acceptodds
Under review as a conference paper at ICLR 2027

Form Generation of Lost Ancient Chinese Characters Using Component Priors by Diffusion Model

Abstract

The form generation of lost ancient Chinese characters is a meaningful but challenging task. The existing methods for form generation of lost ancient Chinese characters do not utilize domain-specific knowledge fully and provide inadequate guidance in the generation process. In this paper, we propose a diffusion-based method to improve the form generation of lost ancient Chinese characters by domain-specific knowledge of components and enhancing guidance in the generation process. In particular, to improve generation quality by component priors, we introduce a component pre-training stage to learn evolutionary patterns of components. To enhance guidance in the generation process, we inject conditional features into the bottleneck of the U-Net to improve the similarity between the generated and target character forms. In addition, to better generate the details of components, we employ a unified patch sampling strategy across stages, ensuring a direct correspondence between local component features. Extensive experiments on the Font-607 dataset show that our method improves the quality of ancient Chinese character generation. For example, our method reduces the LPIPS score from 0.2517 to 0.2346, and increases the cosine similarity from 0.7841 to 0.8094 in the Seal Script generation task. Code is available at https://anonymous.4open.science/r/Anc-Diff-81EA.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Form Generation, Diffusion Model, Component Priors, Ancient Chinese Characters

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_atE6OpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents Anc-Diff, a conditional diffusion framework for generating missing historical forms of Chinese characters from available forms in other scripts. It combines component-level pre-training on paired cross-script components, a separate condition encoder that injects features into the U-Net bottleneck, and a shared patch sampling strategy for component and character training. Evaluation uses Font-607, containing 607 characters across six scripts with training and test sets separated by character identity. Validation includes comparisons with eight baselines, image similarity metrics, expert ratings, ablations of pre-training, injection positions, and patch sampling, and paired significance tests for Seal Script generation. The reported improvements are strongest for Seal and Clerical scripts, while performance on earlier scripts is mixed.

Critical strengths
  1. Anc-Diff improves LPIPS and cosine similarity over all eight baselines for Seal and Clerical Script generation under the Regular Script input setting. Expert ratings also favor the method, and paired tests with Holm correction support the Seal Script improvements. These complementary results provide substantive evidence of better generation quality on the evaluated tasks.
  2. Pre-training on paired cross-script components offers a practical way to exploit reusable character structures when whole-character training data are limited. Table 3 shows improvements for both Seal and Clerical scripts, while the separation of character identities across training and testing, and of component-source characters from Font-607, strengthens the evidence for transfer to unseen characters.
Critical weaknesses

1.Experiments rely on a single dataset containing characters with available forms in all six scripts. It remains unclear whether the improvements extend to characters with incomplete historical records, limiting the evidence for the intended lost-character restoration setting. 2.The pretrained model uses additional component data and a different training schedule from the model trained from scratch. This makes it difficult to isolate the benefit of component-specific priors from additional data and optimization. A matched-budget comparison would strengthen the central contribution.

Questions for authors

Reference Critical Weaknesses.

5 comments
AuthorsOpenReview ↗

Author Response To Reviewer (1/2)

Thank you for the detailed and constructive feedback. Your comments help us improve the experimental comparisons and explain our methodological contributions and limitations more clearly. We have revised the manuscript accordingly and respond to your concerns below.


Q1 & W1. On component-specific gains versus generic cross-script pre-training.

In the original Table 3, the full model uses 300 pre-training and 100 character-training epochs, and the variant trains on characters for 300 epochs. These schedules are selected based on the experimental performance of each variant. This ablation demonstrates the benefit of adding component pre-training, but does not isolate the benefit of component-specific knowledge.

To distinguish component-specific gains from the benefit of additional pre-training, we add two variants.

  • The first removes component pre-training and retains the same 100 epochs of character training as the full model (w/o Pre. (100 epochs)).

  • The second replaces paired components with paired characters for generic cross-script pre-training, followed by the same 100 epochs of character training (Character pre-training). Since the component and character training sets contain 337 and 427 paired samples, respectively, we use 237 epochs of character pre-training based on 300 × 337 / 427 ≈ 237. With the same number of patches sampled per image, this setting approximately matches the full model in both patch counts and optimizer updates.

