepo restructure: each game lives in `games/<name>/` with isolated canon, records, and assets
+- All scripts accept `--game <name>` flag (default: `star-freight`)
+- Existing Star Freight data moved to `games/star-freight/`
+- Translations added (7 languages: ja, zh, es, fr, hi, it, pt-BR)
+
+## [0.5.0] - 2026-04-14
+
+- Identity packet system for waves 27A, 27B, and 28
+- Painterly pipeline for stylized asset generation
+
+## [0.4.0] - 2026-03-28
+
+- Wave 26: station identities
+- Pairwise comparison tooling for quality ranking
+
+## [0.3.0] - 2026-03-15
+
+- Wave 25: alien species generation
+- Negative prompt fix (`human` in negative for alien anatomy)
+- Wave 25b: regeneration pass for failed outputs
+
+## [0.2.0] - 2026-02-20
+
+- Waves 11-24: expanded category coverage
+- Canon binding system linking outputs to design intent
+
+## [0.1.0] - 2026-01-10
+
+- Initial scaffold and ComfyUI integration
+- Waves 1-10: character costumes (gritty space theme)
+- First export: 570 training units in TRL format

package/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 mcp-tool-shop
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

package/README.es.md

@@ -0,0 +1,142 @@
+<p align="center">
+  <a href="README.ja.md">日本語</a> | <a href="README.zh.md">中文</a> | <a href="README.md">English</a> | <a href="README.fr.md">Français</a> | <a href="README.hi.md">हिन्दी</a> | <a href="README.it.md">Italiano</a> | <a href="README.pt-BR.md">Português (BR)</a>
+</p>
+
+<p align="center">
+  <img src="https://raw.githubusercontent.com/mcp-tool-shop-org/brand/main/logos/style-dataset-lab/readme.png" width="400">
+</p>
+
+<p align="center">
+  <a href="https://www.npmjs.com/package/@mcptoolshop/style-dataset-lab"><img src="https://img.shields.io/npm/v/@mcptoolshop/style-dataset-lab" alt="npm"></a>
+  <a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-blue" alt="MIT License"></a>
+</p>
+
+# style-dataset-lab
+
+Factoría de conjuntos de datos visuales: genera, organiza y exporta datos de entrenamiento multimodales para el ajuste fino de modelos de lenguaje visual (VLM).
+
+## ¿Qué es esto?
+
+Un conjunto de herramientas para construir **"verdad visual" entrenable**. Cada elemento contiene tres cosas:
+
+1. **Píxeles de la imagen:** generados por ComfyUI con total trazabilidad (punto de control, LoRA, semilla, muestreador, configuración).
+2. **Explicación canónica:** por qué este elemento está dentro o fuera del estilo, basado en una constitución de estilo.
+3. **Juicio de calidad:** aprobado/rechazado con puntuaciones por dimensión y reglas citadas.
+
+La salida se integra con [`@mcptoolshop/repo-dataset`](https://github.com/mcp-tool-shop-org/repo-dataset) para producir datos de entrenamiento multimodales en 10 formatos: TRL, LLaVA, Qwen2-VL, Axolotl, LLaMA-Factory, ShareGPT, pares RLHF, DPO, subtitulado y clasificación.
+
+## Modelo de seguridad
+
+**Solo local.** style-dataset-lab se comunica con ComfyUI en `localhost:8188` y nunca realiza solicitudes a la red externa. No hay telemetría, ni análisis, ni conexión a servidores externos. La generación de imágenes se realiza completamente en su GPU. Los registros y los datos canónicos permanecen en su sistema de archivos.
