acceptodds
Under review as a conference paper at ICLR 2027

InkFlow: Beyond Style Transfer via Writing Flow Learning for Online Handwriting Generation

Abstract

Stylized handwriting generation aims to synthesize new content with specific writer characteristics from reference handwriting samples. Existing methods typically follow the style transfer paradigm, treating style features disentangled from reference samples as the transferable object and fusing them with target text to generate the final result. This paradigm focuses on the ”style presented in the outcome.” In contrast, we focus on ”how such outcomes are produced”: a specific handwriting result can be viewed as a manifestation of how a writer transforms textual content into trajectories under a given content condition. Accordingly, we redefine the transferable object from static style features to the underlying content-to-trajectory transformation, which we term the , and transfer it to new content generation. Based on this perspective, we propose , a writing-flow learning framework for online handwriting generation. Given a complete reference text–trajectory example, InkFlow models the underlying content-to-trajectory relationship and transfers it to new target text. Specifically, we adopt a latent diffusion framework with a task-aware VAE that learns a structured handwriting latent space with both content- and style-aware properties, and design a reference–target conditioned Diffusion Transformer to implicitly model and transfer writing flows. Furthermore, we introduce a writing-flow alignment objective to enhance the preservation of reference writing characteristics in the generated results. Extensive experiments on the IAM-OnDB and ICDAR-2013 text-line handwriting benchmarks demonstrate that our model achieves strong performance in both content fidelity and writer-style consistency. Code will be publicly available.

Primary area:
generative models
Keywords:
Stylized Handwriting Generation, Sequence Modeling, Latent Diffusion Transformer

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_xMcwOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents InkFlow, a latent diffusion framework for reference-conditioned online handwriting generation. It frames transferable writer characteristics as a “writing flow”: the relationship between reference text and its corresponding pen trajectory. The framework combines a content- and style-aware VAE, a reference–target conditioned Diffusion Transformer with asymmetric attention, and a writing-flow alignment loss using a frozen writer-discrimination head. Evaluation on CASIA-OLHWDB 2.0–2.2, ICDAR-2013, and IAM-OnDB includes recognition-based content metrics, FID, HWD, qualitative comparisons, and a user study. The reported results improve over the compared text-line baselines, particularly on Chinese handwriting, while ablations examine the contributions of latent-space regularization, reference text, attention masking, and alignment.

Critical strengths
  1. On OLHWDB, InkFlow improves AR over DiffInk by 2.03 percentage points and reduces FID from 5.42 to 2.30, with consistent improvements on ICDAR-2013. These results indicate practical value in jointly preserving content and writer style.
  2. The asymmetric attention mask improves content accuracy on both Chinese benchmarks in the ablation study, supporting its usefulness as a reusable design for reference-conditioned sequence generation.
Critical weaknesses

1.The evaluation mainly uses metrics computed from rendered images, leaving stroke order, pen-state behavior, and trajectory continuity unassessed. This limits the evidence supporting the central claim of learning transferable writing flows. 2.The paper does not report variability across repeated runs or confidence intervals. Some improvements on IAM-OnDB are relatively small, making it difficult to assess whether the empirical advantage is consistent and reliable.

Questions for authors

Reference Critical Weaknesses.

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

This paper proposes InkFlow, a writing-flow learning framework for online handwriting generation. Unlike traditional style transfer approaches that extract static style features from reference samples and fuse them with target content, InkFlow focuses on how handwriting is generated rather than how the final output looks. Specifically, the authors redefine the transferable representation of handwriting style as a writer-specific content-to-trajectory transformation pattern, termed writing flow. By learning this underlying transformation from a single reference demonstration, InkFlow aims to transfer writer-specific writing characteristics to new textual content while preserving both content fidelity and writer consistency.

