InfoVis: Unified Slide Image Generation and Editing
InfoVis is a unified 9B model for slide image generation and editing, fine-tuned from FLUX.2-klein-base-9B. One model handles text-to-slide generation, template-conditioned generation, multi-image grounded generation, and instruction-based slide editing.
📄 Paper | 🌐 Project page | 💻 Code | 🚀 Demo | 📊 VisSlideBench | 📦 Training data
This repository contains the InfoVis transformer in diffusers format. The text encoder (Qwen3) and the VAE are loaded from the base model.
Usage
Try InfoVis in the browser in the demo Space, or run it locally with the inference script from our code repository:
git clone https://github.com/VectorSpaceLab/InfoVis.git
cd InfoVis
pip install -r requirements.txt
# Text-to-slide
python inference.py --transformer InfoVis/InfoVis \
--prompt "A clean title slide for a talk on renewable energy, with a wind turbine illustration." \
--out slide.png
# Slide editing
python inference.py --transformer InfoVis/InfoVis \
--prompt "Change the title color to red." \
--images source.png --width 1280 --height 720 --out edited.png
--images takes the reference slide (template-conditioned), one to four assets (multi-image grounded), or the source slide (editing). The script masks prompt padding in attention and uses a 4,096-token text context; the default Flux2KleinPipeline does neither, so use the script to reproduce our results.
Results
| Model | Type | VisSlideBench Overall |
|---|---|---|
| GPT-Image-2 | Proprietary | 8.36 |
| Nano-Banana-2 | Proprietary | 7.99 |
| FLUX.2-dev (32B) | Open-source | 6.37 |
| FLUX.2-klein-base-9B | Open-source | 6.30 |
| InfoVis (9B) | Open-source | 7.94 |
The benchmark is available at InfoVis/VisSlideBench.
License
InfoVis is a derivative of FLUX.2-klein-base-9B and is released under the FLUX Non-Commercial License, for non-commercial use only. See NOTICE for the required attribution.
Citation
@article{wang2026infovis,
title={InfoVis: Unified Slide Image Generation and Editing},
author={Jiahao Wang and Xiyan Jiang and Chenyuan Wu and Haoqiang Zhang and Xin Luo and Yueze Wang and Defu Lian and Bo Zhang and Jiajun Zhang and Zheng Liu},
journal={arXiv preprint arXiv:2610.16784},
year={2026}
}