acceptodds
Under review as a conference paper at ICLR 2027

UC-Font: Unified Content Representation for Few-Shot Font Generation

Abstract

Few-shot font generation (FFG) aims to synthesize stylistically consistent characters from limited reference glyphs. The font content sources are restricted to either standard-font glyph images or component sequences, such as radical sequences and stroke sequences, failing to support image and text modality inputs within a unified paradigm. In this paper, we propose UC-Font, a unified framework for FFG with versatile input compatibility. It contains a Unified-Content Representer (UCR) to unify character content representation through cross-modal alignment at local and global levels, enabling flexible input of both glyph images and structural component sequences. For style aggregation, conventional cross-attention fails to adequately suppress irrelevant features, causing stylistic inconsistencies between generated fonts and target fonts even for identical character components. To tackle this issue, we present a Temperature-Adjusted Style Aggregator (TASA), which leverages component similarity-conditioned attention to strengthen the style consistency of shared character components. UC-Font could be built upon diffusion-based and flow matching-based frameworks, where FFG is formulated as a conditional generation task with fused unified contents and aggregated style features serving as conditions. To enhance the content and style feature interaction through joint attention modeling for FFG, a Dual-Projection Fusion (DPF) module is adopted. Extensive experiments show that UC-Font outperforms state-of-the-art FFG methods.

Primary area:
generative models
Keywords:
font generation, unified content representation, flow matching

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_y3rfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents UC-Font, a few-shot font generation framework that accepts a glyph image, a radical-level ideographic description sequence (IDS-R), or a stroke-level sequence (IDS-S) as content input. Its Unified-Content Representer aligns the three input modalities using local and global contrastive objectives and reconstruction losses. A Temperature-Adjusted Style Aggregator uses target–reference content similarity to adjust attention over reference style features, while a Dual-Projection Fusion module supplies content and aggregated style features to a diffusion or flow-matching generator. The paper evaluates generation on held-out Chinese characters and fonts using quantitative metrics, visual comparisons, ablations, and several generator variants. It reports strong FID results and competitive image-level fidelity, particularly in the seen-font, unseen-character setting.

Critical strengths
  1. The flexible content interface addresses a useful limitation of font generation systems that depend on a single source representation. Supporting glyph images, radical sequences, and stroke sequences within one framework is practically meaningful. Table 4 provides evidence that the unified representation improves performance over separately trained single-modality variants, especially for IDS inputs.

  2. The evaluation examines more than one generator and includes held-out fonts and characters, module ablations, and qualitative examples. The results suggest that the framework can produce visually convincing glyphs and that TASA and DPF each contribute to the reported performance, although the magnitude and cause of the gains require further clarification.

Critical weaknesses
  1. The main quantitative comparison needs clarification. Section 4.1 and Appendix D identify FLUX-based FM as the primary model, yet the SFUC entries in Table 1 match the FLUX-based DM results in Table 5, while the UFUC entries match the FLUX-based FM results in Table 6. This may be a reporting error, but the configuration behind each main result must be clear to interpret the central comparison. Tables 3 and 4 provide useful evidence that individual UC-Font components improve performance within the framework. However, comparisons with external methods using different generators and training procedures do not, on their own, establish how much of the overall advantage is attributable to UC-Font rather than the FLUX backbone.

  2. The evidence supports a benefit from TASA, but does not yet establish its proposed shared-component explanation. In Eqs. (2)–(4), target–reference content similarity is summarized into an attention temperature for each target token; the matched reference-content position is not explicitly used to select a reference style token. Table 3 already shows an improvement over ordinary cross-attention, and Fig. 6 provides illustrative examples. Neither result directly demonstrates that the style attention selects the corresponding shared component. An analysis of the attended reference regions or performance separated by component overlap would make this interpretation more convincing.

Questions for authors

Which generator configuration produced the SFUC and UFUC results in Table 1? The SFUC values match FLUX-based DM in Table 5, whereas the UFUC values match FLUX-based FM in Table 6. Is this a reporting error, or were different generators intentionally used for the two settings? Please clarify the selection and correct the table or description as appropriate.

