InfoEdit

Probing Global Layout Reasoning in Infographic Editing

Cheng Yang1,*, Chufan Shi2,*, Huijuan Wang2,*, Bo Shui3, Yaokang Wu4

Muzi Tao2, Yibo Yan2, Xuezhe Ma2, Taylor Berg-Kirkpatrick1

1University of California San Diego 2University of Southern California

3University of Illinois Urbana-Champaign 4Carnegie Mellon University

*Equal contribution

chy085@ucsd.edu · chufansh@usc.edu · huijuanw@usc.edu

Paper Code Dataset BibTeX

tl;dr

InfoEdit is the first benchmark for infographic editing, and it targets the capability existing benchmarks rarely measure: reflow. We evaluate both pathways: pixel-level image editors, and code-level editing over the HTML/PowerPoint source. Neither works well. The best pixel-level editor reaches 62.0% average success rate and the best code-level system (Kimi-K3) 67.9%, while most pixel-level editors fall below 7%.

A vibe-design workflow: a user asks an image editor to add a missing step to an infographic, and the surrounding nodes and connectors must reflow.
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.
01 — Overview

What InfoEdit measures

Given an original infographic X and an instruction I that names a structural change, an editor must produce X′ that satisfies two predicates at once: the named change is faithfully executed, and every element the instruction did not mention survives — free to reflow to accommodate the change.

CEdit Compliance

The change named by the instruction is executed on its target: the right element, at the right anchor, with its content fully rendered inside the canvas.

PContent Preservation

Every other element of the source survives intact. Repositioning, resizing, re-wrapping and connector rerouting are allowed — that is reflow, not failure.

SRSuccess Rate

An edit counts as a success only when both predicates hold. The decomposition exposes correct edits that silently damage unmentioned neighbors.

Overview of InfoEdit: dataset of eight logical-relation families, four editing tasks, execution pipeline, and MLLM-as-judge evaluation on edit compliance and content preservation.
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).
02 — Benchmark

1,000 infographics, eight logical relations

Every infographic is topically realistic, aesthetically credible, and — crucially — logically structured: it conveys information through visual logical-relation components. Each family imposes its own reflow contract that any valid edit must satisfy. Each infographic ships with an editable, structurally addressable source file: 800 in HTML/CSS and 200 in PowerPoint.

Example of the selected logical-relation family

List

Reflow contract.

Family annotations are not mutually exclusive: 81% of infographics compose two to six families on one canvas, so the shares above sum to more than 100%.

03 — Editing Tasks

Four tasks, increasing perturbation scope

From a text-level change to a canvas-level restructuring: Expand-Text, Insert-Element, Swap-Block and Reshape-Canvas. For each of the 1,000 infographics, annotators write one instruction per task — 4,000 pairs in total — and every instruction is verified to be feasible and reflow-inducing. Each example shows the source infographic, its instruction, and the expected versus an unexpected edited result; click any panel to open it full size.

InfoEdit data example: a Process infographic with its editing instruction and expected/unexpected edits InfoEdit data example: a multi-family city infographic with its editing instruction and expected/unexpected edits InfoEdit data example: a dark-themed infographic with its editing instruction and expected/unexpected edits InfoEdit data example: a Reshape-Canvas instruction with its expected and unexpected edited results
04 — Evaluation Protocol

Reflow-aware judging

For each (X, I, X′) triple we prompt an MLLM with the original image, the edited image and the instruction, using a task-specific rubric that reflects the edit's reflow contract. The judge returns a JSON verdict with binary edit compliance and content preservation fields, plus an overall verdict equal to their conjunction.

05 — Pixel-Level Editing

Eight frontier image editors

All models are queried with the 4,000 (infographic, instruction) pairs under each model's recommended inference settings, and every triple is judged by Gemini-3.1-Pro. EC = edit compliance, CP = content preservation, SR = success rate (their conjunction). All values are percentages; Avg. is the mean SR across tasks.

LowerHigher Shade shows each value's rank within its column. Click a task to sort by its SR.
Model Expand-Text Insert-Element Swap-Block Reshape-Canvas Avg.
EC CP SR EC CP SR EC CP SR EC CP SR
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
Seedream-5.0-Lite 22.1 28.0 9.1 12.3 11.0 1.3 9.3 37.0 5.2 64.1 4.6 4.3 5.0
Open-Weight
HunyuanImage-3.0 21.3 34.8 12.6 2.5 4.3 0.0 0.4 16.4 0.0 1.4 5.2 0.0 3.2
Qwen-Image-Edit 9.2 3.3 1.5 1.2 2.7 0.0 0.0 6.3 0.0 16.5 0.0 0.0 0.4
FLUX.2-klein-base-9B 6.1 7.2 0.0 0.0 21.2 0.0 0.0 77.8 0.0 20.3 0.0 0.0 0.0

Editing performance of eight image-editing models on InfoEdit. A clear capability gap separates the frontier proprietary models from the rest: only GPT-Image-2 clears 60% average SR, and all three open-weight models score 0% SR on Insert-Element, Swap-Block and Reshape-Canvas. High CP alone does not imply success — FLUX.2-klein-base-9B reaches 77.8% CP on Swap-Block with 0.0% EC, because the requested structural edit is never completed.

