acceptodds
Withdrawn from ICLR 2027

One Design, Many Canvases: Synthesizing Layout Pairs to Learn Layered Design Adaptation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

11 views

Abstract

Adapting layered graphic designs to different canvases requires rearranging layered elements while preserving their grouping structure, reading order, and relative sizes. Beyond scaling and repositioning, different element types also require specific adaptations: text may reflow, while photos may be cropped. Learning such layout adaptation is challenging because paired layered designs containing the same elements across different canvas sizes are scarce. We address this limitation with a scalable layout-pair synthesis pipeline that, given a source design and target canvas size, generates an adapted target image and then recovers its layout using a trained VLM. Using this pipeline, we construct ReCanvas-Data, a dataset of 4.22M source–target layout pairs derived from 292K layered designs spanning posters, presentation slides, and infographics. To learn source-to-target adaptation from ReCanvas-Data, we develop ReCanvas, a vision-language model based on our proposed Layout-Structure Chain-of-Thought (LS-CoT) method. LS-CoT first infers a scene tree that captures the nested grouping and reading order of elements in the source design, and then conditions target-layout prediction on the inferred tree. Experiments show that ReCanvas outperforms baselines in layout quality while more faithfully preserving the source layout structure across canvas sizes.

Primary area:
datasets and benchmarks
Keywords:
graphic design, layout, dataset

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_fJAqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies scientific diagram editing from rendered images without access to the original source code. It introduces DiagramEditBench, comprising approximately 41K training examples and 718 human-curated evaluation examples, and proposes DiagramEditor, built on Qwen3.5-9B. The approach combines a Hierarchical Structuring Strategy (HSS) for organizing training programs with GRPO training using image-, code-, and vector-level rewards. Experiments demonstrate substantial improvements in compilation reliability and competitive automatic evaluation results, together with transfer to diagram reconstruction.

Critical strengths

1.A practically relevant problem with a clear task formulation. Editing diagrams when only rendered images are available addresses a common limitation in scientific workflows. Producing executable TikZ programs also supports subsequent modification and high-quality vector rendering.

2.A substantial dataset construction effort. The pipeline includes compilation checks, visual filtering, deduplication, code restructuring, instruction synthesis, and manual benchmark verification. The taxonomy of editing targets and compositional difficulty provides a useful foundation for studying diagram editing.

Critical weaknesses

. The discrepancy between automatic metrics and human evaluation weakens the central quality claims.

Although DiagramEditor leads on most automatic metrics, its human evaluation score is 3.42, below Gemini-3.5-Flash at 4.35 and GPT-5.5 at 4.01. This does not negate the contribution, but it shows that the automatic improvements do not straightforwardly establish superior editing quality.

Furthermore, non-compiling outputs receive minimum image and DVR scores. Consequently, these metrics combine compilation reliability with the quality of successfully rendered outputs. The current results do not sufficiently distinguish improvements from fewer compilation failures from improvements in instruction fulfillment and preservation of unrelated content.

The evaluation should report both overall performance and quality conditional on successful compilation, with the evaluated subsets clearly specified. The human evaluation protocol also needs clarification regarding sample size, sampling procedure, and treatment of compilation failures. Separate measurements of edit correctness and preservation would more directly support the structure-preserving editing claim than overall similarity to a reference image.

  1. The independent contributions of the proposed components are not yet sufficiently established.

The SFT and reward ablations are valuable. However, Table 2 shows that the largest improvement comes from image-reward RL: DreamSim increases from 67.68 to 81.18. Once image and code rewards are included, adding DVR increases DreamSim from 81.93 to 82.45, with gains of 1.34, 0.85, and 1.28 in entity, connection, and attribute scores, respectively.

These results suggest that DVR is useful, but the additional gains are modest and are primarily demonstrated through metrics directly optimized during training. Independent human comparisons or edit-completion measurements are needed to establish whether these gains translate into meaningful improvements in editing quality.

