ShowTable: Unlocking Creative Table Visualization with Collaborative Reflection and Refinement

TL;DR

ShowTable is a CVPR 2026 system that targets creative table visualization — given a tabular data source, produce an infographic that is both aesthetic and quantitatively faithful to the data points. The pipeline synergizes an MLLM as orchestrator with a diffusion model as executor, looping over four steps (Rewriting → Generation → Reflection → Refinement) so that geometric, labeling, and rendering errors are caught and re-instructed by the MLLM rather than baked into a single forward pass. The release ships three automated data-construction pipelines (30K rewriting pairs, 5K RL refinement samples, 30K pairwise preference samples for a reward model) and a new benchmark TableVisBench with 800 instances across 5 evaluation dimensions (Data Accuracy, Text Rendering, Relative Relationship, Additional Info Accuracy, Aesthetic Quality). On Qwen-Image as the executor, RW+REF lifts the composite score from 44.3 → 54.9; on weaker bases (Flux, Bagel, Blip3o-Next, UniWorld-V1, OmniGen2) the relative lift is larger but absolute scores stay far below the reference image’s 84.4.

Key claims

The “creative table visualization” task — generate an infographic that faithfully and aesthetically visualizes a given table — is introduced as a stress test for deep reasoning, planning, and precise data-to-visual mapping beyond general T2I generation [Abstract, Fig. 1].
The ShowTable pipeline is a four-step orchestrator/executor loop: (1) Rewriting — MLLM plans a visual sketch from tabular data; (2) Generation — diffusion model produces an initial figure; (3) Reflection — MLLM audits the output and emits precise editing instructions; (4) Refinement — an editing model executes the corrections [Method, Fig. 2].
Three separate automated data pipelines train three modules: 30K table–image pairs with Gemini-2.5-Pro-generated descriptions + chain-of-thought rationales for the Rewriting module; 5K RL samples filtered by an MLLM assessor comparing initial vs refined candidates for the Refinement module; 30K pairwise preference samples for a specialized Reward Model that supplies scalar rewards during RL [Data Construction Pipeline].
TableVisBench scores along 5 axes: Data Accuracy (every data point represented), Text Rendering (legibility/correctness of text), Relative Relationship (visual proportions match quantitative relationships), Additional Info Accuracy (axes/ticks/gridlines), Aesthetic Quality [TableVisBench].
Across 6 image-generation bases, ShowTable’s RW+REF combination consistently beats both the base model and RW-only: e.g. Flux 29.3 → 32.1 (RW) → 36.4 (RW+REF); Bagel 10.1 → 19.5 → 32.7; Blip3o-Next 10.8 → 14.1 → 34.8; UniWorld-V1 14.8 → 18.6 → 33.5; OmniGen2 14.4 → 21.9 → 29.9; Qwen-Image 44.3 → 54.3 → 54.9 [Leaderboard].
The reference (human-authored) infographic scores 84.4 composite, with the strongest model + ShowTable combo reaching 54.9 — a roughly 30-point gap, with the largest individual-dimension deficits in Relative Relationship and Aesthetic Quality (4.5 vs 4.2 the model actually edges out the reference on AQ, but underperforms on DA: 52.4 vs 97.7) [Leaderboard].
Per-case walkthroughs show the Reflection step diagnoses named, structured error classes: geometric proportionality (bar/arc heights not matching the data ratio), angular precision on donut/pie slices (e.g. 81% should be 291.6°), label correctness (mis-assigned legend entries, language errors like Hebrew leaking in, OCR-style digit corruption “7D” → “72”), and visual-hierarchy ordering across circle areas [Cases 1–9].
The refinement loop typically converges in 1–3 rounds and terminates on either reflection-approval or a max-rounds cap; some cases (e.g. pension-pot pie, mobile-OS bar chart) hit the cap with residual issues, indicating the loop is not yet a guaranteed fixpoint [Cases 7–9, “Max rounds reached. Output approved.”].

