Skip to content
@LiveCompose

构妙 LiveCompose

On-device real-time crop assistant APP power by Reinforcement Learning Model

LiveCompose

AI-Powered On-Device Intelligent Composition Assistant Based on Reinforcement Learning

Hugging Face App Store Code Size License

跳转到中文介绍


LiveCompose helps everyday users take professional photos. Standard cameras use static rule-of-thirds grids. This system analyzes the viewfinder in real time. It uses the device gyroscope to track movement. The system guides the user to move the phone for the best composition. It takes the photo automatically when the framing is perfect.

Projects

Project Description
LiveCompose RL model training, knowledge distillation, CoreML deployment
LiveCapture iOS APP source code

Tech Stack

  • Model Architecture: ResNet50 Backbone + Actor-Critic (PPO)
  • Aesthetic Scoring: NIMA / GAIC dual scorers for reward signals
  • Data Construction: BLIP + Stable Diffusion Inpaint dual-model workflow
  • On-Device Deployment: Knowledge distillation (MobileNetV3-Small) → CoreML → iOS
  • Application Framework: SwiftUI + AVFoundation + CoreMotion + CoreML

Key Features

  • Real-Time Dynamic Guidance: AI analyzes frames to show visual movement guides.
  • Sensor Fusion Tracking: Gyroscope and magnetic snapping enable smooth physical tracking.
  • Aesthetic Score Driven: NIMA and GAIC models ensure professional composition.
  • Closed-Loop Experience: The full process automates detection, tracking, guidance, and shooting.
  • Efficient On-Device Inference: The distilled lightweight model runs in real time on iOS devices.

Models and Datasets

Pre-trained model weights, training datasets, and demos are on Hugging Face:

App

iOS App: LiveCapture — Available on the App Store


LiveCompose — Frame the perfect moment with every shutter.


中文 README


构妙 LiveCompose

基于强化学习的 AI 端侧智能构图辅助系统


构妙 LiveCompose 致力于让每一位普通用户都能轻松拍出专业级构图照片。不同于传统相机的静态九宫格辅助线,我们通过 AI 实时分析取景画面,结合设备陀螺仪实现物理级追踪引导,主动"告诉"用户如何移动手机以获得最佳构图,并在对齐完美构图时自动拍摄。

Projects

项目 说明
LiveCompose RL 模型训练、知识蒸馏、CoreML 部署
LiveCapture iOS APP 源码

技术栈

  • 模型架构: ResNet50 Backbone + Actor-Critic (PPO)
  • 美学评分: NIMA / GAIC 双评分器驱动奖励信号
  • 数据构建: BLIP + Stable Diffusion Inpaint 双模型工作流
  • 端侧部署: 知识蒸馏 (MobileNetV3-Small) → CoreML → iOS
  • 应用框架: SwiftUI + AVFoundation + CoreMotion + CoreML

关键特性

  • 实时动态引导:AI 分析画面,可视化移动指引
  • 传感器融合追踪:陀螺仪 + 磁性吸附,物理级流畅追踪
  • 美学评分驱动:NIMA/GAIC 模型确保构图专业性
  • 闭环体验:检测 → 追踪 → 引导 → 拍摄,全流程自动化
  • 端侧高效推理:蒸馏后的轻量模型在 iOS 设备上实时运行

模型与数据集

预训练模型权重、训练数据集及 Demo 托管在 Hugging Face:

Hugging Face

App

iOS App: 构妙 LiveCapture — 已上架 App Store


构妙 LiveCompose — 让每一次快门,都定格最美的瞬间.

Popular repositories Loading

  1. LiveCompose LiveCompose Public

    LiveCompose is a reinforcement learning-based AI on-device composition assistance system. By analyzing the camera viewfinder in real time, combined with gyroscope tracking and aesthetic score-drive…

    Python 8

  2. LiveCapture LiveCapture Public

    LiveCompose is a reinforcement learning-based AI on-device composition assistance system. By analyzing the camera viewfinder in real time, combined with gyroscope tracking and aesthetic score-drive…

    Swift 3

  3. .github .github Public

Repositories

Showing 3 of 3 repositories

People

This organization has no public members. You must be a member to see who’s a part of this organization.

Top languages

Loading…

Most used topics

Loading…