Similarly, the HSS ablation primarily reports compilation rates. It does not directly validate the claimed benefits for localized editing and preservation of unrelated content. Targeted comparisons against the corresponding ablated models would strengthen the methodological contribution without requiring an extensive expansion of the experiments.

  1. The relationship between DVR scores and semantic structural correctness requires further validation.

The connection reward is based on geometric matching within the connection pool, independently of entity matching. This can assess whether lines appear in similar locations, but the description does not clearly establish whether they connect the correct entities or encode the correct directional relationships.

A related concern applies to text rewards. After text-based matching, unmatched text boxes may receive geometric fallback matches, while the final reward primarily aggregates spatial overlap. It is therefore unclear whether an incorrect label at a similar location could still receive a high score.

These concerns do not imply that DVR is ineffective, but they limit its interpretation as a measure of semantic or topological correctness. A small set of controlled perturbations—such as replacing labels, changing connection endpoints, or reversing arrows—would help establish whether the reward reliably detects important errors. The conversion from PDF primitives to entities and connections also requires a clearer description.

  1. Dataset coverage and cross-task generalization need stronger documentation.

Training and evaluation examples are produced through similar instruction-synthesis and target-code-generation pipelines. Temporal splitting and visual deduplication help reduce overlap, but do not independently establish generalization across instruction styles, diagram categories, and generation patterns. The main paper reports aggregate results without sufficient breakdowns by editing operation, difficulty, or diagram type.

In addition, both the training corpus and DaTikZ v3 involve scientific diagrams and TikZ sources, while the main paper does not clearly document deduplication between all training sources and the reconstruction test set. The reconstruction results provide promising evidence of transfer, but their strength depends on the independence of the evaluation data.

The paper should provide dataset distributions, subgroup results, and a clear account of overlap checks against DaTikZ v3. This is a request for stronger documentation, rather than an allegation of data leakage.

Questions for authors

How many examples were included in the human evaluation, and how were compilation failures handled? What explains the ranking discrepancy between human and automatic evaluation?

Does the full model outperform the image-plus-code-reward variant under independent human judgments or edit-completion measurements?

How does DVR respond to incorrect labels, wrong connection endpoints, and reversed arrows? Does connection matching explicitly depend on entity correspondences?

Were all training sources deduplicated against the DaTikZ v3 test set? Can the complete HSS and DVR implementations be provided?

1 comment
AuthorsOpenReview ↗

Possible Review-Submission Mismatch

Dear Reviewer fJAq,

Thank you for your time. It seems that this review may have been intended for another submission, as it discusses DiagramEditBench, DiagramEditor, HSS and DVR scores for TikZ diagram editing, whereas our submission focuses on adapting layered graphic designs to different canvas sizes. We would appreciate it if you could check this and update the review to reflect your assessment of our submission. We would be happy to clarify any aspects of our work.

Reviewer_6EqQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper focuses on the graphic design layout adaptation problem, which adapts the layout of a source layered graphic design to target canvases of different sizes, while preserving the source layout structure.

The main contributions are: (1) a pipeline for synthesizing source-target layout pairs, where given a source design and a target canvas size, an adapted target design image is first generated and its layout is then extracted using a VLM; (2) a dataset consisting of more than 4M synthetic layout pairs for layout adaption model training and evaluation; (3) a LVM-based layout adaptation method, which is trained on the synthetic data and works by first inferring a scene tree that captures high-level structure in the source design, and then predicting the target layout based on the tree to preserve the source layout structure.

Critical strengths
  1. The paper investigates an important research problem because adapting graphic designs to different canvas sizes is a common but tedious task, which can benefit greatly from automated layout adaptation methods.

  2. The proposed data synthesis pipeline and layout adaptation method are well motivated and designed. The constructed dataset of source-target layout pairs could be valuable for future research on the layout adaptation problem.