+
+## Estadísticas del conjunto de datos
+
+| Métrica | Valor |
+|--------|-------|
+| Registros organizados | 1,182 |
+| Elementos totales | 2.571 (893 aprobados, 887 variantes con estilo pictórico) |
+| Combinaciones de indicaciones | 28 |
+| Categorías visuales | 18 (disfraces, barcos, interiores, equipos, entornos, especies, estaciones, señalización, iluminación, carga, arquitectura, criaturas, superficies, vida cotidiana, planetario, daño/reparación, biología alienígena, detalles realistas) |
+| Comparaciones por pares | 6 creados por humanos + 71 creados sintéticamente |
+| Tipos de rechazo distintos | 30+ |
+| Sistema de paquetes de identidad | Personajes con nombre con ADN de facción, rango y anclajes visuales. |
+| Formatos de exportación | 10 (a través de `@mcptoolshop/repo-dataset`) |
+
+## Instalación
+
+```bash
+npm install -g @mcptoolshop/style-dataset-lab
+```
+
+Luego, clone un proyecto o inicializa un nuevo espacio de trabajo de conjunto de datos:
+
+```bash
+git clone https://github.com/mcp-tool-shop-org/style-dataset-lab my-dataset
+cd my-dataset
+npm install
+```
+
+## Flujo de trabajo
+
+```bash
+# 1. Start ComfyUI
+# (point it at your checkpoint + LoRA setup)
+
+# 2. Generate candidates from a prompt pack
+npm run generate -- inputs/prompts/wave1.json
+npm run generate -- inputs/prompts/wave1.json --dry-run
+
+# 3. Generate identity-packet characters
+npm run generate:identity -- inputs/identity-packets/wave27a-identity-spine.json
+
+# 4. Generate painterly variants of approved assets
+npm run painterly -- <asset_id>
+
+# 5. Curate -- approve, reject, or mark borderline
+npm run curate -- <asset_id> approved "explanation" --scores "silhouette:0.9,palette:0.8"
+npm run curate -- <asset_id> rejected "explanation" --failures "too_clean,wrong_material"
+
+# 6. Bind canon explanations to assets
+npm run canon-bind -- <asset_id>
+
+# 7. Record pairwise comparisons
+npm run compare -- <asset_a> <asset_b> a "A has better faction read because..."
+
+# 8. Export training data via repo-dataset
+npm run export
+npm run inspect
+npm run validate
+```
+
+## Estructura de directorios
+
+```
+canon/                  Style constitution, review rubric, identity gates, species canon
+inputs/
+  prompts/              Prompt packs per wave (JSON: subjects, variations, defaults)
+  references/           IP-Adapter reference images
+  control/              ControlNet control images
+  control-guides/       ControlNet guide overlays
+  identity-packets/     Named character identity spines (faction DNA, rank, visual anchors)
+outputs/
+  candidates/           Raw generations (gitignored)
+  approved/             Curated approved (gitignored)
+  rejected/             Curated rejected (gitignored)
+  borderline/           Curated borderline (gitignored)
+  painterly/            Painterly-style variants (gitignored)
+records/                Per-asset JSON (provenance + judgment + canon binding)
+comparisons/            A-vs-B preference judgments
+exports/                repo-dataset output (gitignored)
+scripts/                generate, curate, compare, canon-bind, painterly, identity gen
+workflows/              Reusable ComfyUI workflow templates
+```
+
+## Configuración de generación
+
+```yaml
+checkpoint: dreamshaperXL_v21TurboDPMSDE.safetensors
+lora: classipeintxl_v21.safetensors (weight: 1.0)
+resolution: 1024x1024
+steps: 8
+cfg: 2.0
+sampler: dpmpp_sde
+scheduler: karras
+speed: ~9s per image (RTX 5080)
+```
+
+Modos de generación adicionales: ControlNet (guiado por pose/profundidad), IP-Adapter (guiado por referencia) y paquetes de identidad (consistencia de personajes con nombre).
+
+## Requisitos
+
+- [ComfyUI](https://github.com/comfyanonymous/ComfyUI) ejecutándose en localhost:8188
+- Punto de control DreamShaper XL Turbo + LoRA ClassipeintXL
+- Node.js 20+
+- [`@mcptoolshop/repo-dataset`](https://github.com/mcp-tool-shop-org/repo-dataset) para la exportación de entrenamiento.
+
+## Licencia
+
+MIT
+
+---
+
+Creado por <a href="https://mcp-tool-shop.github.io/">MCP Tool Shop</a>

package/README.fr.md

@@ -0,0 +1,142 @@
+<p align="center">
+  <a href="README.ja.md">日本語</a> | <a href="README.zh.md">中文</a> | <a href="README.es.md">Español</a> | <a href="README.md">English</a> | <a href="README.hi.md">हिन्दी</a> | <a href="README.it.md">Italiano</a> | <a href="README.pt-BR.md">Português (BR)</a>
+</p>
+
+<p align="center">
+  <img src="https://raw.githubusercontent.com/mcp-tool-shop-org/brand/main/logos/style-dataset-lab/readme.png" width="400">
+</p>
+
+<p align="center">
+  <a href="https://www.npmjs.com/package/@mcptoolshop/style-dataset-lab"><img src="https://img.shields.io/npm/v/@mcptoolshop/style-dataset-lab" alt="npm"></a>
+  <a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-blue" alt="MIT License"></a>
+</p>
+
+# style-dataset-lab
+
+Fabrique de jeux de données visuels : génère, organise et exporte des données d'entraînement multimodales pour l'affinage des modèles de langage visuels (VLM).
