acceptodds
Under review as a conference paper at ICLR 2027

FigStruct: Object Hierarchies in the Loop for Agentic Scientific Figure Redesign

Abstract

Improving the visual quality of scientific figures often requires time-consuming manual editing. Researchers must adjust layouts, labels, and connections while preserving the underlying scientific meaning. Image generation can improve presentation, but further revisions require an editable representation of the result. We introduce FigStruct, an agentic framework that couples visual redesign with object-hierarchy-guided SVG correction. A local open-source controller analyzes the input, retrieves design references, and plans a revised composition to guide an image generator. The generated design is converted to SVG and organized into an object hierarchy. Enclosing groups provide diagnostic context, while object bindings identify the SVG elements selected for revision. These bindings are updated after structural changes to support subsequent corrections. Correction rounds run locally, followed by one paid model call for final global refinement. On 150 scientific figures, our best redesign configuration achieves the highest mean scores across all five model-judged visual-quality dimensions among the compared methods. With the generator fixed, agentic guidance improves all five scores for GPT Image 2.0 and four for Qwen Image 2.0 Pro over direct prompting. Reconstructed SVGs attain 99.39% semantic text preservation against the redesigned image and 96.34% strict native-text editability. On 50 complex figures, manual post-editing to a common design target averages 9 minutes 15 seconds for FigStruct outputs, compared with 45 minutes 19 seconds for AutoFigure-Edit outputs. Including retries, mean paid API cost for SVG reconstruction and refinement is $0.009429 per output. This cost is 96.78% lower than that of AutoFigure-Edit and 99.61% lower than that of CraftEditor.

Primary area:
generative models
Keywords:
Scientific Figure Redesign, Multimodal Agents, Editable SVG Reconstruction, Object Hierarchies, Iterative Self-Correction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.80 over 5 reviews, above 27% of the others.
Reviewer_jMxpOpenReview ↗
Rating
2(+1)
Confidence
3(+1)
Informativeness
1
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper proposes FIGSTRUCT, an agentic framework for scientific figure redesign and editable SVG reconstruction. It integrates multimodal figure understanding, design-reference retrieval, image generation, SVG reconstruction, object hierarchy construction, localized correction, and global refinement. The key idea is an object hierarchy over SVG elements: higher-level groups provide diagnostic context, while object bindings specify editable regions. This hierarchy is maintained throughout iterative correction, enabling localized repairs without altering unrelated figure parts.

Critical strengths
  1. Generating editable scientific figures is a meaningful problem beyond conventional image generation. Balancing visual quality and editability is hard, and the paper addresses a real researcher workflow bottleneck.

  2. Using an object hierarchy as an intermediate representation is reasonable: separating diagnostic context from edit scope enables iterative SVG correction without regenerating the whole figure.

Critical weaknesses
  1. The framework combines existing ideas—multimodal understanding, retrieval-augmented planning, image generation, iterative self-correction, vector graphics reconstruction, and structured editing. The main technical contribution appears to be the object hierarchy for SVG correction, yet the paper does not establish why this representation fundamentally differs from existing structured graphics representations (e.g., SVG DOM editing, programmatic vector representations, or other hierarchical editing frameworks). Ablations show that removing components hurts performance, but do not isolate whether the object hierarchy itself provides unique benefits over simpler structured editing. A stronger evaluation would compare alternative intermediate representations under the same generation and correction pipeline.
  2. Scientific figure redesign is not merely visual improvement; it must preserve scientific meaning, including entities, relationships, causal flows, and methodological structures. Although the paper introduces fidelity metrics (text, unit, relation, order preservation), these rely substantially on model-based annotation. It remains unclear whether redesigned figures preserve the original authors' intended scientific interpretation. For example, a cleaner diagram may still introduce subtle semantic changes in connector direction, module dependency, or hierarchy interpretation. Domain-expert human evaluation would provide stronger evidence that the system improves scientific communication rather than only visual appearance.
  3. The benchmark comprises 150 manually redrawn figures with preserved textual and logical structures. While controlled, it may not reflect real-world workflows, where researchers often work with incomplete drafts, PowerPoint figures, LaTeX/TikZ sources, low-resolution screenshots, or inconsistent/ambiguous content. The current benchmark is closer to controlled redesign/beautification than general scientific figure editing. More diverse real-world inputs would strengthen claims of general applicability.
  4. The paper emphasizes low API expenditure, but the default configuration relies on local inference with Gemma-4 31B on two H100 GPUs. API cost alone therefore does not represent true computational expense. A complete analysis should include GPU inference cost, hardware requirements, energy consumption, and total wall-clock cost. The reported runtime also suggests the system is computationally expensive, which warrants more careful discussion.
Questions for authors
  • The paper would benefit from a clearer distinction between scientific figure generation, figure redesign, and SVG reconstruction, as these tasks have different evaluation criteria.

  • The contribution of Global Refinement appears significant in the ablation study. The authors should clarify whether the main gains come from the hierarchy-guided correction mechanism or from the final refinement stage.

  • The benchmark size (150 figures) is relatively small considering the diversity of scientific visualization styles.

  • The paper should discuss failure cases more extensively, especially cases where the system modifies scientific content or introduces unsupported visual elements.

6 comments
AuthorsOpenReview ↗

(1/5) Response to Reviewer jMxp

Dear Reviewer jMxp,

Thank you for recognizing the importance of editable scientific figures and the value of separating diagnostic context from edit scope. We have added representation controls and expert ratings, and provide fuller cost accounting below.

W1. Object Hierarchies Compared with Simpler Structured Representations

W1: The framework combines existing ideas—multimodal understanding, retrieval-augmented planning, image generation, iterative self-correction, vector graphics reconstruction, and structured editing. The main technical contribution appears to be the object hierarchy for SVG correction, yet the paper does not establish why this representation fundamentally differs from existing structured graphics representations (e.g., SVG DOM editing, programmatic vector representations, or other hierarchical editing frameworks). Ablations show that removing components hurts performance, but do not isolate whether the object hierarchy itself provides unique benefits over simpler structured editing. A stronger evaluation would compare alternative intermediate representations under the same generation and correction pipeline.

A: A dynamically maintained hierarchy improves on a flat object table by 4.91 to 6.09 composite points across the tested correction-round settings. SVG DOMs already support groups and nesting. Our contribution concerns how semantic relationships and SVG bindings are used and updated during correction.

The hierarchy connects the context needed to diagnose an error with the SVG objects that may be changed. Parent, sibling, and neighboring objects provide spatial and semantic context. The repair plan then specifies either an object-focused edit or a coordinated revision within a target module. After merging the replacement fragment, reconciliation updates object membership and SVG bindings for subsequent review.

For example, moving a module label may require adjusting nearby nodes and connectors. The worker receives the complete target fragment for context, but the correction plan determines the allowed changes. This separates the material needed to understand the problem from the elements selected for repair.

We compared four representations under the same generation and correction pipeline. The flat object table retains one level of module-to-object membership. Objects are assigned directly to a module without finer parent-child nesting, and patches replace the corresponding SVG objects. It therefore retains module grouping and object-level updates.

All four settings use the same 150 figures, initial SVGs, controller, and global-refinement configuration. Hardware is two H100 80 GB GPUs with four complete figure jobs in parallel. Composite scores are on a 0-to-100 scale.

Representation2 rounds3 rounds4 rounds5 rounds6 rounds
Dynamic hierarchy85.4785.5884.7985.1085.23
Static hierarchy81.7079.8779.3678.2579.16
No hierarchy79.5480.1376.2178.3377.21
Flat object table80.5680.1279.4779.7879.14

The dynamic hierarchy exceeds the static hierarchy by 3.77 to 6.85 points and the flat table by 4.91 to 6.09 points. The static control tests the value of updating the representation. The flat-table comparison tests finer hierarchical organization and its maintenance beyond module-level grouping.

