cs.CVOct 5, 2026
SaveOne Figure, Every Canvas: Editable Flowchart Relayout via Agentic Pipeline
Organizations: Department of Computer Science, National Yang Ming Chiao Tung University · Siebel School of Computing and Data Science, University of Illinois Urbana-Champaign
Abstract
Pipeline figures in ML papers must be repurposed across many canvases, including 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 formulate aspect-ratio-adaptive flowchart relayout as a distinct task: given a raster flowchart and a target ratio, produce a structurally faithful, hallucination-free, editable 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 feedback so connectivity is explicitly checked and prevented from being silently broken. 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. Project page: https://onefigureeverycanvas.vercel.app/
Figures & tables
Figure 1: Aspect-ratio-adaptive flowchart relayout. Given a single flowchart at one fixed ratio (left) , our method produces structure-preserving relayouts at arbitrary aspect ratios (right) : phone, tablet, social, slide, and banner. Three different source figures (top to bottom) demonstrate generality across architecture diagrams, experimental pipelines, and process flows. Outputs are not re-rendered images but draw.io-editable XML, ready for further manual refinement.
Figure 2: Three categorical failure modes on flowchart relayout. Given input (a) at 1:1 and a request for 3:2, we examine three families. (b) Image-to-image models (e.g., Nano Banana Pro) resize in pixel space, stretching embedded frames and blocks, and they also reject aspect ratios outside 3:1–1:3. (c) Text-to-image agentic systems (e.g., PaperBanana) re-synthesize from a prompt, dropping blocks present in (a) and adding labels that were not. (d) Parse-then-render systems (e.g., AutoFigure-Edit, SciFig) recover components but mis-route edges, breaking the computational graph. Red annotations mark each failure. None achieves structural fidelity, hallucination-freedom, and arbitrary resizing simultaneously. These are the three properties our task requires.
| Method | Style Control | Canvas Resizing | Editable Format |
|---|---|---|---|
| Image-to-image models | |||
| Nano Banana Pro | Image reference | (~3:1 to 1:3) | × |
| GPT Image 2.0 | Image reference | (3:1 to 1:3) | × |
| Text-to-image / agentic systems | |||
| SciFig ( Huang et al., 2026b ) | Prompt (Hard) | × | × |
| PaperBanana ( Zhu et al., 2026a ) | Prompt (Hard) | (~3:1 to 1:3) | × |
Table 1: Capability comparison with prior work. We compare against three families of figure-generation approaches along three axes for paper-figure relayout: how style is specified, what canvas ratios are supported, and whether outputs are editable in a standard tool. Only ours achieves all three: image-reference style control, arbitrary aspect ratios, and draw.io compatibility. = full support, = partial support, ×= none.
Figure 3: Overview of our agentic relayout pipeline. Given input flowchart (raster, here 4:3) and target aspect ratio (here 16:9), our method produces draw.io-editable XML in three stages, each pairing a main agent with a critic that iterates until convergence. (a) Parse converts into structured XML by combining SAM 3 instance segmentation (bounding boxes for every component) with a VLM-based parse agent. (b) Style produces styled XML from and ; we extract dominant colors directly from ’s pixels with PIL, since VLM color perception is unreliable. (c) Layout extracts inter-block relationships from and rearranges components into canvas via DAG-based hierarchical layering; the critic verifies no relationship is broken and no same-level components newly overlap. Output preserves the structure, style, and content of while fitting the new aspect ratio.
Figure 4: Iterative critic refinement across pipeline stages. (a) The Parse Stage combines deterministic checks and VLM feedback to progressively correct out-of-bound components and edge errors, yielding the parsed XML . (b) The Style Stage transfers colors, typography, and corner radius, while critic feedback corrects residual font and shape mismatches. (c) The Layout Stage applies DAG-based hierarchical relayout and iteratively resolves boundary, overlap, and routing issues, producing the final layout .
Figure 5: Editable XML output in draw.io. (a) The generated can be opened directly in draw.io, where each block, label, connector, and style attribute remains editable. (b) Users can modify fonts, borders, colors, and component positions independently, while explicit source/target references preserve connector relationships after editing.
| Correctness (%) | Visual Quality | |||||
| Method | Relationship | Hallu.-free | Content | Layout | Style | Overall |
| Image-to-image | ||||||
| Nano Banana Pro | 49.8 | 69.4 | 41.4 | 2.85 | 1.35 | 71.10 |
| GPT Image 2.0 | 50.2 | 75.4 | 40.2 | 3.17 | 2.08 | 71.10 |
| T2I / agentic | ||||||
| PaperBanana ( Zhu et al., 2026a ) | 38.4 | 18.6 | 11.2 | 2.14 | 4.34 | 39.05 |
Table 2: Quantitative comparison on FlowchartRelayoutBench . Five methods across three families are evaluated under Correctness, Visual Quality, and Overall, a hierarchical score prioritizing Content Fidelity over Visual Quality. Our method achieves the best Correctness metrics and Overall score (81.85, 1.15 the next-best). PaperBanana ranks best on Layout despite weak correctness, while Nano Banana Pro and GPT Image 2.0 rank well on Style by largely preserving the original appearance. Bold = best, underline = second-best, italic = third-best.
