acceptodds
Under review as a conference paper at ICLR 2027

HR-Font: Cross-Script Font Generation via Realization Priors and Appearance Completion

Abstract

We introduce HR-Font, a framework for extending a typeface to new scripts from a few Chinese reference glyphs. The central challenge is to translate a family’s visual language into characters whose structures differ from those observed. HR-Font addresses this challenge through two complementary forms of conditioning. The Character Realization Prior (CRP) represents same-character examples from a font bank as variations relative to neutral content and uses them to condition spatial deformation in the denoiser. Target-character appearance completion reads multi-scale local reference features to predict a spatial appearance representation of the requested glyph, supervised by training-target features. The resulting conditions guide a diffusion generator toward recognizable characters with a coherent family identity. We evaluate paired reconstruction and cross-script reference-style consistency on 47 fonts across Latin, kana, and Bopomofo. In a blinded Latin-character study with 28 designers and domain experts, HR-Font is the highest-ranked generated method, with a 77.5% Top-2 inclusion rate.

Primary area:
generative models
Keywords:
cross-lingual font generation, few-shot font generation, multi-script typeface design, diffusion models, reference-based generation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_qbzqOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper addresses the task of cross-script font generation and proposes the HR-Font method, which aims to generate Latin characters, Japanese kana, and Zhuyin symbols with consistent stylistic attributes based on a small set of Chinese reference characters. The authors argue that existing methods primarily focus on the extraction and transfer of stylistic features while overlooking the geometric structural variations and local appearance adaptation required when generating characters across different writing systems. To address this, the paper introduces two core modules: (1) Character Realization Prior (CRP), which retrieves similar fonts from an existing library using a pre-trained style encoder and provides structural priors for the generation process by analyzing the feature differences between the target character in candidate fonts and the standard glyph; and (2) Target-Character Appearance Completion (TC), which queries multi-scale local features from reference characters based on target character features and supervises the generation of aligned appearance representations using the perceptual features of the target glyph. Experimental results demonstrate that the method achieves superior style consistency and perceptual quality in cross-script font generation tasks, with expert subjective evaluations also confirming its significant advantages.

Critical strengths
  1. The paper addresses the challenge of adapting the geometric structure and local appearance of target characters during cross-script font generation; treating these two aspects as complementary modeling objectives is a sound approach.
  2. The CRP and TC components enhance the generation process by leveraging structural priors and local appearance, respectively, and are integrated into an existing diffusion-based generation framework; the overall technical approach is clear.
  3. Experiments were conducted across multiple target writing systems, accompanied by module ablation studies, comparisons of retrieval strategies, and expert subjective evaluations. The expert evaluation results, in particular, provide evidence supporting the method's effectiveness in maintaining font style consistency.
Critical weaknesses
  1. CRP requires candidate fonts to contain both the input reference characters and the target characters, thereby relying on external font libraries capable of cross-script character coverage. This prerequisite may be difficult to satisfy for writing systems with limited font resources.
  2. A key premise of CRP is that a single font maintains stylistic consistency across different writing systems. While this assumption is intuitively plausible, the paper needs to provide supporting evidence—such as displaying glyph samples of the same font across different scripts or offering analysis from other perspectives. Without this, the design rationale for CRP lacks sufficient justification.
  3. The necessity of the dynamic candidate strategy remains unclear. The paper's retrieval ablation studies show that the "Uniform Top-K" approach performs almost identically to the full method; in fact, retaining only the Top-1 candidate font yielded superior results on reconstruction metrics such as LPIPS, L1, and SSIM.
  4. Font generation tasks typically include an evaluation setting for UFUC (Unseen Fonts, Unseen Characters), yet this paper omits such an experiment. Since all target characters used in testing were present in the training data, the model's generalization capability regarding unseen characters has not yet been verified. Besides, I suggest that the authors set aside a subset of Chinese fonts and characters for evaluation. As I understand it, the strategy proposed in this paper is language-agnostic; evaluating it in this context would strengthen the paper.

The literature review is insufficient; many methods in the field of font generation have been omitted, such as CG-GAN (CVPR 2022), IF-Font (NeurIPS 2024) and numerous recent approaches.

Overall, this is an interesting approach. Unlike CF-Font, which seeks an optimal content representation, this paper achieves style transfer by identifying font examples with similar styles to serve as priors. However, the current submission feels somewhat hasty; if the authors can adequately address my concerns, I would be open to raising my score.

Questions for authors

N/A

7 comments

Thank you for identifying bank coverage, the rationale for cross-script retrieval, and generalization to unseen characters as questions that needed further investigation. We added candidate coverage statistics, a zero-donor fallback, real glyph examples across scripts, retrieval controls, a formal UFUC evaluation, and a Chinese holdout evaluation. The revised paper explicitly reports results for uniform weighting, Top-1, and dynamic routing, centers CRP’s contribution on structural guidance from same-character glyphs, and expands the related work.

Q1: CRP requires candidate fonts to contain both the input reference characters and the target characters, thereby relying on external font libraries capable of cross-script character coverage. This prerequisite may be difficult to satisfy for writing systems with limited font resources.

