---
language:
- en
- ja
license: apache-2.0
tags:
- vision
- vision-language
- multimodal
- aerial
- drone
- vqa
- image-captioning
- defect-detection
pipeline_tag: image-text-to-text
base_model: Qwen-VL-7B
---

# AIRPHA-VLM-7B

<div align="center">
  <img src="airpha-vlm-7b.png" alt="AIRPHA-VLM-7B" width="600"/>
</div>

## Model Overview

AIRPHA-VLM-7B is a 7-billion parameter vision-language model specialized for **aerial scene understanding** from drone and UAV perspectives. The model excels at:

- **Aerial Visual Question Answering (VQA)**: Understanding and answering questions about aerial scenes
- **Aerial Image Captioning**: Generating detailed descriptions of drone/aerial imagery
- **Infrastructure Defect Detection**: Identifying and describing structural defects from aerial inspections

## Model Details

- **Developed by:** AquaAge Inc.
- **Model type:** Vision-Language Model (VLM)
- **Base Model:** Qwen-VL-7B
- **Language(s):** English, Japanese
- **Parameters:** 7B
- **License:** Apache 2.0
- **Version:** v0.1 (Beta)

## Capabilities

### 1. Aerial Visual Question Answering
```python
Q: "What infrastructure is visible in this aerial image?"
A: "The image shows a highway overpass with multiple lanes..."
```

### 2. Aerial Scene Captioning
```python
Input: [Aerial drone image]
Output: "An aerial view of an urban intersection with surrounding buildings..."
```

### 3. Defect Detection
```python
Input: [Infrastructure inspection image]
Output: "Visible cracks on the bridge surface, approximately 2 meters long..."
```

## Quick Start
```python
from transformers import AutoModel, AutoTokenizer
from PIL import Image

# Load model
model = AutoModel.from_pretrained("AquaAge/airpha-VLM-7B", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("AquaAge/airpha-VLM-7B", trust_remote_code=True)

# Load image
image = Image.open("aerial_image.jpg")

# Inference
prompt = "Describe what you see in this aerial image."
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, images=image)
response = tokenizer.decode(outputs[0])
print(response)
```

## Performance

| Task | Metric | Score |
|------|--------|-------|
| Aerial VQA | Accuracy | TBD |
| Captioning | CIDEr | TBD |
| Defect Detection | F1 | TBD |

## Limitations

- Beta version (v0.1) - performance may vary
- Optimized for aerial/drone perspectives
- May require fine-tuning for specific use cases

## Use Cases

- 🚁 Drone-based infrastructure inspection
- 🏗️ Construction site monitoring
- 🌆 Urban planning and analysis
- ⚠️ Disaster assessment
- 🛣️ Transportation infrastructure evaluation

## Citation
```bibtex
@misc{airpha-vlm-7b-2026,
  author = {AquaAge Inc.},
  title = {AIRPHA-VLM-7B: Vision-Language Model for Aerial Scene Understanding},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/AquaAge/airpha-VLM-7B}}
}
```

## Contact

- **Organization**: [AquaAge Inc.](https://huggingface.co/AquaAge)
- **Issues**: Please report issues on our [GitHub](https://github.com/AquaAge-Inc/airpha-model-zoo)

## License

This model is released under the Apache 2.0 License.