Appendix G.5 provides an illustrative flight-control example. A first repair jointly adjusts the title, node arrangement, and connections. A later repair addresses only background-border overflow while retaining the corrected content. The representation supports both coordinated edits and a narrower scope when only a small issue remains.

Independent editability. At the default two rounds:

Hierarchy settingObject SeparabilitySVG EditabilityComposite quality
No hierarchy47.0371.8679.54
Static hierarchy49.6071.9781.70
Dynamic hierarchy52.7875.8186.23

We report these follow-up component scores separately from the component values in submitted Table 4. The dynamic hierarchy improves Object Separability by 3.18 points and SVG Editability by 3.84 points over the static hierarchy. The corresponding gains over no hierarchy are 5.75 and 3.95 points.

Object Separability checks whether a complete required semantic object can be selected or modified independently of other required objects. It carries 50% of the SVG Editability score. Appendices B.4.1 and B.5.5 define the metric and scoring instructions.


AuthorsOpenReview ↗

(2/5) Response to Reviewer jMxp

Q2. Hierarchy-Guided Correction and Global Refinement

Q2: The contribution of Global Refinement appears significant in the ablation study. The authors should clarify whether the main gains come from the hierarchy-guided correction mechanism or from the final refinement stage.

A: Global Refinement is important in the submitted implementation. The new paired control and sequential component analysis also support a contribution from structured correction.

Direct + Luna control. We reused the same 150 Direct Gemma-4 SVG outputs that scored 77.88 in the submitted Table 7. Each received one GPT-5.6 Luna global-refinement call. The style-reference image was produced by GPT image generation from an agent-prepared prompt informed by retrieved design references. Thus, the before-and-after comparison starts from the same Direct outputs.

ConfigurationComposite qualitySource
Direct Gemma-477.88Submitted Table 7
Direct Gemma-4 + one Luna refinement79.23New paired control
FigStruct without Global Refinement75.98Mean of the three metrics in submitted Table 4
Full FigStruct with Gemma-485.47Submitted Table 7

One Luna call improves Direct by 1.35 points. Full FigStruct remains 6.24 points higher. We also agree with the observation about the original ablation: without Global Refinement, the submitted pipeline scores 75.98, below Direct at 77.88. Global Refinement therefore plays an important role in that implementation. The new control shows that adding this call directly to an unstructured output does not recover the full result.

On inspecting the outputs, we found that Direct SVGs often lack clear semantic groups. Refinement then requires interpreting and reorganizing a large amount of code. This provides a possible explanation for the limited gain from a single refinement call.

Sequential component analysis in the optimized implementation. During review, we simplified redundant workflow steps while retaining the object-hierarchy mechanism, reconciliation logic, and evaluation prompts. We evaluated this implementation on 150 generated reference images, adding Region Localization, SVG Structure Conversion, and Global Refinement in that order. Global Refinement still uses one Luna call.

Cumulative configurationComposite qualityIncrementShare of total gain
Starting configuration78.16n/an/a
+ Region Localization80.35+2.1927.14%
+ SVG Structure Conversion85.45+5.1063.21%
+ Global Refinement86.23+0.789.65%
Total gainn/a+8.07100.00%

The starting and final scores come from the new evaluation. The intermediate scores and increments shown here are reconstructed from the total gain of 8.07 and the recorded contribution percentages, then rounded to two decimal places. For example, the Region Localization increment is 8.07 × 27.14%.

Under this implementation and addition order, the first two components account for 90.35% of the observed gain, and Global Refinement accounts for 9.65%. These are sequential contributions, which depend on the preceding components. They should be interpreted within this implementation. The submitted ablation, round-count comparison, and cost analysis retain their corresponding full-system score of 85.47.

Output inspection suggests that the local controller reconstructs most text, icons, and scientific relationships before final refinement. The final call then improves cross-module spacing, content distribution, connector routing, and use of the canvas. The two stages serve complementary roles: structured correction establishes and repairs editable content, and global refinement improves the overall arrangement.

The four-representation control in W1 keeps global refinement fixed and provides a complementary test of hierarchy organization and maintenance.


AuthorsOpenReview ↗

(3/5) Response to Reviewer jMxp

W2 & Q4. Expert Evaluation, Scientific Fidelity, and Failure Cases

W2: Scientific figure redesign is not merely visual improvement; it must preserve scientific meaning, including entities, relationships, causal flows, and methodological structures. Although the paper introduces fidelity metrics (text, unit, relation, order preservation), these rely substantially on model-based annotation. It remains unclear whether redesigned figures preserve the original authors' intended scientific interpretation. For example, a cleaner diagram may still introduce subtle semantic changes in connector direction, module dependency, or hierarchy interpretation. Domain-expert human evaluation would provide stronger evidence that the system improves scientific communication rather than only visual appearance.

Q4: The paper should discuss failure cases more extensively, especially cases where the system modifies scientific content or introduces unsupported visual elements.

A: Three computer-science professors with extensive journal-publication experience each rated all 150 outputs before manual editing. Satisfaction rates were 78%, 82%, and 72%. We also document semantic and graphical failures.

Expert assessment. Experts A, B, and C rated overall satisfaction and required revision:

ExpertSatisfiedModerately satisfied, substantial revision neededUnsatisfied, requires redoingFigures evaluated
A117249150
B1231710150
C1081923150

Across 450 ratings of the same 150 outputs, 348 (77.33%) are satisfied, 60 (13.33%) moderately satisfied, and 42 (9.33%) unsatisfied. These judgments assess overall acceptability and revision needs; the fidelity protocol separately examines scientific relationships.

Scientific fidelity. Full-workflow fidelity uses the original input as reference; Table 2 reconstruction uses the shared Step 1.C design image. Under Appendix B.1, the model records item-level evidence, humans review annotations, and code computes scores.

MetricWhat it checks
UnitPreservation of semantic units, including modules, nodes, labels, and formulas
Text AccuracyCorrect preservation and semantic ownership of text
RelationMatching endpoints, direction, and relation type, with precision, recall, and F1
OrderRequired precedence, stage, numbering, and hierarchical-order constraints

Text evaluation checks omissions, unsupported additions, duplicates, and semantic ownership. Relations receive credit only when both endpoints, direction, and type are correct. Figure 5 reports:

ConfigurationRelationOrder
Direct Gemma-477.8084.39
FigStruct with Gemma-492.0898.33

The gains are 14.28 and 13.94 percentage points. These metrics assess semantics separately from appearance (Appendices B.3.1 and B.5.3).

Post-editing effort. The study covers 50 complex figures and 400 participants. AutoFigure-Edit and Crafter reconstruct the same Step 1.C images. Participants edit in PowerPoint toward a shared design target, with original scientific content overriding any conflicting reference content. Under common acceptance criteria:

MethodMean active post-editing time
FigStruct with Gemma-49 min 15 s
AutoFigure-Edit45 min 19 s
Crafter58 min 10 s

Appendix D and Table 8 describe the protocol. Editing time reflects output quality and editability. The new expert ratings precede these human corrections.

Remaining failures. In Section 4.6, FigStruct with GLM gains 6.03 composite points, but Relation and Order fall by 6.69 and 9.44 percentage points. This shows why fidelity dimensions must accompany the composite score. Further final-output failures include:

LocationObserved errorImplication
Appendix F.2, Figure 9The design image contains a garbled legend. FigStruct infers plausible labels from context.Improved legibility involves an interpretive change. The corrupted reference alone cannot establish that the inferred meaning is correct.
Appendix F.4, Figure 11Some network connections are simplified, and a flowchart arrowhead is oversized.An orderly layout can retain relational and geometric errors.
Appendix E.1, Figure 6The Terra-controlled output simplifies some icons and retains extra arrowheads at several junctions.Preserving the main structure can coexist with inaccurate local symbols.

We will reference these analyses more clearly in the main text. They distinguish visual gains, content preservation, and inference from ambiguity. Appendix G shows successful repairs; the cases above document remaining errors.


