acceptodds
Under review as a conference paper at ICLR 2027

What Information Enables Cross-Script Handwriting Layout Synthesis?

Abstract

Handwriting Layout Synthesis generates the spatial arrangement of handwritten text while preserving writer-specific spacing and layout style, providing the structural foundation for handwriting generation. Current systems achieve strong within-script performance but their ability to generalize to unseen scripts is poorly understood. In this work, we present an investigation of the impact of input token representations on this failure. We conduct a controlled empirical study over the state-of-the-art character-compositional representation as well as script-agnostic, explicit glyph-based, and pretrained visual token representations, introducing the latter three to this task. Conducting our study across three distinct scripts representing the morphosyllabic/logographic, abugida and alphabetic writing systems, we observe that neither character-level vocabulary expansion, nor the removal of script-specific information alone is sufficient to achieve robust cross-script generalization. We find that visual representations substantially outperform character-compositional representations on unseen writing systems, while also achieving modest gains in the in-domain setting. Through representation probing and performance comparisons across representations with increasing information, we show that much of the transfer advantage of pretrained visual representations can be attained using coarse geometric information that is linearly recoverable from the learned embeddings, while visual representations retain a modest in-domain performance advantage. Together, these results provide a principled characterization of how representation choice impacts cross-script robustness in handwriting layout synthesis.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
handwriting synthesis, handwriting layout synthesis, generalisation, token representations, representation analysis

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_9DwnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies which token representations support handwriting layout prediction on an unseen script. It keeps the layout model and training procedure fixed and compares character-compositional embeddings, an expanded-vocabulary control, Unicode token length, explicit rendered-glyph geometry, and frozen ImageNet-pretrained ResNet-18 features. English, Chinese, and Bengali are used in leave-one-script-out experiments. The visual representation improves on the expanded-vocabulary baseline in all three transfer directions. In the main English+Chinese-to-Bengali experiment, a width-only representation achieves cross-script layout loss close to the visual representation, although its in-domain loss is higher. Linear probes show that rendered width is recoverable from the visual features on the held-out script. Reconstruction with ground-truth handwritten crops provides an additional evaluation of the predicted layouts, rather than an end-to-end evaluation with a learned handwriting renderer.

Critical strengths

1.The comparison separates vocabulary coverage from the benefit of rendered visual input. In Table 2a, expanding the character vocabulary reduces cross-script loss from 160.934 to 22.314, and VIS reduces it further to 14.521; these values use the table's scaled units. Table 2b also shows lower VIS loss than EVCC in the other two transfer directions. This is useful evidence that the visual representation offers more than a remedy for missing vocabulary entries. Keeping the layout backbone fixed makes the representation comparison easier to interpret. 2.The simple geometry baselines make the empirical finding more informative. On English+Chinese-to-Bengali, W obtains 14.201 ± 0.498 cross-script loss versus 14.521 ± 0.828 for VIS, while VIS has lower in-domain loss (Table 2a). The width probe also transfers well, with cross-script R² = 0.896 (Table 3). Together, these results identify rendered width as a useful low-dimensional input in this setting, without requiring a new layout architecture. The value is the demonstrated sufficiency of a simple feature for comparable transfer performance; the small difference in means does not establish that W is superior to VIS.

Critical weaknesses

1.The evidence for the geometric explanation is narrower than the cross-script representation claim. W, WH, WHD, and the probe analysis are evaluated only for English+Chinese-to-Bengali. The other two directions compare EVCC with VIS, so they establish the visual representation's advantage over that baseline but do not test whether coarse geometry recovers it. This distinction matters because Chinese uses predominantly character-level tokens, whereas English and Bengali use predominantly word-level tokens (Appendix A.2). A width-based explanation may behave differently when Chinese is the target. The probe establishes that width can be decoded from VIS, not that the layout predictor uses width to obtain its transfer gains. The current evidence supports a useful result for the primary direction, but a broader characterization of the information enabling transfer needs either the corresponding geometry comparisons in the other directions or a more explicit restriction of the conclusion. 2.Inconsistent descriptions and results leave parts of the controlled comparison unresolved. Section 5 states that VIS uses a fixed font supporting all three scripts, thereby avoiding script-specific font differences, whereas Appendix A.3 specifies GoNotoCurrent-Regular.ttf for non-Hanzi/non-Korean tokens and NotoSansSC-Regular.ttf for simplified Chinese. Section 3.2.5 normalizes width and height by their training-set 95th percentiles, but Appendix A.3 uses the constants 384 and 64. These choices directly define the representations being compared. In addition, the three EVCC cross-script entries in Table 9 average to approximately 21.410 in Table 2's units, rather than the 22.314 reported in Table 2a. The seed-1001 pair in Table 9, 0.008956/0.018960, matches the other transfer direction's aggregate entries in Table 2b after rescaling. These could be reporting errors, and they do not by themselves overturn VIS's advantage, but the actual configuration and corrected results are needed to assess the study's control and reproducibility.

