Skip to content

Latest commit

 

History

History
87 lines (61 loc) · 3.27 KB

File metadata and controls

87 lines (61 loc) · 3.27 KB

Mind-Brush: Integrating Agentic Cognitive Search and Reasoning into Image Generation

English | 中文

arXiv Paper GitHub Repo Dataset

📰 新闻

🏆 主要贡献

  • 🧠 Mind-Brush 框架:一种全新的代理范式,将 意图分析多模态搜索知识推理 统一为一个无缝的 “思考-研究-创作” 图像生成工作流。
  • 📊 Mind-Bench:专门设计的基准测试,用于评估生成模型在动态外部知识复杂逻辑推理方面的表现,揭示了当前 SOTA 多模态模型的推理差距。
  • 🏆 卓越性能
    • 在 Mind-Bench 上将 Qwen-Image 基准准确率从 0.02 提升至 0.31
    • WISE+25.8% WiScore)和 RISEBench+27.3% 准确率)上超越了现有基准。

📽️ 演示回放

Search 示例 Search & Reason 示例
case1.mp4
case2.mp4

🚀 快速开始

1. 克隆项目

git clone https://github.com/PicoTrex/Mind-Brush.git
cd Mind-Brush

2. 安装环境

conda create -n mindbrush python=3.12
conda activate mindbrush
pip install -r requirements.txt

3. 配置

填写 config.yaml 中的 [required] 字段(如 API 密钥、路径设置等)。

Note

你可以在 .chainlit/config.toml 中设置你的语言。默认情况下,语言设置为 en-US。你可以通过设置 language = "zh-CN" 将其更改为中文。 [目前我们只支持英语和中文。如果你想支持其他语言,可以在 locales 文件夹中添加相应的语言设置。]

4. 启动运行

使用以下命令运行程序:

chainlit run app.py -w

启动后,通过浏览器访问仪表板:http://localhost:8000

🩷 鸣谢