Reviewer_EfxAOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces FigStruct, an agentic framework for scientific figure redesign that aims to improve visual quality while preserving scientific content and producing editable SVG files. The method first uses a locally deployed controller to understand the input figure, retrieve design references, and plan a new layout. An image generation model then produces a redesigned figure, which is subsequently converted into SVG and iteratively refined through object hierarchies, region localization, local correction, and dynamic structural reconciliation (Merge–Reconcile), followed by global refinement.

The authors evaluate the method on 150 paired scientific figures, assessing visual quality, text fidelity, SVG editability, computational cost, and manual post-editing time. The results show that FigStruct outperforms the evaluated methods on several visual-quality and SVG-reconstruction metrics and substantially reduces manual post-editing time.

Critical strengths
  1. Practical relevance. This paper addresses scientific figure redesign and editable SVG reconstruction, aiming to improve visual quality, preserve scientific content, and facilitate subsequent editing. The task has clear practical value.

  2. A relatively comprehensive system design. The authors integrate figure understanding, design-reference retrieval, object hierarchies, local correction, and global refinement into a unified agentic workflow. In particular, using dynamic object hierarchies to coordinate iterative editing is a promising design idea worth exploring.

  3. Comprehensive experimental evaluation. The authors conduct experiments on 150 paired scientific figures, evaluating visual quality, text fidelity, SVG editability, cost, and runtime. Component ablations and human post-editing studies further assess the method from multiple perspectives.

Critical weaknesses

1. The problem formulation is unclear, and the practical relevance of the task is insufficiently established.

Frankly, I do not think the authors have provided a sufficiently precise definition of the problem or convincingly established that it corresponds to a realistic application scenario. In the Method section, the authors state that the system takes only a single image as input. However, Figure 1 shows that the generated output contains many additional elements, including photographs of dogs, which appear to be pre-existing assets rather than content generated by the model itself. This raises questions about the actual input requirements, the availability and provenance of such assets, and the extent to which the system can operate under the stated input assumptions. Moreover, the caption of Figure 1 and the accompanying discussion are rather superficial and do not adequately explain these aspects. More fundamentally, it remains unclear whether the paper primarily aims to generate better scientific figures, reconstruct editable SVGs, or improve the reliability of iterative editing. These objectives are related but distinct, and the paper does not clearly establish which one constitutes its primary research problem.

2. The dataset construction protocol is insufficiently specified, which further obscures the problem definition.

The authors describe the dataset as consisting of manually redrawn input figures paired with reference figures. However, the criteria and protocol governing this manual redrawing process are not sufficiently clear. For example, what specific instructions and constraints were given to the participants? To what extent were they allowed to modify the layout, graphical elements, and visual representations? Was the difficulty of the resulting figures systematically categorized, and, if so, according to what criteria? These details are important because the characteristics of the input figures directly determine the nature and difficulty of the task being evaluated. Furthermore, it is unclear whether this benchmark construction process reflects how researchers typically work when creating or revising scientific figures in practice. Without a more rigorous description of the data construction protocol and the intended use scenarios, it is difficult to assess the validity and practical relevance of the proposed task.

3. The evaluation protocol does not adequately capture two fundamental characteristics of scientific figure design.

I believe scientific figure design has two particularly important characteristics: (i) the scientific relationships among graphical elements must be represented correctly, and (ii) there is generally no unique correct solution to a figure-design task, as multiple visual arrangements can faithfully communicate the same scientific content. The evaluation metrics selected in this paper do not adequately reflect these two characteristics. In particular, visual similarity to a reference image may not reliably indicate whether the scientific relationships are correctly represented, while evaluating outputs against a single reference may fail to account for the non-uniqueness of valid design solutions. The paper therefore needs a clearer justification for how its metrics assess scientific correctness independently of visual quality and how they accommodate alternative but equally valid figure designs.

4. The central methodological contribution is not sufficiently isolated from the effects of other system components.

FigStruct's central claim is that dynamic object hierarchies improve iterative SVG refinement. However, the current experiments primarily demonstrate the effectiveness of the overall system and the combined contributions of components such as region localization and global refinement, without adequately isolating the contribution of dynamic hierarchy reconciliation itself. Given that related works, including AutoFigure-Edit, CraftEditor, and FigTree, already employ structured representations, localized editing, and iterative validation, this gap directly affects my assessment of the paper's methodological novelty. More targeted experiments are needed to establish whether dynamic hierarchy reconciliation provides a distinct and measurable advantage over existing structured figure-editing approaches.

Questions for authors
  1. Evaluation in realistic usage scenarios: Evaluate the method on real-world scientific drafts, manually simplified figures, and inputs consisting only of text, boxes, and lines. Clarify the provenance of additional visual assets and assess the method's applicability to different input formats.

  2. Difficulty-stratified evaluation and generalization: Establish a reproducible difficulty taxonomy and report performance across different levels of scientific-relation complexity, element count, hierarchy depth, and raster-asset proportion. Further evaluate generalization on scientific figures from other sources.

  3. Evaluation of scientific correctness and design non-uniqueness: Introduce metrics such as scientific-relation preservation and relation error rates. Evaluate multiple visually distinct but scientifically equivalent designs to determine whether the evaluation protocol is overly dependent on a single reference image.

  4. Targeted ablation of dynamic hierarchy reconciliation: Under matched initial SVGs, controller models, and computational budgets, compare dynamic object hierarchies, static object hierarchies, and approaches without object hierarchies. Report non-target element corruption, content loss, and editing success rates to directly establish the independent contribution of the proposed mechanism.

5 comments
AuthorsOpenReview ↗

(1/4) Response to Reviewer EfxA

Dear Reviewer EfxA,

Thank you for recognizing the practical value of the task, the potential of dynamic object hierarchies, and the breadth of the evaluation. We address the corresponding weaknesses and questions together, with particular attention to input requirements and what each evaluation measures.

W1 & Q1. Task, Input Requirements, and Asset Provenance

W1: The problem formulation is unclear, and the practical relevance of the task is insufficiently established. Frankly, I do not think the authors have provided a sufficiently precise definition of the problem or convincingly established that it corresponds to a realistic application scenario. In the Method section, the authors state that the system takes only a single image as input. However, Figure 1 shows that the generated output contains many additional elements, including photographs of dogs, which appear to be pre-existing assets rather than content generated by the model itself. This raises questions about the actual input requirements, the availability and provenance of such assets, and the extent to which the system can operate under the stated input assumptions. Moreover, the caption of Figure 1 and the accompanying discussion are rather superficial and do not adequately explain these aspects. More fundamentally, it remains unclear whether the paper primarily aims to generate better scientific figures, reconstruct editable SVGs, or improve the reliability of iterative editing. These objectives are related but distinct, and the paper does not clearly establish which one constitutes its primary research problem.

Q1: Evaluation in realistic usage scenarios: Evaluate the method on real-world scientific drafts, manually simplified figures, and inputs consisting only of text, boxes, and lines. Clarify the provenance of additional visual assets and assess the method's applicability to different input formats.

A: The primary task is to improve the visual presentation of an existing scientific figure while preserving its content and producing an editable SVG. The dog images in Figure 1 are generated by the image model. Users do not supply additional photographs.

The user provides one figure image. The controller interprets it, retrieves design references, and prepares a prompt covering layout, colors, and visual assets. The image generator produces the redesigned image. SVG reconstruction then converts text and geometric elements into editable components, while complex visual content can remain as separately selectable raster assets. Sections 3.1 to 3.4 describe this workflow.

SVG reconstruction is the output-construction stage, and hierarchy-guided correction is the mechanism for improving that stage. Figure generation from text is a related task, but the present evaluation starts with figure images. We will clarify these roles and the asset provenance in the Figure 1 caption and accompanying discussion.

