acceptodds
Under review as a conference paper at ICLR 2027

One Figure, Every Canvas: Editable Flowchart Relayout via Agentic Pipeline

Abstract

Pipeline figures in ML papers must be repurposed across many canvases, includ- ing paper columns, 16:9 slides, portrait posters, 1:1 social teasers, 9:16 phone previews. Each format imposes a different aspect ratio on the same computational graph, where any silently broken connection misrepresents the method. We for- mulate aspect-ratio-adaptive flowchart relayout as a distinct task: given a raster flowchart and a target ratio, produce a structurally faithful, hallucination-free, ed- itable layout. Existing methods fail characteristically: image-to-image models stretch blocks and reject extreme ratios, text-to-image agentic systems hallucinate content, and parse-then-render systems mis-route edges. We propose an agentic pipeline factored into Parse, Style, and Layout stages, each pairing a main agent with a critic that combines deterministic constraint checks with VLM visual feed- back so connectivity is explicitly checked and prevented from being silently bro- ken. Outputs are draw.io-editable mxGraph XML. On a curated benchmark of 100 flowcharts at five aspect ratios, evaluated by Gemini 3.1 Pro and validated against human judgments, our method reaches 68.6% Content Fidelity versus 11.2–41.4% for prior work.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
flowchart relayout, scientific figure editing, agentic AI, vision-language models, structured diagram understanding, layout generation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Ta7FOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies aspect-ratio-adaptive flowchart relayout: transforming an existing raster diagram to a target canvas while preserving its content, connectivity, and visual style, and producing editable draw.io XML. The proposed pipeline separates parsing, style reconstruction, and layout adaptation, with stage-specific critics combining deterministic checks and visual feedback. The authors introduce FlowchartRelayoutBench, consisting of 100 source diagrams evaluated at five target aspect ratios. The method achieves 68.6% Content Fidelity compared with 41.4% for the strongest evaluated baseline. Ablations and a human evaluation further support the usefulness of the decomposition and iterative refinement.

Critical strengths
  1. The paper addresses a practical problem: adapting scientific flowcharts to different canvases while preserving their meaning. Editable draw.io output further increases its usefulness by supporting subsequent manual refinement.

  2. The Parse–Style–Layout decomposition is well motivated and substantially improves Content Fidelity over the evaluated baselines. Ablations and human evaluation provide supporting evidence for the effectiveness of iterative refinement.

Critical weaknesses
  1. The benchmark contains only 100 source diagrams, with 500 tasks obtained by evaluating each diagram at five aspect ratios; these tasks should not be treated as 500 independent samples. The ablations use just 30 diagrams. This limits confidence in the robustness and generalizability of the reported gains, particularly the smaller differences between ablation variants.

  2. For a benchmark contribution, the coverage of evaluated methods remains limited. Although Table 7 includes Qwen-based experiments, broader comparisons with open-weight code-generation models and general-purpose agent harnesses such as Codex, using comparable tools and inference budgets, are needed to better establish the proposed pipeline’s advantage.

Questions for authors
  1. How were the 30 ablation diagrams selected, and which target aspect ratios were evaluated? Could you expand the ablations and report confidence intervals that account for tasks sharing the same source diagram, together with repeated-run variability, to establish whether the observed differences are stable?

  2. Could you evaluate additional open-weight code-generation models and general-purpose agent harnesses with rendering and iterative revision under comparable budgets? Please also clarify the subset, aspect-ratio coverage, and refinement opportunities used in the existing Qwen experiments.

6 comments
AuthorsOpenReview ↗

Response to Reviewer Ta7F (Part 1/3)

Thank you for recognizing the practical value of editable flowchart relayout and our Parse–Style–Layout design. We address benchmark scale and generalizability, model coverage, and ablation reliability below.

W1: Benchmark Size and Generalizability

Our dataset and generalization coverage leave room for expansion, but our evaluation scale is comparable to recent work. AutoFigure-Edit [1] uses 200 method descriptions, of which 100 each are under the settings with and without a style reference; SlideTailor [2] uses 50 target papers for its main experiments; Paper2Poster [3] is evaluated on a benchmark of 100 paper–author-made poster pairs. For the Parse and Style stages, our evaluation covers 100 distinct source flowcharts; for the Layout stage, we set five target aspect ratios for each diagram, forming 500 relayout tasks with different canvas constraints to test the ability to adapt across ratios. We agree that tasks derived from the same source are dependent, so the number of distinct source diagrams remains 100. At this scale, Table 2 still shows advantages in structure/content preservation and Overall score, while Table 4 supports each module’s contribution.