A1: Thank you for pointing out that low-resource scripts may lack candidates from fonts that cover both the references and the target character.

Bank coverage. Under the current retrieval rule, a glyph is eligible only if its font covers the target character and all reference characters. The test set contains 8,638 font–character pairs, each with 8 references. The median number of eligible candidates is 309, and every request has at least 10. Of the three scripts, Bopomofo has the lowest median at 175 and Latin has the highest at 418.

Target script in test setNumber of pairsAt least 10 eligible candidatesAvailable donor fonts (median)
Overall8,6388,638309
Latin3,5423,542418
Kana4,7264,726309
Bopomofo370370175

Using few or no donors. If the bank contains no eligible candidate, the model skips the CRP deformation residual and continues generation using the neutral target glyph and the reference style. For low-resource scripts, the available structural guidance depends on the coverage and quality of glyphs in the bank. On a fixed set of 256 font–character pairs, we retained 0, 1, 2, 4, or 10 of the top-ranked donors (see Figure 10 in the revised paper for the full donor-count curve). Retrieval Top-1 lowers LPIPS by 4.77% relative to using no donors. Parentheses in the tables indicate changes relative to the corresponding no-donor condition. On the 256-pair panel, a single donor gives the best paired reconstruction, and every tested nonzero donor count achieves a higher mean SC than Top-0. Across the nonzero budgets listed in the table below, L1 and LPIPS are at least 2.33% and 1.59% lower than Top-0, respectively. The examples in Figure 11 suggest that aggregating several donors can support alternative glyph forms that retain the reference style.

Retained donorsL1 ↓SSIM ↑LPIPS ↓SC ↑
00.084940.739970.179880.67724
10.07989(-5.94%)0.75157(+1.57%)0.17131(-4.77%)0.68847(+1.66%)
20.08099(-4.65%)0.74927(+1.26%)0.17437(-3.06%)0.68555(+1.23%)
40.08213(-3.31%)0.74644(+0.88%)0.17494(-2.75%)0.68377(+0.96%)
100.08296(-2.33%)0.74433(+0.59%)0.17701(-1.59%)0.68775(+1.55%)

Cyrillic bank coverage comparison. To evaluate the low-resource setting, we also tested 16 Cyrillic characters in 3 target fonts. These characters appear in neither the training character manifest nor the original bank. We compare the original bank with no donors against a bank expanded only at evaluation time to supply 4 donors. Both conditions use the same 8 Chinese reference glyphs, neutral target glyphs, noise, and sampling settings.

Cyrillic bank coverage comparison (48 pairs)L1 ↓SSIM ↑LPIPS ↓SC ↑
Original bank, no donors0.095880.720770.154050.69865
Expanded bank, 4 donors0.09157(-4.49%)0.73085(+1.40%)0.14345(-6.88%)0.73063(+4.58%)

Relative to the original bank with no donors, adding 4 donors lowers LPIPS by 6.88% and raises SC by 4.58%, with all four mean metrics improving across the three target fonts. This shows that CRP can yield gains even with very few donors.

Q2: A key premise of CRP is that a single font maintains stylistic consistency across different writing systems. While this assumption is intuitively plausible, the paper needs to provide supporting evidence—such as displaying glyph samples of the same font across different scripts or offering analysis from other perspectives. Without this, the design rationale for CRP lacks sufficient justification.

A2: Thank you for asking us to support cross-script retrieval with real glyphs and quantitative evidence.

Real glyphs. Figure 9(a) in the revised paper shows heavy, blocky strokes in row 2, thick–thin contrast in row 4, and pixel-grid outlines in rows 7–8 across Chinese, Latin, kana, and Bopomofo. Figure 9(b) overlays several real forms of each character with a red neutral skeleton in the same coordinate frame, showing variation in the reach of the left- and right-falling strokes of 「永」, the width and crossbar position of A, the loop shape of あ, and the turn and stroke-ending direction of ㄅ. These examples provide qualitative support for CRP’s design. Chinese references help identify fonts with related styles, and donors of the same character provide real examples of structural variation relative to the neutral glyph.

Retrieval versus random selection. To test whether similarities between Chinese reference glyphs carry over to the target scripts, we rank eligible candidates using only 8 Chinese reference glyphs for each of the 8,638 test requests. For each request, we measure the distance between each donor glyph and the ground-truth glyph for the same character. The same rules apply to both retrieval and random selection.

Target scriptRequestsMean LPIPS for retrieved donors ↓Mean LPIPS for 10 randomly selected donors ↓
Overall8,6380.226210.29457
Latin3,5420.214520.28439
Kana4,7260.240410.31071
Bopomofo3700.246950.31237
Test-set metric10 retrieved donors10 randomly selected donors
Nearest-donor LPIPS ↓0.124820.18783
Mean L1 ↓0.089700.11786
Nearest-donor L1 ↓0.054980.07469

Compared with randomly selecting donors from the same candidate set, the retrieved donors have a 23.21% lower overall mean LPIPS distance to the ground truth. By script, the reductions are 24.57% for Latin, 22.63% for kana, and 20.94% for Bopomofo. For retrieved donors, the LPIPS distance from the nearest donor to the ground truth is 33.55% lower, while mean L1 and the nearest donor’s L1 distance to the ground truth are 23.90% and 26.39% lower, respectively. This shows that, for the current bank and evaluated scripts, similarity rankings based on Chinese references are informative for target glyphs. CRP then uses the retrieved glyphs' feature differences relative to the neutral glyph as structural guidance.

