What is BizGen
BizGen is an AI infographic generation tool jointly launched by Tsinghua University and Microsoft Research, focusing on article-level visual text rendering. It can transform long article content into professional-grade infographics and slides with one click, solving the problems of blurry text and chaotic layout in traditional tools when processing long texts. Based on the high-quality dataset Infographics-650K and the advanced "layout-guided cross-attention mechanism," it can decompose long texts into small instructions and precisely inject them into different regions of the image.
Main Features of BizGen
- High-Quality Content Generation: Automatically generates professional-level infographics and slides based on the article content input by the user, solving problems such as blurry text and chaotic layout in traditional tools when processing long articles.
- Multilingual and Style Support: Supports ten different languages and can generate infographics in multiple styles to meet different needs.
- Multi-Layer Transparent Infographics: Performs excellently in generating multi-layer transparent infographics, making information presentation more flexible and diverse.
- High Accuracy and Layout Quality: Text spelling accuracy far exceeds that of other models, and user research shows that its layout quality is more favored.
- Strong Technical Support: Based on the Infographics-650K dataset and introducing the "layout-guided cross-attention mechanism," it ensures that every visual element and text region is finely controlled.
Technical Principles of BizGen
- High-Quality Dataset: The BizGen team built the Infographics-650K dataset, an unprecedented-scale high-quality commercial content dataset containing 650,000 exquisite commercial infographics and slides, each equipped with detailed layout information and descriptions. It lays a solid foundation for the model to learn and understand complex commercial design.
- Layout-Guided Cross-Attention Mechanism: It can decompose long article-level prompts into "small instructions" for different regions, and according to a preset ultra-high-density layout, precisely inject the instructions into different regions of the image. It can ensure that every visual element and text region is finely controlled, avoiding the chaos and errors caused by global processing in traditional methods.
- Layout-Conditioned Controlled Generation: During the inference stage, BizGen uses the "layout-conditioned controlled generation" method, like a strict quality inspector, carefully checking each generated sub-region and promptly correcting possible flaws to ensure the perfect presentation of the final work.
BizGen Project Address
- Project Official Website：https://bizgen-msra.github.io/
- Github Repository：https://github.com/1230young/bizgen
- HuggingFace Model Library：https://huggingface.co/PYY2001/BizGen
- arXiv Technical Paper：https://arxiv.org/pdf/2503.20672
Application Scenarios of BizGen
- Business Reporting: Quickly generate high-quality business reports and presentations.
- Product Display: Create attractive product promotional posters and slides.
- Academic Research: Generate academic reports and presentations.
- Social Media: Create attractive social media content.
- Education: Help teachers quickly create teaching courseware.
- Advertising Design: Automatically generate exquisite advertising posters that fit the theme.