Critical weaknesses
  1. Comparison to existing layout generation methods (such as LaDeCo) is unfair. In the current experiments, the official checkpoints of these methods are directly used, which do not consider the source design and use different training data from the proposed layout adaptation method. For fair comparison, the existing methods should be modified to include the source design or layout as input, and trained on the same dataset as the proposed method.

  2. The necessity of the proposed scene tree is not validated. A variant of the proposed layout adaptation method should be evaluated, which uses the pairwise relationships (used by ReLayout), instead of the scene tree, as the intermediate representation in LS-CoT.

Questions for authors

In the first training stage (in Section 4.3), what is the model input?

1 comment
AuthorsOpenReview ↗

Response to Review

We sincerely thank you for your constructive feedback and recognition of the importance of layout adaptation and the potential value of ReCanvas-Data. We have addressed your concerns about baseline fairness and scene-tree representation, and clarify the input of the first training stage.

Q1. Fair comparison with existing methods

We appreciate the contributions of LaDeCo and PosterCopilot, which have advanced graphic design layout generation and provided valuable foundations for further research. As these methods were originally developed for a different setting, their official checkpoints mainly reflect out-of-the-box generalization, as clarified in Sec. 5.2. Per your suggestion, we have retrained LaDeCo on ReCanvas-Data with the source image and layout as additional inputs, following its official recipe and initializing from the released weights. In addition, we have also implemented LaDeCo on Qwen3.5-9B-α, preserving its five-stage formulation and intermediate rendering while matching our backbone initialization and paired training data. We did not retrain PosterCopilot because its released resources lack sufficient details to faithfully reproduce its multi-stage training.

ModelSource InputPaired TrainingGemini Score ↑Read. ↑Sim. ↑
LaDeCo (Official)NoNo17.8510.0755.35
LaDeCo (Retrained)YesYes72.9455.2090.70
LaDeCo (Qwen3.5-9B-α)YesYes79.7066.0792.03
Qwen3.5-9B-α + LG-SFT + LA-SFTYesYes83.6876.8090.25

Retraining substantially improves LaDeCo (17.85 → 72.94), and adopting Qwen3.5-9B-α further raises its score to 79.70, achieving competitive performance with our LA-SFT model. These results demonstrate the applicability of ReCanvas-Data across different methods and architectures. LaDeCo still shows lower readability, possibly because its stage-wise generation fixes earlier placements and may accumulate errors, whereas end-to-end layout prediction enables more coordinated adaptation.

Q2. Scene tree vs. pairwise relationships as intermediate representations

We appreciate the suggestion to compare LS-CoT with ReLayout-style pairwise relationships. We have trained two variants that first predict source-element relationships (i.e., pairwise position and size relationships) before generating the target layout, using the same initialization, training data, and hyperparameters as Table 5. We consider both dense relationships, containing all pairwise relationships (the same as ReLayout), and sparse relationships, retaining a spanning tree of the relation graph to match the token budget of LS-CoT.

We additionally evaluate structural preservation using three deterministic metrics based on the Gemini-annotated source scene tree. Each non-root internal node defines a semantic group containing its descendant elements, excluding groups with fewer than two elements or covering the entire design:

  • Foreign area increase (↓): For each group, compute the fraction of its enclosing bounding box covered by the union of non-member elements, excluding large background elements. We report the target-minus-source difference, averaged across valid groups and designs.
  • Silhouette (↑): Compute the silhouette coefficient using distances between element bounding boxes, measuring within-group compactness relative to separation from other groups.
  • Reading order (↑): Identify parallel units from the source tree and measure the percentage of designs preserving their relative order.
Intermediate RepresentationAvg. Intermediate TokensVLM Score ↑Read. ↑Sim. ↑Foreign Area Increase ↓Silhouette ↑Reading Order ↑
None (Base)083.6876.8090.252.4818.7793.77
Pairwise Relationships (Dense)1,03783.1674.8090.092.5819.0594.27
Pairwise Relationships (Sparse)23083.0076.3389.822.0319.1395.45
Scene Tree (LS-CoT)22687.3181.1391.711.6719.6196.88

Pairwise relationships can improve structural preservation but offer limited gains in visual quality. Since geometric relationships may change across canvas sizes, scene trees provide a more stable representation of grouping and reading order, leading to better overall adaptation quality.