  • We also retain the original variant without component pre-training, which uses 300 epochs of character training (w/o Pre. (300 epochs)).

The experimental results are presented below and have been added to Section 4.6.1 and Table 3 of the revised manuscript.

TargetMethodCosSim ↑SSIM ↑LPIPS ↓
SSw/o Pre. (100 epochs)0.76320.48670.2729
SSw/o Pre. (300 epochs)0.80250.50620.2470
SSCharacter pre-training0.79990.50540.2468
SSFull model0.80940.50640.2346
CSw/o Pre. (100 epochs)0.83070.42770.2704
CSw/o Pre. (300 epochs)0.83530.44930.2556
CSCharacter pre-training0.83420.45070.2579
CSFull model0.85050.45920.2389

The comparison with w/o Pre. (100 epochs) demonstrates the benefit of adding component pre-training. The comparison with Character pre-training under approximately matched optimizer updates and patch counts further shows that pre-training with paired cross-script components is more effective than generic cross-script pre-training with paired characters.


Q2 & W2. On comparisons with OBSD and the injection-position ablation.

(1) Comparisons with OBSD

We retain OBSD’s original sampling procedure. In the original pipeline, the second-stage refinement converts the font style of the first-stage output to the target style. In our implementation, the diffusion stage is trained directly to generate the target style. We therefore incorporate target-style generation into the diffusion stage and do not apply a separate FontDiffuser refinement stage.

We use the same backbone architecture and character training data as OBSD, with matched training updates to ensure a fair comparison. Our method further incorporates component priors, a sampling strategy, and a decoupled conditioning design. These task-specific designs improve generation quality. The additional component data used for pre-training are explicitly reported. The Character pre-training variant added in Q1 also outperforms OBSD without using additional component data.

(2) injection-position ablation

All variants in Table 4 use the same independent condition encoder, consisting of stacked convolutional layers. At each injection position, the conditional feature map at the corresponding spatial scale is directly concatenated with the target feature map. Concatenation itself introduces no learnable parameters, and no additional projection or attention module is added to favor the bottleneck variant.

The comparison therefore retains the same encoder architecture and fusion operation while changing the injection position. However, feature dimensions differ across U-Net scales, so the convolutional layers following concatenation may have different parameter counts. We have added these implementation details in the revised description of Section 4.6.2.

All variants use the same independent condition encoder. For each injection position, conditional features at the corresponding spatial scale are directly concatenated with the target features.

AuthorsOpenReview ↗

Author Response To Reviewer (2/2)

Q3 & W3. On source-script selection, early-script performance, and genuinely missing historical forms.

(1) Source-script selection.

For each target script, we evaluate generation from the other five source scripts, following the same protocol as Table 1. Table 7 reports each method's results using the source script with the lowest test LPIPS, with CosSim and SSIM evaluated under the same configuration. Source scripts are therefore selected based on test performance and may differ across methods.

(2) Early-script performance and comparison with CycleGAN.

In the Oracle Bone Inscription and Bronze Inscription generation tasks in Table 7, CycleGAN achieves lower LPIPS than our method. However, CycleGAN achieves these favorable scores because of its limited ability to transform the source images, as explained below.

CycleGAN and Anc-Diff use OBI and BI interchangeably as source and target. Direct comparison of the paired original OBI and BI images gives a CosSim of 0.7531 and an LPIPS of 0.3239, which are even better than the corresponding generated results. This indicates that favorable similarity scores can be obtained by leaving the source image unchanged, without performing cross-script translation. CycleGAN performs only limited transformations, which gives it an advantage.

As discussed in Section 4.4 and Figure 4, CycleGAN largely preserves Regular Script characteristics when generating Seal Script. Figure 12 in Appendix A.3 provides source–output–target comparisons for this task, further illustrating this pattern. The comparisons between Oracle Bone and Bronze scripts in Figures 13 and 14 in Appendix A.3 show similar source-preservation patterns.