Figure 6: Qualitative comparison under different target aspect ratios. We compare our method with four baselines at three target ratios: 1:1, 2:3, and 3:2. GPT Image 2.0 typically compresses or crops the original; Nano Banana Pro attempts relayout more often but tends to break relationships. PaperBanana rotates horizontal layouts to fit portrait canvases, and AutoFigure-Edit may introduce hallucinated elements or inconsistent relationships. Ours preserves relationships, avoids hallucinations, uses the canvas effectively, and produces editable XML. Please zoom in for details.
Figure 7: User study results. We report human evaluation results for content-related metrics and visual-quality metrics. Our method achieves the best performance on Relationship Preservation and Hallucination-free Rate, and shows strong pairwise win rates in Layout Quality and Style Similarity.
| Metric | Statistic | Corr. | Agreement | 95% CI |
|---|---|---|---|---|
| Relationship | 0.671 | 83.2% | [0.560, 0.753] | |
| Hallucination | 0.593 | 82.0% | [0.486, 0.693] | |
| Layout | 0.620 | – | [0.513, 0.714] | |
| Style | 0.757 | – | [0.704, 0.807] |
Table 3: Agreement between VLM-as-a-Judge and human evaluation. We report Phi coefficient ( ) and classification agreement for binary Relationship Preservation and Hallucination-free judgments, and Spearman’s rank correlation ( ) for Layout Quality and Style Similarity. All correlations are significant at .
| Method | Rel. | Hallu.-free | Content | Layout | Style | Overall |
|---|---|---|---|---|---|---|
| Direct prompting / decomposition | ||||||
| Direct Prompt Parse Stage | 53.3 | 60.0 | 33.3 | 5.00 | 7.67 | 50.74 |
| Direct Prompt Layout Stage | 70.0 | 93.3 | 66.7 | 4.07 | 5.03 | 69.26 |
| Direct Prompt VLM-to-XML (Gemini 3.1 Pro) | 43.3 | 50.0 | 33.3 | 5.23 | 6.87 | 41.11 |
| Direct Prompt VLM-to-XML (GPT-5.5) | 33.3 | 80.0 | 30.0 | 4.97 | 7.77 | 49.63 |
| Stage / Critic removal | ||||||
Table 4: Quantitative ablation study on 30 flowcharts. We compare direct prompting, Stage/Critic removal, and Graphviz DOT replacement. Layout, Style, and Overall are computed jointly within the ablation and are therefore not directly comparable with Table 2 . Bold = best, underline = second-best, italic = third-best.
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
| Correctness (%) | Visual Quality | ||||||
|---|---|---|---|---|---|---|---|
| Complexity | Method | Relationship | Hallu.-free | Content | Layout | Style | Overall |
| Easy [2pt]Nodes: 10.1 3.4 Edges: 9.4 3.3 | Nano Banana Pro | 62.4 | 77.6 | 56.0 | 2.78 | 1.54 | 69.40 |
| GPT Image 2.0 | 66.4 | 83.2 | 56.8 | 3.01 | 2.09 | 70.00 | |
| PaperBanana | 52.0 | 25.6 | 18.4 | 2.36 | 4.50 | 33.20 | |
| AutoFigure-Edit | 61.6 | 48.8 | 39.2 | 3.62 | 3.88 | 45.00 | |
| Ours | 93.6 | 98.4 | 92.8 | 2.46 | 2.86 | 80.60 | |
Table 5: Performance across source-diagram structural complexity. Results are reported on the easy, medium, and hard subsets of FlowchartRelayoutBench . Overall follows the hierarchical scoring in Table 2 . Our method achieves the highest Content Fidelity and Overall score at all three complexity levels. Bold = best, underline = second-best, and italic = third-best within each level.
| Complexity | Parse | Style | Layout | Total |
|---|---|---|---|---|
| Easy | ||||
| Medium | ||||
| Hard |
Table 6: Refinement Iterations by Structural Complexity. Average refinement iterations (mean std.) for the Parse, Style, and Layout stages and the full pipeline. The maximum iteration budgets are 3, 2, and 5, respectively.
Figure 8: Direct Prompt Layout Stage. Direct prompting leaves substantial unused space and adapts to the target canvas by excessively shrinking components. Our method makes more effective use of the available canvas while largely preserving component sizes and maintaining clear connector flow.
Figure 9: Without Parse Critic. Structural errors from the initial parse remain uncorrected, including incorrect shape types and mismatched connector attachment points. Iterative Critic feedback improves both component geometry and edge anchoring.