Questions for authors

1.Scope of the main finding: Do W, WH, and WHD recover a comparable fraction of VIS's transfer performance in the two other leave-one-script-out directions, especially when Chinese is held out? If these results are unavailable, would you restrict the coarse-geometry conclusion to English+Chinese-to-Bengali? This is a check of the existing central claim, not a request to add new scripts or a new rendering system. 2.Implementation clarification: Which font assignment and width/height normalization produced Tables 2 and 3? Please reconcile Section 5 with Appendix A.3, and Section 3.2.5 with Appendix A.3. If different fonts were used across scripts, what supports the statement that script-specific font differences were avoided? 3.Numerical clarification: Please provide the correct EVCC seed-1001 results for the primary direction and reconcile Table 9 with Table 2a. Also clarify which results outside the primary comparison are single-seed: Table 8 lists only seed 1001 for the other two directions, while Appendix A.5 broadly refers to three seeds. Correcting a table transcription error would substantially reduce this reporting concern; a change to the underlying comparison would require reassessing the results.

3 comments
AuthorsOpenReview ↗

Response To Concerns Regarding Generalisation of Width-Based Features

We thank the reviewer for their insights and appreciate that they recognise the value of our separation of vocabulary coverage and performance advantage of rendered visual input. We address your concerns regarding generalisability below; Reviewer thmM had also raised a similar concern:

The evidence for the geometric explanation is narrower than the cross-script representation claim. W, WH, WHD, and the probe analysis are evaluated only for English+Chinese-to-Bengali. The other two directions compare EVCC with VIS, so they establish the visual representation's advantage over that baseline but do not test whether coarse geometry recovers it. This distinction matters because Chinese uses predominantly character-level tokens, whereas English and Bengali use predominantly word-level tokens (Appendix A.2). A width-based explanation may behave differently when Chinese is the target.

1.Scope of the main finding: Do W, WH, and WHD recover a comparable fraction of VIS's transfer performance in the two other leave-one-script-out directions, especially when Chinese is held out? If these results are unavailable, would you restrict the coarse-geometry conclusion to English+Chinese-to-Bengali? This is a check of the existing central claim, not a request to add new scripts or a new rendering system.

We thank the reviewer for highlighting this limitation. In response, we have extended the evaluation of W, WH, and WHD to the other two leave-one-script-out directions. The results are now presented together in Table 3 (Lines 384–394). As these evaluations were conducted with a single seed (1001) due to their computational cost, we report them as additional empirical evidence of cross-direction generalization, not as statistically conclusive comparisons. The original EN+ZH → BN results retain their multi-seed evaluation.

Mean L1\mathcal{L}_{1} bounding-box loss (×103\times 10^{3}). Primary-direction results (except TL) are averaged over three seeds (mean ±\pm std.), while all other results use a single seed. TL uses a single seed in the primary direction due to structural ambiguity independent of initialization or transfer direction. Best results per transfer direction and In-Domain/Cross-Script setting are made bold.

RepresentationEN+ZH→BN (ID)EN+ZH→BN (CS)BN+EN→ZH (ID)BN+EN→ZH (CS)BN+ZH→EN (ID)BN+ZH→EN (CS)
CC (Wang et al. (2026b))8.042±0.1288.042 \pm 0.128160.934±14.761160.934 \pm 14.76110.30410.30413.93213.9328.9568.95618.96018.960
EVCC8.042±0.1288.042 \pm 0.12822.314±1.27122.314 \pm 1.27110.30410.30417.88817.8888.9568.95618.96018.960
VIS7.739±0.120\mathbf{7.739 \pm 0.120}14.521±0.82814.521 \pm 0.82810.14210.14210.20310.2038.761\mathbf{8.761}14.23114.231
W9.378±0.2189.378 \pm 0.21814.201±0.498\mathbf{14.201 \pm 0.498}10.55610.55611.26011.26010.03810.03814.52414.524
WH8.501±0.1478.501 \pm 0.14715.615±0.34215.615 \pm 0.3429.7619.76110.26010.2609.7009.70013.906\mathbf{13.906}
WHD8.280±0.1518.280 \pm 0.15115.405±0.20115.405 \pm 0.2019.594\mathbf{9.594}8.919\mathbf{8.919}9.3829.38213.94113.941

Our results show that explicit geometric representations indeed generalize across all three transfer directions, although the most effective geometric representation varies by direction.