Our ablation scale is also comparable: SlideTailor [2] randomly selects 30 cases for its quantitative ablation to analyze the contribution of each module. We likewise use 30 flowcharts, selecting 10 each from the three structural complexity levels (easy, medium, and hard), so that every difficulty level has equal weight in the evaluation, in order to quantify the impact of the different modules on the relayout results.

[1] Z. Lin et al., “AutoFigure-Edit: Generating Editable Scientific Illustrations via Reference-Guided Styling,” ACL System Demonstrations, pp. 57–67, 2026.

[2] W. Zeng et al., “SlideTailor: Personalized Presentation Slide Generation for Scientific Papers,” AAAI, 40(41), pp. 34584–34592, 2026.

[3] W. Pang et al., “Paper2Poster: Towards Multimodal Poster Automation from Scientific Papers,” NeurIPS 38, 2025.

W2 & Q2: Additional Comparisons with General-Purpose Agent Harnesses and Open-Weight Code-Generation Models

  1. Open-Weight Code-Generation Models

    Thank you for suggesting broader comparisons with open-weight models. In addition to the existing Qwen3.6-27B experiments, we evaluate GLM-5.3-flash to examine the applicability of our pipeline across models. Both models are evaluated on the same subset of 25 flowcharts, comprising 15 Easy, 7 Medium, and 3 Hard cases. This subset emphasizes easy cases while also covering medium and hard cases. Each flowchart is evaluated at one target aspect ratio, with five cases each for 1:1, 2:3, 3:2, 9:16, and 16:9. We compare one-shot direct XML generation with our three-stage pipeline. Both Qwen and GLM use the same iteration limits as our method, with a maximum of 3, 2, and 5 iterations for the Parse Stage, Style Stage, and Layout Stage, respectively. Rendered outputs and Critic feedback guide refinement until the output passes the checks or reaches the iteration limit.

    MethodRel.↑Hallu.-free↑Content↑Layout↓Style↓Overall↑
    Ours (GLM-5.3-flash)60.072.040.01.281.3666.0
    Direct VLM-to-XML (GLM-5.3-flash)52.044.020.01.641.6434.0

    Using the same GLM model, direct XML generation is more prone to missing connections and failing to faithfully preserve complex shapes and images. Through iterative checking and refinement, our method improves Relationship Preservation from 52.0% to 60.0%, Hallucination-free Rate from 44.0% to 72.0%, and Content Fidelity from 20.0% to 40.0%. Direct generation also more frequently produces inconsistent flow directions, excessive spacing between components, content concentrated on one side, and node or background colors that differ from the source figure. Our method mitigates these issues, improving the average Layout and Style ranks from 1.64 to 1.28 and 1.36, respectively.

    The additional GLM experiments demonstrate that our method can also achieve effective content preservation and aspect-ratio adaptation with an open-weight model. Compared with the full pipeline using Qwen3.6-27B in Table 7, using GLM improves Relationship Preservation from 20.0% to 60.0% and Content Fidelity from 20.0% to 40.0%. Inspection of the outputs shows that Qwen produces better layouts primarily on 1:1 canvases, whereas GLM preserves content more completely and produces reasonable layouts across different target aspect ratios. These results indicate that the lower performance previously observed with Qwen does not represent a general limitation of our method on open-weight models, further supporting its applicability across different open-weight models.

(Continued in Part 2: W2/Q2—General-Purpose Agent Harnesses.)

AuthorsOpenReview ↗

Response to Reviewer Ta7F (Part 2/3)

(Continued from Part 1: W1, and W2/Q2—Open-Weight Code-Generation Models.)

  1. General-Purpose Agent Harnesses

    We further compare our method with general-purpose agent harnesses, using Codex as a representative, on the same 30 flowcharts from the ablation subset, alongside one-shot Gemini 3.1 Pro and one-shot GPT-5.5. Codex uses gpt-6.1-sol with reasoning effort set to medium and has access to exec_command, view_image, apply_patch, and write_stdin for image inspection, code execution, and XML editing. It can also invoke the draw.io Desktop CLI to render the results. A single Codex agent processes all 30 flowcharts in one conversation. The entire batch takes approximately 10.52 minutes, including image inspection, code development, and tool execution, and uses a total of 2,991,501 tokens. Both the runtime and token usage are totals for the full batch. The three baselines use the same task prompt, shown below.