+
+## Qu'est-ce que c'est ?
+
+Une boîte à outils pour créer une **vérité visuelle entraînable**. Chaque élément contient trois informations :
+
+1. **Pixels de l'image** : générés par ComfyUI avec une traçabilité complète (point de contrôle, LoRA, graine, échantillonneur, cfg).
+2. **Explication canonique** : pourquoi cet élément est conforme ou non au style, basé sur une constitution du style.
+3. **Jugement de qualité** : approuvé/refusé avec des scores par dimension et des règles citées.
+
+La sortie est intégrée à [`@mcptoolshop/repo-dataset`](https://github.com/mcp-tool-shop-org/repo-dataset) pour produire des données d'entraînement multimodales dans 10 formats : TRL, LLaVA, Qwen2-VL, Axolotl, LLaMA-Factory, ShareGPT, paires RLHF, DPO, légende et classification.
+
+## Modèle de sécurité
+
+**Uniquement local.** style-dataset-lab communique avec ComfyUI sur `localhost:8188` et ne fait jamais de requêtes réseau externes. Pas de télémétrie, pas d'analyses, pas de communication avec un serveur distant. La génération d'images se fait entièrement sur votre GPU. Les enregistrements et les données canoniques restent sur votre système de fichiers.
+
+## Statistiques du jeu de données
+
+| Métrique | Valeur |
+|--------|-------|
+| Enregistrements organisés | 1,182 |
+| Nombre total d'éléments | 2 571 (893 approuvés, 887 variantes picturales) |
+| Séquences de prompts | 28 |
+| Catégories visuelles | 18 (costumes, bateaux, intérieurs, équipements, environnements, espèces, stations, signalisation, éclairage, cargaison, architecture, créatures, surfaces, vie quotidienne, planétaire, dommages/réparations, biologie extraterrestre, détails de vie) |
+| Comparaisons par paires | 6 réalisées par des humains + 71 synthétiques |
+| Types de refus distincts | 30+ |
+| Système de paquets d'identité | Personnages nommés avec ADN de faction, grade et ancres visuelles. |
+| Formats d'exportation | 10 (via @mcptoolshop/repo-dataset) |
+
+## Installation
+
+```bash
+npm install -g @mcptoolshop/style-dataset-lab
+```
+
+Clonez ensuite un projet ou initialisez un nouvel espace de travail de jeu de données :
+
+```bash
+git clone https://github.com/mcp-tool-shop-org/style-dataset-lab my-dataset
+cd my-dataset
+npm install
+```
+
+## Flux de travail
+
+```bash
+# 1. Start ComfyUI
+# (point it at your checkpoint + LoRA setup)
+
+# 2. Generate candidates from a prompt pack
+npm run generate -- inputs/prompts/wave1.json
+npm run generate -- inputs/prompts/wave1.json --dry-run
+
+# 3. Generate identity-packet characters
+npm run generate:identity -- inputs/identity-packets/wave27a-identity-spine.json
+
+# 4. Generate painterly variants of approved assets
+npm run painterly -- <asset_id>
+
+# 5. Curate -- approve, reject, or mark borderline
+npm run curate -- <asset_id> approved "explanation" --scores "silhouette:0.9,palette:0.8"
+npm run curate -- <asset_id> rejected "explanation" --failures "too_clean,wrong_material"
+
+# 6. Bind canon explanations to assets
+npm run canon-bind -- <asset_id>
+
+# 7. Record pairwise comparisons
+npm run compare -- <asset_a> <asset_b> a "A has better faction read because..."
