侯睿

侯睿

四川大学 人工智能专业

Posts by Collection

portfolio

明厨亮灶管理系统

Published:

明厨亮灶管理系统是四川大学问题求解实战课程团队项目，其中我担任项目负责人，实现了Web端与微信小程序双平台的厨房监控管理系统。

南方财富金融理财平台

Published:

南方财富金融理财平台是基于Flutter框架开发的跨平台金融理财应用，支持Web、移动端等多平台，为用户提供社区交流、行情交易、资产管理等一站式金融服务。我在其中担任主要成员。

水稻图像精准控制生成平台

Published:

基于Swin-ControlNet的精准引导控制水稻图像生成平台，是一个融合了Swin Transformer与ControlNet技术的深度学习图像生成系统，我是该软件著作的第一权利所有人。该平台通过引入Swin Transformer模块构建双流空间变换器架构，利用水稻数据集（CVRP）进行训练，能够根据用户提供的控制提示（红色掩码区域）精准生成高质量的水稻图像。系统提供的Gradio Web界面支持多种参数配置，可广泛应用于农业图像生成、精准农业研究等领域。

基于图像大模型增强的恶劣天气交通目标检测方法研究

Published:

尽管交通目标检测方法在晴朗天气下已取得显著进展，但在恶劣天气条件下往往会失败，这归因于当前的训练数据缺乏包含雾、雨、雪等恶劣天气驾驶图像。我们提出基于图像大模型合成恶劣天气数据集的构建方法，并构建了一套贴近真实恶劣天气下的驾驶场景数据，用于增强后训练交通目标检测模型，大幅提升其在雾、雨、雪等恶劣天气条件下的车辆检测精度。

publications

talks

Efficient Architecture Optimization for Conditional Diffusion Models via Evolutionary Algorithms

Published:

This project is a provincial-level “National Undergraduate Innovation Training Program” project, positioned at the intersection of generative AI and design science. Starting from real-world application scenarios, it systematically examines the deep bottlenecks of existing conditional diffusion models in terms of control precision and generation quality, and conducts a systematic exploration spanning the entire pipeline from theoretical modeling, algorithm design, and experimental validation to engineering implementation and user deployment. The cumulative outputs include high-level conference papers, national invention patents, software copyrights, and a multi-agent AIGC system for design — forming a complete closed loop of “problem-driven → methodological innovation → outcome transformation → applied verification.” The system is currently open to real users, having accumulated genuine user feedback and practical experience, providing technical support and a practical paradigm for building efficient and controllable intelligent design capabilities.

teaching

Understanding Diffusion Models: From DDPM to Score

Undergraduate course, Sichuan University, College of Computer Science (Neural Network Design and Practice), 2026

This tutorial walks through the conceptual bridge from DDPM to score-based generative models — from maximum likelihood and energy-based models, through score matching and Langevin dynamics, to the multi-scale SDE formulation that unifies score-based models and diffusion models.