Our manually simplified inputs already include figures made mainly of text, boxes, and lines. “Manually simplified” describes how they were constructed, whereas “text, boxes, and lines” describes their visual form. They can refer to the same input. We do not report separate results for naturally occurring drafts or alternative source-file formats. The evaluated scenario is the stage where scientific content is established and visual presentation needs work.


AuthorsOpenReview ↗

(2/4) Response to Reviewer EfxA

W2 & Q2. Dataset Construction and Difficulty-Stratified Results

W2: The dataset construction protocol is insufficiently specified, which further obscures the problem definition. The authors describe the dataset as consisting of manually redrawn input figures paired with reference figures. However, the criteria and protocol governing this manual redrawing process are not sufficiently clear. For example, what specific instructions and constraints were given to the participants? To what extent were they allowed to modify the layout, graphical elements, and visual representations? Was the difficulty of the resulting figures systematically categorized, and, if so, according to what criteria? These details are important because the characteristics of the input figures directly determine the nature and difficulty of the task being evaluated. Furthermore, it is unclear whether this benchmark construction process reflects how researchers typically work when creating or revising scientific figures in practice. Without a more rigorous description of the data construction protocol and the intended use scenarios, it is difficult to assess the validity and practical relevance of the proposed task.

Q2: Difficulty-stratified evaluation and generalization: Establish a reproducible difficulty taxonomy and report performance across different levels of scientific-relation complexity, element count, hierarchy depth, and raster- asset proportion. Further evaluate generalization on scientific figures from other sources.

A: The 150 inputs were redrawn by 50 participants, with three figures per participant. They preserved text and logical relationships, and the authors checked every redraw. Appendix A provides the construction protocol, which we will summarize more clearly in the main text.

ItemProtocol
Redrawing50 participants redraw three figures each
Content constraintsPreserve all text and logical relationships
Permitted changesAdjust layout, shapes, font sizes, and spacing while preserving content
Visual simplificationRemove non-semantic color. Use a fixed seven-color scheme when color carries meaning. Remove decorative icons and represent necessary icons with minimal symbols or shapes plus labels.
Participant screeningCheck understanding of scientific diagrams and proficiency with PowerPoint
Quality controlThe authors who selected the reference figures inspect every redraw, return unacceptable versions, and review the replacements

This procedure models a content-complete but visually unfinished figure and provides a verifiable reference for scientific content. All source figures are from Artificial Intelligence Review.

Difficulty-stratified evaluation. We grouped the 150 tasks by relation complexity, element count, hierarchy depth, and raster-asset area in the common task reference images. Asset area estimates include photographs, heatmaps, and textures. Each dimension partitions the same 150 figures.

Difficulty dimensionGroupFiguresDirectFigStruct + LunaGain
OverallAll figures15071.8785.47+13.60
Relation complexityLevel 2: branching or merging3267.3481.49+14.15
Relation complexityLevel 3: feedback, bidirectional, or cross-level links11873.2786.36+13.09
Element countAt most 154370.0887.16+17.08
Element count16 to 303076.2885.67+9.39
Element countMore than 307770.8685.41+14.55
Hierarchy depth10n/an/an/a
Hierarchy depth28674.7289.62+14.90
Hierarchy depth36467.4181.23+13.82
Estimated raster-asset areaAt most 5%10277.3889.84+12.46
Estimated raster-asset areaMore than 5%4857.9376.43+18.50

No sample has hierarchy depth 1. Counts across different difficulty dimensions should not be added. The Fidelity component pools annotated counts and matching credits before computing precision, recall, and F1, as defined in Appendix B.3.1. Consequently, a count-weighted mean of the subgroup composite scores need not reproduce the overall score. Direct at 71.87 is from a separate reevaluation, not the original 77.88 outputs used in the paired Direct + Luna control.

FigStruct improves over Direct in every populated group, by 9.39 to 18.50 points. The 118 figures with feedback, bidirectional, or cross-level links gain 13.09 points. The 77 figures with more than 30 elements gain 14.55 points. Scores are lower for greater hierarchy depth and larger raster-asset area, identifying more demanding subsets within this benchmark.

These strata characterize the current benchmark. They do not include a cross-source evaluation.


AuthorsOpenReview ↗

(3/4) Response to Reviewer EfxA

W3 & Q3. Scientific Relationships and Equivalent Designs

W3: The evaluation protocol does not adequately capture two fundamental characteristics of scientific figure design. I believe scientific figure design has two particularly important characteristics: (i) the scientific relationships among graphical elements must be represented correctly, and (ii) there is generally no unique correct solution to a figure- design task, as multiple visual arrangements can faithfully communicate the same scientific content. The evaluation metrics selected in this paper do not adequately reflect these two characteristics. In particular, visual similarity to a reference image may not reliably indicate whether the scientific relationships are correctly represented, while evaluating outputs against a single reference may fail to account for the non-uniqueness of valid design solutions. The paper therefore needs a clearer justification for how its metrics assess scientific correctness independently of visual quality and how they accommodate alternative but equally valid figure designs.

Q3: Evaluation of scientific correctness and design non-uniqueness: Introduce metrics such as scientific-relation preservation and relation error rates. Evaluate multiple visually distinct but scientifically equivalent designs to determine whether the evaluation protocol is overly dependent on a single reference image.

A: Relation and Order assess scientific relationships separately from visual similarity. Their matching rules allow changes in layout and connector routing when the meaning remains correct.

Full-workflow fidelity is evaluated against the original input figure. The reconstruction comparison in Table 2 instead uses the shared Step 1.C design image as the reconstruction target. Appendix B.1 describes the annotation process: the model records item-level evidence, humans review the annotations, and code computes the scores.

MetricWhat it checks
UnitPreservation of semantic units, including modules, nodes, labels, and formulas
Text AccuracyCorrect preservation and semantic ownership of text
RelationMatching endpoints, direction, and relation type, with precision, recall, and F1
OrderRequired precedence, stage, numbering, and hierarchical-order constraints

Text evaluation checks omissions, unsupported additions, duplicates, and semantic ownership. A relation matches only when both endpoints, direction, and type are correct. Incorrect endpoints, reversed arrows, and wrong relation types receive no partial relation credit. Figure 5 reports:

ConfigurationRelationOrder
Direct Gemma-477.8084.39
FigStruct with Gemma-492.0898.33

The gains are 14.28 and 13.94 percentage points, respectively. These metrics assess relationships and ordering separately from visual appearance. Their definitions and annotation instructions are in Appendices B.3.1 and B.5.3.

Alternative valid designs. The existing annotation instructions explicitly allow equivalent visual arrangements:

  • Appendix B.5.3: “Match by meaning rather than identical position, shape, or number of drawing primitives.”
  • Appendix B.5.3: “Different connector routing or visual layout is acceptable when semantic connections remain correct.”
  • Appendix B.5.4: “A change of position or scale is acceptable if semantic ownership is preserved.”

The evaluator records item-level correspondences and evidence under these rules. Humans review the annotations before scores are computed. Aesthetic quality is evaluated separately through layout, spacing, hierarchy, color, typography, readability, and professionalism. SSIM and related image-similarity metrics characterize reconstruction fidelity in Table 2, rather than serving as the criterion for scientific correctness across alternative designs.

We do not report a controlled test with several deliberately different but scientifically equivalent designs for each input. The rules above describe how our current protocol accommodates such variation. We report relation precision, recall, and F1 through the fidelity protocol, rather than a separately tabulated relation-error rate.

Content requirements in the human study. The submitted human study covers 50 complex figures and 400 participants. AutoFigure-Edit and Crafter receive the same Step 1.C design images for reconstruction. Participants then edit the outputs in PowerPoint toward a common design target. When that target conflicts with the original scientific content, the original content takes precedence. Under the same acceptance criteria, mean active editing times are:

MethodMean active post-editing time
FigStruct with Gemma-49 min 15 s
AutoFigure-Edit45 min 19 s
Crafter58 min 10 s

Appendix D and Table 8 describe the protocol. Editing time reflects both initial output quality and editability.


