acceptodds
Withdrawn from ICLR 2027

CalliMaster: Mastering Page-level Chinese Calligraphy via Layout-guided Spatial Planning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Generating a Chinese calligraphy page requires joint control of character placement, writing style, and stroke connections across characters. These aspects are interdependent. A calligrapher’s style affects character size and spacing, while connecting strokes depend on surrounding characters. Current character-level methods ignore this page context, while existing page-level and general-purpose models often struggle to preserve coherent spatial structure without weakening brushwork fidelity. We present CalliMaster, a unified framework for page-level calligraphy generation and editing that follows the practice of "planning before writing". Given target text rendered in a standard font and a target style, CalliMaster first predicts a page layout and then renders calligraphy under that layout, adapting character rendering to the surrounding composition. A single multimodal diffusion transformer performs both layout planning and layout-guided rendering. It represents layouts and calligraphy pages in a shared visual space. The predicted layout can also be edited. Users can move, resize, insert, or delete characters. The model then replans the surrounding space and regenerates connecting strokes. The same framework further supports restoration under simulated damage and exploratory authenticity analysis. Experiments show that CalliMaster improves generation quality over specialized calligraphy and general-purpose baselines and outperforms a two-model variant on image and layout metrics.

Primary area:
generative models
Keywords:
Chinese Calligraphy, Page-level Generation, Multimodal Diffusion Models, Controllable Generation, Layout-aware Generation, Interactive Editing, Cultural Heritage Preservation

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

The paper presents CalliMaster, a unified diffusion transformer that predicts a style-conditioned page layout and then renders Chinese calligraphy. Rasterized character boxes serve as editable geometric prompts for page regeneration. Validation includes baseline comparisons, image and layout metrics, a human study, and matched-budget ablations, with reported improvements in reference similarity and human ratings. The paper also explores restoration under simulated damage and preliminary reconstruction-based authenticity analysis.

Critical strengths
  1. The layout-first formulation exposes character positions and scales as controllable intermediate variables. The examples demonstrate page regeneration after character replacement, resizing, repositioning, insertion, and deletion.
  2. The unified model improves image and layout metrics over a separate-stage pipeline under a matched training budget. A blinded study with 15 evaluators experienced in calligraphy also reports higher overall quality, style fidelity, and layout ratings.
Critical weaknesses
  1. During training, the standard-font condition image and layout map are both derived from annotated character boxes, so the condition already conveys target geometry. The construction of test-time condition images is unclear, and no baseline directly reuses their layout. Consequently, the experiments do not establish how much the predicted layout contributes beyond the supplied spatial information.
  2. The comparisons do not isolate the proposed method’s contribution. UniCalli is evaluated using pretrained column-level weights with column concatenation, without retraining on the same page-level dataset, while general-purpose baselines receive only text. As acknowledged by the authors, the matched-budget comparison changes regime mixing and box perturbation alongside parameter sharing, so it supports the complete training recipe rather than isolating the proposed mechanisms. A comparison using the same data, conditioning, and budget is needed to substantiate the benefit of sequential planning.
  3. Layout-guided editing and inter-character connections are evaluated mainly through selected examples. Although Layout–Ink Align provides a geometric proxy, no editing-specific evaluation reports edit success or connection correctness, and the human study covers only 10 cases from two calligraphers without a separate connection criterion. The reported metrics therefore provide limited evidence for these central capabilities.
Questions for authors
  1. How are the standard-font condition images constructed for the 50 test samples? Do their positions and scales use reference annotations, or a template independent of the target layout? Were training and test data split by original artwork before cropping and deduplicated across scans?
  2. Can you provide a controlled comparison with joint layout–image generation using the same backbone, data, conditioning, and training budget? In the unified-versus-separate comparison, does the advantage persist when both renderers receive identical box perturbation? These results would clarify the benefit attributable to sequential planning and parameter sharing.
  3. What evidence beyond the displayed examples supports reliable editing and appropriate inter-character connections? Please clarify whether edit success, box recovery, or connection correctness was evaluated separately, and whether the human study included edited outputs or explicit instructions for judging connections.
Reviewer_j9wiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CalliMaster, a page-level Chinese calligraphy generation framework that combines layout planning and image synthesis within a unified diffusion transformer. The method supports style-conditioned generation, layout editing, and additional applications such as restoration. Experiments report improvements in visual similarity, layout consistency, and subjective quality compared with existing approaches.