    You are a visual layout transformation agent specializing in flowchart relayout.
    
    Task:
    Read the input flowchart image and rearrange it to fit the target canvas dimensions (canvasWidth and canvasHeight). Preserve its nodes, shapes, text, connectivity, arrow directions, branching relationships, and visual style.
    Preserve the original reading order, groups, hierarchy, parallel flows, and important relative positions where possible, while adapting the layout to the target aspect ratio.
    
    Output:
    Return only a complete, editable draw.io mxfile XML document. Represent nodes as editable vertices and connections as editable edges. Set pageWidth and pageHeight to the specified canvas dimensions. Do not include Markdown or explanatory text.
    
    [Detailed XML format requirements omitted here for brevity.]
    
    MethodRel.↑Hallu.-free↑Content↑Layout↓Style↓Overall↑
    Ours76.796.776.71.571.2792.78
    One-shot Gemini 3.1 Pro43.350.033.33.572.6735.56
    One-shot GPT-5.533.380.030.03.272.5036.67
    Codex (GPT-6.1 Sol, Medium)46.756.730.02.533.5735.00

    Note: The Layout and Style ranks are computed only among the four methods in this table, so these two metrics, and the Overall score that depends on them, cannot be directly compared with the values in Table 4 of the paper.

    The results show that the Content Fidelity of our method is 76.7%, higher than the 30.0% of Codex and the 30.0% and 33.3% of the two one-shot baselines, which shows that our method preserves the structure and content of the flowchart more completely. From the outputs, Codex is better at cropping and embedding the images in the original diagram, and compared with the two one-shot baselines, fewer cases of images being regenerated or failing to embed occur. However, its nodes and connectors tend to adopt a uniform, template-like style, for example light fill colors, rounded nodes, and solid arrows with rounded bends, so it preserves the individual styles of the original diagram to a limited extent; its average Style Similarity rank is 3.57, behind the other three methods. The three baselines also often produce crossing or overlapping connectors, or connectors passing through nodes; their average Layout Quality rank ranges from 2.53 to 3.57, all behind the 1.57 of our method. The Overall score of our method reaches 92.78, also higher than the 35.00 to 36.67 of the three baselines.

(Continued in Part 3: Q1.)

AuthorsOpenReview ↗

Response to Reviewer Ta7F (Part 3/3)

(Continued from Part 2: W2/Q2—General-Purpose Agent Harnesses.)

Q1: Ablation Design and Statistical Reliability

  1. Subset Selection and Aspect-Ratio Coverage

    The 30 flowcharts used in the ablation study are selected by structural complexity, with 10 diagrams from each difficulty level (easy, medium, and hard), so that every difficulty level has equal weight in the evaluation, in order to fairly compare the contribution of each module across different structural complexities. Each diagram is evaluated at one target aspect ratio, covering 1:1, 2:3, 3:2, 9:16, and 16:9, with 7, 6, 8, 4, and 5 diagrams, respectively, so as to cover square, portrait, and landscape canvases.

  2. Repeated-Run Variability and Statistical Robustness

    To test whether our results depend on a single generation or a single evaluation, we conduct two kinds of repeated experiments on the same 30 ablation flowcharts. Each source diagram in this ablation subset corresponds to only one target aspect-ratio task, so there is no dependence arising from multiple tasks derived from the same source diagram. Each experiment is repeated three times; below we report the mean ± standard deviation, together with 95% confidence intervals computed with the t distribution (df = 2). These intervals reflect the run-to-run variability on a fixed subset and do not cover the uncertainty from data sampling.

    First, we fix the generated outputs of all 10 settings and re-evaluate them independently three times with the VLM-as-a-Judge, to measure the variability of the evaluation itself.

