InfoEdit: Probing Global Layout Reasoning in Infographic Editing

1 Introduction

• Existing image-editing benchmarks focus on natural photographs, leaving infographic editing and its required reflow capability underexplored.Reflow adapts the named element and implicitly affected elements so the image remains balanced, legible, and logically valid.Source: §1 · p. 1
• INFOEDIT covers eight logical-relation families, each with distinct reflow requirements such as rerouting arrows in a process or reparenting a hierarchy.The families are List, Process, Cycle, Hierarchy, Relationship, Matrix, Pyramid, and Picture.Source: §1 · p. 1
• The benchmark pairs 1,000 editable infographics with four editing tasks, yielding 4,000 instruction–infographic pairs that make reflow necessary.Tasks range from text-level changes to canvas-level restructurings.Source: §1 · p. 1
• INFOEDIT reports success only when the requested edit is executed and every other element remains intact.Its evaluation separates edit compliance from content preservation because a correct target edit can still silently damage unmentioned elements.Source: §1 · pp. 1–2
• Only GPT-Image-2 clears 60% average success rate, while no editor exceeds 36% on SWAP-BLOCK even with perfect target localization.Code-level editing reaches 61.6% average success rate, close to GPT-Image-2 at 62.0%, with complementary strengths across tasks.Source: §1 · p. 2
• INFOEDIT is designed to evaluate edits while preserving non-target content through reflow and to diagnose model capabilities and failure modes.The paper identifies placement, neighbor-aware adjustment, and global style and structure preservation as facets of reflow.Source: §1 · p. 2

2.1 Task Definition

• INFOEDIT models an edited infographic as an output that must execute the instructed change while preserving unmentioned elements, which may reflow.Source: §2.1 · p. 2

2.2 Infographics Construction

• The benchmark contains 1,000 editable infographics, built through a five-stage annotation pipeline and implemented in HTML or PowerPoint.The collection includes 800 HTML and 200 PowerPoint files.Source: §2.2 · p. 3, Figure 2 · p. 3
• 81% of infographics combine two to six logical-relation families, while the collection spans all eight families and landscape, portrait, and square canvases.Family annotations can overlap; List structures appear in 88% of infographics.Source: §2.2 · pp. 3–4, Figure 3 · p. 4

2.3 Editing Tasks

• The four tasks progress from lengthening text to inserting elements, swapping blocks, and reshaping the canvas, each requiring corresponding structural reflow.Instructions are checked for spatial coupling, shape-sensitive anchoring, and coverage across regions and relation families.Source: §2.3 · p. 4

2.4 Comparison with Existing Benchmarks

• Unlike natural-image editing and infographic-generation benchmarks, INFOEDIT targets infographic editing organized by logical-relation families and evaluated for reflow.Source: §2.4 · p. 4, Table 1 · p. 5

• Each edit is judged on two binary aspects: Edit Compliance and Content Preservation.Source: §3 · p. 5
• Task-specific rubrics assess compliance, while preservation allows necessary reflow such as repositioning, resizing, realignment, or rerouting connectors.Source: §3 · p. 5
• Success Rate counts an edit only when both Edit Compliance and Content Preservation are judged successful.Source: §3 · p. 5

4.1 Setup

• Eight proprietary and open-source models were evaluated on 4,000 infographic-instruction pairs, with outputs judged by Gemini-3.1-Pro and validated against expert annotators.Source: §4.1 · p. 5

4.2 Main Results

• 85.1–99.0% EC on EXPAND-TEXT and 87.2–96.3% EC on RESHAPE-CANVAS coexist with lower preservation, including 36.4–65.1% CP on RESHAPE-CANVAS.On SWAP-BLOCK, no model exceeds 36.5% EC or 32.0% SR, while CP can remain comparatively high.Source: §4.2 · pp. 5–6
• 62.0% average SR for GPT-Image-2 leads the benchmark, ahead of Nanobanana-Pro at 41.6% and Nanobanana-2 at 38.0%.The remaining five models score at or below 7%.Source: §4.2 · p. 5, Table 2 · p. 6
• 3.2%, 0.4%, and 0.0% average SR for HunyuanImage-3.0, Qwen-Image-Edit, and FLUX.2-klein-base-9B, respectively; all score 0% SR on three structural tasks.Those tasks are INSERT-ELEMENT, SWAP-BLOCK, and RESHAPE-CANVAS.Source: §4.2 · p. 6

