Pith. sign in

REVIEW 4 major objections 6 minor 5 cited by

IGenBench: Benchmarking the Reliability of Text-to-Infographic Generation

T0 review · 4 major / 6 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read IGenBench decomposes infographic reliability into atomic yes/no questions and finds that even the strongest text-to-image model passes only 49% of infographics end-to-end.

desk verdict First real reliability benchmark for text-to-infographic generation with careful curation and interpretable metrics, but evaluator validation is weaker than it looks and the sampling algorithm doesn't produce the stated size. read the letter →

arxiv 2601.04498 v2 pith:4S4KXYUX submitted 2026-01-08 cs.LG cs.CV

Yinghao Tang , Xueding Liu , Boyuan Zhang , Tingfeng Lan , Yupeng Xie , Jiale Lao , Yiyao Wang , Haoxuan Li
show 8 more authors
Tingting Gao Bo Pan Luoxuan Weng Xiuqi Huang Minfeng Zhu Yingchaojie Feng Yuyu Luo Wei Chen
This is my paper
classification cs.LGcs.CV
keywords infographicgenerationtext-to-imagereliabilitybenchmarkatomicquestionevaluationdataencodingcompletenessQ-ACCI-ACC
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

IGenBench introduces the first benchmark for text-to-infographic generation reliability, with 600 curated cases spanning 30 infographic types and 5,259 atomic yes/no verification questions. The paper's central claim is that current text-to-image models are not yet reliable enough for autonomous infographic creation: the best model answers 90% of individual questions correctly but fewer than half of its infographics pass every check at once (I-ACC 0.49), while the average model answers only 39% of questions correctly. The benchmark locates the main failure in data fidelity rather than visual appeal or layout: Data Completeness, Data Ordering, and Data Encoding are the lowest-scoring dimensions across every model tested. If the benchmark is right, it supplies a reusable, interpretable evaluation protocol and a concrete capability hierarchy for guiding future model development.

What carries the argument

The carrying mechanism is the atomic question set. Each of the 600 prompts is decomposed—with expert-informed augmentation for data completeness, data ordering, and data encoding—into self-contained yes/no questions organized under a 10-type taxonomy covering titles, chart type, data marks, axes, legends, annotations, decorative elements, and data-level properties. The evaluator scores each generated infographic with a strict binary rule where ambiguity or partial satisfaction counts as failure, and the two aggregation metrics Q-ACC and I-ACC turn those answers into interpretable reliability scores. The strict binary scoring and the gap between the two metrics are what make reliability measu

What would settle it

Have three or more human annotators answer every atomic question for a random sample of roughly 200 generated infographics (not averaged subsets), then recompute model-level Q-ACC, I-ACC, and dimension-level scores. If the automated evaluator's misses on Data Encoding are not roughly uniform across models, or if human judgments put a second-tier model's I-ACC well above 0.06, the specific tier boundaries and the data-fidelity-bottleneck conclusion would not survive.

Watch

Extended reading notes

Core claim

The central discovery is that infographic reliability can be decomposed into atomic, verifiable constraints and measured by two complementary scores: Q-ACC, the fraction of individual constraints satisfied, and I-ACC, the fraction of infographics that satisfy all constraints simultaneously. Under this decomposition, current text-to-image models form a three-tier hierarchy—Nanobanana-Pro at 0.90 Q-ACC / 0.49 I-ACC, Seedream-4.5 and GPT-Image-1.5 in a middle tier at 0.61 / 0.06 and 0.55 / 0.12, and all remaining models below 0.48 Q-ACC—with data-related dimensions as the universal bottleneck. The paper argues that high component-level accuracy coexists with long-tail failures: a single missed

Load-bearing premise

The benchmark's numbers stand or fall on the assumption that the chosen MLLM evaluator's binary answers to the atomic questions are valid ground truth; the paper's own error analysis shows the largest evaluator-human disagreement is on Data Encoding (12.12%), where the evaluator tends to miss violations, so if those misses are uneven across models the tier ranking and bottleneck ordering could shift.