Interestingly, in EN+BN→ZH, where Chinese is held out, WHD achieves better performance in both the ID and CS settings, suggesting that the degree of dominance of VIS itself may depend on the transfer direction.

We have therefore revised our interpretation to distinguish the general transferability of explicit geometric representations from the stronger claim that width alone is universally sufficient or superior. We believe these additional results substantially strengthen the evidence for the transferability of geometric representations while motivating further investigation into script-dependent geometric properties.

AuthorsOpenReview ↗

Response to Concerns Regarding Implementation Details - Font Use and Token Image Span Normalisation

We thank the reviewer for identifying these inconsistencies:

Implementation clarification: Which font assignment and width/height normalization produced Tables 2 and 3? Please reconcile Section 5 with Appendix A.3, and Section 3.2.5 with Appendix A.3. If different fonts were used across scripts, what supports the statement that script-specific font differences were avoided?

We would like to clarify that all reported experiments use fixed width and height normalization constants of 384 and 64, respectively.

The reference to normalization using the 95th-percentile values of the training set was inadvertently retained from an initial pilot experiment. For all reported results, we fixed the normalization constants to 384 and 64 to maintain consistency across experiments and avoid introducing variation from experiment-specific normalization values.

We have removed the incorrect description from Section 3.3.5 (formerly 3.2.5) (lines 284-285) and updated Appendix A.3 to reflect the actual implementation. (lines 822-827)

We used GoNotoCurrent-Regular.ttf for English and Bengali tokens and NotoSansSC-Regular.ttf for Chinese tokens. Both fonts belong to the Google Noto Font family and were selected to provide coverage for the scripts considered in our experiments.

We acknowledge that using fonts from the same broader family does not guarantee the elimination of script-specific typographic differences. Although the Noto fonts are developed under a common design direction, individual scripts may retain differences in glyph design and metrics. Consequently, our use of these fonts should not be interpreted as fully controlling for all script-specific font characteristics.

We have revised Section 5 (Lines 482–483) to clarify the fixed font family and removed the implication that script-specific font differences are entirely avoided. Appendix A.3 has been updated to discuss the font family selection (lines 818-821). References to a fixed font in Section 3.3 have been qualified to indicate a fixed font family (lines 274-275,lines 283-284).

AuthorsOpenReview ↗

Response to the Numerical Clarifications Regarding EVCC and the number of Random Seeds

We thank the reviewer for highlighting the reporting error in Table 14 (formerly Table 9) of the Appendix and the reference to the number of random seeds in Appendix A.5:

3.Numerical clarification: Please provide the correct EVCC seed-1001 results for the primary direction and reconcile Table 9 with Table 2a. Also clarify which results outside the primary comparison are single-seed: Table 8 lists only seed 1001 for the other two directions, while Appendix A.5 broadly refers to three seeds. Correcting a table transcription error would substantially reduce this reporting concern; a change to the underlying comparison would require reassessing the results.

We would like to clarify that the EVCC results for seed 1001 in the primary transfer direction in the table were inadvertently replaced with the corresponding results from the BN+ZH-to-EN direction. The correct seed-1001 results for the primary direction are:

RepresentationID LossCS Loss
EVCC0.0079110.021671

We have corrected the table accordingly. The discrepancy was caused by a manual transcription error in Table 14 (formerly Table 9) alone. We have verified that the experimental outputs and the aggregate EVCC results reported in Table 3 (formerly Table 2a) and elsewhere are correct.

We also acknowledge that the description in Appendix A.5 was ambiguous regarding the number of random seeds. The three-seed evaluation applies only to the primary transfer direction. The experiments for the other two transfer directions use seed 1001 alone. We have revised Appendix A.5 to state this explicitly (lines 884-887): All reported results for the Primary evaluation direction are obtained using random seeds 1001, 1729, and 2027. All results that use a single seed are obtained using the random seed 1001.

Reviewer_thmMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how token representations influence the synthesis of handwritten layouts across different writing systems. The authors compare various representation methods—including those based on character composition, expanded vocabularies, script-agnostic approaches, explicit geometric features, and pre-trained visual representations—and evaluate their performance across English, Chinese, and Bengali. Final results indicate that vocabulary coverage significantly impacts transfer effectiveness, while visual and geometric representations enhance generalization across writing systems. Furthermore, through linear probing and downstream reconstruction tasks, this paper explores the role played by transferable geometric information.

Critical strengths

Controlled empirical study. By keeping the layout model and training pipeline constant, the authors systematically compare various token representation methods, allowing for a clear analysis of the impact of each representation.

Insightful findings on transferable information. The results demonstrate that, in layout prediction tasks across different writing systems, simple geometric features—specifically token width—can achieve performance comparable to that of pre-trained visual representations. This finding offers valuable insight into the information required for transfer learning.

