ShowTable : Unlocking Creative Table Visualization with Collaborative Reflection and Refinement

Zhihang Liu1*, Xiaoyi Bao2*, Pandeng Li1,7†, Junjie Zhou3, Zhaohe Liao4, Yefei He5, Kaixun Jiang6, Chen-Wei Xie7, Yun Zheng7, Hongtao Xie1
1USTC 2CASIA 3NJU 4SJTU 5ZJU 6FDU 7Tongyi Lab
* Equal contribution, † Project leader
ShowTable Introduction

Figure 1. The illustration of our proposed creative table visualization task and ShowTable pipeline. Given a table about a specific topic, our task requires the model to produce a visualization infographic that is aesthetic and faithful to the data points.

Abstract

While existing generation and unified models excel at general image generation, they struggle with tasks requiring deep reasoning, planning, and precise data-to-visual mapping abilities beyond general scenarios. To push beyond the existing limitations, we introduce a new and challenging task: creative table visualization, requiring the model to generate an infographic that faithfully and aesthetically visualizes the data from a given table. To address this challenge, we propose ShowTable, a pipeline that synergizes MLLMs with diffusion models via a progressive self-correcting process. The MLLM acts as the central orchestrator for reasoning the visual plan and judging visual errors to provide refined instructions, the diffusion execute the commands from MLLM, achieving high-fidelity results. To support this task and our pipeline, we introduce three automated data construction pipelines for training different modules. Furthermore, we introduce TableVisBench, a new benchmark with 800 challenging instances across 5 evaluation dimensions. Experiments demonstrate that our pipeline, instantiated with different models, significantly outperforms baselines, highlighting its effective multi-modal reasoning, generation, and error correction capabilities.

Method

ShowTable Framework

Figure 2. Overview of the ShowTable pipeline.

The ShowTable Pipeline

ShowTable synergizes an MLLM as the central orchestrator with a diffusion model as the executor. The process consists of four key steps:

  • 1
    Rewriting: MLLM reasons over tabular data to plan an aesthetic visual sketch.
  • 2
    Generation: Diffusion model creates an initial figure based on the MLLM's sketch.
  • 3
    Reflection: MLLM assesses the output and provides precise editing instructions.
  • 4
    Refinement: Editing model executes corrections to achieve high fidelity.

Data Construction Pipeline

Data Construction Pipeline

Figure 3. Dataset construction pipeline: (1) Rewriting training data, (2) Refinement training data, and (3) Reward training data.

Rewriting Data

We construct 30K table-image pairs for training the rewriting module. We use Gemini-2.5-pro to generate detailed descriptions and chain-of-thought rationales, enabling the model to learn semantic reasoning and compositional planning.

Refinement Data

We filter 5K challenging samples for reinforcement learning (RL). By comparing initial generations with refined candidates using a powerful MLLM assessor, we select samples that are ideal for training the refinement module.

Reward Data

We construct 30K pairwise preference samples to train a specialized Reward Model (RM). This model is trained to align with human preferences and provide accurate scalar rewards for the RL process.

TableVisBench

We introduce TableVisBench, a benchmark with 800 challenging instances across 5 evaluation dimensions to comprehensively assess creative table visualization.

🎯

Data Accuracy (DA)

Verifies that every data point from the source table is accurately represented in the generated image, ensuring no missing or incorrect data.

🔤

Text Rendering (TR)

Focuses on the legibility and correctness of all textual elements in the image.

📊

Relative Relationship (RR)

Assesses whether the visual proportions (e.g., bar heights, slice angles) correctly reflect the quantitative relationships between data points.

ℹ️

Add. Info Accuracy (AA)

Inspects the accuracy of contextual information added by the model, such as axes, ticks, gridlines, and other artifacts.

🎨

Aesthetic Quality (AQ)

Evaluates the overall visual appeal, including layout, color palette, typography, and design creativity.