+
+# 8. Export training data via repo-dataset
+npm run export
+npm run inspect
+npm run validate
+```
+
+## Structure des répertoires
+
+```
+canon/                  Style constitution, review rubric, identity gates, species canon
+inputs/
+  prompts/              Prompt packs per wave (JSON: subjects, variations, defaults)
+  references/           IP-Adapter reference images
+  control/              ControlNet control images
+  control-guides/       ControlNet guide overlays
+  identity-packets/     Named character identity spines (faction DNA, rank, visual anchors)
+outputs/
+  candidates/           Raw generations (gitignored)
+  approved/             Curated approved (gitignored)
+  rejected/             Curated rejected (gitignored)
+  borderline/           Curated borderline (gitignored)
+  painterly/            Painterly-style variants (gitignored)
+records/                Per-asset JSON (provenance + judgment + canon binding)
+comparisons/            A-vs-B preference judgments
+exports/                repo-dataset output (gitignored)
+scripts/                generate, curate, compare, canon-bind, painterly, identity gen
+workflows/              Reusable ComfyUI workflow templates
+```
+
+## Configuration de la génération
+
+```yaml
+checkpoint: dreamshaperXL_v21TurboDPMSDE.safetensors
+lora: classipeintxl_v21.safetensors (weight: 1.0)
+resolution: 1024x1024
+steps: 8
+cfg: 2.0
+sampler: dpmpp_sde
+scheduler: karras
+speed: ~9s per image (RTX 5080)
+```
+
+Modes de génération supplémentaires : ControlNet (guidé par la pose/la profondeur), IP-Adapter (guidé par une référence) et paquets d'identité (cohérence des personnages nommés).
+
+## Prérequis
+
+- [ComfyUI](https://github.com/comfyanonymous/ComfyUI) en cours d'exécution sur localhost:8188
+- Point de contrôle DreamShaper XL Turbo + LoRA ClassipeintXL
+- Node.js 20+
+- [`@mcptoolshop/repo-dataset`](https://github.com/mcp-tool-shop-org/repo-dataset) pour l'exportation des données d'entraînement
+
+## Licence
+
+MIT
+
+---
+
+Créé par <a href="https://mcp-tool-shop.github.io/">MCP Tool Shop</a>

package/README.hi.md

@@ -0,0 +1,142 @@
+<p align="center">
+  <a href="README.ja.md">日本語</a> | <a href="README.zh.md">中文</a> | <a href="README.es.md">Español</a> | <a href="README.fr.md">Français</a> | <a href="README.md">English</a> | <a href="README.it.md">Italiano</a> | <a href="README.pt-BR.md">Português (BR)</a>
+</p>
+
+<p align="center">
+  <img src="https://raw.githubusercontent.com/mcp-tool-shop-org/brand/main/logos/style-dataset-lab/readme.png" width="400">
+</p>
+
+<p align="center">
+  <a href="https://www.npmjs.com/package/@mcptoolshop/style-dataset-lab"><img src="https://img.shields.io/npm/v/@mcptoolshop/style-dataset-lab" alt="npm"></a>
+  <a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-blue" alt="MIT License"></a>
+</p>
+
+# style-dataset-lab
+
+विज़ुअल डेटासेट फैक्ट्री -- वीएलएम (VLM) के फाइन-ट्यूनिंग के लिए मल्टीमॉडल प्रशिक्षण डेटा उत्पन्न करें, क्यूरेट करें और निर्यात करें।