Tables 3 and 4 assess the proposed components separately. For the image-input setting, was a simple content-and-style conditioning baseline without UCR, TASA, and DPF evaluated using the same pretrained FLUX checkpoint and comparable training and sampling settings? Such a result would help determine the contribution of the complete framework beyond the backbone.

In TASA, how is a target–reference content match linked to the reference style token that receives attention, given that the content-similarity matrix appears to influence the temperature rather than directly specify style-token correspondence? Could the authors clarify whether this correspondence is learned implicitly and provide evidence, such as attention visualizations or results grouped by component overlap, for the claimed shared-component behavior?

2 comments

We are grateful for your recognition of the practical value of our flexible multimodal content interface, which unifies glyph images, radical and stroke sequences in one framework, as well as our comprehensive evaluations across generators, held-out fonts, characters, module ablations and qualitative visualizations. We have addressed all your questions and suggestions below and integrated all revisions and new experimental results into the updated manuscript (highlighted in blue).

Response to Weakness 1 and Question 1: Which generator was used for SFUC and UFUC in Table 1? Is the mismatch between Table 1 and Table 5/6 a reporting error?

Both the SFUC and UFUC results in Table 1 were obtained using the FLUX-based FM model. The numerical results in Table 1 are correct. During the preparation of Table 5, we inadvertently placed the FM results under “FLUX-based DM” and the DM results under “FLUX-based FM.” We will swap these two numerical result blocks. After this correction, the FLUX-based FM results in Table 5 agree with the SFUC results in Table 1.This was a table preparation error, not an intentional use of different generators for SFUC and UFUC.We used the FM variant consistently for both settings in the main comparison, while the alternative generator variants are reported in the appendix. We apologize for the confusion.

Response to Weakness 1 and Question 2: Is there a matched FLUX baseline without UCR, TASA, and DPF to isolate the backbone effect?

A matched baseline without UCR, TASA, and DPF is useful for assessing the contribution of the proposed framework beyond the FLUX backbone. We have therefore trained an image-input baseline initialized from the same pretrained FLUX checkpoint and evaluated it on UFUC using the same training and sampling protocol as UC-Font. This baseline corresponds to setting (a) in the table below. As shown in the table, UC-Font achieves the best results across all metrics, supporting the contribution of UCR, TASA, and DPF beyond the shared FLUX backbone. This experiment and the corresponding analysis are added in Appendix K.

Table. Comparison between the matched FLUX baseline and UC-Font on UFUC.

ModuleImage
UCRTASADPFFID↓FID\downarrowLPIPS↓LPIPS\downarrowSSIM↑SSIM\uparrowL1↓L1\downarrowRMSE↓RMSE\downarrow
(a)22.07260.32040.39720.21090.4239
(b)✓18.51430.29760.41400.19560.4105
(c)✓✓15.96430.29520.42240.18850.4048
(d)✓✓16.74280.29630.41760.19110.4064
(e)✓✓✓15.75330.29220.42480.18570.4001

Response to Weakness 3 and Question 3: How does TASA link content matching to style attention, and is there evidence for the shared-component behavior?

We agree that the target–reference content-similarity matrix in TASA does not directly route a target token to a particular reference-style token. However, for each target query, it converts the distribution of content similarities over reference-content tokens into an entropy-based temperature. When this distribution is concentrated on a small number of reference-content tokens, its entropy is lower, yielding a lower temperature and a sharper subsequent attention distribution over the reference-style features. When the content-similarity distribution is diffuse, the higher temperature produces broader style aggregation. We added attention visualizations in Fig. 4 in the main text. In the first two illustrative examples, where the target regions have visually identifiable counterparts in the reference glyphs, TASA produces more localized attention responses than vanilla cross-attention. In examples with less evident local counterparts, the responses are broader, allowing information to be aggregated from a wider set of reference-style tokens. These visualizations illustrate how the query-dependent temperature modulates style aggregation.We also added temperature distribution in Fig. 5 in the main text. We can observe that at low ranks, the temperatures of the high-overlap group are substantially lower than those of the zero-overlap group, whose temperatures consistently cluster at higher values. This also demonstrates that TASA performs more selective style aggregation for queries with clearer content matches.

