docs/PHASE1E_LORA_GUIDE.md

CompI Phase 1.E: Personal Style Fine-tuning with LoRA

🎯 Overview

Phase 1.E enables you to train personalized artistic styles using LoRA (Low-Rank Adaptation) fine-tuning on Stable Diffusion. This allows you to create AI art that reflects your unique artistic vision or mimics specific artistic styles.

LoRA Benefits:

• ✅ Lightweight: Only trains a small adapter (~10-100MB vs full model ~4GB)
• ✅ Fast: Training takes minutes to hours instead of days
• ✅ Flexible: Can be combined with different base models
• ✅ Efficient: Runs on consumer GPUs (8GB+ VRAM recommended)

🛠️ Tools Provided

1. Dataset Preparation (compi_phase1e_dataset_prep.py)

• Organize and validate your style images
• Generate appropriate training captions
• Resize and format images for optimal training
• Create train/validation splits

2. LoRA Training (compi_phase1e_lora_training.py)

• Full LoRA fine-tuning pipeline with PEFT integration
• Configurable training parameters and monitoring
• Automatic checkpoint saving and validation
• Memory-efficient training with gradient checkpointing

3. Style Generation (compi_phase1e_style_generation.py)

• Generate images using your trained LoRA styles
• Interactive and batch generation modes
• Adjustable style strength and parameters
• Integration with existing CompI pipeline

4. Style Management (compi_phase1e_style_manager.py)

• Manage multiple trained LoRA styles
• Cleanup old checkpoints and organize models
• Export style information and analytics
• Switch between different personal styles

🚀 Quick Start Guide

Step 1: Install Dependencies

# Install LoRA training dependencies
pip install peft datasets bitsandbytes

# Verify installation
python -c "import peft, datasets; print('✅ Dependencies installed')"

Step 2: Prepare Your Style Dataset

# Organize your style images in a folder
mkdir my_artwork
# Copy 10-50 images of your artistic style to my_artwork/

# Prepare dataset for training
python src/generators/compi_phase1e_dataset_prep.py \
    --input-dir my_artwork \
    --style-name "my_art_style" \
    --trigger-word "myart"

Dataset Requirements:

• 10-50 images (more is better, but 20+ is usually sufficient)
• Consistent style across all images
• 512x512 pixels recommended (will be auto-resized)
• High quality images without watermarks or text

Step 3: Train Your LoRA Style

# Start LoRA training
python run_lora_training.py \
    --dataset-dir datasets/my_art_style \
    --epochs 100 \
    --learning-rate 1e-4

# Or with custom settings
python run_lora_training.py \
    --dataset-dir datasets/my_art_style \
    --epochs 200 \
    --batch-size 2 \
    --lora-rank 8 \
    --lora-alpha 32

Training Tips:

• Start with 100 epochs for initial testing
• Increase to 200-500 epochs for stronger style learning
• Monitor validation loss to avoid overfitting
• Use gradient checkpointing if you run out of memory

Step 4: Generate with Your Style

# Generate images with your trained style
python run_style_generation.py \
    --lora-path lora_models/my_art_style/checkpoint-1000 \
    "a cat in myart style" \
    --variations 4

# Interactive mode
python run_style_generation.py \
    --lora-path lora_models/my_art_style/checkpoint-1000 \
    --interactive

📊 Advanced Usage

Training Configuration

# High-quality training (slower but better results)
python run_lora_training.py \
    --dataset-dir datasets/my_style \
    --epochs 300 \
    --learning-rate 5e-5 \
    --lora-rank 16 \
    --lora-alpha 32 \
    --batch-size 1 \
    --gradient-checkpointing

# Fast training (quicker results for testing)
python run_lora_training.py \
    --dataset-dir datasets/my_style \
    --epochs 50 \
    --learning-rate 2e-4 \
    --lora-rank 4 \
    --lora-alpha 16

Style Management

# List all trained styles
python src/generators/compi_phase1e_style_manager.py --list

# Get detailed info about a style
python src/generators/compi_phase1e_style_manager.py --info my_art_style

# Clean up old checkpoints (keep only 3 most recent)
python src/generators/compi_phase1e_style_manager.py --cleanup my_art_style --keep 3