Q3. Input of the first training stage (LG-SFT)

In LG-SFT, the model receives the target canvas specification and a set of image and text elements, and predicts their layout on the specified canvas. Unlike LA-SFT, this stage uses individual designs without a reference source image or layout, following the conventional layout-generation setting. Its purpose is to establish general layout-generation capability before learning source-conditioned adaptation.

We will include these additional experiments and clarifications in the revised manuscript.

Reviewer_s4TwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces a 4.2M-pair synthetic dataset (RECANVAS-DATA) and ReCanvas (an RGBA-extended 9B VLM using structure-guided CoT and multi-reward RL), outperforming frontier models in adapting layered graphic designs across diverse canvas aspect ratios.

Critical strengths
  1. A rigorously audited data pipeline with clear proof that paired supervision matters. The data engine reliably scales 291k unpaired designs into ~896k filtered target layouts, with commendable attention to reproducibility. Most importantly, Table 3 demonstrates the value of paired supervision on objective, judge-independent metrics: on the same Qwen3.5-9B-α\alpha backbone, paired training boosts strict readability from 64.67 to 73.13 and similarity from 87.63 to 90.41 over relation-conditioned baselines.

  2. Convincing human evaluation that grounds the claims beyond automatic judges. The performance claims do not rest solely on an LLM scorer. The human evaluation is well-powered and methodologically sound. ReCanvas clearly dominates with an 83.90% win rate and a Bradley–Terry score of 50.38, demonstrating a much wider gap in practice than the modest 3-point delta in automatic VLM scores.

Critical weaknesses
  1. The component ablation relies entirely on the training judge. In Table 5, the performance gains attributed to LS-CoT (+3.63) and RL (+2.55) are reported exclusively via Gemini-3.6-Flash scores. There is no programmatic metric breakdown, no cross-validation with an independent second judge, no error bounds across random seeds, and no human evaluation for these checkpoints. While the human study in Table 6 is well-executed, it only tests the final model against off-the-shelf VLMs. It entirely omits the ablation variants (w/o LS-CoT, w/o RL, and ReLayout-style).

  2. The central claim of structural preservation is never measured. The abstract and Section 4.2 argue that LS-CoT maintains grouping structure and reading order, yet none of the five similarity metrics assess grouping compactness or sibling hierarchy, even though the predicted scene tree makes both straightforward to compute. Crucially, the paper reports neither test-time tree accuracy against Gemini annotations on ReCanvas-Data nor internal consistency between the emitted tree and the final layout. Furthermore, the paper never tests whether autoregressively generating the tree as a prompt prefix is necessary (e.g., comparing against oracle tree inputs or treating the tree as an auxiliary training loss), leaving it ambiguous whether gains stem from structural reasoning or merely extra supervised token signal.

  3. Missing evaluation of the teacher pipeline. The synthetic data engine itself (GPT-Image-2 + grounding) is never evaluated on RECANVAS-BENCH. Because ReCanvas is effectively trained on distillation from this pipeline, the teacher represents the natural performance ceiling and the only direct measure of post-filtering target quality. Knowing whether ReCanvas approaches, matches, or exceeds its generative source is essential for contextualizing the results.

Questions for authors
  1. For Table 5 (+ LG-SFT →\to + LS-CoT →\to + RL), can the authors share: (a) the programmatic readability/similarity scores, (b) numbers from a neutral judge like GPT-5.6-Sol, and (c) even a small-scale human A/B test between these steps?

  2. What is actually in the internal 5.63M dataset, does it overlap with the public sets, and was it cleanly deduplicated against the 6k held-out test designs? Also, what’s the concrete release plan for code, model weights, the grounding model, benchmark data, and the RL reward weights?

I look forward to the authors' response and would be glad to reconsider my rating.