    VariantOverall Score95% CIContent Fidelity95% CI
    Ours84.8 ± 1.6[80.8, 88.8]70.0 ± 0.0[70.0, 70.0]
    w/o All Critic Agents47.7 ± 2.4[41.7, 53.6]50.0 ± 5.8[35.7, 64.3]
    w/o Parse Critic58.9 ± 2.3[53.3, 64.5]51.1 ± 5.1[38.5, 63.8]
    w/o Layout Critic56.4 ± 3.3[48.1, 64.7]54.4 ± 1.9[49.7, 59.2]
    w/o Style Stage74.7 ± 2.4[68.8, 80.6]68.9 ± 3.8[59.3, 78.5]
    Direct Prompt VLM-to-XML (Gemini 3.1 Pro)34.7 ± 2.1[29.5, 39.9]35.6 ± 3.8[26.0, 45.1]
    Direct Prompt Layout Stage65.4 ± 3.0[58.0, 72.9]61.1 ± 3.8[51.5, 70.7]
    Direct Prompt VLM-to-XML (GPT-5.5)34.3 ± 4.4[23.3, 45.4]23.3 ± 5.8[9.0, 37.7]
    Direct Prompt Parse Stage46.8 ± 4.1[36.7, 56.9]41.1 ± 10.2[15.8, 66.4]
    w/ Graphviz (DOT) Layout11.1 ± 0.6[9.5, 12.7]13.3 ± 0.0[13.3, 13.3]

    The results show that the standard deviation of the Overall Score over the three repeated evaluations ranges from 0.6 to 4.4, and our method still has the highest mean score, indicating that the overall scoring variability of the evaluation is small and that the main ablation conclusions remain supported under repeated evaluation.

    In addition, for our method only, we regenerate three times with the full pipeline and feed each generated output separately into the evaluation, to measure the combined variability of generation and evaluation. The table below places this result alongside the re-scoring result on the fixed outputs above.

    ExperimentOverall Score95% CIContent Fidelity95% CI
    Ours — Re-scoring84.8 ± 1.6[80.8, 88.8]70.0 ± 0.0[70.0, 70.0]
    Ours — Re-generation and Evaluation83.8 ± 3.4[75.4, 92.2]67.8 ± 1.9[63.0, 72.6]

    The results show that after regeneration and evaluation, the standard deviation of the Overall Score of our method is 3.4 and that of Content Fidelity is 1.9 percentage points, and the mean performance is close to that of re-scoring the fixed outputs. Even when the variability of both generation and evaluation is included, the results still fluctuate relatively little, supporting the repeatability of our method's performance on the current subset.

Reviewer_bWX1OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses a practical authoring pain point: a landscape method diagram designed for a two-column paper layout often needs to be relaid out when placed into a portrait poster, a mobile screen, or a square promotional image. Direct resizing makes text too small, stretching distorts shapes, and regenerating the figure with a generative model may drop modules or misconnect arrows. To address this, the authors formulate aspect-ratio-adaptive, structure-preserving flowchart relayout as a distinct task and propose a three-stage Parse–Style–Layout pipeline: Parse converts the raster figure into an editable structure, Style restores the original visual appearance, and Layout rearranges the components for the target canvas ratio while preserving connectivity. Experiments show that the pipeline outperforms existing baselines on structural fidelity, content fidelity, and visual quality, indicating that this task decomposition is effective and well targeted.

Critical strengths
  1. The paper defines aspect-ratio-adaptive flowchart relayout as a distinct task and requires editable draw.io XML output, which is practically useful and clearly differentiated from image-to-image resizing, text-to-image agentic regeneratio, and parse-then-render pipelines. The problem is well motivated, and the output format aligns with real downstream editing workflows.

  2. The paper proposes a sensible and empirically supported pipeline. The three stage (Parse, Style, Layout) decomposition is well motivated, with each stage targeting a specific source of error, and the use of deterministic checks and critic feedback is reasonable. The main experiments and user study consistently support the effectiveness of this design, particularly on content fidelity and relationship preservation.

Critical weaknesses
  1. The paper claims that “connectivity is explicitly checked and prevented from being silently broken” for several times. Table 4 shows that removing all Critic Agents reduces Content Fidelity from 76.7% to 46.7%, supporting the benefit of critic feedback. However, it does not show whether outputs receiving a “passing judgment” are structurally correct. I recommend independently checking both the graph structures before and after critic feedback against source-graph ground truth on a representative subset. Errors can be categorized into incorrect source or target connections, wrong flow directions, duplicate edges, etc. And then report both the reduction in each error category after critic feedback and the remaining errors in the final output. Also, critic-approved outputs should be separated from outputs that terminate only because they reach the maximum iteration limit. Otherwise, an apparent quality gain may partly reflect additional generation and revision budget rather than reliable error detection.

