# Aerial Landscape - Flux Dev LoRA

A LoRA model trained on aerial landscape and overhead shot photography for Flux Dev, specializing in top-down perspectives of natural landscapes, urban environments, and scenic views.

## 📋 Model Details

- **Base Model**: Flux Dev
- **Training Type**: LoRA (Low-Rank Adaptation)
- **Rank**: 64
- **Training Steps**: 7,500
- **Training Resolution**: 1024×1024
- **Dataset Size**: 531 images
- **Hardware**: NVIDIA A40 GPU (48GB VRAM)

## 🎨 Model Capabilities

This LoRA specializes in generating:

- **Aerial landscapes**: Ocean views, beaches, forests, mountains, and natural terrain from above
- **Urban aerial photography**: Cities, buildings, roads, and infrastructure from bird's-eye view
- **Overhead perspectives**: Top-down shots with authentic aerial photography composition
- **Natural scenery**: Water bodies, waves, coastlines, rock formations, vegetation
- **Architectural views**: Buildings and urban structures from elevated angles

### Common Themes
- 80% no_humans scenes (pristine landscape focus)
- Natural elements: water (46%), ocean/beach (29%), trees/nature (23%)
- Urban elements: buildings (17%), roads, vehicles
- Artistic styles: traditional media aesthetics, painting-like qualities

## 🚀 Usage

### Basic Prompt Structure

```
[subject], aerial view, overhead shot, from above, [environment details], [style modifiers]
```

### Example Prompts

**Natural Landscapes:**
```
ocean waves, aerial view, from above, no humans, water, scenery, realistic photography
```

**Urban Scenes:**
```
city street, aerial view, from above, buildings, roads, motor vehicles, urban landscape, no humans
```

**Coastal Views:**
```
beach coastline, overhead shot, from above, ocean, waves, sand, rocks, no humans, natural scenery
```

**Forest/Nature:**
```
dense forest, aerial photography, from above, trees, nature, greenery, no humans, scenic landscape
```

### Recommended Settings

- **LoRA Weight**: 0.6 - 1.0 (adjust based on desired strength)
- **CFG Scale**: 3.5 - 7.0
- **Steps**: 20-30 (Flux Dev standard)
- **Sampler**: Euler, DPM++ 2M, or other Flux-compatible samplers
- **Resolution**: 1024×1024 or higher (model trained at 1024px)

### Key Trigger Words

| Category        | Keywords                                                             |
| --------------- | -------------------------------------------------------------------- |
| **Perspective** | `from_above`, `aerial view`, `overhead shot`, `bird's eye view`      |
| **Environment** | `scenery`, `outdoors`, `nature`, `urban`, `landscape`                |
| **Natural**     | `ocean`, `water`, `waves`, `beach`, `tree`, `forest`, `rock`         |
| **Urban**       | `building`, `city`, `road`, `street`, `architecture`                 |
| **Composition** | `no_humans`, `vehicle_focus`, `watercraft`                           |
| **Style**       | `realistic`, `photography`, `traditional_media`, `painting_(medium)` |

## 💡 Tips for Best Results

1. **Use "from_above" or "aerial view"** to activate the overhead perspective style
2. **Add "no_humans"** for pure landscape shots (primary training focus)
3. **Combine natural + urban elements** for interesting mixed scenes
4. **Adjust LoRA strength**:
   - 0.6-0.8 for subtle aerial influence
   - 0.8-1.0 for strong aerial photography style
5. **Resolution**: Works best at 1024×1024 or higher aspect ratios
6. **Negative prompts**: `ground level, eye level, portrait, close-up` to avoid non-aerial perspectives

## 📊 Training Dataset Statistics

- **Total images**: 531 aerial/overhead photographs
- **Resolution**: 1024×1024 (square format)
- **Content distribution**:
  - Landscapes/nature: ~70%
  - Urban/architecture: ~20%
  - Mixed/other: ~10%
- **Caption format**: Booru-style tags with detailed scene descriptions

### Most Common Tags
```
no_humans (428), traditional_media (369), scenery (312), outdoors (286),
water (244), painting_(medium) (178), ocean (102), waves (94),
from_above (92), building (92), sky (87), tree (120), beach (50)
```

## 🖼️ Sample Images

Sample training images are available in the `1024/` directory, showcasing the variety of aerial perspectives, natural landscapes, and urban scenes used to train this model.

## 📝 Technical Specifications

- **Training Framework**: Likely Kohya/SimpleTuner/AI-Toolkit
- **Optimizer**: AdamW or similar
- **Precision**: Mixed precision (FP16/BF16)
- **Batch Size**: Optimized for 48GB VRAM
- **Learning Rate**: Default LoRA learning rate schedule
- **Rank**: 64 (balanced between quality and file size)

## 🔧 Integration

### ComfyUI
1. Place the `.safetensors` file in `ComfyUI/models/loras/`
2. Add LoRA Loader node
3. Connect to your Flux Dev workflow
4. Set weight between 0.6-1.0

### Automatic1111/Forge (with Flux support)
1. Place in `models/Lora/` directory
2. Use `<lora:aerial-landscape:0.8>` in prompts
3. Adjust weight as needed

### Python (diffusers)
```python
from diffusers import FluxPipeline

pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev")
pipe.load_lora_weights("path/to/aerial-landscape.safetensors")
```

## 📄 License

Please respect the licensing terms of the base Flux Dev model and any applicable dataset licenses.

## 🙏 Acknowledgments

- Base Model: [Flux Dev by Black Forest Labs](https://blackforestlabs.ai/)
- Training Hardware: NVIDIA A40 (48GB)
- Dataset: 531 curated aerial landscape photographs

## 📧 Contact & Updates

For questions, improvements, or dataset inquiries, please refer to the model repository or contact the creator.

---

**Version**: 1.0
**Release Date**: 2025
**Training Steps**: 7,500
**Model Type**: Flux Dev LoRA (Rank 64)