Editorial extensions

If this is right

  • Even the strongest model fails more than half of full infographics, so human verification and post-editing remain necessary before generated infographics can be used in business, education, or journalism.
  • Data fidelity is the universal bottleneck: Data Completeness averages 0.21, Data Ordering 0.27, and Data Encoding 0.26 across all tested models, far below scores for titles, chart types, and decorative elements.
  • Q-ACC alone overstates readiness; I-ACC should be reported alongside it whenever end-to-end correctness matters, because a single missed constraint can invalidate an entire infographic.
  • Infographic reliability is a distinct capability from natural-image generation: model rankings on infographic checks do not always match rankings on general image-quality arenas.
  • The 5,259-question protocol offers a reusable, interpretable evaluation route for future models, with question-level scores identifying exactly which dimension to improve.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: with typically 7–11 checks per infographic, the observed 0.90 Q-ACC / 0.49 I-ACC pair is close to what independent per-question errors would predict (0.9^7 is about 0.48), suggesting the full-infographic failure rate is mostly a compounding-error effect rather than a separate failure mode.
  • Beyond the paper: the evaluator's known blind spot on Data Encoding (12.12% disagreement with humans, mostly missed violations) suggests a hybrid evaluation that adds programmatic geometric checks—comparing detected bar heights, circle areas, or polygon sizes to the stated values—would harden the benchmark and could change model ordering on the very dimension the paper identifies as the bottleneck
  • Beyond the paper: a testable consequence of the bottleneck claim is that training or fine-tuning on data-fidelity objectives should raise I-ACC faster than aesthetic-oriented improvements; comparing two models matched on Q-ACC but differing in encoding-specific training would test this directly.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper introduces IGenBench, a benchmark for evaluating the reliability of text-to-infographic generation. It curates 600 test cases from 42k real infographics across 30 chart types, decomposes reliability verification into 5,259 atomic yes/no questions over a 10-type taxonomy, and proposes Q-ACC and I-ACC metrics. Ten T2I models are evaluated with Gemini-2.5-Pro as the automated verifier. The main claims are a three-tier performance hierarchy (top model Q-ACC 0.90, I-ACC 0.49), universal data-related bottlenecks (Data Completeness 0.21, Data Encoding 0.26, Ordering 0.27), and the overall unreliability of current models for autonomous infographic generation. The paper also reports a human-LLM subset-average correlation of r=0.90 and a leakage sensitivity experiment in the appendix.

Significance. If the evaluator-validity concern is resolved, IGenBench would be a valuable contribution: it is the first dedicated benchmark for text-to-infographic reliability; the dataset construction is careful (real-world provenance, manual verification of extracted data and intent, diversity-aware sampling); the question taxonomy is interpretable; and the protocol is transparent, with prompts and case studies included in the appendix. The leakage sensitivity check in Appendix D.2 is a welcome addition, and the distinction between Q-ACC and I-ACC is well motivated. The reported findings, if robust, would give the community concrete model rankings and a diagnostic tool. However, the central quantitative conclusions currently rest on the validity of a single MLLM evaluator, and the evidence for that validity is not yet strong enough to support the specific tier boundaries and bottleneck ordering claimed in the paper.