To compare donor counts on the same 8,638 font–character pairs, we retain the top 1, 2, 4, and 10 retrieved donors and take the same number from a random sequence sampled without replacement. The variance, minimum, and maximum in the table summarize mean LPIPS across requests, while the nearest-donor column takes the minimum LPIPS for each request and then averages with equal weight per font.

Retained donorsSelection methodMean LPIPS ↓VarianceMinimumMaximumMean nearest-donor LPIPS ↓
1Retrieved using Chinese references0.192940.009820.003690.651150.19294
1Random selection0.294460.012210.013070.718100.29446
2Retrieved using Chinese references0.206890.007540.010860.646500.16196
2Random selection0.295220.008530.051190.698000.24941
4Retrieved using Chinese references0.214880.007030.013890.606250.14232
4Random selection0.294730.006640.059930.684770.21593
10Retrieved using Chinese references0.226210.006960.030310.599740.12482
10Random selection0.294570.005520.081560.657640.18783

Retrieval has lower mean LPIPS than random selection at all four donor counts. As the donor count increases from 1 to 10, the mean LPIPS of the retrieved donors rises from 0.19294 to 0.22621, while the mean LPIPS of the nearest donor falls from 0.19294 to 0.12482. With random selection, the mean LPIPS of the nearest donor also falls, from 0.29446 to 0.18783. These results support using Chinese references to retrieve glyphs for the target typeface.

Q3: The necessity of the dynamic candidate strategy remains unclear. The paper's retrieval ablation studies show that the "Uniform Top-K" approach performs almost identically to the full method; in fact, retaining only the Top-1 candidate font yielded superior results on reconstruction metrics such as LPIPS, L1, and SSIM.

A3: Thank you for highlighting the Uniform Top-K and Top-1 results, which helped us clarify the scope of CRP’s contribution. The core contribution of CRP is to represent structural variation across donor glyphs of the same target character and use it to guide deformation of the denoiser’s skip features. Dynamic weighting is one way to aggregate this structural information and is the strategy used in our default configuration. We have clarified this distinction in the contribution statement in Section 1 and in the candidate analysis in Section 4.5 (Table 7B).

On the full test set, uniform weighting and dynamic routing give similar results on all four metrics. Top-1 performs better on the reconstruction metrics L1, SSIM, and LPIPS, and the two configurations have similar mean SC.

ConditionL1 ↓SSIM ↑LPIPS ↓SC ↑
Full (dynamic routing, Top-10)0.0845490.7396380.1764350.699777
Uniform Top-K0.084568(+0.02%)0.739591(−0.01%)0.176462(+0.02%)0.699714(−0.01%)
Retrieval Top-10.080835(−4.39%)0.748479(+1.20%)0.171600(−2.74%)0.698470(−0.19%)
Random donors0.087100(+3.02%)0.733961(−0.77%)0.180701(+2.42%)0.695751(−0.58%)

Uniform weighting and dynamic routing. The percentages in the table are relative to Full. Uniform Top-K differs from Full by +0.02% in LPIPS and −0.01% in SC. Uniform weighting and dynamic routing yield nearly identical results. One possible explanation is that CRP learns during training to extract useful structural information from donor features, so it can still use those features effectively with uniform weights at inference.

Reconstruction and style with a single donor. Relative to Full, Top-1 lowers L1 and LPIPS by 4.39% and 2.74%, respectively, and raises SSIM by 1.20%. These metrics measure how close the output is to the ground-truth glyph. Style Compatibility (SC) measures consistency with the reference style, with mean values of 0.69847 for Top-1 and 0.69978 for Top-10. In Figure 11 of the revised paper, our visual inspection shows that the angle and length of the upper bar of F vary, while both forms retain the sharp corners and tapered strokes of the Chinese references. The long stroke of ゾ is more curved with Top-10 and straighter with Top-1, and both retain the stroke-width variation of the references. Based on the mean SC and these examples, we observe that using a single donor or aggregating multiple donors can produce different glyph forms while preserving recognizable features of the reference style.

Purpose of the default configuration. We use Top-10 as the default in the main experiments to evaluate how the same model aggregates structural guidance from multiple donors. Top-1 is an option that produces glyphs closer to the ground truth with the same model. The Random donors condition draws the same number of donors at random from the eligible candidate set, yielding LPIPS 2.42% higher and SC 0.58% lower than Full. These results show that CRP's structural guidance can be used with dynamic routing, uniform weighting, or a single donor. CRP's contribution is to represent structural variation through different forms of the same target character and use those variations to modulate the denoiser's skip features.