Critical strengths
  1. The paper clearly defined the problem and the motivation for addressing it is well justified and easy to understand.
  2. The proposed method effectively leverages self-attention and CFG to achieve its intended learning purpose.
  3. The qualitative and quantitative results are strong enough to demonstrate the hypothesis and idea.
Critical weaknesses
  1. The authors used Handwriting Distance (HWD) to evaluate writer consistency, howerver, HWD primarily measures handwriting-style similarity rather than directly verifying whether writer identity is preserved. So, HWD alone may not be sufficient to support the claim of writer consistency.
  2. Although the perceptual module leads separate content and style, both heads operate on the same VAE latent representation. It remains unclear whether the learned representations effectively capture intended content and writer-specific style characteristics. Further analysis would help validate the effectiveness of the proposed perceptual enhancement.
  3. The style-aware head is trained with writer classification based on identities in the training dataset. While this encourages writer-discriminative representations, it is unclear whether such supervision helps capture transferable style characteristics beyond the training writers. Further analysis or ablation studies would be required to clarify its contribution to generalization.
  4. Although the motivation and main idea are clearly explained, the Method section repeats much of the Introduction, making the paper unnecessarily lengthy and difficult to follow. In particular, the Traditional Style Transfer Formulation subsection seems unnecessary, as it does not present the authors' contribution. Instead, the paper would benefit from explaining essential implementation details, such as how the trajectory is defined, why pen states (pen-down/pen-up) are needed, and how handwriting trajectories are processed as input sequences. I suggest reducing the redundancy and moving key architectural details from the Appendix to the main text to improve clarity.
Questions for authors
  1. Figure 3, Could the authors clarify how the DiT is trained using the noisy target latent xtx_t? Including the training objective and loss computation in Figure 3 would help readers better understand the training process.

  2. Figure 1, The outputs of the traditional style transfer model and InkFlow appear visually identical. Are these actual generation results from the respective models, or are they illustrative examples?

Reviewer_6uDPOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper re-examines the problem of stylized handwriting generation by focusing on the question: "What elements should be transferred across different content?" Rather than treating a writer's style as a static representation extracted from existing handwriting samples, the authors define the transferable elements as the "transformation from content to strokes"—a process unique to each writer—which they term "Writing Flow." Based on this perspective, they propose the InkFlow model, designed to learn and transfer these "Writing Flows" within a structured latent space. This concept is realized through a practical generative framework that incorporates task-aware latent modeling, joint diffusion of reference and target samples, and a "Writing Flow" alignment mechanism. Extensive experiments on benchmark datasets of Chinese and English handwritten text lines demonstrate that this approach excels in both content fidelity and the consistency of the writer's style. These results confirm that learning "Writing Flows" serves as an effective alternative for personalized online handwriting generation, outperforming traditional style-transfer methods.

Critical strengths

1. A coherent framework is proposed that fully leverages the correspondence between reference content and writing trajectories.

This work focuses on the relationship between the written content and the corresponding writing trajectories (i.e., the writing process itself), rather than relying solely on isolated style representations. This perspective is particularly significant for personalized online handwriting generation, where the organization of trajectories and the dynamic characteristics of the writing process are just as critical as the final visual appearance. The proposed framework translates this perspective into a concrete conditional generation pipeline.

2. The method achieves superior experimental results, supported by component-level ablation studies.

InkFlow demonstrates excellent performance on Chinese handwriting benchmarks. On the English IAM-OnDB dataset, it also shows consistent—albeit modest—performance gains over DiffInk. Ablation studies confirm that the reference-text conditioning, the asymmetric "Writing-Flow Mask," and the "Writing-Flow Alignment" mechanism each contribute to improved generation quality from different dimensions. Evaluations on both Chinese and English datasets further demonstrate that this approach is not limited to a single language and possesses strong generalizability.

Critical weaknesses
  1. There is insufficient evidence to support the novelty of the "writing-flow" learning concept.

The paper conceptualizes "writing-flow" as a transformation process that converts content into trajectories, specific to an individual writer; however, it fails to clearly distinguish this from existing "reference-conditioned generation" methods. While the ablation studies demonstrate the effectiveness of individual components, they do not isolate the specific contribution of the "writing-flow" concept itself. Conducting comparative experiments against other conditional control mechanisms would help strengthen the argument for its novelty.