Reviewer_dNKuOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

FigStruct is an agentic framework for redesigning scientific figures into editable SVGs. It uses a dynamic object hierarchy to guide localized SVG correction, separating broader diagnostic context from the actual edit scope and updating object bindings as the structure changes. The method is evaluated on 150 figures and shows strong results in visual quality, reconstruction fidelity, editability, and human post-editing efficiency.

Critical strengths
  1. The experiments are very thorough and cover much more than visual quality alone.
  2. The paper combines LLM judging, automatic metrics, manual verification, and a large human study.
  3. The 400-participant editing study is especially strong and makes the practical benefit convincing.
  4. The ablations and matched-model comparisons are well designed and support the main claims.
  5. The paper is clear, well organized, and easy to follow.
Critical weaknesses
  1. The novelty is limited given the recent work on editable scientific figures and structured SVG generation, the strongest new idea is the dynamic object hierarchy and the separation between context for diagnosis and the actual edit scope.
  2. The lack of released code makes the results harder to reproduce and independently verify.
  3. The related-work discussion is not fully complete: a. VCG-Bench: Towards A Unified Visual-Centric Benchmark for Structured Generation and Editing b. DiagramBank: A Quality-Audited Dataset of Scientific Schematic Diagrams with Multi-Level Document Context c. SciFigure2Code: An AI-Reconstructed Benchmark for Scientific Figure-to-Code
Questions for authors

Do you plan to release the code, prompts, benchmark, and evaluation scripts during the review period?

5 comments
AuthorsOpenReview ↗

(1/2) Response to Reviewer dNKu

Dear Reviewer dNKu,

Thank you for recognizing the breadth of the evaluation, the practical evidence from the human study, and the clarity of the paper. We respond to the novelty concern and clarify our release plan below.

W1. Dynamic Hierarchies and the Separation of Context from Edit Scope

W1: The novelty is limited given the recent work on editable scientific figures and structured SVG generation, the strongest new idea is the dynamic object hierarchy and the separation between context for diagnosis and the actual edit scope.

A: We agree that dynamic hierarchy maintenance and the separation of diagnostic context from edit scope are the central methodological contributions.

The hierarchy connects the context needed to diagnose an error with the SVG objects that may be changed. Parent, sibling, and neighboring objects provide spatial and semantic context. The repair plan then specifies either an object-focused edit or a coordinated revision within a target module. After merging the replacement fragment, reconciliation updates object membership and SVG bindings for subsequent review.

For example, moving a module label may require adjusting nearby nodes and connectors. The worker receives the complete target fragment for context, but the correction plan determines the allowed changes. This separates the material needed to understand the problem from the elements selected for repair.

We compared dynamic, static, and no-hierarchy variants on the same 150 figures. The initial SVGs, controller, and global-refinement settings were held fixed. All variants used two H100 80 GB GPUs with four complete figure jobs running concurrently. The table varies the number of outer local-correction rounds.

Correction roundsDynamic hierarchyStatic hierarchyNo hierarchy
285.4781.7079.54
385.5879.8780.13
484.7979.3676.21
585.1078.2578.33
685.2379.1677.21
Range across round settings0.793.453.92

The dynamic hierarchy scores highest at every tested round setting, exceeding the static hierarchy by 3.77 to 6.85 points. Its score range across these settings is 0.79 points, compared with 3.45 for the static hierarchy and 3.92 without a hierarchy. These results support maintaining object relationships and SVG bindings throughout editing. Updating the bindings allows subsequent diagnosis to use the current SVG, while a fixed hierarchy can retain outdated associations after edits.

W2 & Q1. Release of Code, Prompts, Data, and Evaluation Scripts

W2: The lack of released code makes the results harder to reproduce and independently verify.

Q1: Do you plan to release the code, prompts, benchmark, and evaluation scripts during the review period?

A: Yes. We plan to release the code, prompts, benchmark, and evaluation scripts before the ICLR review discussion closes. We will provide an anonymous link once the materials are available, so that reviewers can inspect and reproduce the results.


AuthorsOpenReview ↗

(2/2) Response to Reviewer dNKu

W3. Additional Related Work

W3: The related-work discussion is not fully complete: a. VCG-Bench: Towards A Unified Visual-Centric Benchmark for Structured Generation and Editing b. DiagramBank: A Quality-Audited Dataset of Scientific Schematic Diagrams with Multi-Level Document Context c. SciFigure2Code: An AI-Reconstructed Benchmark for Scientific Figure-to-Code

A: Thank you for identifying these relevant works. We will add their citations and discuss the following connections in the revised related-work section.

WorkPlanned discussion
VCG-Bench: Towards A Unified Visual-Centric Benchmark for Structured Generation and EditingBenchmarks for structured visual generation and editing
DiagramBank: A Quality-Audited Dataset of Scientific Schematic Diagrams with Multi-Level Document ContextScientific-diagram data and associated document context
SciFigure2Code: An AI-Reconstructed Benchmark for Scientific Figure-to-CodeEvaluation of scientific figure-to-code reconstruction

We also plan to assess which data and evaluation settings from these works can support further testing of FigStruct. These evaluations are planned, and we do not include results from them in the present response.

We appreciate your suggestions and would be happy to discuss the mechanism or release materials further.


Thank you for the response. It addresses some of my concerns. However, why weren’t the code, prompts, benchmark, and evaluation scripts released with the submission, even through an anonymous link? They are still unavailable. Were these materials incomplete at the time of submission? Please explain.

Data integrity, valid experiments, and reproducibility are basic requirements for a research paper. If the experiments cannot be reproduced, I will keep my current score or lower it.

Reviewer_MaL3OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes FigStruct, an agentic framework for redesigning existing scientific figures and reconstructing them as editable SVGs. The framework first uses an LLM-based controller to analyze the input figure, retrieve design references, and guide visual redesign. It then constructs an object hierarchy to support localized SVG correction, maintains object bindings across revisions, and performs a final global refinement. Experiments on 150 scientific figures evaluate visual quality, reconstruction fidelity, editability, and computational cost. The authors also conduct a human study with 400 participants to assess the manual editing effort required to refine generated figures.

Critical strengths
  1. Practical Relevance. The paper addresses the practical problem of improving existing scientific figures while preserving editability. The object-hierarchy-guided correction framework provides a structured approach to iterative SVG refinement.
  2. Extensive Empirical Evaluation. The evaluation covers multiple baselines and several aspects of figure quality, including visual fidelity, editability, and efficiency. In particular, the large-scale human study involving 400 participants provides useful evidence of reduced manual post-editing effort, strengthening the practical relevance of the reported results.
Critical weaknesses
  1. Limited Novelty and Task Scope. I find the task setting and novelty of this work somewhat underwhelming. The proposed framework requires an existing raster scientific figure as input and mainly focuses on visual redesign and reconstruction into editable SVG. However, existing systems such as AutoFigure-Edit already support a more comprehensive end-to-end workflow, from scientific text to raster generation, SVG reconstruction, and iterative refinement, while also supporting existing raster inputs. Moreover, LiveFigure: Generating Editable Scientific Illustration with VLM Agents (ICML 2026) has demonstrated the generation of editable scientific illustrations through VLM agents and PowerPoint-based workflows, including human-in-the-loop editing evaluations, but is neither cited nor discussed. Furthermore, modern general-purpose coding agents, such as Codex combined with frontier GPT models, can already perform impressive raster-to-editable-diagram reconstruction and iterative refinement. The paper's Direct LLM baselines do not adequately represent such agents equipped with code execution and visual feedback. Given the capabilities of these existing systems and tools, I am not fully convinced that FigStruct offers a sufficiently substantial methodological or practical advancement.
  2. Unclear Methodological Gains and Efficiency. The ablation results raise questions about what actually drives the improvements. In Table 4, removing Global Refinement causes the largest performance degradation, even though this stage relies on a powerful GPT-5.6 model. This suggests that a substantial portion of the gains may come from strong-model refinement rather than the proposed object-hierarchy mechanisms. The current ablations do not sufficiently isolate the benefits of dynamic hierarchy maintenance and reconciliation from those of additional refinement calls. Moreover, the improvements over strong baselines are relatively modest in several visual-quality dimensions, and the qualitative examples do not demonstrate a particularly striking advantage. The efficiency claims are also questionable: the reported API cost of approximately $0.0094 per figure excludes substantial local computation and the preceding raster-generation stage. The default configuration requires two H100 GPUs and approximately 29.45 minutes per figure for SVG reconstruction and refinement. A comprehensive end-to-end cost analysis and budget-matched comparisons are therefore necessary.
  3. Limited Generalizability. The evaluation dataset raises concerns about generalizability. All 150 examples are drawn from a single journal, Artificial Intelligence Review, and the inputs are manually simplified versions of already published, visually polished figures. This artificial setting may not adequately represent real research drafts, which often contain incomplete layouts, ambiguous structures, and evolving scientific content. The paper lacks sufficient validation across scientific disciplines, figure styles, and naturally occurring research drafts.
Questions for authors
  1. Novelty and Baseline. Given the capabilities of existing systems such as AutoFigure-Edit and LiveFigure, what substantial methodological or practical advantages does FigStruct offer? How does it compare with modern general-purpose coding agents?
  2. Methodological Contribution. Table 4 shows that Global Refinement contributes the largest performance improvement. How much of the gain can be attributed specifically to the proposed object hierarchy and dynamic reconciliation, rather than additional refinement steps or stronger models?
  3. Generalizability. how well does FigStruct generalize to naturally occurring research drafts beyond the manually simplified figures from a single journal?
5 comments
AuthorsOpenReview ↗