Critical weaknesses

To date, research into geometric features has offered only limited validation of their generalization capabilities. Comprehensive comparisons of explicit geometric representations have largely been confined to a single direction of transfer. It remains unclear whether the superior performance observed with representations based on "width" can generalize to other target writing systems.

Evaluating performance across different writing systems is complicated by potential confounding factors. For instance, while English and Bengali datasets primarily utilize word-level annotations, Chinese datasets rely predominantly on character-level annotations. Furthermore, cross-writing-system evaluations—particularly those involving Chinese—are subject to the issue of "dataset shift." These discrepancies complicate the interpretation of any observed gains from transfer. Additionally, because downstream evaluation relies on actual (ground-truth) handwritten samples rather than a learned renderer, the ability to assess end-to-end generation quality is inherently limited.

Questions for authors

1.Is it possible to assess whether the advantages observed from "width-based features" hold consistently across all three directions of transition?

2.To what extent do differences in token granularity and dataset distribution affect performance across different writing systems (cross-script)? Furthermore, is it possible to provide evidence from controlled experiments that disentangles the specific impact of each of these factors?

5 comments
AuthorsOpenReview ↗

Response To Concerns Regarding Generalisation of Width-Based Features

We would like to thank the reviewer for recognizing the value of our controlled representation comparison and the finding that simple geometric features can approach the performance of pretrained visual representations. We address concerns regarding generalisation below:

To date, research into geometric features has offered only limited validation of their generalization capabilities. Comprehensive comparisons of explicit geometric representations have largely been confined to a single direction of transfer. It remains unclear whether the superior performance observed with representations based on "width" can generalize to other target writing systems.

Is it possible to assess whether the advantages observed from "width-based features" hold consistently across all three directions of transition?

We thank the reviewer for highlighting this limitation. In response, we have extended the evaluation of W, WH, and WHD to the other two leave-one-script-out directions. The results are now presented together in Table 3 (Lines 384–394). As these evaluations were conducted with a single seed (1001) due to their computational cost, we report them as additional empirical evidence of cross-direction generalization, not as statistically conclusive comparisons. The original EN+ZH → BN results retain their multi-seed evaluation.

The results are summarized below:

Mean L1\mathcal{L}_{1} bounding-box loss (×103\times 10^{3}). Primary-direction results (except TL) are averaged over three seeds (mean ±\pm std.), while all other results use a single seed. TL uses a single seed in the primary direction due to structural ambiguity independent of initialization or transfer direction. Best results per transfer direction and In-Domain/Cross-Script setting are made bold.

RepresentationEN+ZH→BN (ID)EN+ZH→BN (CS)BN+EN→ZH (ID)BN+EN→ZH (CS)BN+ZH→EN (ID)BN+ZH→EN (CS)
CC (Wang et al. (2026b))8.042±0.1288.042 \pm 0.128160.934±14.761160.934 \pm 14.76110.30410.30413.93213.9328.9568.95618.96018.960
EVCC8.042±0.1288.042 \pm 0.12822.314±1.27122.314 \pm 1.27110.30410.30417.88817.8888.9568.95618.96018.960
VIS7.739±0.120\mathbf{7.739 \pm 0.120}14.521±0.82814.521 \pm 0.82810.14210.14210.20310.2038.761\mathbf{8.761}14.23114.231
W9.378±0.2189.378 \pm 0.21814.201±0.498\mathbf{14.201 \pm 0.498}10.55610.55611.26011.26010.03810.03814.52414.524
WH8.501±0.1478.501 \pm 0.14715.615±0.34215.615 \pm 0.3429.7619.76110.26010.2609.7009.70013.906\mathbf{13.906}
WHD8.280±0.1518.280 \pm 0.15115.405±0.20115.405 \pm 0.2019.594\mathbf{9.594}8.919\mathbf{8.919}9.3829.38213.94113.941

Our results show that explicit geometric representations indeed generalize across all three transfer directions, although the most effective geometric representation varies by direction.

Interestingly, in EN+BN→ZH , where Chinese is held out, WHD achieves better performance in both the ID and CS settings, suggesting that the degree of dominance of VIS itself may depend on the transfer direction.

We have therefore revised our interpretation to distinguish the general transferability of explicit geometric representations from the stronger claim that width alone is universally sufficient or superior (lines 420-426). We believe these additional results substantially strengthen the evidence for the transferability of geometric representations while motivating further investigation into script-dependent geometric properties.

AuthorsOpenReview ↗

Response to Concerns Regarding Token Granularity

We thank the reviewer for their insights on our work and discuss the concern regarding token granularity below:

To what extent do differences in token granularity and dataset distribution affect performance across different writing systems (cross-script)? Furthermore, is it possible to provide evidence from controlled experiments that disentangles the specific impact of each of these factors?

