InfoVis: Unified Slide Image Generation and Editing

News | Quick Start | Training | Citation

English | 简体中文

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.

✨ Highlights

• One model, four slide tasks: input images serve as a style reference, content assets, or the slide to edit.
• Slide-native data and benchmark: over 13M continual-pretraining and 1.9M SFT records, and VisSlideBench, 1,527 manually screened cases with task-specific metrics.
• Strong results: 7.94 overall on VisSlideBench against 7.99 for Nano-Banana-2, leading every evaluated open-source model on all eleven metrics.

🔥 News

• 2026-10: Try InfoVis online in our Hugging Face Space.
• 2026-10: The paper is on arXiv. Code, model weights, VisSlideBench with its evaluation code, and all training data (continual pretraining and SFT) are released.

📖 Introduction

InfoVis is trained in three stages: slide-native continual pretraining, four-task mixed continual pretraining, and task-specific SFT whose branches are merged into one model. VisSlideBench scores generation on content faithfulness, instruction following, and aesthetic quality, and editing on semantic consistency and presentation quality.

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

Per-task results and more baselines are in the paper.

📌 TODO

• Training and inference code.
• InfoVis model weights.
• VisSlideBench and evaluation code.
• SFT training data.
• Continual-pretraining data.

🚀 Quick Start

No GPU at hand? Try InfoVis in the browser on the Hugging Face Space.

🛠️ Environment Setup

conda create -n infovis python=3.12
conda activate infovis

# Choose the command that matches your CUDA version (CUDA 12.6 shown here)
pip install torch==2.7.1 torchvision --extra-index-url https://download.pytorch.org/whl/cu126

git clone https://github.com/VectorSpaceLab/InfoVis.git
cd InfoVis
pip install -r requirements.txt

🌏 For users in Mainland China

pip install torch==2.7.1 torchvision --index-url https://mirror.sjtu.edu.cn/pytorch-wheels/cu126
pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple

🧪 Inference

Download the weights (or pass --transformer InfoVis/InfoVis to load them directly from Hugging Face):

hf download InfoVis/InfoVis --local-dir checkpoints/InfoVis

# Text-to-slide
python inference.py --transformer checkpoints/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 checkpoints/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 base model black-forest-labs/FLUX.2-klein-base-9B is downloaded from Hugging Face on first use.

To run all four VisSlideBench tracks on 8 GPUs:

hf download InfoVis/VisSlideBench --repo-type dataset --local-dir VisSlideBench
bash scripts/infer_visslidebench.sh checkpoints/InfoVis VisSlideBench outputs/infovis

The scoring scripts and judge prompts are in VisSlideBench/eval; see its README.

🏋️ Training

One script runs every stage; each stage is a config in configs/train/:

NNODES=10 NODE_RANK=$RANK MASTER_ADDR=$MASTER bash scripts/train.sh configs/train/02_cpt_mix_a.yaml

The training data of all stages is available at InfoVis/Data. See docs/training.md for the data format, all stage configs, and how to merge the SFT branches.

📜 License

The code is released under the Apache-2.0 License. The InfoVis weights are a derivative of FLUX.2-klein-base-9B and follow the FLUX Non-Commercial License. VisSlideBench and the training data are released under CC BY-NC 4.0; see their dataset cards for details.

🙏 Acknowledgements

InfoVis builds on FLUX.2-klein, diffusers, and OmniGen2.

❤️ Citation

If you find this repository or our work useful, please consider giving a star ⭐ and 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}
}