InfoEdit: Probing Global Layout Reasoning in Infographic Editing

Cheng Yang; Chufan Shi; Huijuan Wang; Bo Shui; Yaokang Wu; Muzi Tao; Yibo Yan; Xuezhe Ma

cs.CVSep 27, 2026

InfoEdit: Probing Global Layout Reasoning in Infographic Editing

Authors: Cheng Yang, Chufan Shi, Huijuan Wang, Bo Shui, Yaokang Wu, Muzi Tao, Yibo Yan, Xuezhe Ma, +1 more

Organizations: University of California San Diego · University of Southern California · University of Illinois Urbana-Champaign · Carnegie Mellon University

Abstract

Multimodal foundation models edit natural photographs at production quality, yet the same models struggle with structured visual content such as infographics. Unlike photographs, infographics encode information through logical relations; editing one element often requires surrounding elements to be adapted. We refer to this global layout reasoning capability as reflow. Existing image-editing benchmarks neither provide a dedicated setting for structured visual content nor evaluate the reflow capability. We introduce InfoEdit, a novel benchmark of 1,000 infographics across eight logical-relation families, paired with 4,000 editing instructions across four editing tasks, and a reflow-aware evaluation protocol. Across eight frontier editors, only GPT-Image-2 clears 60% average success rate; most models fall below 7%, and no editor exceeds 36% on the Swap-Block task even with perfect target localization. We further show that code-level editing can match the strongest pixel-level editor, revealing complementary strengths across tasks. InfoEdit identifies reflow as a central challenge in structured visual content editing and provides a diagnostic benchmark to facilitate future progress.

Figures & tables

Figure 1: A typical “vibe-design” workflow with structured visual content. A user asks an image editor to add a missing step to an infographic. The instruction names one element, but the surrounding nodes and connectors must all reflow for the result to be usable. InfoEdit measures whether current editors can perform this reflow reliably.

Figure 2: Overview of InfoEdit . We curate $1{,}000$ infographics covering eight logical-relation families (left), annotate four editing tasks for each infographic (middle), and evaluate model outputs on two binary aspects, edit compliance and content preservation (right). Data examples are provided in Appx. A.3 .

Figure 3: Distributional summary of InfoEdit . (a) Distinct logical-relation families per infographic;

19\%

contain one,

81\%

contain two to six. (b) Fraction of infographics containing each family; an infographic can contain multiple families. (c) Source aspect-ratio distribution over eight ratios.

Benchmark	Task	Domain	#Samples	Logical Relation	Reflow- Aware	Evaluation Metric
Image Editing
ImgEdit Ye et al. (2025)	Image Editing	Natural	$811$	✗	✗	MLLM-judge
WiseEdit Pan et al. (2025)	Image Editing	Natural	$1{,}220$	✗	✗	MLLM-judge
GEditBench-v2 Jiang et al. (2026)	Image Editing	Natural	$1{,}200$	✗	✗	Pairwise MLLM
Infographics Generation
BizGen Peng et al. (2025)	Text-to-Image	Infographic	$1{,}000$	✗	–	MLLM + OCR

Table 1: Comparison of InfoEdit with related benchmarks. “–” indicates not applicable to text-to-image generation. The

4{,}000

samples come from

1{,}000

infographics with

4

edit tasks each.

	Expand-Text			Insert-Element			Swap-Block			Reshape-Canvas			Avg.
Model	EC	CP	SR	EC	CP	SR	EC	CP	SR	EC	CP	SR	Avg.
Proprietary
GPT-Image-2	99.0	86.7	85.9	85.4	80.7	72.0	29.0	79.2	25.5	96.3	65.1	64.4	62.0
Nanobanana-Pro	87.1	64.0	57.8	58.8	57.8	40.2	36.5	79.8	32.0	87.2	36.9	36.4	41.6
Nanobanana-2	85.1	58.6	52.6	65.3	49.4	39.0	30.0	72.1	24.0	92.0	36.4	36.3	38.0
Nanobanana	23.2	23.8	9.2	3.1	14.6	0.4	1.6	41.8	0.9	25.4	8.4	3.5	3.5

Table 2: Editing performance of eight image-editing models on InfoEdit across four editing tasks. EC = edit compliance, CP = content preservation, SR = success rate, the conjunction of both. All values are percentages. Avg. is the mean SR across tasks.