06 — Code-Level Editing

Editing the source instead of the pixels

Because every InfoEdit infographic ships with HTML or PowerPoint source, we can run code-level editing under the same protocol. Code feeds the model the source alone; Code + Image adds the rendered original as visual grounding. Gemini, GPT and Kimi models serve as code generators, and their edited outputs are rendered back into images.

LowerHigher Shade shows each value's rank within its column. Click a task to sort by its SR.
Model Expand-Text Insert-Element Swap-Block Reshape-Canvas Avg.
EC CP SR EC CP SR EC CP SR EC CP SR
Code
Kimi-K3 89.4 81.5 74.7 86.9 77.2 72.3 66.8 84.2 60.6 77.8 52.7 52.6 65.1
GPT-5.6-Sol 87.1 80.8 73.8 86.8 74.5 69.2 66.1 81.2 56.7 83.7 56.3 55.6 63.8
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
Kimi-K3 91.6 82.5 79.1 90.2 82.4 77.2 69.4 83.2 62.8 76.5 53.2 52.5 67.9
GPT-5.6-Sol 89.1 79.6 74.8 91.8 77.8 74.7 67.2 83.9 60.8 82.7 58.4 58.3 67.2
Gemini-3.1-Pro 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
Gemini-3.5-Flash 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
Gemini-3.1-Flash 80.2 64.5 59.6 64.3 51.4 43.0 56.4 74.0 50.2 38.1 8.7 8.6 40.4
Pixel-level reference
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

Code-level editing on InfoEdit. Kimi-K3 with Code + Image reaches 67.9% average SR, surpassing GPT-Image-2's 62.0% — but the task-level pattern is not uniform. GPT-Image-2 stays stronger on Expand-Text and Reshape-Canvas, while Kimi-K3 is stronger on Insert-Element and much stronger on Swap-Block (EC 29.0 → 69.4, SR 25.5 → 62.8), because source code makes named blocks directly addressable.

07 — Analysis

Where and why editors fail

Difficulty scales as designed

Success rates across easy, medium and hard difficulty levels for Expand-Text, Insert-Element and Swap-Block.
Success rates across difficulty levels. Degradation is mildest for Expand-Text, where top models absorb longer text by shrinking fonts or locally re-wrapping neighbors. Insert-Element drops sharply because adding multiple elements requires redistributing space and maintaining logical-relation consistency. Swap-Block is hardest at every level.

Where the failures concentrate

Distribution of failure types per task, split into edit compliance failures and content preservation failures.
Distribution of failure types per task for GPT-Image-2, the strongest editor evaluated. Misplacement dominates the structurally demanding tasks (79% of Swap-Block failures, 42% of Insert-Element failures); Expand-Text fails almost entirely on preservation (EC failures ≤ 4%, occlusion at 31%); Reshape-Canvas concentrates on global breakdown (47% style change, 38% structural break).

Is localization the bottleneck? No.

Swap-Block failures could stem from two causes: the model cannot execute the structural verb, or it cannot resolve which two blocks the instruction names. To separate them, we annotate the input image with interactive selection hints — each swap pair marked with a bounding box and a matching numeric label — and tell the model the marks are localization guidance that must not appear in the output.

Model Hint EC CP SR ΔSR
GPT-Image-2 w/o 29.0 79.2 25.5 —
w/ 34.2 92.7 34.9 +9.4
Nanobanana-Pro w/o 36.5 79.8 32.0 —
w/ 39.4 83.6 35.2 +3.2
Nanobanana-2 w/o 30.0 72.1 24.0 —
w/ 34.2 79.1 27.5 +3.5

Effect of interactive selection hints on Swap-Block. The gain is consistently modest (+3.2 to +9.4 SR), and even with perfect localization no model exceeds 35% SR. The bottleneck is the reflow operation itself, not target localization.

Original infographic next to a hinted version where the two blocks to swap are marked with magenta bounding boxes labelled (1).
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.
08 — Key Findings

What InfoEdit reveals

01

A capability cliff, not a gradient

Only GPT-Image-2 clears 60% average SR (62.0%); Nanobanana-Pro and Nanobanana-2 reach 41.6% and 38.0%; the remaining five models fall at or below 7%. Many models fail to produce usable infographics at all.

02

Swap-Block is essentially unsolved

No editor exceeds 32.0% SR on Swap-Block, and marking the swap targets with bounding boxes lifts SR by only +3.2 to +9.4 points: the bottleneck is reflow, not grounding.

03

Correct edits that silently break the layout

On Reshape-Canvas the top three models achieve 87–96% EC but only 36–65% CP. On Expand-Text, EC failures are ≤ 4% while most errors are preservation failures. Grading a single overall verdict would hide this entirely.

04

Code-level editing is a complementary pathway

Kimi-K3 over source code plus the rendered image reaches 67.9% average SR, surpassing GPT-Image-2's 62.0% — and more than doubles Swap-Block SR (25.5 → 62.8). The two pathways have different strengths, pointing toward hybrid systems.

Citation

BibTeX

If you find InfoEdit useful in your research, please consider citing:

@article{yang2026infoedit,
  title   = {InfoEdit: Probing Global Layout Reasoning in Infographic Editing},
  author  = {Yang, Cheng and Shi, Chufan and Wang, Huijuan and Shui, Bo and
             Wu, Yaokang and Tao, Muzi and Yan, Yibo and Ma, Xuezhe and
             Berg-Kirkpatrick, Taylor},
  journal = {arXiv preprint arXiv:2609.33286},
  year    = {2026}
}