Dear Reviewer,

Thank you for your review. Our article has been updated significantly, and we hope that we addressed all of your concerns. Since the discussion period is concluding, if there is anything else you would like us to clarify, please let us know. Thank you for your time and efforts in reviewing our paper.

Thanks,

Authors

Reviewer_Vy6hOpenReview ↗
Rating
3(+1)
Confidence
4(+3)
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

UC-Font addresses few-shot font generation with interchangeable content inputs: glyph images, radical-level IDS, and stroke-level IDS. Its Unified-Content Representer (UCR) aligns these modalities through global and local contrastive learning and reconstruction objectives. The Temperature-Adjusted Style Aggregator (TASA) uses the entropy of content-token similarities to adjust style-attention temperature. A Dual-Projection Fusion module integrates content and style conditions into a diffusion or flow-matching generator.

Critical strengths
  • Supporting multiple content modalities within a shared generation framework is useful and clearly motivated.

  • The paper addresses a relevant problem: preserving local style consistency between shared components in reference and generated glyphs.

  • Experiments cover unseen characters and fonts, multiple input modalities, module ablations, and alternative generation configurations. The reported results suggest benefits from unified training and the proposed aggregation module.

Critical weaknesses
  • The contribution of UCR is insufficiently isolated. Image–IDS alignment and image–stroke cross-modal learning have precedents in CCR-CLIP and XMP-Font. UCR adds interchangeable inputs and fixed-length local representations, but Table 4 does not separate the effects of additional modality supervision, global alignment, local alignment, and reconstruction. A matched global-alignment baseline is needed to establish the specific value of local alignment.

  • Component-level semantics remain unverified. The local contrastive objective aligns same-index tokens across modalities for the same character. This does not explicitly establish correspondence between shared components across different characters. Since TASA relies on these representations to estimate local correspondence, component-level retrieval, localization, or intervention experiments would strengthen the proposed explanation.

  • TASA requires stronger comparisons and mechanism validation. FS-Font, VQ-Font/SSEM, and IF-Font/SSA already investigate local or structure-guided style aggregation. TASA’s specific addition is entropy-conditioned attention temperature. Comparison against vanilla attention alone does not establish whether the benefit comes from content-dependent temperature, a better fixed temperature, or the learnable global scale.

  • The temperature formulation needs clarification. A positive per-query temperature changes attention concentration but preserves the ranking of attention logits. Thus, improved component correspondence does not follow directly from the formulation. Moreover, if unscaled cosine similarities are passed directly to softmax as written, the resulting entropy has a restricted range. This raises a question about the magnitude of the adaptive effect, rather than demonstrating that the mechanism is ineffective.

  • Some reporting inconsistencies affect interpretation. The SFUC results in Table 1 appear to match the FLUX-based diffusion configuration in the appendix, although the main text identifies flow matching as the primary configuration. Some loss-ablation results also do not support the stated uniformly progressive improvements. These discrepancies should be clarified. Variability across independent training runs is not reported.

Questions for authors
  • Which generation configuration produces each part of Table 1, and how were configurations selected?

  • Does the implementation rescale cosine similarities before computing entropy? What are the observed temperature distributions?

  • Does the vanilla-attention baseline retain the learnable scale parameter and include temperature tuning?

  • What evidence indicates that UCR tokens represent corresponding local components across characters?

  • How are contrastive batches constructed, and how are similarities and temperatures computed with multiple references?

7 comments

We sincerely thank the reviewer for the positive assessment and recognition of our multi-modality shared framework, local style consistency design, and comprehensive experiments. We respond to all concerns below and have incorporated all corresponding revisions in the updated manuscript (highlighted in blue). 