Reviewer_nvCmOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

HR-Font generates Latin, kana, and Bopomofo glyphs from a few Chinese glyphs in the desired style. It extends a FontDiffuser-based model with two modules. The Character Realization Prior retrieves similar training fonts using the Chinese references, then uses their renderings of the requested character to guide spatial deformation during denoising. Target-Character Appearance Completion uses cross-attention over multi-scale reference features to predict target-aligned appearance features. The paper evaluates 8,638 glyphs from 47 held-out fonts, reports improvements in style compatibility and perceptual similarity, and includes ablations of both modules. In a blinded Latin-glyph study with 28 designers and domain experts, HR-Font ranks highest among the generated methods.

Critical strengths

(1) CRP addresses a genuine and underexplored ambiguity in cross-script generation. Chinese reference glyphs do not determine how a Latin, kana, or Bopomofo character should be geometrically realized. Using corresponding characters from compatible donor fonts as a prior is therefore well motivated. Representing donor glyphs as deviations from neutral content and dynamically routing these deviations across spatial locations and diffusion steps is technically interesting and more flexible than static basis-feature averaging. This is the paper’s strongest contribution and could be consequential if its advantage over CF-Font- and ABW-Font-style alternatives is established.

(2) The generated glyphs are visually compelling. The qualitative and human-evaluation results provide credible evidence that the overall system is effective.

Critical weaknesses

(1) The key contribution lacks a comparison with the closest prior work. CRP uses Chinese reference glyphs to select similar donor fonts, then uses those fonts’ renderings of the requested target character to guide generation. CF-Font [1] and ABW-Font [2] also use corresponding glyphs from other fonts as generation bases, yet the paper compares against neither method and does not cite ABW-Font. A controlled comparison using the same backbone and donor bank should test CF-Font-style basis fusion, ABW-Font-style adaptive basis fusion, and CRP. Without it, the ablations show that CRP helps the authors’ baseline, but leave its advantage over the closest prior work unresolved.

(2) Limited novelty beyond CRP. TC uses content-derived queries to attend to local features from reference glyphs and form a target-aligned representation. This core design closely resembles FSFont [3]’s content-to-reference attention; multi-scale local and global style extraction also has precedent in FTransGAN [4] and VQ-Font [5]. TC adds target-feature supervision and calibration when injecting its tokens into the diffusion decoder, but the paper does not ablate these choices separately or compare against a standard local-attention module. It is therefore unclear which part of TC accounts for its measured improvement, or whether TC warrants treatment as a separate major methodological contribution.

(3) Retrieved donors are not shown. The visual results omit the retrieved fonts and their renderings of the requested character. Since these glyphs directly guide generation, readers cannot assess whether outputs synthesize new designs or closely reproduce a donor. Donor visualizations, output-to-donor comparisons, and a near-duplicate audit would address this concern.

References: [1] CF-Font: Content Fusion for Few-Shot Font Generation. CVPR, 2023. [2] Learning Adaptive Basis Fonts to Fuse Content Features for Few-Shot Font Generation (ABW-Font). Computational Visual Media Conference, 2025. [3] Few-Shot Font Generation by Learning Fine-Grained Local Styles (FSFont). CVPR, 2022. [4] Few-Shot Font Style Transfer Between Different Languages (FTransGAN). WACV, 2021. [5] Few Shot Font Generation Via Transferring Similarity Guided Global Style and Quantization Local Style (VQ-Font). ICCV, 2023.

Questions for authors

(1) Can you compare CRP with their basis-fusion mechanisms using the same backbone, donor bank, data, and reference budget? This would clarify whether CRP improves on existing uses of corresponding donor glyphs.

(2) For the examples in the qualitative figures, can you show the retrieved fonts’ target-character glyphs and report how similar each generated output is to its nearest donor?

5 comments

Thank you for identifying the closest basis-fusion methods and the specific questions about TC and donor reuse that needed further investigation. We implemented CF-style and ABW-style controls on the HR-Font backbone and compared two local attention controls and separately ablated completion supervision and injection calibration. Figure 8 in the revised paper shows the retrieved donor glyphs, generated outputs, and ground-truth glyphs for 16 audit examples. Our response to Q3 reports the LPIPS distance from each output to its nearest donor.

Q1: Can you compare CRP with their basis-fusion mechanisms using the same backbone, donor bank, data, and reference budget? This would clarify whether CRP improves on existing uses of corresponding donor glyphs.

The key contribution lacks a comparison with the closest prior work. CRP uses Chinese reference glyphs to select similar donor fonts, then uses those fonts’ renderings of the requested target character to guide generation. CF-Font [1] and ABW-Font [2] also use corresponding glyphs from other fonts as generation bases, yet the paper compares against neither method and does not cite ABW-Font. A controlled comparison using the same backbone and donor bank should test CF-Font-style basis fusion, ABW-Font-style adaptive basis fusion, and CRP. Without it, the ablations show that CRP helps the authors’ baseline, but leave its advantage over the closest prior work unresolved.

A1: Thank you for suggesting a comparison between CRP and the two closest basis-fusion methods under a common setup.

