data/lora/README.md

# LoRA Configuration - Safetensors Only

## Directory Structure
```
data/lora/
├── memo-scene-lora.safetensors    # Main LoRA weights
├── readme.md                      # This file
└── versions/                      # Versioned LoRA files
    ├── v1.0/
    └── v1.1/
```

## LoRA File Requirements

### Security Requirements
- **ONLY .safetensors files** - No .bin, .ckpt, or other formats allowed
- **Model signatures required** - All LoRA files must have proper signatures
- **Version tracking** - Each version must be clearly identified

### Technical Requirements
- **Format**: PyTorch safetensors
- **Precision**: FP16 recommended for memory efficiency
- **Compression**: Quantized versions for faster loading
- **Metadata**: Include training information and compatibility notes

## Loading LoRA Weights

### Basic Loading
```python
from models.image.sd_generator import get_generator

generator = get_generator(lora_path="data/lora")
```

### Version-Specific Loading
```python
generator = get_generator(lora_path="data/lora/versions/v1.1")
```

### Multiple LoRA Support
```python
# Load multiple LoRA files
lora_paths = [
    "data/lora/memo-scene-lora.safetensors",
    "data/lora/style-lora.safetensors"
]

for lora_path in lora_paths:
    generator.pipe.load_lora_weights(
        os.path.dirname(lora_path),
        weight_name=os.path.basename(lora_path)
    )
```

## LoRA Training Configuration

### Recommended Settings
- **Base Model**: stabilityai/stable-diffusion-xl-base-1.0
- **LoRA Rank**: 16-64 (higher rank = more capacity)
- **Alpha**: 32-128 (typically 2x the rank)
- **Dropout**: 0.1-0.2 for regularization
- **Precision**: FP16 for training, FP16 inference

### Training Script Usage
```bash
python scripts/train_scene_lora.py \
    --base_model "stabilityai/stable-diffusion-xl-base-1.0" \
    --output_dir "data/lora/versions/v1.2" \
    --rank 32 \
    --alpha 64 \
    --epochs 5
```

## Model Tier Configuration

### Free Tier
- Base model only (no LoRA)
- Lower inference steps (15-20)
- Standard resolution (512x512)

### Pro Tier  
- Base + scene LoRA
- Higher inference steps (25-30)
- Higher resolution (768x768 or 1024x1024)
- LCM acceleration

### Enterprise Tier
- Base + multiple LoRAs
- Highest quality settings
- Custom resolution
- Priority processing

## Security Notes

1. **Never load .bin files** - Use only safetensors
2. **Verify signatures** - Check LoRA file integrity
3. **Isolate environments** - Separate model loading contexts
4. **Audit logs** - Track all LoRA loading operations
5. **Version pinning** - Lock specific LoRA versions for production

## Performance Notes

1. **Memory optimization** - Use quantized LoRA when possible
2. **Preloading** - Load frequently used LoRA files at startup
3. **Caching** - Cache LoRA states for faster switching
4. **Cold start** - Minimize initial LoRA loading time
5. **Dynamic loading** - Load LoRA on-demand for different scenes