Response to Weakness 1: Table 4 does not separate the effects of additional modality supervision, global alignment, local alignment, and reconstruction.

We supplement experiments on reconstruction losses in Table 2 of the paper. We observe that the reconstruction loss Lrec\mathcal{L}_{rec} yields a particularly large individual improvement when comparing results in (a) and (d), (g) and (h). We also add experimental results and analysis by separating the global and local alignment loss in Table 7 in Appendix. The results demonstrate that combining the global contrastive loss and local contrastive loss yields the best overall performance, while using either contrastive loss alone also brings performance improvements.

Table. Ablation study on local and global contrastive loss.

$\mathcal{L}_{c} $Image
$\mathcal{L}_{g} $mathcalL_l\\mathcal{L}\_{l} FID\\downarrow\LPIPS↓LPIPS\downarrowSSIM↑SSIM\uparrowL1↓L1\downarrowRMSE↓RMSE\downarrow
(a)✓✓15.75330.29220.42480.18570.4001
(b)✓16.14000.30110.41670.19170.4021
(c)✓16.07560.30140.41720.19190.4015

Response to Weakness 2 and Question 4: Component-level semantics remain unverified and What evidence indicates that UCR tokens represent corresponding local components across characters?

The local contrastive objective establishes cross-modal local consistency and ensures each token is different, while cross-character similarities are computed from these learned representations within TASA. Importantly, the representations and style aggregation are jointly optimized through the flow-matching and perceptual objectives, providing task-level supervision beyond contrastive losses. Table 3 demonstrates the benefit of TASA over vanilla cross-attention, and Appendix F shows improved style consistency in shared regions. Furthermore, we have added Fig. 4 in the main paper, where the first two column shows that TASA yields more concentrated attention on characters with shared components given a component-level query token. The results demonstrate our model can establish correspondence between shared components across different characters.

Response to Weakness 3: TASA requires stronger comparisons and mechanism validation.

The unique component of our TASA is the entropy-conditioned adaptive attention temperature. To disentangle whether the performance gain comes from an improved fixed temperature or our content-dependent dynamic temperature, we add an ablation study comparing multiple fixed temperature values τ=0.5,1,5\tau=0.5, 1, 5, learning scale and our entropy-conditioned adaptive τ\tau, as shown in the table below. A fixed temperature is a static hyperparameter that strikes a compromise across all samples. In contrast, our entropy-conditioned temperature dynamically adjusts the attention smoothness according to the entropy of each content feature. To control for the effect of an additional global scale, we also evaluated a matched variant that retains the same vanilla attention architecture but replaces the entropy-conditioned temperature with a learnable global scalar, λ=α/ds\lambda=\alpha/\sqrt{d_s}. Finally the learned α=0.9845\alpha=0.9845, close to 1. This control variant remains close to the standard scaled dot-product attention. This indicates that the learned global rescaling remains modest. We have added this ablation study in Appendix I.

Table. Ablation of attention temperature (fixed vs entropy-conditioned adaptive)

Attention TemperatureFID↓FID\downarrowLPIPS↓LPIPS\downarrowSSIM↑SSIM\uparrowL1↓L1\downarrowRMSE↓RMSE\downarrow
Fixed τ=0.5\tau=0.516.6894 0.2756 0.4182 0.1880 0.4059
Fixed τ=1\tau=116.7428 0.2763 0.41760.1884 0.4064
Fixed τ=5\tau=516.4687 0.2752 0.4180 0.1879 0.4059
(λ=α/ds\lambda=\alpha/\sqrt{d_s}), τ=1\tau=116.57620.29730.42230.18960.4030
Entropy-conditioned adaptive τ\tau15.75330.29220.42480.18570.4001

Response to Weakness 4: A positive per-query temperature changes attention concentration but preserves the ranking of attention logits.

