ramanv-image-foundation

A production-grade, highly-scalable dataset engineering platform designed for training and fine-tuning text-to-image models (e.g., FLUX, SDXL, and image editing models).

The repository is built to process datasets scaling up to 10M+ images, supporting multi-GPU captioning, OCR extraction, CLIP alignment scoring, aesthetic analysis, deduplication checks, quality filters, and automated uploading to the Hugging Face Hub.

---

📂 Folder Layout

ramanv-image-foundation/
├── README.md                 # Dataset Card & Guide
├── LICENSE                   # Apache-2.0 License
├── pyproject.toml            # Project configurations and packages
├── requirements.txt          # Python dependency list
├── dataset_infos.json        # HF Dataset features schema
│
├── configs/                  # Pipeline configurations (YAML)
│   ├── captions.yaml
│   ├── ocr.yaml
│   ├── clip.yaml
│   ├── aesthetic.yaml
│   ├── deduplicate.yaml
│   ├── quality.yaml
│   ├── splits.yaml
│   └── upload.yaml
│
├── docs/                     # Detailed guides
│   ├── architecture.md
│   ├── metadata.md
│   ├── pipelines.md
│   ├── quality.md
│   └── hf_upload.md
│
├── images/                   # Image folders grouped into 86 categories
│   ├── marketing/
│   ├── social_media/
│   ├── advertisements/
│   └── ... (83 other folders)
│
├── captions/                 # Raw descriptive text files
├── metadata/                 # Schema-compliant JSON files (49 fields)
├── manifests/                # Dataset manifest files (jsonl)
│   ├── train.jsonl
│   ├── validation.jsonl
│   ├── test.jsonl
│   └── all.jsonl
│
├── quality/                  # Quality metrics and reports
├── splits/                   # Split metrics and records
├── embeddings/               # Precomputed CLIP/SigLIP feature matrices (.npy)
├── thumbnails/               # Generated preview thumbnails
├── logs/                     # Rotation log output files
└── src/                      # Core utility codebase

---

📊 Image Categories

The dataset organizes assets into 86 distinct category folders under images/, including:

• Design & Layout: marketing, branding, poster, flyer, banner, business_cards, advertisements, logos, icons, illustrations, 3d, packaging, mockups, social_media.
• Commerce: products, electronics, fashion, cosmetics, food, restaurants.
• Transport: vehicles, cars, motorcycles, buses, trains, aircraft, boats.
• Spaces: architecture, real_estate, interiors, homes, office, workspace, travel, hotel.
• Nature: landscape, forest, mountains, beaches, rivers, space, science, technology, medical, education, sports.
• Culture & People: festival, temple, wedding, indian, villages, cities, street, humans, portraits.
• Documents & Text: documents, receipts, forms, books, menus, newspapers, typography, signboards, certificates, whiteboards, handwriting.

---

🛠️ Getting Started

1. Clone & Install Dependencies:

   pip install -r requirements.txt
   

2. Initialize Folders: On Linux:

   bash scripts/create_repo.sh
   

   On Windows (PowerShell):

   # Run the PowerShell folder setup commands
   

3. Place Source Images: Add your raw .jpg/.png images to their respective categories under the images/ directory.

4. Run Pipelines:

   # 1. Generate captions using VLMs
   python scripts/generate_captions.py
   
   # 2. Extract OCR texts
   python scripts/generate_ocr.py
   
   # 3. Calculate CLIP alignments and extract embeddings
   python scripts/compute_clip_scores.py
   
   # 4. Compute LAION aesthetic ratings
   python scripts/compute_aesthetic.py
   
   # 5. Filter out low-quality/blurry images and deduplicate
   python scripts/deduplicate.py
   python scripts/quality_filter.py
   
   # 6. Partition train/validation/test splits
   python scripts/create_splits.py
   
   # 7. Validate schemas
   python scripts/validate_dataset.py
   

---

🚀 Model Training Guide

1. Fine-tuning FLUX Models

• Long Captions: FLUX models leverage T5-XXL text encoders, which excel at processing highly-detailed descriptions. Use the long_caption metadata field to provide full scene composition, color palettes, styles, lighting, and textures.
• Aspect Ratio Bucketing: Group image training items by their computed aspect_ratio to avoid distortion during crop training.
• Aesthetic Filtering: Exclude training images with aesthetic_score < 6.0 to preserve the visual fidelity of output weights.

2. Training SDXL Models

• CLIP Scores: Filter out records with clip_score < 0.22 to ensure proper visual concept alignment.
• Tags: Use the computed tags array joined as comma-separated values for the conditioning prompt.

3. Training Editing / ControlNet Models

• Use the OCR labels (contains_text: true) to isolate text-heavy graphics for training layout-aware control weights (e.g., ControlNet-OCR).
• Filter using contains_people or people_count to selectively build portrait or posture-conditioned model adapters.

---

📄 License

This repository is licensed under the Apache License 2.0. See LICENSE for more details.