(1/4) Response to Reviewer MaL3

Dear Reviewer MaL3,

Thank you for recognizing the practical relevance of editable figure redesign and the value of the human editing study. Your comments help us clarify the methodological contribution and distinguish it from the capabilities of the component models. We address each concern below.

W1 & Q1. Methodological Contribution and Relevant Baselines

W1: Limited Novelty and Task Scope. I find the task setting and novelty of this work somewhat underwhelming. The proposed framework requires an existing raster scientific figure as input and mainly focuses on visual redesign and reconstruction into editable SVG. However, existing systems such as AutoFigure-Edit already support a more comprehensive end-to-end workflow, from scientific text to raster generation, SVG reconstruction, and iterative refinement, while also supporting existing raster inputs. Moreover, LiveFigure: Generating Editable Scientific Illustration with VLM Agents (ICML 2026) has demonstrated the generation of editable scientific illustrations through VLM agents and PowerPoint-based workflows, including human-in-the-loop editing evaluations, but is neither cited nor discussed. Furthermore, modern general-purpose coding agents, such as Codex combined with frontier GPT models, can already perform impressive raster-to-editable-diagram reconstruction and iterative refinement. The paper's Direct LLM baselines do not adequately represent such agents equipped with code execution and visual feedback. Given the capabilities of these existing systems and tools, I am not fully convinced that FigStruct offers a sufficiently substantial methodological or practical advancement.

Q1: Novelty and Baseline. Given the capabilities of existing systems such as AutoFigure-Edit and LiveFigure, what substantial methodological or practical advantages does FigStruct offer? How does it compare with modern general-purpose coding agents?

A: FigStruct studies how explicit object relationships and updated bindings can support iterative SVG correction with a local controller.

The hierarchy connects the context needed to diagnose an error with the SVG objects that may be changed. Parent, sibling, and neighboring objects provide spatial and semantic context. The repair plan then specifies either an object-focused edit or a coordinated revision within a target module. After merging the replacement fragment, reconciliation updates object membership and SVG bindings for subsequent review.

For example, moving a module label may require adjusting nearby nodes and connectors. The worker receives the complete target fragment for context, but the correction plan determines the allowed changes. This separates the material needed to understand the problem from the elements selected for repair.

AutoFigure-Edit already supports a broad editable-figure workflow, and LiveFigure uses VLM agents to generate editable PowerPoint figures through executable scripts and visual diagnostics. We will add the missing LiveFigure citation and discuss its relationship to our work. Our methodological focus is the hierarchy-guided correction mechanism.

In the new controller-substitution experiments, AutoFigure-Edit and Crafter score 71.49 and 70.37 with Gemma-4 on 150 figures each. FigStruct scores 85.47 with Gemma-4 and one Luna refinement call. This tests how the frameworks operate with the same main controller. The dynamic, static, and no-hierarchy controls below more directly examine the proposed mechanism.

We do not report a quantitative comparison with Codex or LiveFigure. Direct LLM generation is a single-pass control and does not represent a coding agent with execution and visual feedback. The visual-redesign stage can also use supplied textual content, although the reported evaluation starts with images. The local controller offers a deployment option for interpretation and repeated SVG repair. The evaluated full pipeline still calls external models for image generation and final refinement.


AuthorsOpenReview ↗

(3/4) Response to Reviewer MaL3

W2 & Q2, continued. Hierarchy Maintenance and Practical Gains

Dynamic maintenance. We compared dynamic, static, and no-hierarchy variants on the same 150 figures. The initial SVGs, controller, and global-refinement settings were held fixed. All variants used two H100 80 GB GPUs with four complete figure jobs running concurrently. The table varies the number of outer local-correction rounds.

Correction roundsDynamic hierarchyStatic hierarchyNo hierarchy
285.4781.7079.54
385.5879.8780.13
484.7979.3676.21
585.1078.2578.33
685.2379.1677.21
Range across round settings0.793.453.92

The dynamic hierarchy scores highest at every tested round setting, exceeding the static hierarchy by 3.77 to 6.85 points. Its score range across these settings is 0.79 points, compared with 3.45 for the static hierarchy and 3.92 without a hierarchy. These results support maintaining object relationships and SVG bindings throughout editing. Updating the bindings allows subsequent diagnosis to use the current SVG, while a fixed hierarchy can retain outdated associations after edits.

The best no-hierarchy score across two to six rounds is 80.13, below 85.47 for the dynamic hierarchy with two rounds. Increasing the dynamic variant from two to three rounds adds only 0.11 points, with no sustained improvement thereafter. This supports our choice of two rounds. These comparisons match the tested round setting and controller, but do not impose an equal token, GPU-time, or monetary budget.

Practical gains. Table 2 reports a 6.20-percentage-point increase in text containment and an approximately 81.9% reduction in the text-overlap metric relative to Crafter. The submitted human study covers 50 complex figures and 400 participants. AutoFigure-Edit and Crafter receive the same Step 1.C design images for reconstruction. Participants then edit the outputs in PowerPoint toward a common design target. When that target conflicts with the original scientific content, the original content takes precedence. Under the same acceptance criteria, mean active editing times are:

MethodMean active post-editing time
FigStruct with Gemma-49 min 15 s
AutoFigure-Edit45 min 19 s
Crafter58 min 10 s

Appendix D and Table 8 describe the protocol. This measures the work needed to correct and finish the outputs, reflecting both initial output quality and editability.

AuthorsOpenReview ↗

(3/4) Response to Reviewer MaL3

W2 & Q2, continued. GPU Cost and End-to-End Latency

Hardware and allocation rule. We used two H100 80 GB PCIe GPUs, 16 vCPUs, and 256 GiB of RAM, with four complete figure jobs running concurrently. We allocate GPU cost using the Runpod rate recorded on November 10, 2026: USD 2.89 per GPU-hour. The calculation is:

GPU cost per figure = runtime in minutes / 60 × 2 GPUs × USD 2.89 / 4 concurrent figures.

This is a rental-rate estimate allocated across concurrent jobs. The original USD 0.009429 covers paid API calls for SVG reconstruction and refinement. The following table adds the GPU estimate and the new controller-substitution runs.