Comparison setup. All three configurations share the initialization, TC implementation, training data, and bank. They use the same test set and sampling settings, covering 47 fonts and 8,638 font–character pairs, with the same 8 reference glyphs for each pair.

How the conditions differ. We implemented both basis-fusion controls on the HR-Font backbone. CF-style uses fixed bases and weights them according to the feature distances between the bases and the reference glyphs. ABW-style updates its bases using a Gaussian mixture model and a winner–competitor rule, with weights determined by the distance from the reference features to the centers. Both controls feed the fused absolute content features of the same character into the content pathway.

CF-Font and ABW-Font both supplement the content condition with features of the same target character from basis fonts. CRP also uses same-character donors, computing their feature differences relative to the neutral glyph in the content and style encoder spaces and using those differences to deform the decoder’s skip features. Different forms of the same target character thus provide examples of how its structure can vary.

L1 ↓SSIM ↑LPIPS ↓SC ↑
CRP0.0845490.7396380.1764350.699777
CF-style0.088233(+4.36%)0.731323(-1.12%)0.185930(+5.38%)0.696626(-0.45%)
ABW-style0.089451(+5.80%)0.728815(-1.46%)0.190866(+8.18%)0.695799(-0.57%)

Relative to CRP, CF-style and ABW-style have LPIPS values that are 5.38% and 8.18% higher, respectively. At the same training budget, CRP also achieves lower LPIPS than both CF-style and ABW-style, supporting the use of donor-conditioned deformation for paired reconstruction.

Q2: Limited novelty beyond CRP. TC uses content-derived queries to attend to local features from reference glyphs and form a target-aligned representation. This core design closely resembles FSFont [3]’s content-to-reference attention; multi-scale local and global style extraction also has precedent in FTransGAN [4] and VQ-Font [5]. TC adds target-feature supervision and calibration when injecting its tokens into the diffusion decoder, but the paper does not ablate these choices separately or compare against a standard local-attention module. It is therefore unclear which part of TC accounts for its measured improvement, or whether TC warrants treatment as a separate major methodological contribution.

A2: Thank you for pointing out TC's connections to FSFont, FTransGAN, and VQ-Font and suggesting that we test the added design choices separately.

FSFont uses content queries to aggregate local reference styles, and FTransGAN and VQ-Font combine local and global style information. TC makes the local appearance of the requested glyph an explicit prediction before decoder conditioning. It uses queries from the neutral target glyph to organize features from the Chinese references into a shared spatial representation. One projection predicts spatial features extracted from the paired target glyph by a frozen VGG network. A second projection produces local appearance tokens that the decoder reads through cross-attention with calibrated strength.

We retained CRP and evaluated all configurations on the same test pairs with eight references. We used the same initialization and training recipe.

L1 ↓SSIM ↑LPIPS ↓SC ↑
Full TC0.0845490.7396380.1764350.699777
Standard local attention0.086823(+2.69%)0.734814(-0.65%)0.186727(+5.83%)0.690264(-1.36%)
Direct local attention0.087406(+3.38%)0.733341(-0.85%)0.188609(+6.90%)0.692529(-1.04%)
Without completion0.086138(+1.88%)0.735750(-0.53%)0.181146(+2.67%)0.695821(-0.57%)
Fixed κ=10.086138(+1.88%)0.736492(-0.43%)0.181101(+2.64%)0.693684(-0.87%)

Completion supervision and injection calibration. Completion supervision trains the shared target-aligned representation to predict frozen VGG features of the paired target glyph. Adaptive κ adjusts the strength of the local decoder condition based on the relative magnitudes of the global and local attention outputs.

Local attention controls. Both controls retain the decoder’s global and local cross-attention branches and add their outputs after scaling the local branch. Standard local attention keeps the neutral-target queries and the TC reader, switches the reader to standard scaled dot-product attention, removes completion supervision, and fixes κ=1. Direct local attention bypasses the reader and passes pooled and projected two-scale reference features directly to the decoder’s local cross-attention. It removes completion supervision and retains adaptive injection. Relative to the Full TC row, these controls increase LPIPS by 5.83% and 6.90%, respectively, and reduce mean SC by 1.36% and 1.04%. The lower LPIPS of Full TC against both controls supports its combined design for predicting and injecting target-aligned local appearance.

Q3: For the examples in the qualitative figures, can you show the retrieved fonts’ target-character glyphs and report how similar each generated output is to its nearest donor?

Retrieved donors are not shown. The visual results omit the retrieved fonts and their renderings of the requested character. Since these glyphs directly guide generation, readers cannot assess whether outputs synthesize new designs or closely reproduce a donor. Donor visualizations, output-to-donor comparisons, and a near-duplicate audit would address this concern.

A3: Thank you for suggesting that we show the actual retrieved donors and measure the distance from each generated glyph to its nearest donor.

Retrieved glyphs. Figure 8 in the revised paper shows 16 examples, with one row per example containing 8 Chinese references, the neutral input, 10 donors in retrieval order, the HR-Font output, and the ground-truth glyph.