3 comments
AuthorsOpenReview ↗

Response to Review (1/3)

We appreciate your constructive feedback and your recognition of the value of ReCanvas-Data and our human evaluation. We address your concerns with additional judge-independent ablations, direct structural-preservation metrics, LS-CoT analysis, teacher-pipeline evaluation, and clarifications on data overlap and reproducibility.

Q1. Independent validation of component ablations

We extend Table 5 with programmatic metrics, an independent VLM judge, and human evaluation.

(a) Programmatic metrics and independent judge. We report overall readability (Read.), source similarity (Sim.), and scores from GPT-6.1-Sol using the same evaluation rubrics. Results are averaged over three evaluations with randomized input ordering (mean ± SD).

ModelRead. ↑Sim. ↑Gemini-3.6-Flash Eval ↑GPT-6.1-Sol Eval ↑
LA-SFT71.09±0.7589.90±0.3479.33±0.2882.15±0.01
+ RGBA74.53±1.2290.45±0.1082.74±0.1884.18±0.19
+ LG-SFT77.07±0.2490.06±0.1983.86±0.2184.74±0.18
+ LS-CoT80.71±0.8591.64±0.2487.09±0.1987.13±0.10
+ RL84.77±0.6392.60±0.1989.56±0.2989.21±0.07

The improvements from LS-CoT and RL are consistent across programmatic metrics and the independent judge. GPT-6.1-Sol confirms gains of +2.39 [1.80, 3.06] and +2.08 [1.66, 2.51] (95% CI), respectively.

(b) Human evaluation. Given the limited rebuttal time, we focused on three key comparisons rather than repeating the broader pairwise evaluation in the paper. For each comparison, we sampled 100 design pairs, stratified by their Gemini score differences to approximate the score-gap distribution of the full benchmark while balancing the five data sources. Three designers independently evaluated these pairs in a blinded A/B study, without access to model identities or automatic scores.

Comparison (A → B)Prefer A (%)Tie (%)Prefer B (%)
w/o LS-CoT → w/ LS-CoT27.3318.3354.33
w/o RL → w/ RL35.3310.0054.67
ReLayout-style → Paired LA-SFT20.331.6778.00

Human preferences consistently favor the stronger variants, further supporting the contributions of paired supervision, LS-CoT, and RL beyond VLM-based evaluation.

AuthorsOpenReview ↗

Response to Review (2/3)

Q2. Direct evaluation of structural preservation and LS-CoT

(a) Structural-preservation metrics. While the original evaluation considers source-layout relationships and perceptual structural preservation, we agree that direct grouping and reading-order measurements provide stronger evidence. We introduce three deterministic metrics based on the annotated source scene tree. Each non-root internal node defines a group containing its descendant elements. We exclude groups with fewer than two elements or covering the entire design.

  • Foreign area increase (↓): For each group, compute the fraction of its enclosing bounding box covered by the union of non-member elements, excluding large background elements. We report the target-minus-source difference, averaged across valid groups and designs.
  • Silhouette (↑): Compute the silhouette coefficient using distances between element bounding boxes, measuring within-group compactness relative to separation from other groups.
  • Reading order (↑): Identify parallel units from the source tree and measure the percentage of designs preserving their relative order.
ModelForeign Area Increase ↓Silhouette ↑Reading Order ↑
Source layout (reference)0.0022.14100.00
LG-SFT + LA-SFT (Base)2.56±0.2918.88±0.1493.65±0.27
+ LS-CoT1.58±0.0819.79±0.1896.44±0.38
+ RL (ReCanvas)1.00±0.1322.62±0.1197.04±0.34
Gemini-3.6-Flash (High)1.58±0.1021.58±0.1898.77±0.14
GPT-5.6-Sol (High)1.49±0.1422.61±0.1698.48±0.27

LS-CoT improves all three metrics over the base model, reducing foreign area increase by 38% and improving reading-order preservation by 2.79 percentage points. ReCanvas further achieves the lowest foreign area increase and the highest silhouette score among the evaluated models, while proprietary VLMs achieve slightly higher reading-order preservation.