MethodComposite qualityAPI cost, USD/figureAllocated GPU cost, USD/figureAccounted subtotal, USD/figureMean runtime
FigStruct, Gemma-4 + one Luna call85.470.0094290.70930.718729 min 27 s
AutoFigure-Edit, original configuration82.570.2924Not itemized0.29245 min 44 s
AutoFigure-Edit, Gemma-4 controller71.4900.18790.18797 min 48 s
Crafter, original configuration79.002.439Not itemized2.43927 min 26 s
Crafter, Gemma-4 controller70.370.0631*1.29451.357653 min 45 s

Both controller-substitution experiments cover all 150 figures. The original configurations use the submitted results. FigStruct retains one Luna global-refinement call, so the comparison matches the main controller, not every model call in the system. In these tables, Crafter denotes the editable-output baseline implemented through CraftEditor.

The Crafter + Gemma-4 API item is an estimate for one GPT image-generation call per figure, using the USD 0.0631 mean reported for the FigStruct generator configuration in Appendix C.2, Table 6. Its subtotal combines this estimate with allocated GPU cost. “Not itemized” does not mean zero: local GPU charges for the original baseline configurations are not separately reported. This includes the short local SAM 3.1 extraction step in AutoFigure-Edit.

End-to-end cost and latency. Adding input interpretation, reference retrieval, and raster generation gives the following breakdown.

StageAPI cost, USD/figureAllocated GPU cost, USD/figureAccounted subtotal, USD/figureMean latency
Input interpretation, retrieval, and visual redesign0.06310.24570.308810 min 12 s
SVG reconstruction and global refinement0.0094290.70930.718729 min 27 s
End to end0.0725290.95491.027439 min 39 s

Totals use unrounded values. The first-stage API item is the image-generation cost from Table 6. Retrieval is controlled by local Gemma-4 without a paid external retrieval service, and its computation is included in the GPU allocation. API costs include billable retries, and latency includes retries and reconnection waits. The GPU estimate uses the dated rate above from Runpod GPU Pricing.

Replacing the baseline controllers with Gemma-4 lowers composite quality by 11.08 points for AutoFigure-Edit and 8.63 points for Crafter. Crafter also encounters output-format parsing, SVG validity, and context-adaptation problems. These results show that replacing the controller alone does not preserve the original performance.

FigStruct scores 85.47, exceeding the original AutoFigure-Edit and Crafter configurations by 2.90 and 6.47 points. Including allocated GPU cost raises its reconstruction subtotal to USD 0.7187, above the USD 0.2924 accounted for the original AutoFigure-Edit configuration. We therefore report API expenditure, local compute, output quality, and latency separately. Together, these results describe the quality, cost, and latency tradeoff of the locally controlled pipeline.

Reviewer_jKsqOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes FigStruct which is a pipeline for converting a given source image to being more aesthetic along with making it editable. The authors convert the image to an SVG artifact and iteratively refine it using guidance from a set of parallel VLM agents (with Gemma-4 31B as region workers) followed by a global refinement phase. The evaluation is done on a set of 150 samples and evaluates visual fidelity across 5 dimensions along with text preservation, editability, geometry and API costs. A human evaluation (across 400 participants) for measuring editing time required by different methods is also shown. FigStruct claims a cheaper method (in terms of cost) for generating higher fidelity and editable scientific images.

Critical strengths
  1. One of the most interesting results from the paper is the editablity experiment across 400 participants. The paper evaluates the editable output with these participants editing in PowerPoint and not only with image metrics. Against different methods, FigStruct's 9m 15s vs 45m 19s (AutoFigure) and 58m 10s (Crafter) is a meaningful practical result.

  2. Text geometry metrics (TBR, TPVR, overlap) clearly improve over Crafter and AutoFigure-Edit baselines, which does support better layout of the output image.

Critical weaknesses
  1. From Table 7, we see that the composite quality for the direct Gemma baseline is 77.88, whereas the composite quality from the ablation table (Table 4) without global refinement is 75.98. This means that the whole pipeline (design image, retrieval, hierarchy, about 90 local calls, 29 min on 2 H100s) before the global refinement scores below one direct Gemma call, until the global refinement by Luna is added. This directly underscores the main contribution of the paper regarding hierarchy-guided SVG correction. Additionally, we need a Direct + Luna refinement baseline as well to concretize this and get the effect of Luna refinement.

  2. The cost analysis is incomplete/misleading. Table 7 shows that when FigStruct uses API controllers its cost is no longer low -> 0.21 USD with GLM (about AutoFigure's 0.29 USD) and 3.70 USD with Terra (1.5× Crafter). So the saving comes from using a free local model and not from the method. This is an unfair comparison, as authors didn't run AutoFigure or Crafter on a cheap or local model, so we can't tell how much of the saving belongs to the method. Moreover, GPU costs are being left out totally. Please mention total cost including GPU time at a stated rate.

  3. The contributions are incremental over existing editable scientific figure pipelines like FigTree [1] which does something very similar to the proposed pipeline (its critic reviews both the render and the SVG code which traces defects to specific elements and then the generator repairs only those elements).

References ->

[1] Liu et al. 2026, Figures as Programs (FigTree), arXiv 2609.01006

Questions for authors
  1. Please add a "Direct + Global Refinement with Luna" row for addressing Weakness 1
  2. Please report total cost including image generation, web search, and GPU time at a stated price, with a price table and pricing date. Also run AutoFigure and Crafter with the same model as FigStruct (for addressing weakness 2). Please also report end to end latency including Step 1.
  3. Please state for the human study, exactly which input (original figure or Step 1.C image) each of AutoFigure and Crafter received.

Minor Mistakes

  1. Figure 2 has spelling errors - obejct -> object, Strcucture -> Structure
4 comments
AuthorsOpenReview ↗

(1/3) Response to Reviewer jKsq

Dear Reviewer jKsq,

Thank you for the thorough review and for recognizing the practical value of the human editing study and the improved text geometry. We have run additional controls to examine global refinement, local computation costs, and hierarchy maintenance. We address your concerns below.

W1 & Q1. Direct Generation Followed by Luna Refinement

W1: From Table 7, we see that the composite quality for the direct Gemma baseline is 77.88, whereas the composite quality from the ablation table (Table 4) without global refinement is 75.98. This means that the whole pipeline (design image, retrieval, hierarchy, about 90 local calls, 29 min on 2 H100s) before the global refinement scores below one direct Gemma call, until the global refinement by Luna is added. This directly underscores the main contribution of the paper regarding hierarchy-guided SVG correction. Additionally, we need a Direct + Luna refinement baseline as well to concretize this and get the effect of Luna refinement.

Q1: Please add a "Direct + Global Refinement with Luna" row for addressing Weakness 1

A: Adding one Luna call improves Direct Gemma-4 from 77.88 to 79.23. Full FigStruct scores 85.47, a further 6.24 points higher.

Direct + Luna control. We reused the same 150 Direct Gemma-4 SVG outputs that scored 77.88 in the submitted Table 7. Each received one GPT-5.6 Luna global-refinement call. The style-reference image was produced by GPT image generation from an agent-prepared prompt informed by retrieved design references. Thus, the before-and-after comparison starts from the same Direct outputs.

ConfigurationComposite qualitySource
Direct Gemma-477.88Submitted Table 7
Direct Gemma-4 + one Luna refinement79.23New paired control
FigStruct without Global Refinement75.98Mean of the three metrics in submitted Table 4
Full FigStruct with Gemma-485.47Submitted Table 7

One Luna call improves Direct by 1.35 points. Full FigStruct remains 6.24 points higher. We also agree with the observation about the original ablation: without Global Refinement, the submitted pipeline scores 75.98, below Direct at 77.88. Global Refinement therefore plays an important role in that implementation. The new control shows that adding this call directly to an unstructured output does not recover the full result.

On inspecting the outputs, we found that Direct SVGs often lack clear semantic groups. Refinement then requires interpreting and reorganizing a large amount of code. This provides a possible explanation for the limited gain from a single refinement call.


