ShowTable: Unlocking Creative Table Visualization with Collaborative Reflection and Refinement

• Creative table visualization requires both aesthetic graphic reasoning and high-fidelity mapping from table data to visuals.Source: §1 · pp. 1–2
• ShowTable pairs an MLLM planner and error assessor with a diffusion generator and refiner in a progressive self-correcting loop.Source: §1 · p. 2
• The authors construct 30K rewriting examples and use a reward model with reinforcement learning on 5K filtered samples to train refinement.Source: §1 · p. 2
• TableVisBench contains 800 challenging, manually verified table instances assessed across five evaluation dimensions.The dimensions are data accuracy, text rendering, relative relationship, additional information accuracy, and aesthetic quality.Source: §1 · p. 2
• ShowTable significantly boosts all base models on TableVisBench.Source: §1 · p. 2

• General image generation includes work on accurate text rendering and aesthetic poster layout, while creative table visualization also requires precise data-to-visual mapping.Source: §2 · p. 2
• MLLM reasoning has been used for image understanding and text-based generation planning, while emerging reflection methods refine general-purpose image generation.Source: §2 · p. 2
• Creative table visualization poses a greater challenge because high information density and structural constraints demand combined reasoning, planning, and reflection.Source: §2 · p. 2

3.1. ShowTable Pipeline

• ShowTable structures creative table visualization as a Rewriting → Generation → Reflection → Refinement self-correction loop.Source: §3.1 · pp. 2–3, Figure 2 · p. 3
• An MLLM turns dense table structure into an executable descriptive prompt because raw markdown prompts may make generators render the table instead of visualizing its data.Source: §3.1 · p. 3
• Initial text-to-image generation may misstate bar heights or labels; an MLLM audits the image against the table and gives actionable edits to an image editor.Source: §3.1 · p. 3

3.2. Training Details

• The rewrite module is fine-tuned from Qwen3-8B on 30K constructed examples because general LLMs may plan poorly on complex tables.Source: §3.2 · p. 3
• Across correction rounds, Qwen-Image-Edit degrades while Wan2.5-I2I-Preview improves, indicating refinement-model capability constrains the loop.Source: §3.2 · p. 4, Figure 4 · p. 4
• A Qwen2.5-VL-3B reward model learns pairwise preferences and combines with ImageReward to guide RL refinement training on 5K samples.Source: §3.2 · p. 4

4.1. Data Collection and Filtering

• 30K table-image pairs are collected from public datasets, filtered for resolution and text, annotated independently by two MLLMs, and retained when annotations agree.Images under 200 × 200 or lacking text are discarded before MLLM filtering and annotation.Source: §4.1 · pp. 4–5

4.2. Training Data Construction

• For rewriting, Gemini-2.5-pro generates image descriptions and conversion rationales, producing 30K {table, rationale} → {description} pairs for fine-tuning.Source: §4.2 · p. 5
• Refinement training filters candidate edits to 5K challenging samples, while reward training uses model-voted preferences to create 30K positive-negative image pairs.Samples are excluded when all five refinement attempts are judged either worse or better than the initial image.Source: §4.2 · p. 5

4.3. Benchmark Design

• TableVisBench contains 800 manually verified instances and evaluates data accuracy, text rendering, relative relationships, additional-information accuracy, and aesthetics.The first four dimensions range from 0 to 100, while aesthetic quality ranges from 0 to 10; the final score averages the first four with 10×AQ.Source: §4.3 · pp. 5–6
• Collected ground-truth images score highly on the benchmark dimensions, supporting the authors’ claim that its metrics align with human-annotated quality.Source: §4.3 · p. 6

5.1. Implementation Setup

• The default setup uses Qwen3-8B for rewriting, distilled Qwen-Image-Edit-2509 for refinement, GPT-5 for reflection, and at most three correction rounds.Reflection may terminate the process early when it judges an image satisfactory.Source: §5.1 · p. 6

5.3. Ablation Studies

• GPT-5 scores 54.9 with the trained refiner, while Gemini-2.5-pro scores 51.2 with the base refiner.Source: Table 3 · p. 7, §5.3 · pp. 7–8

5.2. Main Results

• 44.3→54.9 overall for Qwen-Image; RW+Base+REF achieves the best overall score across all tested base models.For Blip3o-Next, the full pipeline raises DA from 0.5 to 21.3 and TR from 14.5 to 63.9.Source: §5.2 · p. 7

5.3. Ablation Studies

• 49.4→54.9 (+5.5) overall; trained Qwen-Image-Edit-2509* improves over its base model, while Wan2.5 reaches 63.4.The trained model gains +9.8 DA and +9.3 RR over its base model.Source: Table 4 · p. 7, §5.3 · p. 8
• 54.3 overall; the fine-tuned Qwen3-8B* leads rewriting strategies and attains 51.2 DA, exceeding the reference-caption DA of 50.3.Its DA and RR gains over the baseline are +3.7 and +24.0, respectively, while TR decreases slightly.Source: §5.3 · p. 7

5.2. Main Results

• Figure 6 shows reflection and refinement correcting text, numerical, proportional, and visual-element errors, with some examples requiring no refinement.Source: §5.2 · p. 7, Figure 6 · p. 8

5.3. Ablation Studies

• Across rounds, the base editor declines from 54.3 to 49.4, whereas the trained editor improves from 53.7 to 54.9 and Wan2.5 from 61.3 to 63.4.Source: Table 5 · p. 8, §5.3 · p. 8

6. Conclusion

• Creative table visualization requires both aesthetic graphic reasoning and high-fidelity mapping from data to visuals; ShowTable combines MLLMs and diffusion models through iterative reflection and refinement.Source: §6 · p. 8
• The paper also presents three automated training-data construction pipelines and TableVisBench, with five evaluation dimensions, as a foundation for future research in multimodal reasoning and visual synthesis.Source: §6 · p. 9