Since these two scripts already share similar forms, limited changes to the source images may partly explain CycleGAN’s favorable LPIPS. On tasks with larger differences between source and target scripts, such as generating Seal Script from Regular Script, CycleGAN performs substantially worse than our method.

(3) Variation across training runs and statistical significance.

The source-preservation patterns discussed above help explain CycleGAN’s favorable LPIPS, but do not establish the statistical significance of the differences on Oracle Bone and Bronze scripts. We therefore do not claim statistically significant improvements on these two tasks.

For Seal and Clerical script generation, we report the mean and standard deviation over five independent training runs for Anc-Diff, BBDM, FontDiffuser, and OBSD. These results support the stability of our method and indicate that the performance gains are consistent across runs rather than specific to a particular training run.

TargetSourceMethodCosSim ↑SSIM ↑LPIPS ↓
SSRSBBDM0.7617±0.00560.5110±0.01230.2831±0.0049
SSRSFontDiffuser0.7555±0.00670.4662±0.00950.3140±0.0069
SSRSOBSD0.7697±0.00290.5031±0.00270.2814±0.0022
SSRSAnc-Diff (Ours)0.8087±0.00110.5092±0.00190.2348±0.0005
CSRSBBDM0.8016±0.00970.4650±0.00330.2922±0.0114
CSRSFontDiffuser0.8155±0.01410.4501±0.00730.2758±0.0107
CSRSOBSD0.8176±0.00270.4528±0.00180.2694±0.0013
CSRSAnc-Diff (Ours)0.8511±0.00160.4590±0.00100.2394±0.0010

We also report paired significance tests using scores for all 180 individual test characters in Seal Script generation in Appendix A.5. Although we have not conducted significance tests for Oracle Bone and Bronze script generation, the repeated experiments and significance tests above provide additional evidence of the stability and effectiveness of our method.

(4) Evaluation of genuinely missing historical forms.

Regarding genuinely missing historical forms, we have explored such cases and discussed the generated forms with ancient Chinese character experts. However, without surviving historical evidence, these examples can be assessed for plausibility but cannot be verified as historically authentic forms. Our quantitative evaluation therefore uses characters with all six script forms available to provide ground truth for controlled evaluation.


We hope our responses and the revisions to the manuscript address all your concerns. If anything remains unclear or insufficiently addressed, please let us know so that we can respond further.

AuthorsOpenReview ↗

Author Response To Reviewer On Q3

We have previously explained why CycleGAN achieves relatively strong results on early scripts (Appendix A.3) and presented multi-run experimental results for some methods on the SS and CS generation tasks, demonstrating that our method's performance is not attributable to randomness. Here, we provide the mean and standard deviation over five independent training runs of our method across all tasks, as shown in the table below.

TargetSourceMethodCosSim ↑SSIM ↑LPIPS ↓
OBIBIAnc-Diff0.7236±0.00230.5119±0.00180.3489±0.0008
BIOBIAnc-Diff0.7238±0.00290.4421±0.00230.3615±0.0028
SSWSSAnc-Diff0.7935±0.00180.3584±0.00150.3320±0.0013
SSRSAnc-Diff0.8087±0.00110.5092±0.00190.2348±0.0005
CSRSAnc-Diff0.8511±0.00160.4590±0.00100.2394±0.0010

For each task, we tested five source scripts and reported the best result for each method. Even under this favorable setting, all baselines except CycleGAN perform worse than ours. Using the same source script would further degrade the baselines' performance. Therefore, under the same source inputs, our results remain statistically significantly superior to those of the baseline methods across both Oracle Bone and Bronze Script tasks. As shown in the table below.

