ShowTable

[🍎 Project Page] [📖 ArXiv Paper] [🤗 HF Dataset]

🎉 Our paper has been accepted by CVPR 2026!

ShowTable is a novel pipeline that synergizes Multimodal Large Language Models (MLLMs) with diffusion models via a progressive self-correcting process for Creative Table Visualization. This task challenges models to generate infographics that are both aesthetically pleasing and faithful to the data points in a given table.

This repository contains the official inference and evaluation implementation of the paper:

ShowTable: Unlocking Creative Table Visualization with Collaborative Reflection and Refinement

🚀 Repository Overview

The ShowTable pipeline consists of four key stages, we implement all inference process via API or Sever Calling.

1. Rewriting: The MLLM reasons over the tabular data to plan an aesthetic visual sketch and rewrites the user prompt. You can find the code in stages/rewrite_stage.py and clients/rewrite_client.py.
2. Generation: The diffusion model creates an initial figure based on the MLLM's sketch. You can find the code in stages/image_gen_stage.py and clients/image_gen_client.py.
3. Reflection: The MLLM assesses the generated output, identifying inconsistencies and formulating precise editing instructions. You can find the code in stages/judge_stage.py and clients/judge_client.py.
4. Refinement: The diffusion model edits the figure based on the reflective feedback to produce the final high-fidelity visualization. You can find the code in stages/edit_stage.py and clients/edit_client.py.

The repository also supports the evaluation on our proposed TableVisBench. You can find the code in stages/eval_stage.py and clients/eval_client.py.

🛠️ Getting Started

1. Installation

Clone the repository and install the required dependencies:

git clone https://github.com/ali-vilab/ShowTable.git
cd ShowTable
pip install -r requirements.txt

2. Model Server Preparation

The pipeline requires several models to be running. You can start them using the provided scripts in the script/ directory.

For example, to start the Qwen-Image service:

bash script/start_qwen_image.sh

Ensure you have the necessary model weights and environment configurations set up as per the scripts.

3. Configuration

Configuration is managed via YAML files. A sample configuration is provided in configs/example.yaml.

Example configs/example.yaml modification:

data:
  prompts_file: "data/eval_data.jsonl"
  out_dir: "work_dirs/eval_outputs"

rewrite:
  enable: true
  provider: "server"
  server:
    - model: "qwen3"
      api_base: "http://your-server-ip:port/v1"

image_gen:
  provider: "qwen_image_server"
  qwen_image_server:
    - api_base: "http://your-server-ip:port"

4. Running the Pipeline

To run the pipeline, use the main.py script with your configuration file:

python main.py -c configs/example.yaml

📊 Benchmark (TableVisBench)

The repository supports evaluation on TableVisBench. Ensure your data.prompts_file points to the benchmark dataset and run.mode is set to eval. The evaluation metrics include:

• DA: Data Accuracy
• TR: Text Rendering
• RR: Relative Relationship
• AA: Additional Information Accuracy
• AQ: Aesthetic Quality

📝 Citation

If you find this work useful, please consider citing our paper:

@article{liu2025showtable,
  title={ShowTable: Unlocking Creative Table Visualization with Collaborative Reflection and Refinement},
  author={Liu, Zhihang and Bao, Xiaoyi and Li, Pandeng and Zhou, Junjie and Liao, Zhaohe and He, Yefei and Jiang, Kaixun and Xie, Chen-Wei and Zheng, Yun and Xie, Hongtao},
  journal={arXiv preprint arXiv:2512.13303},
  year={2025}
}