Distance to the nearest donor. For each example, the table reports the donor with the lowest LPIPS to the generated glyph, along with the LPIPS between the generated glyph and the ground truth.

No.Target fontTarget characterNearest donor (LPIPS)Output → ground truth LPIPS
1FZHanBTJW_Cud (U+0064)FZYaSSJW-B (0.16816)0.13762
2FZHanBTJW_Cuǎ (U+01CE)FZFengRSTJW-B (0.19070)0.07025
3FZHanBTJW_Cuア (U+30A2)FZBingMYTJW (0.14253)0.18805
4FZLangYJW-M6 (U+0036)FZDaLTJW (0.16537)0.07236
5FZLangYJW-Mè (U+00E8)FZDaLTJW (0.13598)0.07085
6FZLangYJW-Mò (U+00F2)FZZuoZJHJW-T (0.11126)0.06359
7FZQianWTJW_Xiネ (U+30CD)FZAiSJW_Xian (0.20021)0.11276
8FZQianWTJW_XiP (U+0050)FZAiSJW_Xian (0.12679)0.04931
9FZQianWTJW_XiB (U+0042)FZAiSJW_Xian (0.16448)0.06164
10FZVDLJingQSJ-Rざ (U+3056)FZFWZhuZiHeiD (0.14973)0.09499
11FZVDLJingQSJ-RM (U+004D)FZShiTSJW (0.12066)0.14510
12FZVDLJingQSJ-Rd (U+0064)FZVariable-YouSongJ (0.07641)0.09807
13FZVDLJingQSJ-Rボ (U+30DC)FZYouSJJW_509R (0.16085)0.12897
14FZJiHXTJW_XiG (U+0047)FZHengWHJ-L (0.33908)0.16911
15FZJiHXTJW_Xiü (U+00FC)FZLiYHJW-EL (0.20016)0.08842
16FZJiHXTJW_Xi9 (U+0039)FZLiYHJW-EL (0.27956)0.12151
Median LPIPS
Output → nearest donor0.16266
Output → ground truth0.09653
Ground truth → nearest donor0.16638
Pixel-identical output and donor0/160

For P (U+0050), the LPIPS distances from the generated glyph to the nearest donor and to the ground truth are 0.12679 and 0.04931, respectively, and Figure 8 shows the visual differences. Across the 16 examples, the median LPIPS from the generated glyph to the ground truth is 0.09653, compared with 0.16266 to the nearest donor. Across 160 comparisons, no output matched a donor pixel for pixel, and in 13 of the 16 examples the output was closer to the ground truth than to its nearest donor, supporting the interpretation that the retrieved glyphs guide generation. These results support CRP’s use of donor glyphs as structural priors for the target character, guiding synthesis without directly reproducing the donor glyphs.

Reviewer_6mhhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HR-Font, a diffusion-based framework for cross-script font generation from a few Chinese reference glyphs. The framework introduces two complementary modules: Character Realization Prior (CRP) and Target-Character Appearance Completion (TC). CRP retrieves candidate glyphs from a font bank to provide structural priors, while TC employs cross-attention to predict target-aligned appearance features from reference characters. Experiments across Latin, kana, and Bopomofo demonstrate improvements in style compatibility and perceptual quality over existing methods, supported by ablation studies and human evaluations.

Critical strengths

1.The paper addresses an interesting and practical problem in cross-script font generation. The proposed framework considers both structural variations and local appearance characteristics, providing a reasonable motivation for the two complementary modules. 2.The experimental evaluation is relatively comprehensive, including comparisons with existing methods, ablation studies, and expert evaluations. The results demonstrate the effectiveness of the proposed framework in improving style consistency and perceptual quality.

Critical weaknesses

The methodological novelty of the proposed framework appears relatively limited, as its core components mainly build upon established techniques. Specifically, CRP extends retrieval-based structural conditioning, while TC relies on cross-attention for local appearance modeling. Although these components are adapted to cross-script font generation, their contributions appear to be largely incremental over existing approaches. The paper would benefit from a clearer justification of the distinctive technical contributions and advantages of the proposed designs.

Questions for authors
  1. All target characters used in the evaluation have already appeared in the generator's training data. Therefore, the generalization capability of the proposed method to completely unseen target characters remains unverified. Furthermore, CRP requires candidate fonts to contain both the target character and all reference characters. How does the framework handle cases where no eligible candidate fonts are available?
  2. The CRP module relies on an external candidate font bank and the availability of corresponding target glyph renderings. However, the sensitivity of the proposed method to the completeness of the font bank and the quality of candidate retrieval remains insufficiently explored. Can the framework maintain stable performance when the candidate pool is reduced, certain glyphs are missing, or the reference characters are corrupted by noise?
  3. The overall framework illustration (Figure 2) could be further improved in terms of visual clarity and presentation. In particular, the information flow between CRP and TC could be illustrated more clearly to improve the readability and understanding of the proposed architecture.
  4. The TC module primarily relies on cross-attention for local appearance modeling of target characters. However, its incremental contribution over existing cross-attention-based font style transfer approaches remains unclear. Could the authors further clarify the distinctive technical contributions of TC compared with existing methods?
7 comments

Thank you for pointing out where the paper needed more direct evidence on generalization to unseen characters, bank coverage, and the technical contribution of TC.

