---
title: README
emoji: 📉
colorFrom: indigo
colorTo: indigo
sdk: static
pinned: false
---

# LiveCompose

**构妙 LiveCompose — AI-Powered On-Device Composition Assistant**

LiveCompose explores intelligent image composition, outpainting, and automatic cropping for mobile deployment.

[![GitHub](https://img.shields.io/badge/GitHub-LiveCompose-black?logo=github)](https://github.com/LiveCompose)
[![App Store](https://img.shields.io/badge/App_Store-构妙_LiveCapture-blue)](https://apps.apple.com/cn/app/%E6%9E%84%E5%A6%99/id6754213088)

---

LiveCompose is an AI-powered intelligent composition assistant that helps everyday users capture professionally-composed photos. Using reinforcement learning with aesthetic scoring, the system analyzes the viewfinder in real-time and actively guides users to the optimal composition — going far beyond traditional static grid lines.

## What We Build

We develop and host the AI models, training datasets, and evaluation benchmarks behind the LiveCompose intelligent photography system.

AdaCrop uses a two-stage cropping pipeline: a BBox head predicts an initial crop region, then an Actor policy refines the crop through sequential actions such as move, zoom, and stop.

## What We Hosts

- **Spaces**: the AdaCrop image cropping demo.
- **Models**: original AdaCrop PyTorch checkpoints, distilled MobileNetV3 student models, and Core ML exports for iOS.
- **Datasets**: outpainted image data with crop bounding boxes for training.

---

*LiveCompose — making every snapshot a masterpiece.*