README.md

---
title: LocateAnything Detection
emoji: 🔍
colorFrom: blue
colorTo: green
sdk: docker
pinned: false
license: mit
---

# LocateAnything Detection API

Serves [NVIDIA LocateAnything-3B](https://huggingface.co/nvidia/LocateAnything-3B) via [locate-anything.cpp](https://github.com/mudler/locate-anything.cpp) as an HTTP API.

Open-vocabulary object detection: give it an image + text prompt, get back labeled bounding boxes.

> Also see **Modal** backend config (`la_backend = "modal"`) for GPU-accelerated vLLM via Modal's serverless platform.

## API

### `GET /health`

```json
{"status": "ok", "model_loaded": true, "model_path": "/app/models/locate-anything-q4_k.gguf"}
```

### `POST /detect`

Multipart form:
- `image` — JPEG/PNG file
- `prompt` — text prompt (e.g. `"Locate the person."`)

Response:
```json
{
  "success": true,
  "detections": [
    {"label": "person", "box": [0.1, 0.2, 0.5, 0.8]}
  ],
  "elapsed_s": 5.2
}
```

Boxes are normalized `[x1, y1, x2, y2]` in 0–1 range.

## Deploy

1. Create a new [Hugging Face Space](https://huggingface.co/new-space) with **Docker** SDK
2. Set **Space hardware** to CPU (free) or upgrade to CPU-upgraded for faster inference
3. Upload the files from this `hf_space/` directory to the Space root
4. The Space auto-builds and starts — check `/health` after ~10 minutes

## Resources

- Model: [mudler/locate-anything.cpp-gguf](https://huggingface.co/mudler/locate-anything.cpp-gguf)
- C++ engine: [mudler/locate-anything.cpp](https://github.com/mudler/locate-anything.cpp)
- Official model: [nvidia/LocateAnything-3B](https://huggingface.co/nvidia/LocateAnything-3B)