CRP uses feature differences between same-character donors and a neutral glyph to guide deformation of the denoiser’s skip features. TC predicts local appearance on the target glyph’s spatial grid before decoder conditioning, with paired-target feature supervision and calibrated injection.

We added a formal UFUC evaluation, a zero-donor fallback, and bank perturbation experiments, and redrew Figure 2 to show where CRP and TC take their inputs and inject their outputs.

Q1: All target characters used in the evaluation have already appeared in the generator's training data. Therefore, the generalization capability of the proposed method to completely unseen target characters remains unverified. Furthermore, CRP requires candidate fonts to contain both the target character and all reference characters. How does the framework handle cases where no eligible candidate fonts are available?

A1: Thank you for asking us to clarify generalization to unseen target characters and how the model handles an empty candidate set.

Unseen target characters. The original evaluation held out fonts, but the target characters appeared in the training data. Our UFUC evaluation holds out both, covering 45 fonts and 67 distinct target characters.

MethodReferencesSC ↑LPIPS ↓L1 ↓SSIM ↑
FTransGAN60.6865090.2368440.0691420.568418
FCA-GAN60.6888950.2281620.0678310.584072
GAR-Font80.6644090.2466690.0839100.431497
FontDiffuser10.7222780.1923590.0900600.452980
HR-Font10.7331270.1823240.0854750.457901
HR-Font80.7540410.1771060.0848660.467802

With one reference glyph for both methods, HR-Font improves SC by 1.50% and reduces LPIPS by 5.22% relative to FontDiffuser, while also improving L1 and SSIM. Across the reference counts reported in Table 4, HR-Font with eight references achieves the highest SC (0.754041) and lowest LPIPS (0.177106). FCA-GAN achieves the best L1 and SSIM, which measure paired reconstruction but do not fully capture reference-style consistency. These rankings reflect different aspects of quality: L1 and SSIM measure pixel and structural agreement with the paired target, LPIPS measures perceptual similarity to that target, and SC measures compatibility with the reference style. Figure 4 shows how HR-Font carries pixel textures and parallel strokes from the Chinese references into unseen target glyphs.

A candidate is eligible only if its font covers the target character and every reference character. If the bank contains no eligible candidate, the model skips the CRP deformation residual and continues generation using the neutral target glyph and the reference style.

Since CRP and TC do not rely on script-specific rules, we also tested Chinese-to-Chinese generation. Figure 5 shows that the generated glyphs retain parallel strokes, tapered terminals, and pixel-grid patterns from the references.

Q2: The CRP module relies on an external candidate font bank and the availability of corresponding target glyph renderings. However, the sensitivity of the proposed method to the completeness of the font bank and the quality of candidate retrieval remains insufficiently explored. Can the framework maintain stable performance when the candidate pool is reduced, certain glyphs are missing, or the reference characters are corrupted by noise?

A2: Thank you for suggesting these three concrete robustness tests.

We evaluated 8-shot robustness on the full main test set, using clean Top-10 inference as the baseline.

Available donor fonts (median)Donors usedL1 ↓SSIM ↑LPIPS ↓SC ↑
Clean baseline (Top-10)309100.0845490.7396380.1764350.699777
Reduced font families in bank163100.08357(-1.16%)0.74343(+0.51%)0.17818(+0.99%)0.68255(-2.46%)
Masked glyphs in bank121100.08369(-1.01%)0.74265(+0.41%)0.17785(+0.80%)0.68698(-1.83%)
Reference image noise309100.08470(+0.18%)0.74070(+0.14%)0.17972(+1.86%)0.68030(-2.78%)

Bank reduction. We retained approximately half of the font families in the bank, keeping all fonts within each retained family.

Missing glyphs. We made approximately 10% of the glyph entries in the bank unavailable and excluded a font from retrieval whenever a required reference or target glyph was unavailable.

Reference image noise. We added Gaussian noise with σ=0.05 to the reference images, shared across RGB channels, and used the perturbed references for both retrieval and generation.

Both bank perturbations reduced the number of available donor fonts while leaving enough to fill the retrieval budget, and generation performance remained stable. Reference image noise had a larger effect, particularly on SC.

Effect of using few donors. For the donor-count sweep, we fixed the bank and retrieval ranking on 256 font–character pairs and varied only the number of retained donors. Top-0 skips the CRP deformation residual and continues generation using the neutral target glyph and the reference style. The changes in parentheses in the table below are relative to Top-0, and Figure 10 in the revised paper shows the full donor-count curve.