Leaderboard

Methods DA TR RR AA AQ Score
Reference Image 97.7 99.5 86.4 96.6 4.2 84.4
Flux 12.1 46.7 28.9 18.7 4.0 29.3
+ Rewrite (RW) 12.0 52.3 27.0 25.3 4.4 32.1
+ ShowTable (RW+REF) 20.3 63.1 31.8 24.0 4.3 36.4
Bagel 0.1 1.6 14.2 7.7 2.7 10.1
+ Rewrite (RW) 3.4 18.3 28.9 13.0 3.4 19.5
+ ShowTable (RW+REF) 18.3 54.8 36.7 15.9 3.8 32.7
Blip3o-Next 0.4 18.0 4.4 6.2 2.5 10.8
+ Rewrite (RW) 0.5 14.5 19.1 7.6 2.9 14.1
+ ShowTable (RW+REF) 21.3 63.9 33.4 19.2 3.6 34.8
UniWorld-V1 3.0 18.3 14.7 2.9 3.5 14.8
+ Rewrite (RW) 4.0 20.8 23.7 11.6 3.3 18.6
+ ShowTable (RW+REF) 18.7 54.6 37.6 18.8 3.8 33.5
OmniGen2 3.1 17.8 13.5 2.6 3.5 14.4
+ Rewrite (RW) 4.0 32.1 25.0 9.5 3.9 21.9
+ ShowTable (RW+REF) 16.2 49.8 30.6 13.8 3.9 29.9
Qwen-Image 47.5 90.9 26.1 14.1 4.3 44.3
+ Rewrite (RW) 51.2 83.1 50.1 40.9 4.6 54.3
+ ShowTable (RW+REF) 52.4 82.9 54.3 40.0 4.5 54.9

Qualitative Results

Detailed pipeline visualizations demonstrating collaborative reflection and refinement.

Advertising Effectiveness by Medium

Medium Percentage
Magazine/Newspaper 35%
Billboard 13%
Final 3D Visualization
Click to View Process

Survey Results: Yes vs No

Response Percentage
Yes 81%
No 19%
Final 3D Donut Chart
Click to View Process

Advertiser Outlook for Paid Social Advertising Budgets

Outlook Percentage
Increase 64%
Stay the Same 34%
Decrease 2%
Final Advertiser Outlook Visualization
Click to View Process

Smartphone & AI Usage Statistics

Category Number
Smartphone Users 169M
AI Assistant Users 72M
Final Smartphone & AI Usage Visualization
Click to View Process

Financial Year Comparison

Financial Year Value % Change
18/19 247.5 -
19/20 211.5 -14.5%
Final Financial Year Visualization
Click to View Process

Social Media Impact Survey

Effect Number
Limits attention span when completing assignments 75
Makes me proud to realize I am known online 25
I hide behind my social media 10
Decreases my mental health and creates toxicity 35
Final Social Media Impact Visualization
Click to View Process

Mobile OS Sales Forecast (Millions of Units)

Year Symbian Android
2009 80 10
2010 115 40
2014 260 250
(Data includes 2009-2014 sales for multiple OS)
Final Mobile OS Sales Chart
Click to View Process

Breakdown of Pension Pot

Category Amount
Some of your tax-free cash £10,000
Equivalent sum to drawdown £30,000
Remaining pension pot untouched £60,000
Final Pension Distribution Chart
Click to View Process

Vulnerability Severity Distribution and Scores

Score Category Percentage
Scores of 9 or higher 28%
Scores of 7 or higher 84%
Severity Distribution:
CRITICAL 5.1%
HIGH 65.3%
Final Vulnerability Dashboard
Click to View Process

Public Priorities for Scientific Advancements

Priority Support
Finding cures for diseases 79%
Reducing poverty and hunger 61%
Improving education 56%
Developing clean energy 55%
(Survey includes 11 categories total)
Final Scientific Priorities Visualization
Click to View Process

BibTeX

@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}
}