major comments (4)
  1. [§5.3, Appendix D.3, Table 1] The evaluator is validated only through correlation of 100 subset averages (r=0.90). This does not validate individual yes/no judgments and can mask category- or model-dependent bias. Appendix D.3 reports the largest disagreement is on Data Encoding (12.12%), predominantly over-positive errors where Gemini-2.5-Pro misses encoding violations. Data Encoding is one of the headline bottlenecks in Table 1 (average 0.26), so such bias could change the ordering of data-related bottlenecks and could inflate Q-ACC gaps if the false-positive rate varies by model. Please report per-question confusion matrices by question category and by model, with the number of human-annotated questions, and show the sensitivity of Table 1's rankings to a correction based on this confusion matrix.
  2. [Table 1, §5.2] Table 1 reports Q-ACC and I-ACC as point estimates with no confidence intervals, standard errors, or significance tests. The three-tier hierarchy separates models by gaps as small as 0.07 (Seedream-4.5 0.61 vs GPT-Image-1.5 0.55 vs Nanobanana 0.48), and many I-ACC values are exactly 0.00. Because the central claims depend on these tier boundaries, I-ACC's sensitivity to a single erroneous question, and the absence of uncertainty are load-bearing. Add bootstrap 95% CIs over test cases and, where tier distinctions are claimed, paired model comparisons or at least overlap checks.
  3. [Appendix D.2, Table 1] The leakage experiment shows GPT-Image-1.5's Q-ACC drops from 0.52 on the original benchmark to 0.29 on infographics published after December 2025, a 44% relative drop. Yet Table 1 and §5.2 report GPT-Image-1.5 as second tier (Q-ACC 0.55, I-ACC 0.12) with no caveat. Since the model hierarchy is a central claim, either exclude or clearly flag leakage-affected scores in the main results and state which tier conclusions are robust to this contamination.
  4. [Appendix D.3, Figure 10] The error analysis reports disagreement rates (e.g., 12.12% for Data Encoding, 0% for Ordering) but does not state the sample size, the number of questions or images annotated, or whether the disagreements are computed on all 5,259 questions or a subset. The claim that 'categories such as Title & Subtitle, Data Completeness, and Ordering demonstrate near-perfect agreement' may be an artifact of a small denominator. Please provide the exact counts and, if the analysis is on a subset, the sampling procedure.
minor comments (6)
  1. [Table 1] The header row of Table 1 appears mangled (e.g., 'Comp./code', 'Enc./char', 'char◎-line', 'list-ul'). It should be replaced with explicit, readable question-type names (Data Completeness, Data Encoding, Data Ordering, etc.).
  2. [§5.2] The sentence 'Data Completeness shows the lowest average performance at 0.21, followed by Ordering (0.27) and Data Encoding (0.26)' is ordered incorrectly; the second-lowest is Data Encoding (0.26), then Ordering (0.27).
  3. [§2] The text refers to 'MatplotBench (Yang et al., 2024b)', but the cited reference is 'MatPlotAgent: Method and evaluation for LLM-based agentic scientific data visualization'. Please clarify whether MatplotBench is the same artifact or add the correct citation.
  4. [§4.1] The paper states that prompt-derived and expert-informed questions are generated by MLLMs but does not say whether the resulting question set is manually verified. Since the evaluation metrics depend entirely on these questions, please state whether human verification was performed on the generated questions and, if not, acknowledge this as a limitation.
  5. [Appendix D.2, Figure 9] In Figure 9, the y-axis label is not visible in the text/description, and the caption does not explain what the green and orange dots represent beyond 'original IGENBENCH' and 'recent 100 samples'. Clarify the axis and point semantics.
  6. [Full text title] The running title on page 1 reads 'THERELIABILITY' / 'TEXT-TO-INFOGRAPHICGENERATION'; these appear to be missing spaces due to typesetting. Fix for camera-ready.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: benchmark ground truth is fixed before evaluation and the MLLM evaluator is externally validated against human judgments.

full rationale

IGenBench's derivation chain is not circular. Ground truth is fixed before any T2I output is inspected: prompts are built from human-verified design descriptions and data tables (§3.2), and atomic questions are decomposed from those prompts plus fixed expert-informed seeds for Data Completeness, Ordering, and Encoding (§4.1). The reported Q-ACC and I-ACC are literal aggregations of binary evaluations of independently generated images against those pre-specified questions (Eqs. in §4.2); no quantity used in the metrics is fitted from the model outputs it is used to rank. The only same-family-LLM concern is that MLLMs synthesize questions and also answer them, but the paper breaks that loop with an external human-alignment check: Gemini-2.5-Pro answers are validated against human annotations on the same atomic questions (§5.3), and the evaluator is chosen among 12 MLLMs by correlation with human judgments (§5.4). That selection-on-validation procedure can be statistically optimistic, and Appendix D.3's 12.12% disagreement on Data Encoding is a validity caveat, but it is not a definitional or fitted-input circularity: human judgments are outside the pipeline, and the headline bottleneck findings are not encoded into the evaluator by construction. Related-work self-citations are contextual and do not carry the central argument. No step quoted in the paper reduces by construction to its own input.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

No new physical or conceptual entities are postulated. The free constants are benchmark-construction and validation hyperparameters; the axioms are the methodological presuppositions that a question set plus MLLM/human visual judgment measures ground-truth infographic correctness.