	Expand-Text			Insert-Element			Swap-Block			Reshape-Canvas			Avg.
Model	EC	CP	SR	EC	CP	SR	EC	CP	SR	EC	CP	SR	Avg.
Code
Gemini-3.1-Pro	85.7	72.3	67.4	85.6	70.7	68.0	68.8	77.6	59.6	69.4	45.2	45.1	60.0
Gemini-3.5-Flash	83.3	68.1	64.1	84.1	64.5	62.0	60.1	79.2	54.8	70.4	42.4	42.2	55.8
Gemini-3.1-Flash	79.6	65.5	60.6	63.4	49.1	40.2	56.1	72.7	47.1	38.0	5.3	5.2	38.3
Code + Image

Table 3: Code-level editing on InfoEdit . Code feeds the source code alone; code + image adds the rendered image. EC = edit compliance, CP = content preservation, SR = the conjunction of both. All values are percentages. Avg. is the mean SR across tasks.

Model	Hint	EC	CP	SR	$\Delta$ SR
GPT-Image-2	w/o	29.0	79.2	25.5	–
GPT-Image-2	w/	34.2	92.7	32.9	+7.4
Nanobanana-Pro	w/o	36.5	79.8	32.0	–
Nanobanana-Pro	w/	39.4	83.6	35.2	+3.2
Nanobanana-2	w/o	30.0	72.1	24.0	–
Nanobanana-2	w/	34.2	79.1	27.5	+3.5

Table 4: Effect of interactive selection hints on Swap-Block performance.

Figure 4: Success rates across difficulty levels for edit tasks: Expand-Text , Insert-Element , and Swap-Block .

Figure 5: Distribution of failure types per task. We categorize unsuccessful cases by the diagnostic sub-field they violate, separated into Edit Compliance (blue) and Content Preservation (purple). Detailed error cases in Appx. K .

Appendix figures & tables42 assets

Supplementary material from the paper’s appendix.

Appendix

Family	#Templates	Family	#Templates
List	16	Matrix	6
Process	16	Pyramid	6
Relationship	15	Hierarchy	4
Cycle	7	Picture	9
Total			79

Appendix

Table 5: Templates per logical-relation family in the logical template library. Each template is implemented in both HTML and PowerPoint.

Figure 6: List : Tab List and Trapezoid List.

Figure 7: Process : Timeline and Step Down Process.

Figure 8: Cycle : Basic Cycle and Basic Radial.

Figure 9: Hierarchy : Top-Down Hierarchy and Horizontal Hierarchy.

Figure 10: Relationship : Basic Venn and Nested Target.

Figure 11: Matrix : Basic Matrix and Titled Matrix.

Figure 12: Pyramid : Pyramid List and Inverted Pyramid.

Figure 13: Picture : Circle Picture Grid and Vertical Picture List.

Figure 14: Distribution of infographic topics across 31 categories in InfoEdit , demonstrating broad coverage of real-world domains.

Task	Magnitude Bucket	#Samples	Ratio	Difficulty
Expand-Text	5–10 words	285	28.50%	Easy
	11–15 words	302	30.20%	Medium
	16–20 words	205	20.50%	Medium
	21–25 words	99	9.90%	Hard
	26–30 words	109	10.90%	Hard
Insert-Element	1 element	170	17.00%	Easy

Appendix

Table 6: Difficulty-level statistics for editing tasks with explicit magnitude axes. Counts are reported over

1{,}000

instructions per task.

Figure 15: Example of the Expand-Text task in InfoEdit . The editor must lengthen the body text of a specified card while preserving its content. The expected edit reflows nearby elements to maintain readability and layout consistency, whereas the unexpected edit inserts the longer text without properly adjusting the local layout.

Figure 16: Example of the Insert-Element task in InfoEdit . The editor must insert new nodes into an existing logical structure. The expected edit places them at the specified positions and adjusts the surrounding structure, whereas the unexpected edit adds the nodes without sufficient reflow, resulting in an invalid local layout.