5.1 Pixel-Level versus Code-Level Editing

• 61.6% average SR from Gemini-3.1-Pro with Code + Image nearly matches GPT-Image-2’s 62.0%, with complementary strengths across editing tasks.GPT-Image-2 is stronger on EXPAND-TEXT and RESHAPE-CANVAS; Gemini-3.1-Pro is much stronger on SWAP-BLOCK, raising EC from 29.0% to 69.0% and SR from 25.5% to 60.2%.Source: §5.1 · p. 6

5.2 Interactive Selection Hint

• Even with perfect localization, no model exceeds 36% SR on SWAP-BLOCK, and selection hints add only 3.2–7.4 SR across the three strongest editors.The modest gains indicate that target localization alone does not resolve the reflow operation’s difficulty.Source: §5.2 · pp. 6–7

5.3 Different Difficulty Levels

• Success generally falls as edit difficulty rises, with SWAP-BLOCK hardest even at the easy level and performance dropping as more block pairs are exchanged.EXPAND-TEXT degrades least; INSERT-ELEMENT drops more sharply, especially for weaker models, because it requires redistributing space while maintaining logical relations.Source: §5.3 · p. 7

5.4 Human Agreement and Judge Robustness

• Gemini-3.1-Pro achieves 0.920–0.981 agreement with expert majority verdicts across outputs from GPT-Image-2, Nanobanana-2, and HunyuanImage-3.0.Agreement remains high across task, difficulty, and predicted-outcome strata, and is stable under a cross-vendor judge and repeated evaluation.Source: §5.4 · p. 7

5.5 Error Analysis

• Misplacement accounts for 79% of SWAP-BLOCK failures and 42% of INSERT-ELEMENT failures, while EXPAND-TEXT errors mainly reflect preservation failures despite edit-compliance failures of ≤4%.RESHAPE-CANVAS failures concentrate on style change (47%) and structural break (38%); occlusion is the most common preservation failure on EXPAND-TEXT at 31%, but is 0% on SWAP-BLOCK and RESHAPE-CANVAS.Source: §5.5 · pp. 7–8

Image Editing Benchmarks

• Existing image-editing benchmarks evaluate natural-image edits or cognition- and creativity-informed edits, but generally preserve unmentioned regions unchanged.That pixel-identity assumption is reasonable for photographs but breaks down when structured content edits require layout changes, leaving reflow unevaluated.Source: §6 · p. 8

Structured Visual Content Generation

• BizGen, IGenBench, and BizGenEval assess structured-visual generation, whereas INFOEDIT studies editing existing infographics while preserving structure through reflow.The cited generation benchmarks address visual text rendering, text-to-infographic reliability, and commercial document generation, respectively.Source: §6 · p. 8

Conclusion

• INFOEDIT covers 1,000 infographics across eight logical-relation families and 4,000 editing tasks, with evaluation decomposing verdicts into edit compliance and content preservation.Across eight frontier editors, only GPT-Image-2 clears 60% average success rate.Source: §7 · p. 8
• No model exceeds 36% on SWAP-BLOCK even with perfect target localization, identifying reflow as the dominant bottleneck.The paper also examines code-level editing and conducts error analyses to provide diagnostic insights.Source: §7 · p. 9

Limitations

• INFOEDIT evaluates single-turn instructions, while real-world design workflows may involve iterative multi-turn changes to content, layout, and style.Extending reflow evaluation to interactive settings is identified as a promising direction.Source: Limitations · p. 9