+
+## यह क्या है
+
+यह एक टूलकिट है जिसका उपयोग "**प्रशिक्षण योग्य दृश्य सत्य**" बनाने के लिए किया जाता है। प्रत्येक एसेट में तीन चीजें होती हैं:
+
+1. **इमेज पिक्सेल** -- ComfyUI द्वारा उत्पन्न, जिसमें पूर्ण जानकारी (चेकपॉइंट, LoRA, सीड, सैंपलर, cfg) शामिल है।
+2. **मानक व्याख्या** -- यह क्यों 'ऑन-स्टाइल' या 'ऑफ-स्टाइल' है, जो एक 'स्टाइल संविधान' पर आधारित है।
+3. **गुणवत्ता मूल्यांकन** -- स्वीकृत/अस्वीकृत, जिसमें प्रत्येक आयाम के लिए स्कोर और संदर्भित नियम शामिल हैं।
+
+आउटपुट [`@mcptoolshop/repo-dataset`](https://github.com/mcp-tool-shop-org/repo-dataset) में जाता है ताकि 10 प्रारूपों में मल्टीमॉडल प्रशिक्षण डेटा उत्पन्न किया जा सके: TRL, LLaVA, Qwen2-VL, Axolotl, LLaMA-Factory, ShareGPT, RLHF जोड़े, DPO, कैप्शनिंग और वर्गीकरण।
+
+## सुरक्षा मॉडल
+
+**केवल स्थानीय।** style-dataset-lab, `localhost:8188` पर ComfyUI से संवाद करता है और कभी भी बाहरी नेटवर्क अनुरोध नहीं करता है। कोई टेलीमेट्री नहीं, कोई विश्लेषण नहीं, कोई डेटा संग्रह नहीं। इमेज जेनरेशन पूरी तरह से आपके GPU पर होता है। रिकॉर्ड और अन्य डेटा आपके फ़ाइल सिस्टम पर ही रहते हैं।
+
+## डेटासेट आँकड़े
+
+| मेट्रिक | मान |
+|--------|-------|
+| क्यूरेट किए गए रिकॉर्ड | 1,182 |
+| कुल एसेट | 2,571 (893 स्वीकृत, 887 'पेंटरली' वेरिएंट) |
+| प्रॉम्प्ट वेव्स | 28 |
+| विज़ुअल श्रेणियां | 18 (पोशाक, जहाज, आंतरिक भाग, उपकरण, वातावरण, प्रजातियां, स्टेशन, संकेत, प्रकाश, माल, आर्किटेक्चर, जीव, सतह, दैनिक जीवन, ग्रह, क्षति/मरम्मत, एलियन जीव विज्ञान, वास्तविक विवरण) |
+| जोड़े में तुलनाएं | 6 मानव + 71 सिंथेटिक |
+| विभिन्न अस्वीकृति प्रकार | 30+ |
+| पहचान पैकेज सिस्टम | नाम वाले पात्र, जिनमें गुट (faction) की जानकारी, रैंक और दृश्य संबंधी विशेषताएं शामिल हैं। |
+| निर्यात प्रारूप | 10 ( `@mcptoolshop/repo-dataset` के माध्यम से) |
+
+## इंस्टॉल करें
+
+```bash
+npm install -g @mcptoolshop/style-dataset-lab
+```
+
+फिर एक प्रोजेक्ट को क्लोन करें या एक नया डेटासेट वर्कस्पेस इनिशियलाइज़ करें:
+
+```bash
+git clone https://github.com/mcp-tool-shop-org/style-dataset-lab my-dataset
+cd my-dataset
+npm install
+```
+
+## वर्कफ़्लो
+
+```bash
+# 1. Start ComfyUI
+# (point it at your checkpoint + LoRA setup)
+
+# 2. Generate candidates from a prompt pack
+npm run generate -- inputs/prompts/wave1.json
+npm run generate -- inputs/prompts/wave1.json --dry-run
+
+# 3. Generate identity-packet characters
+npm run generate:identity -- inputs/identity-packets/wave27a-identity-spine.json
+
+# 4. Generate painterly variants of approved assets
+npm run painterly -- <asset_id>
+
+# 5. Curate -- approve, reject, or mark borderline
+npm run curate -- <asset_id> approved "explanation" --scores "silhouette:0.9,palette:0.8"
+npm run curate -- <asset_id> rejected "explanation" --failures "too_clean,wrong_material"
+
+# 6. Bind canon explanations to assets
+npm run canon-bind -- <asset_id>
+
+# 7. Record pairwise comparisons
+npm run compare -- <asset_a> <asset_b> a "A has better faction read because..."