free parameters (5)
  • K (samples per chart type) = 5
    Appendix C.1 sets K=5; controls dataset size and diversity, but is inconsistent with the claimed 600 total across 30 types.
  • C (clusters per type) = 10
    Appendix C.1 sets C=10; used in k-means clustering before medoid selection.
  • Human-alignment subset size = 25
    Section 5.3 samples 25 generated infographics per resampling iteration to compute automatic-vs-human correlation.
  • Resampling iterations = 100
    Section 5.3 repeats the 25-sample evaluation 100 times before computing Pearson r.
  • Evaluator selection threshold = Pearson r > 0.8
    Section 5.4 uses this hand-set threshold to select Gemini-2.5-Pro as the automated evaluator.
assumptions (5)
  • domain assumption Atomic yes/no questions derived from prompt constraints and four seed requirements cover all aspects of infographic reliability.
    Section 4.1 defines Q(p)=Qp(p) ∪ Qe(p) and claims this taxonomy captures reliability; any untested failure mode is invisible to Q-ACC/I-ACC.
  • domain assumption The manually corrected MLLM-extracted data tables are correct ground-truth data for each prompt.
    Section 3.2 embeds 'The given data is: {data}' into each prompt; if extraction is wrong, generated images are judged against wrong specifications.
  • domain assumption Yes/no judgments of visual correctness can be made from static images by humans and MLLMs.
    Section 4.1 and Appendix D.3; the 12.12% disagreement on Data Encoding shows this assumption is only partially satisfied.
  • domain assumption Human annotator judgments are an unbiased reference for automatic-evaluator validation.
    Section 5.3 and Appendix B: three undergraduate CS students were recruited; no inter-annotator agreement is reported.
  • domain assumption Single-run API evaluations are stable enough to rank models without error bars.
    Table 1 reports one point estimate per model with no confidence intervals, API version snapshots, or reruns.

how reviews work

0 comments
Cite this review

Pith. "Pith review of IGenBench: Benchmarking the Reliability of Text-to-Infographic Generation." pith.science (2026). https://pith.science/paper/4S4KXYUX

@misc{pith2026260104498,
  author       = {Pith},
  title        = {Pith review of: IGenBench: Benchmarking the Reliability of Text-to-Infographic Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4S4KXYUX}},
  note         = {Machine review of arXiv:2601.04498}
}
read the original abstract

Infographics are composite visual artifacts that combine data visualizations with textual and illustrative elements to communicate information. While recent text-to-image (T2I) models can generate aesthetically appealing images, their reliability in generating infographics remains unclear. Generated infographics may appear correct at first glance but contain easily overlooked issues, such as distorted data encoding or incorrect textual content. We present IGENBENCH, the first benchmark for evaluating the reliability of text-to-infographic generation, comprising 600 curated test cases spanning 30 infographic types. We design an automated evaluation framework that decomposes reliability verification into atomic yes/no questions based on a taxonomy of 10 question types. We employ multimodal large language models (MLLMs) to verify each question, yielding question-level accuracy (Q-ACC) and infographic-level accuracy (I-ACC). We comprehensively evaluate 10 state-of-the-art T2I models on IGENBENCH. Our systematic analysis reveals key insights for future model development: (i) a three-tier performance hierarchy with the top model achieving Q-ACC of 0.90 but I-ACC of only 0.49; (ii) data-related dimensions emerging as universal bottlenecks (e.g., Data Completeness: 0.21); and (iii) the challenge of achieving end-to-end correctness across all models. We release IGENBENCH at https://igen-bench.vercel.app/.

Figures

Figures reproduced from arXiv: 2601.04498 by the authors.