TASA is intended to control the style aggregation degree, rather than reorder attention logits. As stated in the introduction, shared components should encourage faithful transfer of reference style, while unshared components require broader style aggregation. Preserving their ranking does not preserve their contributions, but sharpening an attention distribution can substantially reduce the influence of lower-scoring, unrelated style features. This is validated by our newly added Fig. 4 in the main text. The reference and target characters in the first two columns show explicit shared-component correspondences, where TASA produces more focused attention on relevant regions. By contrast, the examples in the final two columns lack direct component-level alignment, leading to broader attention under TASA. These visualizations demonstrate that TASA sharpens attention toward shared components while maintaining reasonable global coverage for dissimilar characters.

Response to Weakness 4 and Question 2: Does the implementation rescale cosine similarities before computing entropy?

Before computing entropy, the cosine similarities are multiplied by a fixed factor of 16. This scaling factor was omitted in the original manuscript. Thus, entropy is computed from scaled cosine similarities, rather than unscaled values. This expands the softmax logit range and allows more concentrated similarity distributions and greater query-dependent entropy variation. The actual variation depends on the learned representations. We have added this equation and description in the revised paper (highlighted in blue).

Response to Weakness 5 and Question 1: Inconsistencies in experimental reporting and the generation configuration used for Table 1.

All three content-input variants (Image, IDS-R, and IDS-S) in Table 1 use the FLUX-based flow-matching (FM) generator for both SFUC and UFUC. We use this variant as the primary configuration for the main comparison, while FLUX-based DM and PixelDiT-based FM are evaluated separately in the appendix to examine the applicability of our framework across generation formulations and architectures. The apparent inconsistency resulted from inadvertently swapping the FM and DM result blocks in Table 5. We have corrected Table 5 and the numerical results in Tables 1 and 6 remain unchanged. This was a writing error on our part and has been corrected in the manuscript (highlighted in blue).

Response to Weakness 5 : Some loss-ablation results also do not support the stated uniformly progressive improvements.

Adding LcL_c to the LpL_p setting improves SSIM, L1, and RMSE for both IDS-R and image input, and improves LPIPS for image input. Although FID decreases, this non-monotonic behavior does not imply that the contrastive objective or its implementation is malfunctioning. LpL_p supervises perceptual similarity of decoded outputs, but LcL_c supervises the UCR representation space by aligning global features across modalities and aligning same-index local features for the same character. It does not directly optimize FID or the perceptual distance of every generated image. Adding it changes the conditioning representations seen by the generator, and joint optimization of these objectives does not guarantee monotonic improvement on every metric. This is also consistent with the different evaluation scopes. FID compares feature distributions over a set of generated images, whereas LPIPS, SSIM, L1, and RMSE evaluate paired outputs. Consequently, improved paired structural/reconstruction metrics can coexist with a less favorable FID.

Response to Weakness 5 : Variability across independent training runs is not reported.

We add Table 12 in Appendix N, which presents the ablation studies on different training losses under the two-stage training mode. Built upon the baseline trained with only LCFMCFG\mathcal{L}_{CFM}^{CFG}, sequentially adding perceptual, contrastive, and reconstruction losses achieves persistent improvements. These consistent improvements confirm the robustness and effectiveness of training losses.

Response to Question 2: What are the observed temperature distributions?

We present the temperature distributions for similar and dissimilar target–reference character pairs in Fig. 5 in the main paper. The overlap ratio, used to measure similarity, is defined as the proportion of the target’s radical-level IDS components present in the reference, counting repeated components. Character pairs with an overlap ratio no less than 0.4 are assigned to the high-overlap similar group, and pairs with zero overlap to the zero-overlap dissimilar group. We collect temperature statistics for 200 pairs per group, with representative cases deferred to the appendix. For each sample, we sort its 72 local-token temperatures and compute the average value across samples at the same rank. We can see the high-overlap group exhibits substantially lower temperatures for low ranks compared with the zero-overlap group. The high-overlap group contains many low-temperature entries, whereas temperatures in the zero-overlap group cluster at higher values. Lower temperatures yield sharper style-attention distributions, whereas higher temperatures result in more diffuse attention. These observations also reflect that component similarity can effectively modulate the degree of attention.

