InfoVis: Unified Slide Image Generation and Editing

InfoVis. Unified Slide Image Generation and Editing. By Jiahao Wang and nine others, from the Institute of Automation of the Chinese Academy of Sciences, the Beijing Academy of Artificial Intelligence, and two other institutions. Published on the twelfth of October. Its code repository is public and real: eleven source files.

A presentation slide asks a lot of one canvas. Text has to be rendered accurately, instructions followed, content visually grounded, the design has to look good, and edits have to land precisely. Existing work covers only some of these capabilities. It also lacks two things at scale: supervision that is native to slides, and evaluation that can diagnose where a model falls short.

The authors answer with a whole stack built around slides. They formulate four slide-centric tasks, build the data and the benchmark to go with them, and then train a single model, InfoVis, that handles both generating and editing slides.

The data comes from a scalable engine producing over thirteen million records for continual pretraining and one point nine million records for supervised fine-tuning. The benchmark is VisSlideBench: one thousand five hundred and twenty-seven manually screened instances covering all four tasks, each scored with task-specific, multidimensional evaluation. Its automatic rankings agree with blinded human rankings made dimension by dimension. InfoVis itself is a unified nine billion parameter model, starting from FLUX.2-klein-base-9B and given slide-specific post-training.

On VisSlideBench, InfoVis leads every evaluated open-source model on all eleven metrics. Its overall score is seven point nine four, against seven point nine nine for Nano-Banana-2. The slide-specific post-training adds one point six points overall to the base model, with the largest generation gains in content faithfulness.

The remaining gap to proprietary models has a pattern. It is smallest in content faithfulness and largest in aesthetic quality. Outside their own benchmark, a document-to-slide workflow built on InfoVis achieves the second-highest Aesthetics score on SlidesGen-Bench.

The open-source lead is broad, and the distance to the closed model is a few hundredths of a point overall. Where that distance remains is informative: getting the content right is nearly solved here, and making it beautiful is the harder half. The authors say they will release the model, code, data, and benchmark.