A.1. Chart Generation with Agentic Tools

• Chart-generation approaches include direct structured specifications, executable plotting code, and template retrieval, each constrained by grammar, artistic flexibility, or template diversity and alignment.Source: §A.1 · p. 13
• Code-based methods achieve higher factual accuracy but lower Aesthetic Quality (AQ: 4.1 vs. our 4.9) than ShowTable.The comparison uses Gemini-2.5-Pro-generated Matplotlib code; Figure A1 also shows the qualitative comparison.Source: Table A1 · p. 13, Figure A1 · p. 13, §A.1 · p. 14
• ShowTable is not intended to replace code rendering when absolute numerical precision is the sole priority; it aims to retain generative flexibility while mitigating data-fidelity weaknesses.Source: §A.1 · p. 14

A.2. Reinforcement Learning for Image Generation

• Image-generation reinforcement learning progressed from policy-gradient guidance of denoising to preference alignment and GRPO-based adaptations.Source: §A.2 · p. 14

B.1. Prompts in Pipeline

• The MLLM rewrites table data into a diffusion-ready visual prompt, then audits the generated image and supplies actionable editing instructions; prompts remain consistent across tested models.Source: §B.1 · p. 14

B.2. Rewriting Training Details

• The Qwen3-8B rewriting module is fine-tuned on 30K SFT examples; its reasoning supports improved prompts and data integrity in the generation pipeline.Training uses table-visualization captions and includes both reasoning and rewriting results.Source: §B.2 · p. 14

B.3. Refinement Training Details

• Refinement uses on-policy GRPO with rendering accuracy as its reward; inconsistent VLM scoring can destabilize training, and VLM latency hampers the on-policy loop.The reward combines the trained assessor with ImageReward, while the assessor is trained on preference pairs.Source: §B.3 · pp. 14–16
• The trained reward model yields stable, consistent performance gains over raw LLM scores in refinement training.Source: Figure A4 · p. 16, §B.3 · p. 16

C.1. Training Data Format

• Rewriting data pairs annotated markdown tables with reasoning and image descriptions; refinement data pairs erroneous initial images with correction instructions and filters out too-easy or too-hard cases.Source: §C.1 · p. 16
• Reward-model data uses judged positive–negative image pairs drawn from refined versus initial, stronger versus weaker, and ground-truth versus generated comparisons.The output labels each candidate as Winner or Loser.Source: §C.1 · pp. 16–17

C.2. Benchmark Statistics

• TableVisBench covers tables from fewer than 5 rows to over 40 data points, with most instances containing 5–15 points and topics spanning business, sociology, and technology.Source: §C.2 · p. 17

C.3. Benchmark Evaluation Details

• Data Accuracy counts incorrect or missing table values; Text Rendering scores character correctness; Relative Relationship checks whether visual proportions match the data.Data Accuracy and Relative Relationship use the fraction of correctly rendered or proportioned data points; Text Rendering uses the fraction of correct characters.Source: §C.3 · p. 17
• Additional Information Accuracy evaluates contextual elements such as axes, ticks, and extra annotations by averaging up to three applicable submetrics; Aesthetic Quality uses a pretrained model score from 0 to 10 for visual appeal.When no additional information is present, Additional Information Accuracy is excluded from the calculation.Source: §C.3 · pp. 17–19

D.1. Results of Wan2.5-Preview

• With Wan2.5-T2I-Preview, rewriting lifts overall score 62.5→72.1 (+9.6), while trained editing recovers 59.5→68.1 (+8.6) after base-editor degradation.Rewriting also raises Data Accuracy (62.8→73.3) and Text Rendering (82.5→92.9); Wan2.5-I2I-Preview reaches Data Accuracy 76.9 and Relative Relationship 74.3.Source: Table A2 · p. 18, §D.1 · p. 20

D.2. More Visualization Results

• Without rewriting, unified baselines produce chaotic layouts and Qwen-Image and Flux fail to render data-point structure; rewriting yields coherent layouts but can leave logical errors.Figure A10 contrasts outputs without rewriting in its first column with ShowTable pipeline results in the second.Source: §D.2 · p. 20, Figure A10 · p. 21

D.3. User Study

• 73.75% human preference for ShowTable in Data Accuracy; base Qwen-Image receives 38.50% in Text Rendering and 2.25% in Data Accuracy.Twenty evaluators independently judged outputs for 20 randomly selected tables across five dimensions.Source: Table A3 · p. 20, §D.3 · pp. 20–22

D.2. More Visualization Results

• Iterative refinement repairs data segmentation, garbled text, and layout misalignment; high-quality Wan2.5-Preview outputs trigger early stopping to avoid unnecessary computation.The pipeline visualizations show stepwise instructions, including geometric corrections to bar-height ratios and pie-slice angles.Source: §D.2 · p. 20

Full-pipeline training

• Current training primarily covers rewriting and refinement; generation and reflection still rely on off-the-shelf models.The authors propose extending supervised fine-tuning or reinforcement learning to all components and jointly optimizing the pipeline.Source: E. Limitations and Future Work · pp. 22, 24

Towards a unified model

• The four-stage cascade increases deployment complexity and inference latency; a unified end-to-end multimodal architecture is proposed as a future direction.The four tasks are rewriting, generation, reflection, and refinement.Source: E. Limitations and Future Work · p. 24

Dependency on foundation models

• Pipeline performance is constrained by foundation-model capabilities, and refinement precision often determines the upper bound of visualization fidelity.Source: E. Limitations and Future Work · p. 24