We agree that token granularity is an important consideration when interpreting cross-script layout prediction results. We intentionally chose to retain the annotation conventions of the dataset of each script rather than impose a uniform character-level segmentation.

Imposing a uniform character-level segmentation across scripts would not necessarily provide a controlled or task-equivalent comparison. Chinese characters are commonly realized as independent spatial units, whereas English letters in cursive handwriting are often connected within words, and Bengali characters within a word are visually connected through the horizontal bar at the top of each glyph, called the shirorekha. Splitting these words into individual characters could therefore introduce segmentation boundaries that are not meaningful in the corresponding handwriting, and introduce segmentation assumptions that are not present in the original annotations. We therefore retain the original annotations and qualify our conclusions accordingly: the experiments compare representation performance under the datasets' native annotation protocols, rather than under a unified segmentation protocol.

We now report in Table 7 the mean and standard deviation of ground-truth token bounding-box widths and heights for each script both normalised using the paragraph width (the same normalisation scheme used in our dataset preprocessing), along with the average number of annotated tokens per line, to make differences in token geometry and annotation granularity explicit. We have now extended the discussion of the Tokenisation in Appendix A.2 (lines 765-795).

ScriptNormalized width mean ± stdNormalized height mean ± stdTokens per line
BN0.093020 ± 0.0492150.041086 ± 0.0130737.548
ZH0.030705 ± 0.0124050.037557 ± 0.01382525.888
EN0.082678 ± 0.0575300.036947 ± 0.0169778.503
Widths and heights are normalized by paragraph width; standard deviations are population values.
AuthorsOpenReview ↗

Response to Concerns Regarding the Dataset Shift In the ICDAR and CASIA Test-Sets

We thank the reviewer for their concerns regarding the dataset shift. We address them below:

cross-writing-system evaluations—particularly those involving Chinese—are subject to the issue of "dataset shift." These discrepancies complicate the interpretation of any observed gains from transfer.

We agree that the Chinese evaluation combines script shift and a dataset shift, complicating the attribution of performance differences exclusively to script shift. We had acknowledged this limitation in Section 5.

We retain the ICDAR2013 evaluation protocol for comparability with Wang et al. 2026(b). The ICDAR2013 dataset shares no writers with the released CASIA-HWDB/OLHWDB databases, thus also assessing generalization to previously unseen writers.

**To provide a reference unaffected by the ICDAR2013 dataset shift, we now additionally report results on the CASIA validation set (CASIA 2.0–2.2), whose writers are disjoint from the training set in Appendix B.2, Tables 8–10. **

Across all three directions, the CASIA validation results yield lower L1 losses than the corresponding ICDAR2013 results, retaining the broader patterns discussed in Section 4.

The validation sets of the seen-scripts were used only for model selection and early stopping, not gradient-based training. The unseen script's validation set played no role at all in either process. We acknowledge that since the validation set results influenced checkpoint selection, the results may be optimistic and should not be treated as an independent test estimate. Nevertheless, they provide an in-domain comparison with ICDAR2013 under the CASIA distribution.

Table 8: EN+ZH → BN

Mean L1\mathcal{L}_{1} bounding-box loss (×103)(\times 10^{3}) for EN+ZH→BN. Since Chinese is in-domain, we report the Chinese evaluation using both the ICDAR2013 test set and the CASIA validation set; the CS evaluation on Bengali is unchanged. Primary-direction results (except TL) are averaged over three seeds (mean ± std.); best results in each column are bold.

RepresentationID(ICDAR Testset)ID(CASIA ValidationSet)CS
CC8.042±0.1287.610±0.132160.934±14.761
EVCC8.042±0.1287.610±0.13222.314±1.271
VIS7.739±0.1207.513±0.07714.521±0.828
TL11.31610.85917.337
W9.378±0.2188.986±0.21214.201±0.498
WH8.501±0.1478.033±0.28015.615±0.342
WHD8.280±0.1517.904±0.17515.405±0.201

Table 9: BN+EN → ZH

Mean L1\mathcal{L}_{1} bounding-box loss (×103)(\times 10^{3}) for BN+EN→ZH. Since Chinese is the held-out target script, we report the CS evaluation using both the ICDAR2013 test set and the CASIA validation set; the ID evaluation is unchanged. Results use a single seed; best results in each column are bold.

