acceptodds
Withdrawn from ICLR 2027

From Character Priors to Controllable Chinese Glyph Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Pretrained image editors can render complex Chinese characters, yet this capability alone does not ensure transfer of a typeface from a few examples or control over its weight. We present a unified framework that turns pretrained character priors into few-shot Chinese font style transfer and explicit weight control. We first post-train Qwen-Image-Edit-2511 on font-transfer data to synthesize text-specified characters from two other-character references, without per-font tuning. We then extend the learned generator with signed weight control while keeping its generation parameters fixed. This enables font style transfer and weight control in one model and a single generation run, with the reference glyphs held fixed. On 24 adaptation-held-out targets, adaptation reduces LPIPS from 0.6832 to 0.1925 and raises foreground IoU from 0.2107 to 0.5490 relative to the unadapted base editor, with the largest gain where neither the font nor the character was seen; it also improves both metrics over the evaluated configurations of dedicated font generation models MX-Font++ and FontDiffuser. Fixed-reference control experiments change stroke weight in every requested direction and recover ordinary generation exactly at zero requested weight change, and a subspace-constrained residual improves target-weight matching, perceptual quality, and spatial fidelity together, where an unconstrained residual improves only the first. These results connect pretrained character knowledge with reference-guided, controllable glyph synthesis.

Primary area:
generative models
Keywords:
Diffusion Model, Chinese Font Generation, Cross-Character Style Transfer

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_txNzOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper investigates the problem of few-shot Chinese font generation and stroke-weight control. The authors argue that while pre-trained image models possess Chinese character generation capabilities, they struggle to directly achieve font style transfer based on a limited number of reference characters; furthermore, adjusting stroke weight often compromises the original font structure and style. To address this, the authors propose a two-stage LoRA adaptation method based on Qwen-Image-Edit-2511: first, LoRA fine-tuning is employed to generate fonts using two reference characters; subsequently, an additional LoRA module is introduced to control stroke weight, utilizing soft subspace regularization to minimize interference with the model's original generation capabilities. Experimental results demonstrate that this method outperforms various baseline approaches in font generation tasks, enables continuous stroke-weight control, and effectively maintains generation quality.

Critical strengths
  1. Qwen-Image-Edit is adapted for few-shot font generation via LoRA fine-tuning; the overall framework is streamlined, requiring no specialized font encoders or character structure modeling modules.
  2. By constraining the parameters associated with stroke-weight control to minimize overlap with those of the font-generation LoRA, the method mitigates interference between tasks, thereby improving stroke-weight control while maintaining generation quality.
  3. This approach outperforms various baseline models in font generation tasks, and ablation studies confirm the effectiveness of soft subspace regularization for stroke-weight control.
Critical weaknesses
  1. The authors have not sufficiently justified the necessity of introducing an additional stroke-weight control module into the font generation model, nor have they clearly demonstrated its practical advantages over traditional font editing methods—an issue that does not arise in vector font generation tasks.
  2. The font generation component relies primarily on existing Qwen-Image-Edit and LoRA fine-tuning techniques, while the stroke-weight control is implemented via additional LoRA layers and subspace regularization; the paper lacks novel perspectives or contributions regarding reusable techniques.
  3. Although the paper compares various LoRA control strategies, it fails to benchmark the proposed learning-based stroke-weight control method against traditional techniques such as morphological dilation, erosion, and contour adjustment, making it difficult to fully substantiate the method's necessity and advantages.
  4. While the paper evaluates stroke-weight control performance using Ink Error, LPIPS, and IoU, Ink Error measures only the proportion of foreground pixels and does not directly reflect actual stroke width; furthermore, there is a lack of independent quantitative analysis regarding the consistency of font style and the preservation of character structure following stroke-weight adjustments.
Questions for authors

N/A

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

This work trains a model to support Chinese font style transfer and fine-grained font weight control.

The model is trained on Qwen-Image-Edit-2511 in a two-stages manner. First, it learns a LoRA for general font style transfer ability given two other-character references. Then, it introduces a second LoRA to support explicit font weight control (lighter or heavier) while keeps the first learned LoRA frozen. To avoid conflict of two LoRAs, soft subspace regularization is used during training of the 2nd LoRA.

Experiments demonstrate substantial improvements over the base Qwen-Image-Edit model and favorable performance against several specialized font generation baselines.

Critical strengths
  • Font generation is valuable and expensive in real-world applications. The paper demonstrates that a general-purpose image editing model can be effectively adapted to few-shot font style transfer.
  • The weight LoRA provides an elegant mechanism for controlling stroke weight without modifying reference glyphs or introducing a separate editing stage.
  • The proposed soft subspace regularization addresses an important challenge in composing multiple adapters: introducing new controllability without degrading previously learned generation capabilities.
Critical weaknesses
  • Only Chinese font generation is discussed in this work.
  • Lack of comparison with recent SOTA image generation models, such as GPT-Image-2.5. These models reveal better reference-based generation ability.
  • Limited assessment of perceptual font quality and character correctness. The evaluation primarily relies on LPIPS, IoU, SSIM, and foreground ink coverage. Human evaluation or a reliable character-accuracy benchmark would be helpful.
  • Joint e2e training a single or two LoRAs is not discussed. Such comparisons would help establish whether the two-stage formulation is necessary.
  • Insufficient details on dataset construction. For example, how the font-transfer examples are sampled from the 1,000-font corpus, how reference and target characters are selected. Furthermore, the construction of the weight-control dataset requires additional clarification, especially regarding the availability and consistency of weight labels across different font families.
  • The proposed method lacks robustness to varying numbers of reference images and fails to benefit from additional style references.
  • Several Figures are not well-structured and not easy to read.
Questions for authors
  • Appendix shows substantial degradation when increasing the number of reference images. Could the authors clarify whether this degradation is caused by the mismatch between training and inference reference counts?
  • Is it possible to provide more details of how to construct training dataset? Is it scalable?
Reviewer_Hd2YOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a few-shot Chinese font generation framework with explicit stroke-weight control. A generation LoRA is first trained on font-transfer data using Qwen-Image-Edit-2511. The resulting generator is then frozen, and a second LoRA is trained on reference–target pairs with different font weights. This control update is scaled by the requested weight displacement. An overlap regularizer penalizes its overlap with the generation update’s output subspace to reduce interference.

Critical strengths

The framework addresses a practical need: generating characters with different stroke weights while retaining the reference typeface. The design is straightforward, and the reported results demonstrate improved font generation against the evaluated baselines and effective weight control. The soft-constraint results show improvements in both target-weight matching and image-quality metrics.

Critical weaknesses
  • The control-interface comparisons require clearer justification. According to the reported protocol, prompt-only introduces weight requests at inference, whereas the proposed controller receives dedicated cross-weight training. This comparison therefore does not isolate the benefit of the conditioning mechanism from that of additional supervision. Please clarify whether a text- or numeric-conditioned baseline receives matched training data and training budget.

  • The numeric-conditioned interface is insufficiently specified: its encoding, injection mechanism, and trainable components are unclear. The learned gate also needs a more explicit implementation description.

  • Since the overlap regularizer is a central contribution, the comparison between unconstrained, soft-constrained, and hard-projected updates in Table 21 should be included in the main paper.

Questions for authors

Please analyze how the glyph scale and layout used here relate to the model's original distribution. Could lower reference resolution, different input/output glyph scales, or multiple reference characters arranged within one image improve quality or efficiency?

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

Related papers

Discussion (0)

Sign in to comment.