+
+# 8. Export training data via repo-dataset
+npm run export
+npm run inspect
+npm run validate
+```
+
+## डायरेक्टरी संरचना
+
+```
+canon/                  Style constitution, review rubric, identity gates, species canon
+inputs/
+  prompts/              Prompt packs per wave (JSON: subjects, variations, defaults)
+  references/           IP-Adapter reference images
+  control/              ControlNet control images
+  control-guides/       ControlNet guide overlays
+  identity-packets/     Named character identity spines (faction DNA, rank, visual anchors)
+outputs/
+  candidates/           Raw generations (gitignored)
+  approved/             Curated approved (gitignored)
+  rejected/             Curated rejected (gitignored)
+  borderline/           Curated borderline (gitignored)
+  painterly/            Painterly-style variants (gitignored)
+records/                Per-asset JSON (provenance + judgment + canon binding)
+comparisons/            A-vs-B preference judgments
+exports/                repo-dataset output (gitignored)
+scripts/                generate, curate, compare, canon-bind, painterly, identity gen
+workflows/              Reusable ComfyUI workflow templates
+```
+
+## जेनरेशन सेटअप
+
+```yaml
+checkpoint: dreamshaperXL_v21TurboDPMSDE.safetensors
+lora: classipeintxl_v21.safetensors (weight: 1.0)
+resolution: 1024x1024
+steps: 8
+cfg: 2.0
+sampler: dpmpp_sde
+scheduler: karras
+speed: ~9s per image (RTX 5080)
+```
+
+अतिरिक्त जेनरेशन मोड: ControlNet (पोज़/गहराई-निर्देशित), IP-Adapter (संदर्भ-निर्देशित), और पहचान पैकेज (नाम वाले पात्रों की निरंतरता)।
+
+## आवश्यकताएं
+
+- `localhost:8188` पर चलने वाला [ComfyUI](https://github.com/comfyanonymous/ComfyUI)
+- DreamShaper XL Turbo चेकपॉइंट + ClassipeintXL LoRA
+- Node.js 20+
+- प्रशिक्षण के लिए निर्यात के लिए [`@mcptoolshop/repo-dataset`](https://github.com/mcp-tool-shop-org/repo-dataset)
+
+## लाइसेंस
+
+MIT
+
+---
+
+<a href="https://mcp-tool-shop.github.io/">MCP Tool Shop</a> द्वारा निर्मित।

package/README.it.md

@@ -0,0 +1,142 @@
+<p align="center">
+  <a href="README.ja.md">日本語</a> | <a href="README.zh.md">中文</a> | <a href="README.es.md">Español</a> | <a href="README.fr.md">Français</a> | <a href="README.hi.md">हिन्दी</a> | <a href="README.md">English</a> | <a href="README.pt-BR.md">Português (BR)</a>
+</p>
+
+<p align="center">
+  <img src="https://raw.githubusercontent.com/mcp-tool-shop-org/brand/main/logos/style-dataset-lab/readme.png" width="400">
+</p>
+
+<p align="center">
+  <a href="https://www.npmjs.com/package/@mcptoolshop/style-dataset-lab"><img src="https://img.shields.io/npm/v/@mcptoolshop/style-dataset-lab" alt="npm"></a>
+  <a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-blue" alt="MIT License"></a>
+</p>
+
+# style-dataset-lab
+
+Strumento per la creazione di dataset visivi: genera, organizza ed esporta dati di training multimodali per l'affinamento di modelli di visione e linguaggio (VLM).
+
+## Cos'è questo
+
+Un toolkit per creare una "verità visiva" addestrabile. Ogni elemento contiene tre informazioni:
+
+1. **Pixel dell'immagine** -- generati da ComfyUI con informazioni complete sulla provenienza (checkpoint, LoRA, seed, sampler, cfg).
+2. **Spiegazione canonica** -- perché questo elemento è conforme o non conforme allo stile, basata su una "costituzione" dello stile.
+3. **Valutazione della qualità** -- approvato/rifiutato con punteggi per ogni dimensione e regole citate.
+
+L'output viene utilizzato da [`@mcptoolshop/repo-dataset`](https://github.com/mcp-tool-shop-org/repo-dataset) per creare dati di training multimodali in 10 formati: TRL, LLaVA, Qwen2-VL, Axolotl, LLaMA-Factory, ShareGPT, coppie RLHF, DPO, didascalie e classificazione.
+
+## Modello di sicurezza
+
+**Solo locale.** style-dataset-lab comunica con ComfyUI su `localhost:8188` e non effettua richieste alla rete esterna. Nessuna telemetria, nessuna analisi, nessuna trasmissione di dati. La generazione delle immagini avviene interamente sulla tua GPU. I dati e le informazioni canoniche rimangono sul tuo filesystem.