  2. After replacing only the Graphviz layout module, substantial decreases are observed across not only Layout and Overall Score but also other metrics, so the degradation is not limited to layout-specific metrics. This result supports the conclusion that the proposed layout works better than the Graphviz replacement within the current full pipeline, but it does not show that the layout algorithm alone caused the performance drop. To strengthen the attribution, I recommend comparing different layout algorithms after fixing the parsed nodes, edges, labels, containers, and canvas constraints, and reporting where structural, content, and rendering errors are introduced.

Questions for authors

See Critical Weaknesses

6 comments
AuthorsOpenReview ↗

Response to Reviewer bWX1 (Part 1/3)

We thank the reviewer for the constructive review and for recognizing our task formulation, the editable draw.io XML output, and the well-motivated Parse–Style–Layout decomposition supported by our experiments and user study. Below, we address the concerns regarding the reliability of Critic feedback and connectivity verification, and the controlled comparison of layout algorithms.

W1: Reliability of Critic Feedback and Connectivity Verification

  1. Error Reduction and Critic Approval Analysis

    To examine whether Critic feedback effectively corrects errors, and whether outputs that pass the check truly preserve the structure of the original diagram, we use the original image as the reference and manually check the issues and corrections across the rounds of outputs for the 30 flowcharts in the ablation study. According to the Critic stopping status of the Parse Stage and the Layout Stage, we count, separately for the outputs that passed the check and those that reached the iteration limit, the number of cases with fully correct node relationships and with remaining errors.

    StageCritic StatusCasesRelationship CorrectRelationship Incorrect
    ParsePassed26260
    ParseIteration Limit413
    LayoutPassed20182
    LayoutIteration Limit1073

    The results show that for all 26 diagrams that passed the Parse Stage check, the node relationships were manually confirmed to be correct; among the 20 diagrams that passed the Layout Stage, 18 were fully correct. Notably, the relationship error rates of outputs that passed the Critic check were 0% and 10% for the Parse and Layout Stages, respectively, compared to 75% and 30% for those that reached the iteration limit, suggesting that the Critic effectively identifies outputs with a higher likelihood of containing errors. Among the outputs that reached the iteration limit, 1 diagram in the Parse Stage and 7 diagrams in the Layout Stage have correct node relationships, which shows that failing the check may also stem from overlap, alignment, or readability issues. In addition, 2 diagrams passed the Layout Stage Critic check but still contain relationship errors compared with the original image, because the Parse Stage had already parsed the relationships incorrectly and the Layout Stage preserved them. Since the Layout Stage uses the result of the Parse Stage as its reference for checking, these pre-existing errors are not judged as relationship changes caused by relayout.

    We also count the issues reported by the Critics of the Parse Stage and the Layout Stage during generation for the 30 ablation flowcharts, and manually compare the outputs before and after correction to confirm which issues were corrected and which remain. The numbers in the table below are issue counts; the same issue appearing in multiple rounds is counted only once.

    StageIssue CategoryReported IssuesCorrected IssuesRemaining Issues
    ParseEndpoint Mismatch26242
    ParseReversed Flow Direction220
    ParseMissing or Extra Branches990
    ParseContent and Visual Representation22211
    ParseElement Geometry and Edge Routing847113
    LayoutElement Overlap and Readability766412
    LayoutEdge Routing and Arrow Placement867016
    LayoutElement Arrangement and Canvas Utilization33285

    The results show that, of the 338 issues reported by the Critics, 289 were manually confirmed as corrected, covering edge endpoints, flow direction, branches, element overlap, and edge routing, which shows that the Critics' targeted feedback helps correct many types of errors during generation.

(Continued in Part 2: W1—Generation and Revision Budget Controls.)

AuthorsOpenReview ↗

Response to Reviewer bWX1 (Part 2/3)

(Continued from Part 1: W1—Error Reduction and Critic Approval Analysis.)

  1. Additional Generation and Revision Budget Controls

    To clarify whether the quality improvement comes only from additional generation and revision opportunities, we add two comparison settings that do not use Critic feedback. The first is Best-of-N, which generates 3, 2, and 5 candidate results at the Parse, Style, and Layout stages, respectively, and then lets a vision-language model select the result that best meets the task requirements of that stage to pass to the next stage. In addition to candidate generation, this setting also includes additional candidate-selection calls. The second is sequential revision without Critic, which runs 3, 2, and 5 rounds for Parse, Style, and Layout, respectively, matching the iteration limit of each stage in the full method; in each round, the result of the previous round is provided, and the generation model decides on its own whether a modification is needed, without any additional Critic feedback.

