FluxGym Enhanced

Enhanced FluxGym with pretrained adapter support, bulk captioning, and optimized large dataset handling.

🚀 New Features

✨ Pretrained Adapter Support

• Load existing LoRA adapters for continued training
• Continue training from checkpoints
• Fine-tune existing models with new concepts

🤖 Bulk AI Captioning

• Generate captions for 300+ images using Google Gemini
• Automatic caption creation with trigger words
• Structured caption format for optimal training

📁 Large Dataset Optimization

• Handle 300+ images efficiently
• Direct dataset folder support (bypass slow uploads)
• Optimized training settings for large datasets

🛠️ Enhanced Tools

• create_captions.py - Bulk caption creation script
• Auto-detection of existing images in dataset folders
• Improved UI with helpful guidance

📋 Quick Start for Large Datasets

1. Setup Dataset Folder

# Copy your images directly to:
datasets/your-lora-name/

2. Auto-Caption with AI

• Open FluxGym web interface
• Enter LoRA name (matching folder name)
• Leave upload empty
• Click "Add AI captions with Gemini"

3. Optimized Training Settings for 300+ Images

• Repeats: 1-2 (instead of 10)
• Epochs: 8-12 (instead of 16)
• Learning Rate: 4e-4 or 2e-4 (lower)
• LoRA Rank: 8-16 (higher)

🔧 Installation

git clone https://github.com/mgalore/fluxgym-enhanced.git
cd fluxgym-enhanced
pip install -r requirements.txt
python app.py

📚 Enhanced Features

Pretrained Adapter Training

# Modified gen_sh function now supports:
--network_weights /path/to/existing/adapter.safetensors

Bulk Caption Creation

# Create captions for all images in dataset folder
python create_captions.py datasets/my-lora my_trigger_word

Large Dataset Workflow

1. Copy 300+ images to datasets/lora-name/
2. Use AI captioning (no upload needed)
3. Train with optimized settings
4. Monitor progress with frequent checkpoints

⚙️ Configuration

Enhanced config.json with:

• Configurable max image limits
• Path customization
• API key management

🎯 Best Practices

For 300+ Images:

• Use 2 repeats maximum
• Lower learning rate (4e-4)
• Higher LoRA rank (12-16)
• Save checkpoints every 2-4 epochs

Caption Quality:

• Use descriptive trigger words
• Include style, pose, lighting details
• Consistent caption structure

🐛 Troubleshooting

Slow Uploads

• Use direct dataset folder method
• Copy images via file explorer
• Skip web interface uploads

Memory Issues

• Reduce batch size
• Use 12G VRAM settings
• Enable gradient checkpointing

📖 Documentation

• Original FluxGym
• LoRA Training Guide
• Large Dataset Tips

🤝 Contributing

Feel free to submit issues and enhancement requests!

📄 License

Same as original FluxGym project.

---

Enhanced by AI Assistant with focus on production-ready large dataset training.