RepresentationIDCS (ICDAR Testset)CS (CASIA Validation Set)
CC10.30413.93212.935
EVCC10.30417.88816.657
VIS10.14210.2039.672
TL---
W10.55611.26010.621
WH9.76110.2609.646
WHD9.5948.9198.136

Table 10: BN+ZH → EN

Mean L1\mathcal{L}_{1} bounding-box loss (×103)(\times 10^{3}) for BN+ZH→EN. Since Chinese is in-domain, we report the Chinese evaluation using both the ICDAR2013 test set and the CASIA validation set; the CS evaluation on English is unchanged. Results use a single seed; best results in each column are bold.

RepresentationID (ICDAR Testset)ID (CASIA Validation Set)CS
CC8.9568.51718.960
EVCC8.9568.51718.960
VIS8.7618.40414.231
TL---
W10.0389.69314.524
WH9.7009.35813.906
WHD9.3829.02813.941

We therefore clarify that the main Chinese results in Table 3 reflect performance under combined cross-script and cross-dataset differences, rather than an isolated measurement of script shift.

Reviewer_K3DwOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates different token representations for cross-script handwriting layout synthesis (HLS). Specifically, character-compositional, expanded-vocabulary control, script-agnostic token-length, explicit glyph-based, and frozen visual representations are extensively compared and discussed using four datasets in three languages.

Critical strengths

None

Critical weaknesses
  1. The novelty and contribution of this paper are substantially limited. The authors merely compared some well-established or superficial representations for HLS without introducing any new ideas. The experiments do not include any comparison with existing SOTA methods, which is quite unconvincing.
  2. The presentation of this paper is poor. Most technical details are unclear, for example, the overall pipeline, how EVCC works without training representations for held-out characters, the definition of ff function in Eq. 8, etc. I suggest thoroughly polishing the paper before submitting it to a top conference.
Questions for authors

None

4 comments
AuthorsOpenReview ↗

Response to Concerns Regarding Novelty and Contributions

We thank the reviewer for their critical assessment of our novelty. We have carefully revised our manuscript taking into account your concerns. Please find our response and specific changes below:

The novelty and contribution of this paper are substantially limited. The authors merely compared some well-established or superficial representations for HLS without introducing any new ideas

Our work addresses a setting distinct from conventional HLS: cross-script generalisation, where target-script samples are entirely held out during training. To our knowledge, prior HLS layout works evaluate only scripts seen during training, leaving unseen-script generalisation unexplored. Better transfer could reduce the script-specific supervision needed to adapt to new scripts, helping to address the common lack-of-data issue in conventional HLS.

Our objective is not to propose a new, custom HLS method, but to investigate a related open question: what token information aids HLS, and to what extent, even on unseen scripts? This is not equivalent to simply comparing alternative token encodings, as character identity, visual appearance, and glyph footprint all expose different types of information to the layout model, and their usefulness under script shift is not self-evident.

Our results reveal several findings that go beyond reporting the best-performing representation. We first show that representation choice substantially affects layout prediction, especially cross-script transfer. Furthermore, across all three directions in both seen- and cross-script settings, non-character-compositional representations outperform the character-compositional ones (CC and EVCC), indicating that character-compositional representations are not necessary for strong HLS performance.

We further demonstrate that more complex representations are not necessarily more performant: in the BN+EN→ZH direction, WHD moderately outperforms VIS in both the CS and ID settings. Interestingly, the best-performing model differs by transfer direction, suggesting that the information most useful for cross-script HLS may vary depending on the direction of transfer. However, some lower-dimensional representations yield worse seen-script performance than either CC or EVCC in some transfer directions, whereas VIS consistently outperforms CC and EVCC. This indicates that seen-script performance is not a reliable indicator of unseen-script performance.

Taken together, our findings demonstrate the efficacy of HLS without learned character-level embeddings, challenging the prevailing use of character-compositional representations in the existing literature, and suggest that visual features provide a more transferable basis for HLS, providing one direction for building HLS systems that better generalise to unseen scripts.

To make the underlying comparisons explicit, we have extended our work to report CC, EVCC, VIS, W, WH, and WHD results across all three transfer directions in one consolidated table (Table 3), replacing the earlier Tables 2a and 2b.

The results are below:

Mean L1\mathcal{L}_{1} bounding-box loss (×103\times 10^{3}). The best results for each direction and setting are shown in bold.