    MethodRel.↑Hallu.-free↑Content↑Layout↓Style↓Overall↑
    Ours (Full Pipeline)76.796.776.71.601.7066.67
    Best-of-N Selection (w/o Critic Feedback)73.386.770.01.871.7752.50
    Sequential Self-Revision (w/o Critic Feedback)73.376.760.02.532.5330.83

    The results show that the Content Fidelity of the full method is 76.7%, higher than the 70.0% of Best-of-N and the 60.0% of sequential revision without Critic; its Hallucination-free Rate is also higher by 10 and 20 percentage points, respectively. The full method also achieves better Relationship Preservation, Layout Quality, and Style Similarity, and its Overall score reaches 66.67, higher than the 52.50 and 30.83 of the two comparison settings. These results show that, under this setting, merely increasing the number of candidates or providing sequential revision opportunities still does not reach the performance of the full method, which supports the additional benefit of the Critic's targeted feedback.

(Continued in Part 3: W2.)

AuthorsOpenReview ↗

Response to Reviewer bWX1 (Part 3/3)

(Continued from Part 2: W1—Generation and Revision Budget Controls.)

W2: Controlled Comparison of Layout Algorithms

Thank you for pointing out the substantial decline in Graphviz DOT’s content-related metrics, which prompted us to re-examine our evaluation procedure. Since the outputs of these layout algorithms almost never matched the specified target aspect ratios, we centered the results and added padding to fit the target canvases. However, the padded outputs occupied only a small portion of the canvas, making text, symbols, connections, and style details difficult to distinguish during evaluation. To reduce this effect, we now provide the unpadded layout outputs alongside the source flowcharts to the VLM-as-a-Judge for evaluating Relationship Preservation, Hallucination-free Rate, and Style, and update Content Fidelity and Overall Score accordingly. Layout is still evaluated on the outputs placed within the specified target canvases to reflect actual canvas utilization and layout quality.

To further compare layout algorithms, we additionally evaluate ELK Layered from the Eclipse Layout Kernel framework. We reuse the same Parse Stage and Style Stage outputs, fixing the input nodes, edges, labels, containers, and target canvas constraints, and replace only the Layout Stage. All comparisons use the same 30 flowcharts from the ablation subset. All three methods follow the evaluation procedure described above, and Style ranks are recomputed within this comparison.

MethodRel.↑Hallu.-free↑Content↑Layout↓Style↓Overall↑
Ours76.796.776.71.401.0395.0
w/ Graphviz (DOT) Layout53.383.350.02.372.8331.7
w/ ELK Layered Layout56.780.053.32.432.0745.0

The lower Relationship Preservation of both alternatives primarily stems from overlapping, crossing, or excessively routed connections after relayout, which obscure the relationships between nodes. In addition, related elements within containers, or cuboids composed of multiple faces, may be separated during relayout, disrupting their original grouping and visual meaning. The separated components appear as extraneous content, while the original complete objects are no longer preserved, resulting in lower Hallucination-free Rates. For Style, Graphviz DOT requires mapping XML styles to DOT attributes, but some shape and style details cannot be represented directly and equivalently. ELK Layered retains the original XML styles and maps the layout results back into the XML, but separating composite elements can still alter their overall appearance. The Layout gap mainly arises from adaptation to the target canvas. Both alternatives almost never produced layouts matching the specified target aspect ratios, leaving substantial whitespace after padding.

In contrast, our method directly edits XML containing position, grouping, and style information, and inspects rendered outputs for iterative refinement. It can therefore better preserve composite elements, arrange connections, and utilize the target canvas. Our method achieves an average Layout rank of 1.40, outperforming Graphviz DOT at 2.37 and ELK Layered at 2.43, supporting its advantage in adapting layouts to specified canvases.

Reviewer_7LofOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper formulates aspect-ratio-adaptive flowchart relayout as a distinct task for adapting pipeline figures across different canvases. The authors propose an agentic pipeline with Parse, Style, and Layout stages, each guarded by a critic combining deterministic checks and VLM feedback to preserve structural connectivity.