  1. The evaluation details and robustness analysis are inadequate.

The user study protocol lacks critical information, such as the number of participants, evaluation criteria, and statistical significance. Furthermore, the performance improvement over DiffInk on the IAM-OnDB dataset is quite limited. Additionally, the paper does not provide results from multiple independent training runs, nor does it include an assessment of uncertainty. These omissions make it difficult to evaluate the robustness and generalizability of the claimed improvements.

Questions for authors
  1. Can the authors provide the complete user study protocol and the corresponding statistical evidence?

Specifically, what was the number of participants? What criteria were used to evaluate the generated handwriting? How were the comparative results presented and aggregated? Are the observed differences statistically significant? Providing this information would help establish the reliability of the user study results reported in Table 1.

  1. What evidence is there to demonstrate that the benefits derived from "writing-flow learning" exceed those provided by its individual architectural components or loss function terms alone?

Could the authors clarify how their proposed approach differs, in practical terms, from existing "reference-conditioned diffusion methods"? Furthermore, would it be possible to conduct controlled experiments comparing this method against other reference-conditioning mechanisms, while keeping the backbone network and training configuration constant? In particular, it would be highly valuable to disentangle the contribution of the "correspondence between reference text and trajectories" from the contributions of the "Writing-Flow Mask" and "Writing-Flow Alignment." Such evidence would significantly strengthen the paper’s claims regarding its core innovations.

Reviewer_P8r3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes InkFlow, a framework for online handwriting generation that introduces writing-flow learning as an alternative to conventional style transfer. Built upon a latent diffusion framework, the method combines a task-aware VAE, reference–target conditioned Diffusion Transformer, and writing-flow alignment objective. Experiments on Chinese and English handwriting benchmarks show improvements in content accuracy and style consistency.

Critical strengths
  1. The paper addresses an interesting problem of generating personalized handwriting from a single reference example, with attention to both content accuracy and writer-style consistency.

  2. The method achieves competitive results on Chinese and English benchmarks, with improvements over DiffInk in several quantitative metrics.

Critical weaknesses
  1. The methodological novelty is limited, particularly compared with DiffInk. InkFlow follows a similar latent diffusion framework, including VAE-based trajectory encoding and a conditional Diffusion Transformer. The main differences involve incorporating reference text, introducing an asymmetric attention mask, and adding auxiliary alignment objectives. These appear to be incremental modifications rather than substantial methodological advances. More importantly, while the paper introduces writing-flow learning as a new concept, it remains unclear how this formulation fundamentally differs from conventional reference-conditioned style transfer.

  2. The advantages of the proposed writing-flow learning are not sufficiently demonstrated. Although InkFlow achieves better quantitative results than DiffInk, it is unclear whether these improvements come from the proposed writing-flow formulation or from additional conditioning information and architectural modifications. The current experiments do not adequately distinguish these factors. Furthermore, comparisons with existing methods are relatively limited, particularly on ICDAR-2013 and IAM-OnDB. The qualitative analysis also mainly focuses on final generation results, without clearly illustrating what writing-flow learning offers beyond existing approaches.

Questions for authors
  1. Could the authors further clarify the main technical differences between InkFlow and DiffInk, particularly given their similar latent diffusion architectures?

  2. Although InkFlow improves several quantitative metrics, how much of the improvement can be attributed specifically to writing-flow learning rather than additional reference conditioning or other architectural changes?

  3. Could the authors provide more direct evidence that the model learns meaningful content-to-trajectory relationships, rather than simply using the reference handwriting and text as additional conditioning information?

  4. The comparisons on ICDAR-2013 and IAM-OnDB involve relatively few methods. Could the authors include more relevant baselines to better establish the advantages of InkFlow?

  5. Could the authors provide more targeted qualitative analysis to illustrate the specific benefits of writing-flow learning, rather than mainly comparing the final generated handwriting?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.