Figure 17: Example of the Swap-Block task in InfoEdit . The instruction asks the editor to exchange multiple logical blocks on the canvas. The expected edit swaps the specified blocks as whole units while preserving their internal content and adapting the surrounding layout; the unexpected edit only partially satisfies the requested swaps or damages the original layout structure.

Figure 18: Example of the Reshape-Canvas task in InfoEdit . The instruction asks the editor to re-render a portrait infographic as a landscape infographic. The expected edit globally reflows the content to fit the target aspect ratio while preserving the original information and visual style; the unexpected edit changes the canvas shape but fails to perform a complete global re-layout.

Figure 19: Interactive selection hint for Swap-Block . Matching numeric labels indicate the two blocks to be exchanged. The boxes are used only for target localization and should not appear in the edited output.

Output model	Samples	Agreement
GPT-Image-2	800	0.920
Nanobanana-2	800	0.934
HunyuanImage-3.0	800	0.981

Appendix

Table 7: Agreement between Gemini-3.1-Pro and the human-majority verdict across output models.

Axis	Subset	Agreement
Task	Expand-Text	0.920
	Insert-Element	0.905
	Swap-Block	0.925
	Reshape-Canvas	0.930
Difficulty	Easy	0.943
	Medium	0.912

Appendix

Table 8: Gemini-3.1-Pro agreement with the human majority on GPT-Image-2 outputs, stratified by task, difficulty, and predicted outcome.

Judge	GPT-Image-2	Nanobanana-2
Gemini-3.1-Pro	0.920	0.934
Claude-Opus-4.7	0.914	0.928

Appendix

Table 9: Agreement with the human-majority verdict for two MLLM judges.

Output model	Agreement over five runs
GPT-Image-2	$0.924\pm 0.007$
Nanobanana-2	$0.936\pm 0.008$

Appendix

Table 10: Repeated Gemini-3.1-Pro evaluation against human-majority labels.

Family	GPT-Image-2	Nanobanana-Pro	Nanobanana-2
List	78.9	45.3	40.5
Matrix	75.2	37.0	34.5
Process	77.1	47.3	45.0
Picture	75.5	36.5	35.2
Cycle	49.3	20.6	31.6
Pyramid	72.7	49.5	46.5

Appendix

Table 11: Insert-Element SR by host logical-relation family. All values are percentages.

		Expand-Text			Insert-Element			Swap-Block			Reshape-Canvas			Avg.
Model	Source	EC	CP	SR	EC	CP	SR	EC	CP	SR	EC	CP	SR	SR
GPT-Image-2	Overall	99.0	86.7	85.9	85.4	80.7	72.0	29.0	79.2	25.5	96.3	65.1	64.4	62.0
	HTML	99.1	86.3	85.5	85.8	80.3	71.9	29.5	78.9	25.9	96.6	64.6	64.0	61.8
	PPT	98.5	88.5	87.5	84.0	82.5	72.5	27.0	80.5	24.0	95.0	67.0	66.0	62.5
Nanobanana-2	Overall	85.1	58.6	52.6	65.3	49.4	39.0	30.0	72.1	24.0	92.0	36.4	36.3	38.0
	HTML	84.9	58.1	52.1	65.9	48.9	38.9	30.5	71.8	24.4	92.3	35.9	35.9	37.8

Appendix

Table 12: Pixel-level editing results by source format.

		Expand-Text			Insert-Element			Swap-Block			Reshape-Canvas			Avg.
Model	Source	EC	CP	SR	EC	CP	SR	EC	CP	SR	EC	CP	SR	SR
Gemini-3.1-Pro	Overall	88.2	72.0	68.5	85.3	72.2	71.0	69.0	78.4	60.2	71.8	47.2	46.8	61.6
	HTML	96.5	79.6	77.6	92.6	80.9	79.5	75.6	85.4	68.0	75.9	53.4	52.9	69.5
	PPT	55.0	41.5	32.0	56.0	37.5	37.0	42.5	50.5	29.0	55.5	22.5	22.5	30.1
Gemini-3.5-Flash	Overall	85.1	69.5	65.8	85.4	67.2	63.2	60.9	79.8	56.8	71.2	46.6	43.1	57.2
	HTML	93.1	76.9	74.5	92.8	75.6	70.8	66.8	86.9	64.1	75.1	52.8	48.8	64.5

Appendix

Table 13: Code-level editing with code and image input, broken down by source format.