Figure 10: Without Style Stage. The reconstructed diagram exhibits mismatched block colors, missing or simplified containers, inconsistent connector styles, and typography differences. The Style Stage substantially improves visual correspondence while preserving the recovered structure.
Figure 11: Without Layout Critic. The single-pass result exhibits weaker space utilization, less natural arrow routing, and poorer visual balance. Iterative deterministic and visual feedback improves placement, compactness, and readability while preserving connectivity.
Figure 12: Replacing Layout Stage with Graphviz DOT. The topology-driven layout often leaves substantial unused space and may fail to preserve relationships involving container structures. Our Layout Stage better utilizes the target canvas while maintaining these structural relationships.
| Method | XML Fail. | Rel. | Hallu.-free | Content | Layout | Style | Overall |
|---|---|---|---|---|---|---|---|
| Direct VLM-to-XML (Qwen3.6-27B) | 16.0 | 8.0 | 12.0 | 0.0 | 2.36 | 2.28 | 38.0 |
| Direct I2I (Qwen-Image-Edit-2511) | 0.0 | 24.0 | 8.0 | 8.0 | 2.16 | 2.20 | 40.0 |
| Ours (Qwen3.6-27B) | 0.0 | 20.0 | 52.0 | 20.0 | 1.40 | 1.44 | 90.0 |
Table 7: Transfer to an open-weight VLM. We replace Gemini 3.1 Pro Preview with Qwen3.6-27B while keeping the proposed pipeline unchanged. Although absolute structural fidelity decreases with the open-weight model, our pipeline substantially improves over direct XML generation using the same VLM and eliminates XML-format failures. Overall is computed within this three-method comparison and is not directly comparable with Overall scores in Tables 2 and 4 . Bold = best, underline = second-best, italic = third-best.
Figure 13: Qualitative comparison using open-weight Qwen models under the 1:1 target aspect ratio. Direct VLM-to-XML introduces structural and connectivity errors, including overlapping blocks. Direct I2I preserves most visual elements but exhibits missing components, distorted text, and substantial geometric stretching. Our method instead produces a coherent relayout with less distortion while better preserving the original structure and connectivity.
Figure 14: Style transfer results. Given a completed relayout (a) and an unrelated reference (b), our Style Transfer extension produces (c) that adopts ’s fill colors, typography, corner style, and connectors while preserving (a)’s block connectivity and semantic content. Colors are transferred via semantic role assignment (background, primary fill, connector) rather than pixel matching, since (a) and share no spatial correspondence.
Figure 15: Additional qualitative comparison under the 1:1 target aspect ratio. In the left example, Nano Banana Pro and GPT Image 2.0 introduce relationship errors when attempting relayout. In the middle and right examples, only our method performs clear relayout, while Nano Banana Pro, GPT Image 2.0, and AutoFigure-Edit mainly rely on stretching, compression, or direct reuse of the original image.
Figure 16: Additional qualitative comparison under the 2:3 target aspect ratio. In the left example, Nano Banana Pro and GPT Image 2.0 mostly stretch the original image without meaningful relayout. In the middle example, all baselines introduce relationship errors after relayout. In the right example, GPT Image 2.0 and AutoFigure-Edit leave large unused space, while PaperBanana rotates a horizontal layout to fit the portrait canvas.
Figure 17: Additional qualitative comparison under the 3:2 target aspect ratio. In the left example, Nano Banana Pro and PaperBanana introduce many hallucinated elements, while only our method performs clear relayout. In the middle example, Nano Banana Pro and PaperBanana also hallucinate structures, and AutoFigure-Edit and GPT Image 2.0 leave large unused space. In the right example, Nano Banana Pro, GPT Image 2.0, and AutoFigure-Edit mostly stretch the original image without meaningful relayout.
Figure 18: Additional qualitative comparison under the 9:16 target aspect ratio. In the left example, all four baselines introduce relationship errors, while our method preserves the correct flow structure. In the middle example, Nano Banana Pro introduces relationship errors after relayout, GPT Image 2.0 mainly stretches the original image, and AutoFigure-Edit leaves large unused space. In the right example, Nano Banana Pro, GPT Image 2.0, and AutoFigure-Edit all produce relationship errors, while PaperBanana rotates a horizontal layout to fit the portrait canvas.
Figure 19: Additional qualitative comparison under the 16:9 target aspect ratio. In the left example, Nano Banana Pro and GPT Image 2.0 directly stretch the original image, causing the embedded image blocks to deform, while PaperBanana and AutoFigure-Edit introduce hallucinated elements. In the middle example, Nano Banana Pro, GPT Image 2.0, and AutoFigure-Edit leave large blank regions, while PaperBanana produces many hallucinated structures. In the right example, Nano Banana Pro, GPT Image 2.0, and PaperBanana largely keep the original layout without meaningful relayout, and AutoFigure-Edit leaves large unused space. Only our method performs clear relayout to better adapt the layout to the target canvas.