AuthorsOpenReview ↗

(2/3) Response to Reviewer jKsq

W2 & Q2. GPU Cost, End-to-End Latency, and Controller Substitution

W2: The cost analysis is incomplete/misleading. Table 7 shows that when FigStruct uses API controllers its cost is no longer low -> 0.21 USD with GLM (about AutoFigure's 0.29 USD) and 3.70 USD with Terra (1.5× Crafter). So the saving comes from using a free local model and not from the method. This is an unfair comparison, as authors didn't run AutoFigure or Crafter on a cheap or local model, so we can't tell how much of the saving belongs to the method. Moreover, GPU costs are being left out totally. Please mention total cost including GPU time at a stated rate.

Q2: Please report total cost including image generation, web search, and GPU time at a stated price, with a price table and pricing date. Also run AutoFigure and Crafter with the same model as FigStruct (for addressing weakness 2). Please also report end to end latency including Step 1.

A: Including visual redesign and allocated GPU rental cost gives an estimated USD 1.0274 per figure and a mean end-to-end latency of 39 min 39 s. We also ran AutoFigure-Edit and Crafter with Gemma-4 controllers on all 150 figures.

Hardware and allocation rule. We used two H100 80 GB PCIe GPUs, 16 vCPUs, and 256 GiB of RAM, with four complete figure jobs running concurrently. We allocate GPU cost using the Runpod rate recorded on November 10, 2026: USD 2.89 per GPU-hour. The calculation is:

GPU cost per figure = runtime in minutes / 60 × 2 GPUs × USD 2.89 / 4 concurrent figures.

This is a rental-rate estimate allocated across concurrent jobs. The original USD 0.009429 covers paid API calls for SVG reconstruction and refinement. The following table adds the GPU estimate and the new controller-substitution runs.

MethodComposite qualityAPI cost, USD/figureAllocated GPU cost, USD/figureAccounted subtotal, USD/figureMean runtime
FigStruct, Gemma-4 + one Luna call85.470.0094290.70930.718729 min 27 s
AutoFigure-Edit, original configuration82.570.2924Not itemized0.29245 min 44 s
AutoFigure-Edit, Gemma-4 controller71.4900.18790.18797 min 48 s
Crafter, original configuration79.002.439Not itemized2.43927 min 26 s
Crafter, Gemma-4 controller70.370.0631*1.29451.357653 min 45 s

Both controller-substitution experiments cover all 150 figures. The original configurations use the submitted results. FigStruct retains one Luna global-refinement call, so the comparison matches the main controller, not every model call in the system. In these tables, Crafter denotes the editable-output baseline implemented through CraftEditor.

The Crafter + Gemma-4 API item is an estimate for one GPT image-generation call per figure, using the USD 0.0631 mean reported for the FigStruct generator configuration in Appendix C.2, Table 6. Its subtotal combines this estimate with allocated GPU cost. “Not itemized” does not mean zero: local GPU charges for the original baseline configurations are not separately reported. This includes the short local SAM 3.1 extraction step in AutoFigure-Edit.

End-to-end cost and latency. Adding input interpretation, reference retrieval, and raster generation gives the following breakdown.

StageAPI cost, USD/figureAllocated GPU cost, USD/figureAccounted subtotal, USD/figureMean latency
Input interpretation, retrieval, and visual redesign0.06310.24570.308810 min 12 s
SVG reconstruction and global refinement0.0094290.70930.718729 min 27 s
End to end0.0725290.95491.027439 min 39 s

Totals use unrounded values. The first-stage API item is the image-generation cost from Table 6. Retrieval is controlled by local Gemma-4 without a paid external retrieval service, and its computation is included in the GPU allocation. API costs include billable retries, and latency includes retries and reconnection waits. The GPU estimate uses the dated rate above from Runpod GPU Pricing.

Replacing the baseline controllers with Gemma-4 lowers composite quality by 11.08 points for AutoFigure-Edit and 8.63 points for Crafter. Crafter also encounters output-format parsing, SVG validity, and context-adaptation problems. These results show that replacing the controller alone does not preserve the original performance.

FigStruct scores 85.47, exceeding the original AutoFigure-Edit and Crafter configurations by 2.90 and 6.47 points. Including allocated GPU cost raises its reconstruction subtotal to USD 0.7187, above the USD 0.2924 accounted for the original AutoFigure-Edit configuration. We therefore report API expenditure, local compute, output quality, and latency separately. Together, these results describe the quality, cost, and latency tradeoff of the locally controlled pipeline.


AuthorsOpenReview ↗

(3/3) Response to Reviewer jKsq

W3. Relationship to FigTree and the Role of Dynamic Maintenance

W3: The contributions are incremental over existing editable scientific figure pipelines like FigTree [1] which does something very similar to the proposed pipeline (its critic reviews both the render and the SVG code which traces defects to specific elements and then the generator repairs only those elements).

A: Our contribution concerns how an object hierarchy is used and maintained to coordinate repeated repairs. We agree that hierarchical decomposition and localized repair are already present in related work. FigTree, Section 3.4, reviews both SVG fragments and their renders, anchors defects to code elements, and repairs those elements locally. We cite this work in Section 2.

The hierarchy connects the context needed to diagnose an error with the SVG objects that may be changed. Parent, sibling, and neighboring objects provide spatial and semantic context. The repair plan then specifies either an object-focused edit or a coordinated revision within a target module. After merging the replacement fragment, reconciliation updates object membership and SVG bindings for subsequent review.

For example, moving a module label may require adjusting nearby nodes and connectors. The worker receives the complete target fragment for context, but the correction plan determines the allowed changes. This separates the material needed to understand the problem from the elements selected for repair.

Dynamic-maintenance control. We compared dynamic, static, and no-hierarchy variants on the same 150 figures. The initial SVGs, controller, and global-refinement settings were held fixed. All variants used two H100 80 GB GPUs with four complete figure jobs running concurrently. The table varies the number of outer local-correction rounds.

Correction roundsDynamic hierarchyStatic hierarchyNo hierarchy
285.4781.7079.54
385.5879.8780.13
484.7979.3676.21
585.1078.2578.33
685.2379.1677.21
Range across round settings0.793.453.92

The dynamic hierarchy scores highest at every tested round setting, exceeding the static hierarchy by 3.77 to 6.85 points. Its score range across these settings is 0.79 points, compared with 3.45 for the static hierarchy and 3.92 without a hierarchy. These results support maintaining object relationships and SVG bindings throughout editing. Updating the bindings allows subsequent diagnosis to use the current SVG, while a fixed hierarchy can retain outdated associations after edits.

The control tests maintenance within FigStruct. A direct quantitative comparison with FigTree remains planned.

Q3. Inputs to the Human Editing Study

Q3: Please state for the human study, exactly which input (original figure or Step 1.C image) each of AutoFigure and Crafter received.

A: Both AutoFigure-Edit and Crafter received the redesigned images from Step 1.C. They reconstructed the same design images into SVGs before the PowerPoint editing study. The shared design image supplied the visual target, while the original input supplied the scientific content. If the two conflicted, the original scientific content took precedence. We will make this input distinction explicit in the main text, with a reference to Appendix D.2.

Minor. Spelling in Figure 2

Figure 2 has spelling errors - obejct -> object, Strcucture -> Structure

A: Thank you for pointing these out. We will correct them to object and Structure in the revised figure.

We appreciate your constructive feedback and would be happy to discuss any remaining concerns.


I thank the authors for their detailed responses. The newly added cost analysis shows that the cost advantage holds wrt Crafter but not against Autofigure-Edit and the quality seems comparable. Moreover, the claim made in the abstract about their cost being 96.78% lower than AutoFigure-Edit no longer holds. This also limits the novelty of their contributions and because of that I would like to keep my score of Weak Reject. My other concerns regarding Direct+Luna baseline have been addressed. I also advise the authors to add the inputs of the human study clearly in the main paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.