# Russian Blue Cat - Flux Dev LoRA v1.0

**Initial Release** - Photorealistic Russian Blue cat generation model

## Version Information

- **Version**: 1.0
- **Release Date**: 2025-10-16
- **Status**: Stable
- **Model Type**: Flux Dev LoRA

## What's New in v1.0

🎉 **Initial Release**
- First public release of the Russian Blue cat LoRA model
- 1000 training images across multiple resolutions
- Optimized for photorealistic output
- No trigger word required - natural language prompting

## Model Specifications

### Architecture
- **Base Model**: Flux Dev
- **Model Type**: LoRA
- **Transformer**: float8
- **Text Encoder**: float8
- **Linear Rank**: 64
- **Data Type**: bf16

### Training Configuration
- **Training Steps**: 7000
- **Gradient Accumulation**: 1
- **Batch Size**: 1
- **Timestep Type**: Sigmoid
- **Timestep Bias**: Balanced
- **Optimizer**: AdamW8bit

### Dataset
- **Total Images**: 1000
- **Resolutions**: 1024×1024, 768×768, 512×512
- **Caption Format**: Individual `.txt` files per image

## Usage Guide

### Basic Usage

**No trigger word required** - Simply describe what you want:

```
russian blue cat, photorealistic photography
Russian blue cat portrait, professional lighting, detailed fur
Russian blue cat sitting on windowsill, natural light, shallow depth of field
```

### Recommended Tags
```
russian blue, cat, photorealistic, photography, detailed fur, green eyes
```

### Optimal Settings for v1.0

| Parameter | Recommended Value | Range |
| ------------------ | ----------------- | ------------------------- |
| **LoRA Weight** | 0.8-0.9 | 0.7-1.0 |
| **CFG Scale** | 5.0 | 3.5-7.0 |
| **Sampling Steps** | 25 | 20-30 |
| **Sampler** | Euler a | Euler a / DPM++ 2M Karras |

### Example Prompts

**Portrait Style**:
```
russian blue cat portrait, photorealistic, professional photography,
detailed blue-grey fur, bright green eyes, studio lighting, sharp focus
```

**Environmental Shots**:
```
russian blue cat sitting on a velvet cushion, natural window light,
photorealistic, elegant pose, detailed whiskers
```

**Close-up Detail**:
```
russian blue cat face close-up, photorealistic photography,
detailed fur texture, green eyes, soft lighting, shallow depth of field
```

## Version-Specific Notes

### Strengths
- ✅ Excellent photorealistic quality
- ✅ Accurate breed characteristics (blue-grey coat, green eyes)
- ✅ Good detail in fur texture
- ✅ Works well with natural lighting descriptions
- ✅ Stable across different resolutions

### Known Characteristics
- 📌 Best results at LoRA weight 0.8-0.9
- 📌 Responds well to photography-style prompts
- 📌 Optimized for realistic lighting scenarios
- 📌 Works best with descriptive prompts about pose and environment

### Tips for Best Results
1. **Lighting**: Include lighting descriptions (natural light, studio lighting, etc.)
2. **Detail Level**: Mention specific details (fur texture, eye color, whiskers)
3. **Photography Terms**: Use photography vocabulary (shallow depth of field, bokeh, sharp focus)
4. **Environment**: Describe the setting or background
5. **Weight**: Start at 0.8 and adjust based on your needs

## Technical Details

This v1.0 release uses float8 precision for both transformer and text encoder to optimize memory usage while maintaining high-quality output. The model was trained over 7000 steps using the AdamW8bit optimizer with a balanced sigmoid timestep bias for efficient convergence and stable results.

The diverse dataset spanning three resolution tiers (1024px, 768px, 512px) ensures the model performs well across different output sizes.

## Compatibility

- **Base Model**: Flux Dev
- **Recommended**: Use with Flux Dev base model for best results
- **ComfyUI**: Compatible with standard LoRA loader nodes
- **A1111/Forge**: Compatible with built-in LoRA support

## File Information

- **File Size**: ~TBD MB
- **Format**: SafeTensors
- **Precision**: bf16

## Future Versions

This is the initial release. Future versions may include:
- Fine-tuning based on community feedback
- Additional training on specific poses or scenarios
- Optimization for different art styles
- Extended dataset for improved variety

## Support & Feedback

Found an issue or have suggestions for v2.0? Please provide feedback to help improve future releases.

## License

Please specify your license here.

## Credits

**Model Training**: wangkanai
**Release Date**: 2025-10-16
**Version**: 1.0

---

For general model information and specifications, see [readme-model.md](readme-model.md)