Figure 1
Figure 1. IGENBENCH overview. et al., 2023; Deo et al.; Dibia, 2023) explore em￾bedding semantic context into infographics for bet￾ter aesthetics, while BizGen (Peng et al., 2025) advances text-rich infographics generation using layout-guided cross-attention mechanisms. With the rapid advancement of increasingly powerful T2I models like Nanobanana-Pro (Google, 2025), public interest in using them for infographic cre￾ation has… view at source ↗
Figure 2
Figure 2. A three-stage pipeline for constructing IG [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Statistical analysis of IGENBENCH. 4 Evaluation Protocol As shown in Stage 3 of [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (43 more)
Figure 4
Figure 4. Figure 4: (a) Correlation between automatic evaluation [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Case of Proportional Area Chart. between the two rankings and obtain ρ = 0.78 with p = 0.04. This moderate-to-strong positive correlation suggests that models with strong per￾formance on natural image generation often also perform well on infographic generation. At the…
Figure 6
Figure 6. Figure 6: Pearson correlation between different MLLMs’ automatic scores and human judgments. 5.5 Case Study As shown in [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Representative bad cases identified during manual quality review. Examples of samples excluded from [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: Performance breakdown by chart type. ination in the benchmark. We acknowledge that some degree of data leakage may exist for certain models, particularly those with more recent train￾ing cutoffs. To address this limitation, we plan to evolve IGENBENCH into a live bench…
Figure 10
Figure 10. Figure 10: Agreement and disagreement rates between [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: Case of Alluvial Diagram. Reference Nanobanana-pro Seedream Qwen-image 1.00 0.44 Area Chart PROMPT: Create an infographic that features a large title, "Foreign-Born", at the top center, with a subtitle, "SHARE OF U.S. POP￾ULATION", directly below it. The main visual i…
Figure 12
Figure 12. Figure 12: Case of Area Chart. Reference Nanobanana-pro Seedream Qwen-image 0.92 0.23 0.08 Bonus PROMPT: Create an infographic that features the title "AMERICA'S RICHEST SUBURBS" at the top, with an illustration of curren￾cy bills behind the word "RICHEST". Below the title is a …
Figure 13
Figure 13. Figure 13: Case of Bonus. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_13.png]
Figure 14
Figure 14. Figure 14: Case of Bubble Chart. Reference Nanobanana-pro Seedream Qwen-image 0.91 0.36 Bump Chart PROMPT: Create an infographic that displays a banner at the top left with the word 'RANKED' above the main title, 'A Decade of Major Currency Returns'. To the right of the main tit…
Figure 15
Figure 15. Figure 15: Case of Bump Chart. PROMPT: Create an infographic that compares two categories of investments, "Black Assets" and "Green Assets," using a mirrored horizontal bar chart layout split by a central vertical axis. The main title "Sovereign Wealth Funds" sits at the top, ab…
Figure 16
Figure 16. Figure 16: Case of Diverging Bar Chart. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_16.png]
Figure 17
Figure 17. Figure 17: Case of Donut Chart. PROMPT: Create an infographic that features a title and subtitle at the top, with a vertical bar element to the left of the title. Below the subtitle, a two-item legend with circular markers indicates two data series: 'By my own country' and 'By o…
Figure 18
Figure 18. Figure 18: Case of Dot Chart. PROMPT: Create an infographic that displays the main title "How Americans' FOREIGN POLICY PRIORITIES" and a subtitle "Changed Since 2018" at the top left, with an icon of a globe located in the upper right. The layout is divided into two main hori￾z…
Figure 19
Figure 19. Figure 19: Case of Dumbbell Plot. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_19.png]
Figure 20
Figure 20. Figure 20: Case of Grouped Bar Chart. PROMPT: Create an infographic that features the title 'Global Happiness Index 2023' and subtitle 'THE MOST & LEAST HAPPY COUNTRIES IN THE WORLD' at the top. A horizontal scale below the title is labeled 'LEAST' to 'MOST' with sad and happy e…
Figure 21
Figure 21. Figure 21: Case of Heatmap. PROMPT: Create an infographic that features the title 'THE BIGGEST SINGLE-DAY STOCK DECLINES' in large, stacked text at the top left. Below the title is a large downward-pointing arrow, followed by a line drawing of a crowd with speech bubbles con￾tai…
Figure 22
Figure 22. Figure 22: Case of Horizontal Bar Chart. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_22.png]
Figure 23
Figure 23. Figure 23: Case of Layered Area Chart. PROMPT: Create an infographic that features the title "Decoding the AI Economy Series" with "Part 3 of 3" at the top right. The main visual is a large line chart titled "U.S. vs. EU" comparing private AI investment, with a vertical axis for…
Figure 24