Critical strengths
  1. Addressing the practical need for cross-canvas figure adaptation with editable XML output provides immediate value to the community, distinguishing this work from generic image generation tasks.
  2. Combining deterministic constraint checks with VLM reasoning in the critic effectively prevents silent connectivity failures, which is the core limitation of prior pure-VLM approaches.
Critical weaknesses
  1. Heavy reliance on VLM raises potential bias risks; if human validation was limited, the reported performance gap may overstate true superiority over baselines.
  2. Testing only on polished paper figures limits confidence in generalizability to messy drafts, non-standard notations, or domain-specific diagrams encountered in real-world use.
Questions for authors
  1. What are the exact sample size and correlation statistics between human judgments and VLM scores, specifically for connectivity correctness?
  2. How does the system handle non-standard visual elements or parsing failures outside the curated benchmark distribution?
  3. What is the average token cost per relayout?
5 comments
AuthorsOpenReview ↗

Response to Reviewer 7Lof (Part 1/2)

We thank the reviewer for the constructive review and for recognizing the practical value of editable XML output for cross-canvas figure adaptation, and our Critic design that combines deterministic checks with VLM reasoning to explicitly check connectivity and provide targeted feedback. Below, we address the concerns regarding VLM evaluation bias and human validation, generalization to non-standard inputs and parsing failures, human–VLM agreement on connectivity correctness, and token cost.

W1: VLM Evaluation Bias and Human Validation

  1. Human Validation Protocol

    We agree that our current user study does not yet cover all backgrounds and generalization scenarios, but its scale is comparable to recent work. PaperBanana [1] likewise validates its VLM-as-a-Judge on 50 cases. The human evaluation in our work covers 50 cases, each with five candidate outputs, for a total of 250 outputs. Relationship Preservation and Hallucination-free Rate are first annotated independently by two annotators; if their judgments disagree, a third annotator adjudicates. In addition, we recruited 30 participants with computer science backgrounds, who completed a total of 600 Layout Quality and Style Similarity evaluations, and we aggregated the rankings by averaging.

  2. Strictness of VLM Judgments

    To check whether the VLM is more lenient than human annotators, we compare the Relationship Preservation judgments method by method. In the table below, "VLM Stricter" means the human judged pass but the VLM judged fail; "VLM More Lenient" means the human judged fail but the VLM judged pass.

    MethodAgreementVLM StricterVLM More LenientAgreement or Stricter
    Ours3613149/50 (98%)
    Nano Banana Pro454149/50 (98%)
    GPT Image 2.0418149/50 (98%)
    AutoFigure-Edit423545/50 (90%)
    PaperBanana442446/50 (92%)
    All Methods2083012238/250 (95.2%)

    Among the 250 outputs, the VLM agrees with or is stricter than the human judgment in 95.2% of cases, and only 12 outputs (4.8%) are judged more leniently by the VLM, of which only one belongs to our method. For Hallucination-free Rate, the proportion of agreement or stricter is 90.8% (227/250), and 100% (50/50) for our method. These results suggest that, on this evaluation set, the VLM does not show a tendency to be more lenient toward our method.

  3. Human–VLM Comparison of Performance Gaps

    To directly test whether the VLM overestimates our advantage over the baselines, we compare the human and VLM pass rates of our method and the strongest baseline, Nano Banana Pro, on the same 50 cases.

    MetricMethodHuman Pass RateVLM Pass Rate
    RelationshipOurs74%50%
    RelationshipNano Banana Pro50%44%
    HallucinationOurs100%98%
    HallucinationNano Banana Pro80%74%

    In the human evaluation, our method outperforms Nano Banana Pro by 24 percentage points in Relationship pass rate, compared with only 6 percentage points in the VLM evaluation. For Hallucination, the gaps are 20 and 24 percentage points, respectively. Both evaluations favor our method over Nano Banana Pro. Specifically, for connectivity correctness—the focus of your concern—the VLM reports a smaller performance gap than human evaluators, rather than overstating our advantage.

[1] D. Zhu, R. Meng, Y. Song, X. Wei, S. Li, T. Pfister, and J. Yoon, “PaperBanana: Automating Academic Illustration for AI Scientists,” arXiv preprint arXiv:2601.23265, 2026.

(Continued in Part 2: W2/Q2, Q1, and Q3.)