(b) Scene-tree accuracy and consistency. Since multiple scene trees may be valid, we use independent Gemini annotations as an inter-annotation reference to evaluate predicted trees.

ComparisonGroup F1 ↑Cophenetic Corr. ↑Ordering Agreement ↑
Gemini vs. Gemini73.90±0.9891.58±0.2597.83±0.56
Predicted vs. Gemini70.19±1.6390.23±0.6397.62±0.37

The predicted trees achieve agreement close to that between independent Gemini annotations, particularly in hierarchical structure and element ordering. We further examine whether the predicted trees are consistent with the generated layouts. Unlike the source-relative metric above, foreign area here measures the absolute non-member occupancy, without subtracting the source value.

ComparisonForeign Area ↓Silhouette ↑Tree–Layout Order Agreement ↑
Annotated tree vs. source layout15.75±0.1823.16±0.8098.83±0.30
Predicted tree vs. generated layout15.79±0.1824.72±0.2299.50±0.09

The generated layouts exhibit tree–layout consistency comparable to that of the original designs.

(c) Necessity of the tree prefix. We further compare the predicted-tree prefix against oracle-tree conditioning and an auxiliary-supervision variant that predicts the tree after the layout, using matched training settings.

SettingGemini Score ↑Read. ↑Sim. ↑
No tree83.86±0.2177.07±0.2490.06±0.19
Tree after layout83.77 ± 0.2976.62 ± 0.6690.53 ± 0.16
Predicted tree prefix (LS-CoT)87.09±0.1980.71±0.8591.64±0.24
Oracle tree prefix87.25±0.1880.91±0.4091.89±0.05

Generating the tree before the layout outperforms both alternatives. Oracle conditioning yields similar overall quality with further structural improvements. These findings support the benefit of explicit structural conditioning, rather than additional supervised tokens alone.

AuthorsOpenReview ↗

Response to Review (3/3)

Q3. Evaluation of the synthetic teacher pipeline

We evaluated the complete GPT-Image-2 and grounding pipeline on all 1,500 ReCanvas-Bench tasks.

MethodEditableGemini Score ↑
GPT-Image-2 (generated image)No91.20
+ GroundingYes85.24
+ Quality filteringYes89.28
ReCanvas w/o RLYes87.09±0.19
ReCanvas w/ RLYes89.56±0.29

The original generated images achieve high perceptual quality but are not editable. Grounding recovers editable layouts at some quality cost, while filtering retains higher-quality training targets. ReCanvas learns from these synthesized pairs and achieves strong performance across the full benchmark. The filtered teacher score is reported on the retained subset, reflecting the quality of the synthesized training data after filtering.

Q4. Internal dataset, data overlap, and release plan

Internal dataset. The 5.63M designs come from a separate proprietary library of designer-created layered templates, including rendered pages, element images, geometry, and text. They are used for grounding-model training and first-stage layout-generation SFT (LG-SFT), without contributing cross-canvas adaptation pairs.

Data overlap. We checked the internal dataset against the held-out evaluation designs using source identifiers, normalized text matching, perceptual hashing, image-similarity reranking, and manual inspection of high-scoring candidates. No duplicated designs were found.

Release plan. To facilitate future research, we plan to release ReCanvas-Data (including scene-tree annotations), ReCanvas-Bench, and the model weights and inference code for both ReCanvas and the element grounding model by the camera-ready stage. This will enable future work to directly reuse our models and evaluate them on a common benchmark. The training procedures and RL reward design are detailed in Sec. 4.3 and Appendix D, with complete prompts provided in Appendix G. We will also specify the RL reward weights in the revised manuscript. Data will be released subject to the upstream licensing restrictions discussed in Appendix B.1.

We will include these additional experiments and clarifications in the revised manuscript.

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

Related papers

Discussion (0)

Sign in to comment.