Critical strengths
  1. The paper addresses an interesting and challenging task of page-level Chinese calligraphy generation, integrating character content, artistic style, and spatial layout.

  2. The proposed framework supports flexible layout editing and demonstrates promising visual quality through quantitative comparisons and user studies.

Critical weaknesses
  1. The methodological novelty and effectiveness of the proposed framework are not sufficiently established. CalliMaster mainly integrates existing techniques, including pretrained diffusion transformers, rectified flow, and diffusion forcing, with layout-conditioned generation. The technical novelty over these approaches is not clearly demonstrated. Furthermore, the training and inference pipelines differ, as the rendering stage is trained primarily with ground-truth layouts but receives predicted layouts during inference. Although layout perturbation is introduced, its effectiveness in addressing this discrepancy is not sufficiently validated. More comprehensive ablations are needed to establish the contributions of the individual design choices.

  2. The experimental evaluation does not sufficiently support the claimed controllability and generalization. In Table 1, the reported character recognition accuracy is relatively low, while the authors acknowledge that the OCR-based metric is unreliable for cursive calligraphy. Consequently, the correctness of the generated characters remains insufficiently verified. CLIP-Score and GPT-Score also provide limited evidence of actual style fidelity. Although layout-related metrics and a user study are provided, more direct quantitative evaluations of character correctness and layout control are needed. Moreover, the experiments lack systematic evaluation on unseen content combinations and more complex layouts, making it difficult to assess generalization beyond the training distribution. Differences in conditioning inputs across baselines also raise concerns about the fairness of the comparisons.

Questions for authors
  1. During training, the rendering stage primarily uses ground-truth layouts, whereas inference relies on predicted layouts. Could the authors further explain how this discrepancy affects generation quality, especially when the predicted layouts contain errors? Is the introduced box perturbation sufficient to address this issue?

  2. Table 1 reports relatively low CharsAcc, even for real calligraphy. Could the authors further clarify how the correctness of generated characters can be reliably evaluated? Also, are CLIP-Score and GPT-Score sufficient to assess style fidelity? It would be helpful to consider more specialized metrics, such as handwriting-style distances.

  3. Although Layout EMD and related metrics are reported, they mainly measure geometric similarity. Could the authors provide more direct quantitative evaluations of character-level layout accuracy, such as bounding-box matching, positioning accuracy, and adherence to edited layouts?

  4. The user study evaluates only 10 cases involving two calligraphers. Could the authors discuss how representative these results are across different styles and layouts? Reporting inter-rater agreement or additional objective evaluations would also help strengthen the conclusions.

  5. Could the authors clarify how the training and test sets are separated in terms of character content, layout patterns, and calligraphic styles? Additional experiments on more complex layouts, longer text sequences, and unseen content–layout combinations would help establish the generalization ability of the proposed method.

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

CalliMaster generates full pages of Chinese calligraphy. It plans the character layout first, then renders the brushwork using the same diffusion model. Users can edit the boxes to change the page composition.

Critical strengths
  1. The layout is a useful control interface. The editing examples show changes to placement and character size, with the surrounding strokes regenerated along with the page. Also, the style-conditioned examples show changes in spacing and composition.

  2. The calligraphy user study is encouraging. The unified model also compares well with separate planning and rendering models.

Critical weaknesses
  1. The advance over existing guidance-based methods e.g. TextDiffuser-2 seems modest. Layout planning followed by diffusion rendering, with user control over the layout, is already established. CalliMaster trains both stages in one DiT, using a shared visual representation and separate noise schedules.
Questions for authors
  1. How is the standard-font condition c prepared for testing? Does its layout come from the reference image or from a separate initial arrangement?

  2. Were all crops and scans of each original work kept in the same split?

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

Related papers

Discussion (0)

Sign in to comment.