Figure 24. Figure 24: Case of Line Graph. Reference Nanobanana-pro Seedream Qwen-image 0.90 0.40 0.30 Lollipop Chart PROMPT: Create an infographic that displays information on a horizontal timeline axis positioned at the bottom. The main title, '2020's Biggest TECH Acquisitions', is locate…
Figure 25
Figure 25. Figure 25: Case of Lollipop Chart. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_25.png]
Figure 26
Figure 26. Figure 26: Case of Pictorial Chart. Reference Nanobanana-pro Seedream Qwen-image 0.50 0.88 0.50 Pie Chart PROMPT: Create an infographic that features the title 'Healthy Eating Tips' inside a triangle at the top center. The main visual is a large circle representing a plate, whic…
Figure 27
Figure 27. Figure 27: Case of Pie Chart. Reference Nanobanana-pro Seedream Qwen-image 0.64 0.36 0.18 Proportional Area Chart PROMPT: Create an infographic that features a title on the left reading 'The Most Visited Websites in the World' with a subtitle 'As of June 2024'. The main visualiz…
Figure 28
Figure 28. Figure 28: Case of Proportional Area Chart. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_28.png]
Figure 29
Figure 29. Figure 29: Case of Pyramid Chart. PROMPT: Create an infographic that features a central title block with the text 'TOP U.S. SUPPLIERS OF', followed by 'ALUMI￾NUM' and '& STEEL' on separate lines below. The infographic is structured around two opposing semi-circular radial charts…
Figure 30
Figure 30. Figure 30: Case of Radar Chart. PROMPT: Create an infographic that features a large, centered title, "Share holding pattern of AXIS BANK", at the top. Below the title is a semi-circular chart composed of five wedges of varying sizes, positioned in the lower half of the frame. Ea…
Figure 31
Figure 31. Figure 31: Case of Semicircle Donut Chart. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_31.png]
Figure 32
Figure 32. Figure 32: Case of Slope Chart. PROMPT: Create an infographic that displays a large title, "DEBT-TO-GDP RATIO", on the right side, with the text "of Advanced Economies" and a framed "2000 vs 2024E" below it. The main visual element is a horizontal bar chart on the left, listing …
Figure 33
Figure 33. Figure 33: Case of Span Chart. PROMPT: Create an infographic that features a large title, 'GERMANY'S GAS PIVOT', at the top left, with the subtitle 'Natural gas imports to Germany by source' positioned directly below it. To the top right, there is a text block that reads: 'The m…
Figure 34
Figure 34. Figure 34: Case of Stacked Area Chart. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_34.png]
Figure 35
Figure 35. Figure 35: Case of Stacked Bar Chart. PROMPT: Create an infographic that features a title, 'Das verdienen Bundestagsabgeordnete', at the top left, with a subtitle 'Abgeord￾netenentsch‰digungen der Bundestagsabgeordneten (in Euro)*' directly below. A centered legend below the sub…
Figure 36
Figure 36. Figure 36: Case of Stepped Line Graph. PROMPT: Create an infographic that features a large title and subtitle at the top left, with a vertical bar placed to the left of the title. The main visual element is a treemap chart that occupies the area below the header. This chart is c…
Figure 37
Figure 37. Figure 37: Case of Treemap. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_37.png]
Figure 38
Figure 38. Figure 38: Case of Vertical Bar Chart. PROMPT: Create an infographic that displays a large, central treemap-style area chart with the title 'Foreign Holders OF U.S. DEBT' in the upper left corner. Each segment of the treemap is sized proportionally to its value and contains a sm…
Figure 39
Figure 39. Figure 39: Case of Voronoi Treemap. PROMPT: Create an infographic that has a title 'Emmanuel Macron: Sieg ohne Triumph' and a subtitle 'Stimmanzahl beim zweiten Wahlgang der Pr‰sidentschaftswahl in Frankreich 2017' at the top. The main visual is a large waffle chart composed of …
Figure 40
Figure 40. Figure 40: Case of Waffle Chart. 28 [PITH_FULL_IMAGE:figures/full_fig_p028_40.png]
Figure 41
Figure 41. Figure 41: Instructions provided to human evaluators. [PITH_FULL_IMAGE:figures/full_fig_p029_41.png]
Figure 42
Figure 42. Figure 42: Prompt for chart type detection. 30 [PITH_FULL_IMAGE:figures/full_fig_p030_42.png]
Figure 43
Figure 43. Figure 43: Prompt used in T2I prompt construction. 31 [PITH_FULL_IMAGE:figures/full_fig_p031_43.png]
Figure 44
Figure 44. Figure 44: Prompt for question generation. 32 [PITH_FULL_IMAGE:figures/full_fig_p032_44.png]
Figure 45
Figure 45. Figure 45: Prompt for question type classification. [PITH_FULL_IMAGE:figures/full_fig_p033_45.png]
Figure 46
Figure 46. Figure 46: Prompt used for question augmentation from seed questions. [PITH_FULL_IMAGE:figures/full_fig_p034_46.png]
Figure 47
Figure 47. Figure 47: Prompt used in LLM evaluation. 35 [PITH_FULL_IMAGE:figures/full_fig_p035_47.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 5 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. ArtChart: Faithful Artistic Chart Generation with Integrated Text Rendering