Method

ShowTable instantiates the MLLM-orchestrates-diffusion-executor pattern with a closed feedback loop. The MLLM (used in two roles — Rewriter and Reflector) reasons over the input table to plan a visual layout, the diffusion model generates an initial image conditioned on the rewritten prompt, the MLLM re-inspects the rendered image against the original data and emits structured editing instructions in natural language (“shorten the smaller bar so its angle is approximately 85.5% of the larger arc’s; move the endpoint counter-clockwise to ~3:30”), and an editing model executes the correction. The loop iterates until the Reflector approves or a max-round budget is reached.

Training data is built in three parallel synthetic pipelines: (1) 30K (table, image, description, CoT-rationale) tuples for Rewriting, with Gemini-2.5-Pro doing the description + rationale generation; (2) 5K RL samples for Refinement, filtered by an MLLM assessor that pairwise-ranks initial vs refined candidates; (3) 30K pairwise preference samples for a Reward Model that supplies the scalar reward during the RL stage. The reward model is trained specifically to track human preferences on infographic faithfulness, distinct from generic image aesthetics scorers.

Results

Composite TableVisBench scores (DA / TR / RR / AA / AQ → composite): reference image 97.7 / 99.5 / 86.4 / 96.6 / 4.2 → 84.4; Qwen-Image + RW+REF (best system) 52.4 / 82.9 / 54.3 / 40.0 / 4.5 → 54.9; Qwen-Image baseline 47.5 / 90.9 / 26.1 / 14.1 / 4.3 → 44.3 [Leaderboard].
Largest relative lift from REF on top of RW is on weak bases that lack innate text-rendering: Blip3o-Next 14.1 → 34.8 (+20.7) and Bagel 19.5 → 32.7 (+13.2), suggesting refinement compensates most where the base diffusion is weakest on text/data alignment [Leaderboard].
RR (Relative Relationship) — the axis closest to “data faithfulness” — improves dramatically: Qwen-Image 26.1 → 54.3, UniWorld-V1 14.7 → 37.6, Bagel 14.2 → 36.7, demonstrating that the geometric-correction loop is the dominant contributor on this dimension [Leaderboard].
DA (Data Accuracy) lags far behind reference (best 52.4 vs 97.7) — even with the loop, models still miss or misrepresent data points roughly half the time [Leaderboard].
Qualitative case studies (e.g. Mobile OS Sales, Case 7) reveal the loop catching multiple distinct error classes serially: round 1 fixes legend typos and duplicates; round 2 corrects bar values across years; round 3 fixes color-to-OS mapping — a curriculum the model cannot accomplish in a single forward pass [Case 7].

Why it’s interesting

This is the natural complement to IGenBench: Benchmarking the Reliability of Text-to-Infographic Generation — same problem (infographics from tabular data), same era (both flagged on the team’s radar this week), but ShowTable proposes a system where IGenBench proposes a benchmark. The rubric structures are strikingly aligned: TableVisBench’s five axes (DA / TR / RR / AA / AQ) are essentially the union of IGenBench’s data-encoding bottleneck dimensions (Completeness, Encoding, Ordering, Data Marks) plus aesthetic quality, and both papers identify the same failure mode: T2I models render decorations and text competently but cannot bind numerical values to visual mark sizes. ShowTable’s contribution is to show the gap is partly closable by a Reflection-in-the-loop scaffold rather than only by a stronger base generator. The pipeline is also a textbook instance of the VLM-as-Evaluator paradigm at training/inference time both: the MLLM acts as judge (Reflection step), as data-curation gate (the RL sample filter compares initial vs refined candidates), and as reward source (the trained RM is also MLLM-based) — three of the four shapes the concept page already tracks, integrated into one pipeline. Finally, the three-pipeline data-construction recipe (30K + 5K + 30K, each gated by a different MLLM-driven filter) is another instance of the Synthetic Training Data thesis that a verification gate is the load-bearing piece, applied here to a structured-rendering domain that prior synthetic-data work hadn’t touched.