# Export styles information to CSV
python src/generators/compi_phase1e_style_manager.py --export my_styles_report.csv

Generation Parameters

# Adjust style strength
python run_style_generation.py \
    --lora-path lora_models/my_style/checkpoint-1000 \
    --lora-scale 0.8 \
    "portrait in myart style"

# High-quality generation
python run_style_generation.py \
    --lora-path lora_models/my_style/checkpoint-1000 \
    --steps 50 \
    --guidance 8.0 \
    --width 768 \
    --height 768 \
    "landscape in myart style"

🎨 Best Practices

Dataset Preparation

1. Consistent Style: All images should represent the same artistic style
2. Quality over Quantity: 20 high-quality images > 100 low-quality ones
3. Diverse Subjects: Include various subjects (people, objects, landscapes)
4. Clean Images: Remove watermarks, text, and irrelevant elements
5. Proper Captions: Use consistent trigger words in captions

Training Tips

1. Start Small: Begin with 50-100 epochs to test
2. Monitor Progress: Check validation loss and sample generations
3. Adjust Learning Rate: Lower if loss oscillates, higher if learning is slow
4. Use Checkpoints: Save frequently to avoid losing progress
5. Experiment with LoRA Rank: Higher rank = more capacity but slower training

Generation Guidelines

1. Include Trigger Words: Always use your trigger word in prompts
2. Adjust Style Strength: Use --lora-scale to control style intensity
3. Combine with Techniques: Mix with existing CompI style/mood systems
4. Iterate and Refine: Generate multiple variations and select best results

🔧 Troubleshooting

Common Issues

Out of Memory Error:

# Reduce batch size and enable gradient checkpointing
python run_lora_training.py \
    --dataset-dir datasets/my_style \
    --batch-size 1 \
    --gradient-checkpointing \
    --mixed-precision

Style Not Learning:

• Increase epochs (try 200-500)
• Check dataset consistency
• Increase LoRA rank (try 8 or 16)
• Lower learning rate (try 5e-5)

Generated Images Don't Match Style:

• Include trigger word in prompts
• Increase LoRA scale (try 1.2-1.5)
• Train for more epochs
• Check dataset quality

Training Too Slow:

• Reduce image resolution to 512x512
• Use mixed precision training
• Enable gradient checkpointing
• Reduce LoRA rank to 4

📁 File Structure

Project CompI/
├── datasets/                    # Prepared training datasets
│   └── my_art_style/
│       ├── train/              # Training images
│       ├── validation/         # Validation images
│       ├── train_captions.txt  # Training captions
│       └── dataset_info.json   # Dataset metadata
├── lora_models/                # Trained LoRA models
│   └── my_art_style/
│       ├── checkpoint-100/     # Training checkpoints
│       ├── checkpoint-200/
│       └── training_info.json  # Training metadata
├── src/generators/
│   ├── compi_phase1e_dataset_prep.py     # Dataset preparation
│   ├── compi_phase1e_lora_training.py    # LoRA training
│   ├── compi_phase1e_style_generation.py # Style generation
│   └── compi_phase1e_style_manager.py    # Style management
├── run_lora_training.py        # Training launcher
└── run_style_generation.py     # Generation launcher

🎯 Integration with CompI

Phase 1.E integrates seamlessly with existing CompI tools:

1. Combine with Phase 1.B: Use LoRA styles alongside predefined styles
2. Evaluate with Phase 1.D: Assess your LoRA-generated images systematically
3. UI Integration: Add LoRA styles to Streamlit/Gradio interfaces
4. Batch Processing: Generate multiple variations for evaluation

🚀 Next Steps

After mastering Phase 1.E:

1. Experiment with Multiple Styles: Train different LoRA adapters for various artistic approaches
2. Style Mixing: Combine multiple LoRA styles for unique effects
3. Advanced Techniques: Explore Textual Inversion, DreamBooth, or ControlNet integration
4. Community Sharing: Share your trained styles with the CompI community
5. Phase 2 Preparation: Use personal styles as foundation for multimodal integration

---

Happy Style Training! 🎨✨

Phase 1.E opens up endless possibilities for personalized AI art generation. With LoRA fine-tuning, you can teach the AI to understand and replicate your unique artistic vision, creating truly personalized creative content.