A.1 Infographics Construction

• Existing infographic resources often lack editable sources and logical-relation diversity, limiting addressable editing tasks and systematic reflow evaluation.INFOEDIT therefore uses logical-relation templates and editable sources to support feasible, non-trivial edits that induce reflow.Source: §A.1 · p. 13

A.1.1 Stage 1: Topic Sourcing

• Topic sourcing covers 31 categories and yields 6,200 candidate topics through sequential contributions from Claude-Opus-4.7, Gemini-3.1-Pro, and GPT-5.4.The resulting lists are deduplicated within categories and shuffled before the next stage.Source: §A.1.1 · p. 13

A.1.2 Stage 2: Visual Reference Collection

• Two image generators produce reference infographics, and expert annotators select 1,000 for coherent palettes and topic-faithful content.The references guide color palette rather than supplying content or a layout to reproduce.Source: §A.1.2 · p. 13

A.1.3 Stage 3: Logical Template Construction

• The template library contains 79 templates across eight logical-relation families, each implemented in HTML and PowerPoint.Templates specify underlying logical structure, then receive infographic-specific style, color, spacing, and content.Source: §A.1.3 · pp. 13–14

A.1.4 Stage 4: Infographic Composition

• The 800-infographic HTML track uses model-assisted styling and content passes followed by annotator post-editing, while 200 PowerPoint infographics are authored manually.References serve as palette cues; instructions target infographic-level reflow rather than chart-internal facts such as values, axes, or trends.Source: §A.1.4 · p. 14

A.1.5 Stage 5: Quality Control

• A second annotator reviews each infographic for structural integrity, stylistic faithfulness, and content completeness.The checks include overflow or broken connectors, palette matching, and readable, coherent text.Source: §A.1.5 · p. 14

A.1.6 Topic Distribution

• The 1,000 infographics span 31 topic categories, and difficulty is based on edit magnitude for three tasks or aspect-ratio transformation for RESHAPE-CANVAS.Magnitude is measured by appended words, inserted elements, or exchanged block pairs for EXPAND-TEXT, INSERT-ELEMENT, and SWAP-BLOCK, respectively.Source: §A.1.6 · p. 14, §A.2.1 · p. 14

A.2.2 Addressable Tasks

• Instructions must identify feasible structural changes without removing content or breaking logical relations; reviewers also check clarity and non-trivial reflow.Task-specific constraints require expandable text, insertions that preserve semantics, movable whole-block swap targets, and canvas changes feasible through global re-layout.Source: §A.2.2 · pp. 15–17

A.3 Data Examples

• The four task examples contrast valid edits that preserve readability and logical validity with failures to complete the edit or reflow surrounding content.The examples cover expanding text, inserting elements, swapping blocks, and reshaping the canvas.Source: §A.3 · p. 18, Figure 15 · p. 19, Figure 16 · p. 20

B Additional Related Work

• Structured visual editing offers code-level and pixel-level pathways, with INFOEDIT comparing source-available and source-free editing rather than treating them as a like-for-like competition.The pathways show comparable overall performance but complementary task strengths: editable sources aid structural addressability, while pixel-level editing works when source files are unavailable.Source: B Additional Related Work · pp. 18–19

C Details on Interactive Selection Hint

• SWAP-BLOCK hints mark whole logical modules with matching numeric labels, helping models locate the pair to exchange without changing the editing instruction.The boxes and labels are auxiliary input and should not appear in the edited output.Source: C Details on Interactive Selection Hint · pp. 19–20, Figure 19 · p. 23

D Details on Code Generation Baselines

• For code-level editing, models modify infographic source from the instruction, then render the modified source to PNG for evaluation under the image editors’ MLLM-as-judge protocol.The code + image setting additionally supplies the original rendered infographic for visual grounding.Source: D Details on Code Generation Baselines · p. 20

E.1 Human Agreement across Output Models

• Gemini-3.1-Pro is compared with human-majority judgments on 800 outputs per model, with 200 examples per task and three expert annotators.The high agreement on HunyuanImage-3.0 accompanies outputs that are mostly clear failures for both annotators and judge.Source: §E.1 · p. 20, Table 7 · p. 22