ConditionL1 ↓SSIM ↑LPIPS ↓SC ↑
Top-00.0849390.7399660.1798830.677235
Top-10.079891(−5.94%)0.751571(+1.57%)0.171305(−4.77%)0.688466(+1.66%)
Top-20.080989(−4.65%)0.749274(+1.26%)0.174371(−3.06%)0.685551(+1.23%)
Top-40.082127(−3.31%)0.746441(+0.88%)0.174937(−2.75%)0.683767(+0.96%)
Top-80.082916(−2.38%)0.744455(+0.61%)0.175725(−2.31%)0.684713(+1.10%)
Top-100.082962(−2.33%)0.744330(+0.59%)0.177014(−1.59%)0.687750(+1.55%)
Top-150.083196(−2.05%)0.744020(+0.55%)0.176796(−1.72%)0.683758(+0.96%)
Top-200.083666(−1.50%)0.743021(+0.41%)0.177395(−1.38%)0.682769(+0.82%)

A single donor already provides useful structural information. Relative to Top-0, Top-1 reduces L1 and LPIPS by 5.94% and 4.77%, respectively, and increases mean SC by 1.66%. Expanding the donor set changes the size of these gains. For example, Top-10 and Top-1 differ on paired reconstruction, while their mean SC values are close at 0.68775 and 0.68847, respectively.

Figure 11 in the revised paper shows how these differences appear in the glyphs. Both settings generate a g with a more pronounced hook than the ground-truth glyph while preserving the heavy strokes and rightward slant of the Chinese references. Top-10 adds top and bottom bars to I, and both versions retain a slender, slightly tapered main stroke. The two kana examples also vary in curvature and stroke endings, while both variants retain the stroke-width contrast and brush movement of the references. Our visual inspection shows that a target character can have multiple forms that fit the reference style, with different distances to the ground-truth glyph.

AuthorsOpenReview ↗

Answer 3&4

Q3: The overall framework illustration (Figure 2) could be further improved in terms of visual clarity and presentation. In particular, the information flow between CRP and TC could be illustrated more clearly to improve the readability and understanding of the proposed architecture.

A3: Thank you for pointing out that the conditioning paths were unclear in the original Figure 2.

We redrew Figure 2 to label the inputs, outputs, and injection points of the two parallel pathways.

CRP. CRP obtains a retrieval prior from the Chinese references, represents donor features for the same character as differences from those of the neutral glyph, and predicts spatial deformations of skip features within the denoising blocks.

TC. TC derives queries from the neutral target glyph, reads multiscale reference features, and injects the predicted local appearance tokens into the decoder’s cross-attention before combining the local and global attention outputs. The figure uses dashed lines for training-only supervision from VGG features of the ground-truth glyph and distinguishes the donors and TC tokens fixed before sampling from CRP routing and local injection calibration updated at each denoising step.

Q4: The TC module primarily relies on cross-attention for local appearance modeling of target characters. However, its incremental contribution over existing cross-attention-based font style transfer approaches remains unclear. Could the authors further clarify the distinctive technical contributions of TC compared with existing methods?

A4: Thank you for asking us to clarify what TC contributes beyond existing local attention methods.

FontDiffuser uses intermediate denoiser features as queries to inject reference-style embeddings through cross-attention. FSFont uses content queries to aggregate local reference styles. TC makes the appearance of the requested character an explicit prediction before decoder conditioning. It uses the neutral target glyph to organize local evidence from the Chinese references into a shared spatial representation. One projection predicts spatial features extracted from the paired target glyph by a frozen VGG network, and a second projection produces local appearance tokens for the decoder. The decoder reads these tokens in parallel with the global style condition and calibrates the local contribution before adding it to the global output.

Control design. All controls retain CRP and use the same test pairs and eight references. Both local-attention controls retain the decoder’s global and local cross-attention branches and add their outputs after scaling the local branch. Standard local attention keeps the neutral-target queries and the TC reader. It uses standard scaled dot-product attention within the reader, removes completion supervision, and fixes κ=1. Direct local attention bypasses the reader and sends pooled and projected two-scale reference features directly to the decoder’s local cross-attention. It removes completion supervision and retains adaptive injection. Without completion removes only the target-feature loss. Fixed κ=1 keeps completion supervision and sets the adaptive calibration factor to one.

L1 ↓SSIM ↑LPIPS ↓SC ↑
Full TC0.0845490.7396380.1764350.699777
Standard local attention0.086823(+2.69%)0.734814(-0.65%)0.186727(+5.83%)0.690264(-1.36%)
Direct local attention0.087406(+3.38%)0.733341(-0.85%)0.188609(+6.90%)0.692529(-1.04%)
Without completion0.086138(+1.88%)0.735750(-0.53%)0.181146(+2.67%)0.695821(-0.57%)
Fixed κ=10.086138(+1.88%)0.736492(-0.43%)0.181101(+2.64%)0.693684(-0.87%)

Local attention controls. Relative to the Full TC row in the table, Standard and Direct local attention increase LPIPS by 5.83% and 6.90% and decrease mean SC by 1.36% and 1.04%, respectively. The lower LPIPS of Full TC against both controls supports its combined design for predicting and injecting target-aligned local appearance.

Completion supervision and injection calibration. Completion supervision trains the shared target-aligned representation to predict frozen VGG features of the paired target glyph. Adaptive κ adjusts the strength of the local decoder condition based on the relative magnitudes of the global and local attention outputs.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.