+
+## Statistiche del dataset
+
+| Metrica | Valore |
+|--------|-------|
+| Elementi curati | 1,182 |
+| Elementi totali | 2.571 (893 approvati, 887 varianti pittoriche) |
+| Prompt utilizzati | 28 |
+| Categorie visive | 18 (costumi, navi, interni, attrezzature, ambienti, specie, stazioni, segnaletica, illuminazione, carico, architettura, creature, superfici, vita quotidiana, pianeti, danni/riparazioni, biologia aliena, dettagli di ambienti vissuti) |
+| Confronti a coppie | 6 creati da umani + 71 creati sinteticamente |
+| Tipi di rifiuto distinti | 30+ |
+| Sistema di "identity packet" | Personaggi nominati con DNA della fazione, grado e riferimenti visivi. |
+| Formati di esportazione | 10 (tramite @mcptoolshop/repo-dataset) |
+
+## Installazione
+
+```bash
+npm install -g @mcptoolshop/style-dataset-lab
+```
+
+Successivamente, clona un progetto o inizializza un nuovo spazio di lavoro per il dataset:
+
+```bash
+git clone https://github.com/mcp-tool-shop-org/style-dataset-lab my-dataset
+cd my-dataset
+npm install
+```
+
+## Flusso di lavoro
+
+```bash
+# 1. Start ComfyUI
+# (point it at your checkpoint + LoRA setup)
+
+# 2. Generate candidates from a prompt pack
+npm run generate -- inputs/prompts/wave1.json
+npm run generate -- inputs/prompts/wave1.json --dry-run
+
+# 3. Generate identity-packet characters
+npm run generate:identity -- inputs/identity-packets/wave27a-identity-spine.json
+
+# 4. Generate painterly variants of approved assets
+npm run painterly -- <asset_id>
+
+# 5. Curate -- approve, reject, or mark borderline
+npm run curate -- <asset_id> approved "explanation" --scores "silhouette:0.9,palette:0.8"
+npm run curate -- <asset_id> rejected "explanation" --failures "too_clean,wrong_material"
+
+# 6. Bind canon explanations to assets
+npm run canon-bind -- <asset_id>
+
+# 7. Record pairwise comparisons
+npm run compare -- <asset_a> <asset_b> a "A has better faction read because..."
+
+# 8. Export training data via repo-dataset
+npm run export
+npm run inspect
+npm run validate
+```
+
+## Struttura delle directory
+
+```
+canon/                  Style constitution, review rubric, identity gates, species canon
+inputs/
+  prompts/              Prompt packs per wave (JSON: subjects, variations, defaults)
+  references/           IP-Adapter reference images
+  control/              ControlNet control images
+  control-guides/       ControlNet guide overlays
+  identity-packets/     Named character identity spines (faction DNA, rank, visual anchors)
+outputs/
+  candidates/           Raw generations (gitignored)
+  approved/             Curated approved (gitignored)
+  rejected/             Curated rejected (gitignored)
+  borderline/           Curated borderline (gitignored)
+  painterly/            Painterly-style variants (gitignored)
+records/                Per-asset JSON (provenance + judgment + canon binding)
+comparisons/            A-vs-B preference judgments
+exports/                repo-dataset output (gitignored)
+scripts/                generate, curate, compare, canon-bind, painterly, identity gen
+workflows/              Reusable ComfyUI workflow templates
+```
+
+## Configurazione della generazione
+
+```yaml
+checkpoint: dreamshaperXL_v21TurboDPMSDE.safetensors
+lora: classipeintxl_v21.safetensors (weight: 1.0)
+resolution: 1024x1024
+steps: 8
+cfg: 2.0
+sampler: dpmpp_sde
+scheduler: karras
+speed: ~9s per image (RTX 5080)
+```
+
+Modalità di generazione aggiuntive: ControlNet (guidato da posa/profondità), IP-Adapter (guidato da riferimento) e "identity packets" (coerenza dei personaggi nominati).
+
+## Requisiti
+
+- [ComfyUI](https://github.com/comfyanonymous/ComfyUI) in esecuzione su localhost:8188
+- Checkpoint DreamShaper XL Turbo + LoRA ClassipeintXL
+- Node.js 20+
+- [`@mcptoolshop/repo-dataset`](https://github.com/mcp-tool-shop-org/repo-dataset) per l'esportazione dei dati di training
+
+## Licenza
+
+MIT
+
+---
+
+Creato da <a href="https://mcp-tool-shop.github.io/">MCP Tool Shop</a