Model	P $\rightarrow$ L	L $\rightarrow$ P	S $\rightarrow$ L	S $\rightarrow$ P
GPT-Image-2	69.1	59.4	72.5	67.5
Nanobanana-Pro	36.7	31.8	65.0	60.0
Nanobanana-2	41.2	32.0	35.0	37.5

Appendix

Table 14: Reshape-Canvas SR by transformation direction. P, L, and S denote portrait, landscape, and square.

		Expand-Text			Insert-Element			Swap-Block			Reshape-Canvas			Avg.
Model	Setting	EC	CP	SR	EC	CP	SR	EC	CP	SR	EC	CP	SR	SR
GPT-Image-2	Original	99.0	86.7	85.9	85.4	80.7	72.0	29.0	79.2	25.5	96.3	65.1	64.4	62.0
GPT-Image-2	Color-Change	98.2	86.2	84.8	84.2	81.3	74.2	31.0	81.2	27.2	98.2	67.1	66.8	63.3
Nanobanana-2	Original	85.1	58.6	52.6	65.3	49.4	39.0	30.0	72.1	24.0	92.0	36.4	36.3	38.0
Nanobanana-2	Color-Change	86.2	57.2	51.7	63.2	49.0	37.2	32.1	73.2	24.8	92.3	35.2	34.9	37.2

Appendix

Table 15: Color-palette control. All values are percentages.

		Expand-Text			Insert-Element			Swap-Block			Reshape-Canvas			Avg.
Model	Setting	EC	CP	SR	EC	CP	SR	EC	CP	SR	EC	CP	SR	SR
GPT-Image-2	Original	99.0	86.7	85.9	85.4	80.7	72.0	29.0	79.2	25.5	96.3	65.1	64.4	62.0
GPT-Image-2	Chinese	93.0	84.5	82.1	80.4	74.2	67.2	25.2	75.2	24.1	93.2	61.2	57.2	57.7
Nanobanana-2	Original	85.1	58.6	52.6	65.3	49.4	39.0	30.0	72.1	24.0	92.0	36.4	36.3	38.0
Nanobanana-2	Chinese	86.2	62.1	56.4	42.5	46.2	25.3	22.6	67.2	15.2	86.2	28.2	26.3	30.8

Appendix

Table 16: Preliminary Chinese translation pilot. All values are percentages.

Model Name	Type	Inference Interface
Proprietary models
GPT-Image-2	API	gpt-image-2
GPT-Image-1.5	API	gpt-image-1.5
Nanobanana-Pro	API	gemini-3-pro-image-preview
Nanobanana-2	API	gemini-3.1-flash-image-preview
Nanobanana	API	gemini-2.5-flash-image

Appendix

Table 17: Models used in our experiments and their inference interfaces. For proprietary models, we report the API model identifier used in our calls; for open-weight models, we report the corresponding Hugging Face repository.

Figure 20: Misplacement after Insert-Element. The three new circles were inserted into the Core Demographics radial diagram, but “Marine” is placed in the wrong position. It is incorrectly positioned on the opposite side of the diagram between “Military” and “Space”.

Figure 21: Style change after Insert-Element. The five new segments were inserted between “Cloud Ops” and “Mobile UI”, but they do not match the style of the existing segments. Instead of being integrated as arc segments within the main ring, they are rendered as smaller circular nodes branching outward below the cycle.

Figure 22: Incomplete content after Insert-Element. The Core Economic Drivers row is incomplete. There are only three of the four requested cards were inserted, while “Advanced Manufacturing” is missing between “Cleantech Energy” and “Silicon Engineering”.

Figure 23: Element loss after Expand-Text. The “Bulk pricing” was successfully expanded with the longer description, but the adjacent “Site delivery” card in the top-right of the Chapter 04 “Core strategy” grid has been removed. It makes the grid incomplete and disrupts the original 3×2 layout.

Figure 24: Style change after Expand-Text. The “4.5 Stories” caption was correctly expanded, but the “Broadcasting Spire” label at the top of “Structural Hierarchy” pyramid has changed from black text on a light background to white text. This breaks visual consistency with the original styling.