RepresentationEN+ZH→BN (ID)EN+ZH→BN (CS)BN+EN→ZH (ID)BN+EN→ZH (CS)BN+ZH→EN (ID)BN+ZH→EN (CS)
CC (Wang et al. (2026b))8.042±0.1288.042 \pm 0.128160.934±14.761160.934 \pm 14.76110.30413.9328.95618.960
EVCC8.042±0.1288.042 \pm 0.12822.314±1.27122.314 \pm 1.27110.30417.8888.95618.960
VIS7.739 ±\pm 0.12014.521±0.82814.521 \pm 0.82810.14210.2038.76114.231
W9.378±0.2189.378 \pm 0.21814.201 ±\pm 0.49810.55611.26010.03814.524
WH8.501±0.1478.501 \pm 0.14715.615±0.34215.615 \pm 0.3429.76110.2609.70013.906
WHD8.280±0.1518.280 \pm 0.15115.405±0.20115.405 \pm 0.2019.5948.9199.38213.941

We have now clarified our findings and contributions in Section 1 (lines 70 to 85). We have also added Table 1 in Section 2 (lines 108-114) to clarify the novelty and positioning of our work within the existing literature:

WorkHLS Formulation / Layout ModelToken Representation StudyMulti-Script EvaluationUnseen-Target-Script Evaluation
Joint handwriting synthesis (Nikolaidou et al., 2023; Ding et al., 2023)Layout jointly modelled during image generationNoNoNo
Pippi et al. (2025)Layout jointly modelled during image generationNoNo (extension discussed, not evaluated)No (proposed extension requires retraining)
Ren et al. (2025)Online HLS; LSTMNoNoNo
Wang et al. (2026a)Offline HLS; diffusion modelNoNoNo
Wang et al. (2026b)Offline HLS; Conditional Flow MatchingNoYes (English and Chinese)No
OursOffline HLS; fixed layout architectureYesYes (English, Chinese, and Bengali)Yes; leave-one-script-out
AuthorsOpenReview ↗

Response To Concerns Regarding Comparisons With State-of-The-Art Methods

We thank the reviewer for their insights and address their concern regarding SOTA comparisons below:

The experiments do not include any comparison with existing SOTA methods, which is quite unconvincing.

We agree that comparison with existing HLS methods is important. Our comparison is between representations, not architectures. We include a controlled comparison against the character-compositional representation (CC) used by Wang et al. (2026b), who, to our knowledge, report the strongest published performance among works that explicitly evaluate HLS as a distinct task. We retain Wang et al. (2026b)'s SoTA HLS architecture from their public codebase to minimise confounds, and fix the evaluation protocol and dataset split when analysing the impact of the input representations.

Under this baseline, we show that the previously unexplored visual representation (VIS) outperforms CC across all three transfer directions within CC's own native architectural framework.

We also considered other relevant handwriting-generation and HLS methods. Some methods jointly model layout and content (Nikolaidou et al., 2023; Ding et al., 2023; Pippi et al., 2025), making it difficult to isolate the effect of token representation on layout prediction. Other works explicitly study HLS but evaluate within-script generalisation (Ren et al., 2025; Wang et al., 2026a), rather than transfer to scripts absent from the training data. We therefore use Wang et al. (2026b) as the architectural basis for our controlled comparison, while positioning these other works in our discussion of the related literature.

We now clearly mark out that CC is the baseline representation adopted from Wang et al (2026b) in table 3 (line 386) and figure 2 (lines 432-448).

AuthorsOpenReview ↗

Response to Concerns Regarding Presentation and Technical Clarity

We thank the reviewer for their constructive feedback regarding the clarity and presentation of our work, and address their concerns below:

The presentation of this paper is poor. Most technical details are unclear, for example, the overall pipeline, how EVCC works without training representations for held-out characters, the definition of function in Eq. 8, etc. I suggest thoroughly polishing the paper before submitting it to a top conference.

We thank the reviewer for highlighting areas where the presentation could be improved and where our methodology was unclear. We have now added Section 3.1 to elucidate how the overall pipeline functions (lines 155-182):

Given a text sequence and a style reference, the input text is first tokenized according to the conventions of the writing system. In each experiment, all tokens are then mapped using the same token representation scheme, selected from the alternatives investigated in this work. These representations are projected into a common dimensionality through a fully connected layer and provided to the layout prediction model together with the style reference. Following Wang et al. 2026(b), the model predicts the four bounding-box parameters for each token. The predicted layout can subsequently be used to arrange handwritten token crops according to the predicted positions and dimensions. Across our experiments, we hold fixed the architecture and training procedure, varying only the token representation.

We acknowledge that there was ambiguity in how EVCC functions. We have included Algorithm 1 (lines 216-226) to detail precisely how EVCC is produced. We would like to clarify that the embedding entries for held out characters indeed remain untrained. This is done in order to determine whether simply letting the underlying model differentiate between characters of the target script, even in the absence of training data from the target script, improves performance. Our results show that the impact is highly dependent on transfer direction. We discuss this briefly in Section 4.2.1 (lines 371-401) and elaborate further in Appendix B.1 (lines 915-942).