E.2 Stratified Agreement

• GPT-Image-2 outputs have 0.940 inter-annotator agreement, and Gemini-3.1-Pro’s judge–human agreement remains high across task, difficulty, and predicted-outcome strata.The difficulty analysis excludes RESHAPE-CANVAS because it has no easy/medium/hard magnitude label.Source: §E.2 · p. 22, Table 8 · p. 22

E.3 Alternative Judge

• Gemini-3.1-Pro and Claude-Opus-4.7 obtain similar agreement with human-majority labels when evaluating the same model outputs under the same rubric and prompt.Source: §E.3 · p. 23, Table 9 · p. 23

E.4 Repeated Evaluation

• Repeated Gemini-3.1-Pro evaluations of the same 800 outputs per model show small agreement variation across five runs, which the authors say does not affect the reported comparison.Source: §E.4 · p. 23, Table 10 · p. 23

F.1 Logical-Relation Families

• GPT-Image-2 leads INSERT-ELEMENT across all host logical-relation families, while Cycle is hardest for all three editors.The authors relate Cycle’s difficulty to rerouting and reclosing a loop of connectors after insertion.Source: §F.1 · p. 23, Table 11 · p. 23

F.2 HTML versus PowerPoint Sources

• Pixel-level editors perform similarly on HTML- and PowerPoint-authored infographics, with their model ranking unchanged by source format.Source: §F.2 · p. 23, Table 12 · p. 24
• Both code-level models perform substantially better on HTML than PowerPoint when given source code and the rendered image.Source: §F.2 · pp. 23–24

F.3 Canvas Transformation Types

• All three evaluated models are weakest on landscape-to-portrait RESHAPE-CANVAS transformations.Source: §F.3 · p. 24

G.1 Color-Palette Sensitivity

• Replacing each infographic’s palette while preserving layout, structure, and content produces small, opposite-direction average-SR changes for the two models, without changing their ranking.Annotators verified that only the color scheme changed.Source: §G.1 · p. 24

G.2 Multilingual Pilot

• In a preliminary Chinese-language pilot, both models have lower average SR while their ranking remains unchanged.The authors describe the pilot as demonstrating extensibility, not as a complete multilingual evaluation.Source: §G.2 · p. 24

H Ethical Statement

• INFOEDIT is for research use; samples use publicly available information and are manually reviewed to avoid private, sensitive, defamatory, or misleading claims.Source: H Ethical Statement · p. 24

I Reproducibility

• The authors state that INFOEDIT and code to reproduce results will be released upon publication, and report exact experiment model identifiers for reproducibility.Source: I Reproducibility · p. 24

J Models

• The model set focuses on recent baselines that can render complete, legible text, excluding earlier image-editing models that do not meet this requirement.Source: J Models · p. 24

K Error Analysis

• The analysis tracks incomplete content, misplacement, element loss, occlusion, rewritten text, and style changes as distinct editing failure modes.These categories cover incomplete target rendering, misplaced elements, missing or cropped originals, layout collisions, altered non-target text, and changed visual styling.Source: K Error Analysis · p. 25
• Structural breaks include redirected process connectors, broken cycle closure, or skipped sequence indices; no re-layout treats a new canvas as a flat image.The analysis also categorizes changes to swapped-block content as Content Altered.Source: K Error Analysis · p. 26
• Qualitative examples show failures including misplaced insertions, missing elements, style changes, broken structure, and rewritten text.The examples span INSERT-ELEMENT, EXPAND-TEXT, SWAP-BLOCK, and RESHAPE-CANVAS.Source: K Error Analysis · p. 26, Figure 20 · p. 27, Figure 21 · p. 27

L.1 MLLM-as-a-Judge Evaluation Prompts

• The appendix provides separate MLLM-as-a-judge evaluation prompts for EXPAND-TEXT, INSERT-ELEMENT, SWAP-BLOCK, and RESHAPE-CANVAS.Source: §L.1 · p. 26, Figure 31 · p. 36, Figure 32 · p. 37