AuthorsOpenReview ↗

Response to Reviewer 7Lof (Part 2/2)

(Continued from Part 1: W1.)

W2 & Q2: Generalization to Non-standard Inputs and Parsing Failures

The main goal of this work is to relayout already-drawn scientific flowcharts to different canvas aspect ratios, so our evaluation data mainly consist of diagrams that fit this use case. Recognizing and reconstructing messy sketches is not the primary goal of this work, and we have not designed a dedicated mechanism for such inputs. For non-standard symbols and domain-specific visual elements that appear as images or complex icons, the Parse Stage crops and stores them, lets them participate in the layout as IMAGE placeholders, and re-embeds them after the layout is completed to preserve the original visual content.

For parsing errors, the Parse Stage is equipped with a Critic Agent that checks the source image, XML, and rendered output for consistency and provides feedback for iterative refinement until the output passes the checks or reaches the iteration limit. On the current benchmark, Relationship Preservation reaches 74.2%, meaning that 74.2% of the final outputs fully preserve all node relationships in the entire diagram. In the failure cases we observed, the problems usually involve only a few local connections. They mainly stem from blocks in the original image lacking a clear arrow direction or ordering, and from many small, overlapping arrows whose start and end points are harder to identify, which leads to local mis-connections or wrong directions in the parsing result. Moreover, our output is editable draw.io XML, so users can correct errors directly in the GUI. Since most failure cases involve only a few local connections, they can usually be fixed in a short time.

Q1: Sample Size and Human–VLM Agreement on Connectivity Correctness

For the agreement analysis between human and VLM judgments of connectivity correctness, we use 50 cases, each with five candidate outputs, for a total of 250 outputs. Each output is first judged independently by two annotators against the original image on whether all node relationships are fully preserved; if their judgments disagree, a third annotator adjudicates. This forms the final human annotation, which is then compared with the VLM's binary judgment.

Section 4.4 and Table 3 of the paper already report the correlation statistics between human and VLM binary judgments of Relationship Preservation, including the Phi coefficient φ = 0.671, the 95% confidence interval [0.560, 0.753], p < 0.001, and an agreement rate of 83.2% (208/250). To present the details of the judgments, we add the following cross-tabulation of human and VLM judgments.

VLM: relationships correctVLM: relationships incorrect
Human: relationships correct9430
Human: relationships incorrect12114

Q3: Average Token Cost per Relayout

We select 5 flowcharts from each of the three structural complexity levels (easy, medium, and hard), 15 in total, to cover different numbers of nodes and structural complexity. A complete run on each diagram requires 15 API calls on average and takes 21 minutes 57 seconds.

StageMean TimeStd DevAvg. API CallsAvg. TokensAvg. Cost (USD)
Parse Stage5m 18.2s3m 7.8s3.265,959$0.4610
Style Stage3m 50.4s2m 33.8s3.658,756$0.343
Layout Stage12m 48.4s5m 1.1s8.2150,731$1.1403
Overall21m 57s7m 30.8s15.0275,446$1.9452

The cases in which the Layout Stage takes longer are flowcharts with many grouping relations, or flowcharts whose logical groups are determined only by brackets or semantics. The Style Stage takes longer when there are many blocks or gradient colors. The Parse Stage is very fast on text-only flowcharts, but takes longer on flowcharts with rotated text or many groups. Overall, flowcharts that are not text-only, that contain many semantically grouped blocks, or that have unexpected text directions and gradient-colored objects are the factors that affect the overall relayout efficiency.

AuthorsOpenReview ↗

Follow-up: Our Response to Your Review

Dear Reviewer 7Lof,

As the discussion period ends soon, we wanted to follow up. Our response addresses your concerns via (1) a human-VLM agreement analysis on 250 outputs (φ = 0.671, 83.2% agreement, full cross-tabulation), showing that the VLM is stricter than human annotators on our outputs (13 stricter vs. 1 more lenient) and reports a smaller Relationship Preservation gap over the strongest baseline than humans do (6 vs. 24 points) (Response Part 1/2; Part 2/2: Q1); and (2) a per-stage breakdown of token cost, API calls, and runtime per relayout (about 275K tokens, $1.95, and 15 API calls on average; Part 2/2: Q3). We also discuss handling of non-standard elements and parsing failures (Part 2/2: W2/Q2). Please let us know if anything remains unclear. Thank you!

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.