We thank the reviewer for highlighting the redundant function notation in Equation 8. We have rewritten the Equation to make clear that the representations are simply tuples of the properties wiw_i, hih_i and did_i (lines 289-291):

rW(ci)=[wi]r_{W}(c_i) = [w_i]

rWH(ci)=[wi,hi],r_{WH}(c_i) = [w_i,h_i],

rWHD(ci)=[wi,hi,di]r_{WHD}(c_i) = [w_i,h_i,d_i]

W⊂WH⊂WHD\mathrm{W} \subset \mathrm{WH} \subset \mathrm{WHD}

We have also provided extensive technical information to aid reproducibility specifically in Appendix A.

A number of typographical issues and latex-typos have now been resolved. We thank the reviewer again for their constructive feedback, which has helped us improve the clarity and presentation of the paper.

Thank you for your detailed responses.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Thanking AC for taking over and summary of Review-Rebuttal Phase

We sincerely thank the Area Chair for overseeing this submission. To facilitate your assessment, we summarise our contributions, the strengths identified by reviewers, and how we addressed their principal concerns.

Summary of Contributions

We present a systematic study of input representations for Cross-Script Handwriting Layout Synthesis (HLS), evaluating zero-shot transfer across English (Alphabetic), Chinese (Logographic/Morphosyllabic), and Bengali (Abugida). Using a fixed architecture and training regime, we compare character-compositional, pretrained visual, and low-dimensional glyph-based representations under leave-one-script-out evaluation.

Our key findings are:

  • Non-character-compositional representations outperform character-compositional approaches across all settings evaluated in our work, indicating that Character-Compositional representations are not necessary for robust HLS, challenging the prevailing use of Character-compositional representations in prior work.

  • Simpler geometric representations can achieve performance comparable to complex visual representations, showing that greater representational complexity does not necessarily improve transfer.

  • The optimal representation varies across transfer directions, suggesting that effective generalisation depends on the source and target scripts.

Our findings establish visual and geometric representations as promising alternatives to character-compositional representations for HLS.

Strengths Identified by Reviewers

Reviewers thmM and 9Dwn appreciated the controlled experimental design, particularly the fixed architecture and training regime, which enable interpretable comparisons of representations. Reviewer thmM highlighted the finding that simple geometric features can approach the performance of complex visual representations. Reviewer 9Dwn recognised the potential of low-dimensional features to improve transfer without requiring architectural modifications.

Summary of and Responses to Reviewers' Main Concerns

1. Methodological Novelty (K3Dw)

We clarified that our contribution lies in the cross-script generalisation setting and its empirical findings, rather than a completely new architecture. We strengthened the Introduction and added a comparison with prior HLS work (lines 72–88 and 110–115).

2. Comparisons with State-of-the-Art Methods (K3Dw)

We clarified that our work compares token representations rather than architectures, using Wang et al. (2026b) as a controlled baseline. We explained that alternative HLS methods either jointly model layout and content or focus on single-script evaluation, making them unsuitable for isolating representation effects in cross-script transfer.

3. Generalisability of Geometric Representations (9Dwn, thmM)

We extended the evaluation of W, WH, and WHD to all three transfer directions. The additional directions use seed 1001. We revised our conclusions to distinguish the general transferability of geometric representations from the stronger claim that width alone is universally sufficient or superior.

4. Token Granularity Across Scripts (thmM)

We explained why enforcing uniform token granularity would introduce artificial boundaries absent from the original handwriting. We reported token footprint statistics (Table 7) and expanded the discussion of script-specific tokenisation in Appendix A.2 (lines 765–795).

5. CASIA-ICDAR Dataset Shift (thmM)

We additionally reported results on the chinese CASIA validation set, which is not subject to the dataset shift and retains the writer-disjoint properties of the test sets. The performance patterns observed on the test set is also evident in the validation results.

6. Font Choice and Implementation Clarifications (9Dwn)

We clarified that English and Bengali use GoNotoCurrent-Regular, while Chinese uses NotoSansSC-Regular, both from the Google Noto font family. We corrected the normalisation descriptions, confirming fixed width and height constants of 384 and 64 across experiments. We also qualified our claims to acknowledge that a shared font family does not eliminate all script-specific typographic differences. The normalisation details are clarified in Section 3.3.5 and Appendix A.3.

7. Numerical and Presentation Corrections (K3Dw, 9Dwn)

We added an overview of the experimental pipeline in Section 3.1 (lines 155–182), included Algorithm 1 for EVCC construction, and revised Equation 8 (lines 289–292). We corrected the EVCC seed-1001 transcription error in Appendix Table 14 and clarified that three seeds were used only for the primary transfer direction, while the other two directions use seed 1001. We also reconciled implementation descriptions and corrected reporting inconsistencies throughout the manuscript.

We believe these revisions and additional experiments address the reviewers' principal concerns and strengthen the empirical evidence supporting our central findings.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.