Figure 25: Element loss after Swap-Block. The swap was only partially executed. The “Transfer Cycle” was moved to the middle-left position previously occupied by “Core Academic Pillars” but the “Core Academic Pillars” is missing entirely instead of being relocated to the bottom-middle.

Figure 26: Style change after Swap-Block. The “Strategic” and “Global Integration” modules were swapped in position, but their original styling was not preserved. The ascending size progression of the roadmap steps is lost.

Figure 27: Structural break after Reshape-Canvas. The “Supply ecology” network diagram has been structurally altered. The connection topology between the nodes no longer matches the original.

Figure 28: Structural break after Reshape-Canvas. All original content was preserved when rendered as a portrait infographic, but the block ordering does not follow a coherent reading flow.

Figure 29: Style change after Reshape-Canvas. The infographic was re-rendered into a landscape canvas with all original content preserved, but the card styling has been altered.

Figure 30: Text rewritten after Swap-Block using Qwen-Image-Edit. The model partially changes the layout, but many original texts, labels, and chart annotations are rewritten or corrupted instead of being preserved. This shows that the model fails to maintain non-target content.

Figure 31: Prompt for MLLM-as-a-Judge evaluation on Expand-Text editing tasks.

Figure 32: Prompt for MLLM-as-a-Judge evaluation on Insert-Element editing tasks.

Figure 33: Prompt for MLLM-as-a-Judge evaluation on Swap-Block editing tasks.

Figure 34: Prompt for MLLM-as-a-Judge evaluation on Reshape-Canvas editing tasks.

Explore similar work

Is This Edit Correct? A Multi-Dimensional Benchmark for Reasoning-Aware Image Editing
Apr 16, 2026Yixuan Ding, Wei Huang, Ruijie Quan +2Image Editing Edit
Zhejiang University · The University of Hong Kong
PaintBench: Deterministic Evaluation of Precise Visual Editing
May 29, 2026Kai Xu, Ellis Brown, Shrikar Madhu +3Multi-Image Editing Benchmarks Image Editing
New York University
DrawAI: Agentic Benchmark and Workflow for Making Raster Images Editable
Aug 1, 2026Pu Cao, Qingye Kong, Xuedan Yin +5Agentic Workflow Design Agentic Benchmarks
Beijing University of Posts and Telecommunications · Tsinghua University · Beijing Institute of Technology

Apr 16, 2026, cs.HCJ/K move · Enter open · S save

Is This Edit Correct? A Multi-Dimensional Benchmark for Reasoning-Aware Image Editing

Yixuan Ding, Wei Huang, Ruijie Quan, Xiaojuan Qi+1

Zhejiang University · The University of Hong Kong

Image Editing Edit Multi-Turn Benchmark Visual Fidelity

Open paper AskPDF arXiv

Diffusion-based image editing has achieved strong visual fidelity under natural language instructions, yet most existing systems still operate at the level of surface instruction following, without reasoning about the implicit contextual constraints embedded in real user requests. This often leads to visually plausible but logically inconsistent edits. In this work, we introduce RE-Edit, a benchmark for REasoning-aware image Editing that evaluates image editing systems across five complementary reasoning dimensions: physical, environmental, cultural, causal, and referential. RE-Edit comprises 1,000 carefully curated samples, each designed such that visual plausibility alone is insufficient and correct editing requires satisfying implicit logical constraints. To support fine-grained analysis, we establish dimension-aligned evaluation criteria and conduct a comprehensive study of ten open-source and two commercial image editing models. Our results show that even advanced systems frequently struggle with implicit multi-dimensional reasoning despite producing high-quality visuals. We further present a lightweight reasoning-guided post-edit baseline as an initial exploration, illustrating how inserting explicit reasoning can help mitigate such failures in a model-agnostic manner.

Abstract

Explore similar work

Is This Edit Correct? A Multi-Dimensional Benchmark for Reasoning-Aware Image Editing

PaintBench: Deterministic Evaluation of Precise Visual Editing

DrawAI: Agentic Benchmark and Workflow for Making Raster Images Editable