    cs.CV 2026-07 conditional novelty 6.0 of 10

    ArtChart, a ControlNet + GRPO + multi-expert distillation system, achieves about 9.1/10 math, 9.5/10 text, and 7.7/10 layout on a new 2K bilingual artistic-chart benchmark, well above open baselines.

  2. sketch-plot: Progressive Editing for Text-to-Image Academic Figures

    cs.HC 2026-06 unverdicted novelty 6.0 of 10

    sketch-plot introduces a three-layer progressive editing pipeline with human-in-the-loop refinement for targeted modifications to text-to-image academic figures.

  3. DataMagic: Transforming Tabular Data into Data Insight Video

    cs.HC 2026-06 unverdicted novelty 5.0 of 10

    DataMagic generates narrative data videos from tabular data and queries via DVSpec declarative bindings and a Generate-then-Orchestrate multi-agent pipeline.

  4. Demonstrating chart-plot: Closing the Last Mile of Academic Chart Generation

    cs.HC 2026-06 unverdicted novelty 5.0 of 10

    chart-plot is an agentic harness using style-aware code generation from venue figures, a LaTeX-aware render-and-revise loop, and structured edit handles to produce top-venue-ready academic charts.

  5. SenseNova-U1: Unifying Multimodal Understanding and Generation with NEO-unify Architecture

    cs.CV 2026-05 unverdicted novelty 5.0 of 10

    SenseNova-U1 presents native unified multimodal models that match top understanding VLMs while delivering strong performance in image generation, infographics, and interleaved tasks via the NEO-unify architecture.

Reference graph

Works this paper leans on

9 extracted references · 3 linked inside Pith · cited by 5 Pith papers

  1. [1]

    Overall layout (e.g., title position, chart placement, grouping)

  2. [2]

    Chart type (e.g., pie chart, horizontal bar chart)

  3. [3]

    Data encoding (e.g., bars represent values, labels placed beside elements)

  4. [4]

    Text placement (titles, subtitles, labels, annotations)

  5. [5]

    an icon of a hand with coins on the right side

    Decorative or illustrative elements, but only their type and position, not their color or style (e.g., “an icon of a hand with coins on the right side” — without stylistic details) Output (JSON ONLY) Return a JSON object in the exact form: {{ "t2i_prompt": "<one- paragraph structural design description>" }} T2I Prompt Construction Figure 43: Prompt used i...

  6. [2018]

    Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi

    The effect of an infographic promotion on research dissemination and readership: a randomized controlled trial.Canadian Journal of Emergency Medicine, 20(6):826–833. Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. 2024. Playground v2. 5: Three insights towards enhancing aesthetic quality in text-to-image generation.arXiv p...

    arXiv 2024

  7. [2021]

    Infographicvqa.Preprint, arXiv:2104.12756. Meta AI. 2025. Llama 4: Multimodal in- telligence. https://ai.meta.com/blog/ llama-4-multimodal-intelligence/. MinimaxI. 2025. Image-01. https://www.minimaxi. com/news/image-01. Mistral AI. 2025. Mistral small 3.2. https://docs.mistral.ai/models/ mistral-small-3-2-25-06. OpenAI. 2025. Gpt-5. https://openai.com/gp...

    arXiv 2025

  8. [2023]

    In Proceedings of the IEEE/CVF International Confer- ence on Computer Vision, pages 20406–20417

    Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. In Proceedings of the IEEE/CVF International Confer- ence on Computer Vision, pages 20406–20417. Kaiyi Huang, Chengqi Duan, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. 2025. T2i- compbench++: An enhanced and comprehensive benchmark for compositional text-t...

    arXiv 2025

Show all 9 references
  1. [2024]

    Does the infographic feature the title ‘Years in MLS and Average Game Attendance, 2017’ at the top left?

    Haichart: Human and ai paired visualiza- tion system.Proceedings of the VLDB Endowment, 17(11):3178–3191. Yupeng Xie, Zhiyang Zhang, Yifan Wu, Sirong Lu, Jiayi Zhang, Zhaoyang Yu, Jinlin Wang, Sirui Hong, Bang Liu, Chenglin Wu, and Yuyu Luo. 2025b. VisJudge- Bench: Aesthetics ...

    2023

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.