Response to Question 3: Does the vanilla-attention baseline retain the learnable scale parameter and include temperature tuning?

The vanilla cross-attention baseline retains the standard fixed scaling factor 1/dk1/\sqrt{d_k} and learnable Q/K/VQ/K/V projections, but does not include an additional learnable scalar or query-dependent temperature adjustment. To control for the effect of an additional global scale, we further conduct experiments on a variant that keeps the vanilla attention architecture and introduces a learnable global scalar (λ=α/ds\lambda=\alpha/\sqrt{d_s}). The learned value is α=0.9845\alpha=0.9845, which is close to 1. This demonstrates that the learned global rescaling exerts only a modest influence. Results are reported in Appendix I.

Response to Question 5: How are contrastive batches constructed?

In our training setup, the contrastive batch consists of 40 font–character samples. For each target character, UCR receives two content views: its IDS representation and its glyph-image representation. The global contrastive loss treats representations with the same character identity as positives and different identities as negatives. The local loss treats same-character tokens with the same query index across modalities as positives and tokens with different query indices within the same character as negatives.

Response to Question 5: How are similarities and temperatures computed with multiple references?

With kk reference glyphs, we encode each reference separately and concatenate their content features along the token dimension. Since each glyph has 73 content tokens, the target-to-reference similarity matrix has size 73×73k73\times73k. For each target query, we compute cosine similarities to all reference-content tokens, scale them by 16, and apply softmax jointly over the concatenated reference-token dimension. The entropy of each resulting distribution determines a query-specific temperature. We also concatenate the encoded reference-style features. Each target-content query then attends over the full reference-style sequence, with its temperature modulating the concentration of the style-attention distribution. Thus, the 73 target queries yield 73 temperatures, regardless of the number of references. The procedure is identical to the single-reference case, with normalization performed over a larger reference-token set. We have added this description in the Appendix G.

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

This paper studies how to generate the images of style transferred Chinese fonts, and formulate this problem as a conditioned generative model of images. The input to the diffusion or flow matching model is the template structured input of the font, the styled images of this font, and the source structured input of the output target. The goal is to generate the styled font of the target. Since the input structure representation of the form has different forms, this paper first uses a designed encoder to embed all these different categories of form representations and the input style images into the embedding space, and then inserts this latent into the diffusion model cross-attentions. There are quite some design in the hybrid encoding and the joint embedding of source input and the styled reference images. The model is trained in a stnadard diffusion model style. Qualitative results show the improvement and effectiveness of this method.

Critical strengths
  • The qualative results and comparison is convincing and suggests the good performance.
  • Using diffusion image generative model in the font transferring problem is interesting and have direct practical application setting.
Critical weaknesses
  • The technical contribution seems to be minor given it is a straight combination of multiple simple techniques of diffusion model and conditioning in a novel application.
  • There are too many short terms of the modules and is not very easy to follow, especially in the main Fig.2
  • THe UFUC comparison seems not showing very good improvement over baselines.
Questions for authors
  • In the failure cases, it's interesting to see that the method is changing the reference to some similar characters, Further indicate a deep question: how to ensure the output image of the font is valid and correct, and identical to the reference character.
  • Can the author discuss the performance difference with baselines on UFUC?
2 comments

We sincerely appreciate your careful reading, positive feedback on our qualitative results and practical potential, as well as your constructive criticisms and insightful questions. Below we respond point by point to your concerns.

Response to Weakness 1: The technical contribution of our method.

We respectfully argue that our work is not merely a straightforward combination of existing diffusion conditioning techniques. First, our proposed UCR module supports glyph input from diverse modalities, a capability absent in all other existing FFG networks. Second, we introduce a local token alignment loss, which achieves cross-modal alignment while ensuring that each token corresponds to a local region rather than entangled global features. Beyond the encoder and alignment loss, we further develop the TASA module motivated by human visual perception: when the target and reference characters share common components, those shared strokes should inherit exactly the same style and when they have few or no overlapping components, generation allows reasonable stylistic diversity while remaining consistent with the reference. TASA explicitly links component similarity to the temperature parameter of the Softmax function to adaptively smooth attention maps. As validated in Fig. 4, this design aligns the attention behavior with human perceptual expectations. Together, these dedicated, task-specific designs go far beyond off-the-shelf conditioning and constitute the core technical contributions of our method.

Response to Weakness 2: There are too many short terms of the modules and is not very easy to follow, especially in the main Fig. 2.

Fig. 2(a) presents the overall framework, while Figs. 2(b)–(e) detail UCR, TASA, DPF, and the multi-granularity contrastive objective, respectively. Their colors match the corresponding modules or objective in Fig. 2(a), providing visual links between the overview and the detailed illustrations. To further alleviate this concern, we have added a notation clarification paragraph in the main text (marked in blue) and revised part of Fig. 2 for better readability. Specifically, we state: To facilitate understanding of Fig. 2, we abbreviate Unified-Content Representer as UCR, Temperature-Adjusted Style Aggregator as TASA, and Dual-Projection Fusion module as DPF. EVE_V, ERE_R, and ESE_S denote the encoders for glyph images, radical-level IDS, and stroke-level IDS, respectively. IDS-R and IDS-S refer to radical-level and stroke-level Ideographic Description Sequences, respectively. CA denotes cross-attention. In Fig. 2(a), “generative model” refers to the diffusion- or flow-matching-based backbone. The figure is intended to be read hierarchically: Fig. 2(a) shows how information flows through the framework, while the remaining panels explain the individual operations. We hope these clarifications help readers navigate the figure.

Response to Weakness 3 and Question 2: The performance difference with baselines on UFUC.

In Table 1, UC-Font with image input achieves the best FID and LPIPS among all compared methods. Its FID of 15.7533 represents a 14.55% reduction over FSTDiff (18.4361), the strongest external baseline on this metric. Its LPIPS of 0.2922 also surpasses IF-Font (0.2943), indicating closer perceptual similarity to the target glyphs. Both IDS-based variants likewise achieve lower FID than all external baselines. Beyond these leading results, our image-input model ranks second in L1 and third in SSIM and RMSE. Although FSTDiff achieves stronger structural and pixel-level reconstruction scores, this advantage does not extend to distributional fidelity or perceptual similarity. Overall, UC-Font offers leading distributional and perceptual fidelity alongside competitive structural and pixelwise accuracy.

Response to Question 1:  How to ensure the output image of the font is valid and correct, and identical to the reference character.

The reference glyph is generally a different character from the target. The objective is therefore not to reproduce the reference character’s shape, but to preserve the target character’s identity and structure while transferring the reference font’s appearance. This behavior is learned through paired supervision. During training, the target glyph and its style-reference glyphs are sampled from the same target font, while the content condition specifies the target character. The generator is trained with the Flow Matching objective Lgen\mathcal{L}{\mathrm{gen}} denoted $\mathcal{L}_{\mathrm{CFM}}^{\mathrm{CFG}} $ in the paper, which supervises the predicted velocity along the path to the ground-truth target glyph. Consequently, the supervision is on the complete target glyph, rather than on the reference glyph itself. The content representation identifies the character to generate, and the same-font reference conditions the model on the desired appearance. We also use a perceptual loss Lpercept\mathcal{L}{\mathrm{percept}}, computed between the reconstructed prediction and the ground-truth target, to complement the Flow Matching objective with feature-level image supervision. In addition, UCR’s alignment and reconstruction objectives encourage the content representation to retain character-level semantic and structural information. Together, these objectives train the model to produce the correct target character in the reference font’s style, rather than copying the reference character.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.