MethodTarget: OBI (Source)OBI CosSim↑OBI SSIM↑OBI LPIPS↓Target: BI (Source)BI CosSim↑BI SSIM↑BI LPIPS↓
CycleGANBI0.72170.47370.3377OBI0.72030.44030.3501
WSS0.70290.45620.3728WSS0.71940.39770.3712
SS0.70930.49530.3431SS0.72180.43630.3538
CS0.69710.43970.4017CS0.70530.36200.3968
RS0.71070.47270.3620RS0.71890.39480.3676
Pix2pixBI0.65800.54890.3551OBI0.64980.47540.3897
WSS0.67200.54500.3551WSS0.66470.47990.3822
SS0.65100.54650.3606SS0.66880.48120.3820
CS0.65280.53890.3605CS0.65980.47850.3924
RS0.68180.54390.3510RS0.68880.48920.3775
DRIT++BI0.70290.45200.3545OBI0.66780.50100.4234
WSS0.65580.56770.4294WSS0.70360.46190.3820
SS0.69190.47310.3586SS0.69920.44150.3717
CS0.68390.47610.4082CS0.68960.40910.4130
RS0.71020.46930.3588RS0.68500.47090.3985
BBDMBI0.70110.53340.3606OBI0.69630.47860.3839
WSS0.64280.55580.4117WSS0.70360.47490.3765
SS0.69420.51830.3677SS0.60050.51640.4552
CS0.68640.51640.3966CS0.67600.48900.4079
RS0.66290.53050.4087RS0.67200.49330.3856
FontDiffuserBI0.71440.47010.3656OBI0.68030.51140.4017
WSS0.69350.44060.3823WSS0.68400.48310.3869
SS0.71030.44160.3757SS0.70260.48760.3932
CS0.69740.45670.3870CS0.67950.48640.4063
RS0.69810.43910.3804RS0.69270.47540.3871
OBSDBI0.69070.53480.3846OBI0.69540.46740.3940
WSS0.65820.54680.4223WSS0.66970.47770.4199
SS0.67760.54500.3903SS0.68500.48410.3937
CS0.67150.53640.4091CS0.68890.46360.4005
RS0.68960.52690.3870RS0.70260.45250.3948
Anc-DiffBI0.72230.51070.3486OBI0.72590.44020.3647
WSS0.70730.51390.3626WSS0.71650.44410.3784
SS0.72100.50130.3536SS0.72130.44260.3620
CS0.70490.50180.3664CS0.71650.43210.3787
RS0.71420.50080.3607RS0.72410.41960.3805
Reviewer_NqFMOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces Anc-Diff to generate candidate ancient character forms from known forms in other scripts. It first pre-trains a diffusion model on corresponding components, then trains it on complete characters using patches. A separate encoder extracts features from the conditioning image and feeds them into the U-Net bottleneck. During generation, the model produces overlapping patches and combines them into a character image.

The experiments use Font-607, which contains 607 characters in six scripts, with 427 characters for training and 180 for testing. The authors simulate missing forms by withholding known targets, since genuinely lost forms have no ground truth. They compare against eight baselines and provide ablations, expert ratings, and paired statistical tests. I see convincing improvements for Seal and Clerical script generation, although the advantages are less consistent for earlier scripts. The evaluation shows better generation on this benchmark; it does not verify the historical accuracy of genuinely lost forms.

Critical strengths
  1. I find the problem worthwhile. Generating ancient character forms can involve substantial changes in structure, so ordinary font style transfer is not always enough. Pre-training on corresponding components is a reasonable way to bring domain knowledge into the model. I also appreciate the character-disjoint train/test split, the removal of duplicate forms, and the use of separate source characters for component pre-training. These choices reduce the risk of evaluating on overlapping examples.
  2. The authors provide more than a few selected visual examples. They include comparisons with eight baselines, ablations, ratings from three paleography experts, and paired statistical tests. For Regular-to-Seal generation, LPIPS improves from 0.2517 for the strongest baseline to 0.2346, and CosSim increases from 0.7841 to 0.8094. Appendix A.5 reports significant improvements in these two metrics against all eight baselines for the Seal task. The expert ratings provide additional support for the quality of the generated characters.
Critical weaknesses
  1. My main reservation is that the added modeling contribution seems limited as presented. The method keeps the diffusion and U-Net framework and combines component pre-training, bottleneck conditioning, and patch sampling. I would like a clearer explanation of what separates this approach from related methods such as OBSD. Moving conditioning from the input stage to the bottleneck improves CosSim from 0.8027 to 0.8094 and LPIPS from 0.2423 to 0.2346 in Table 4, but this is a modest gain. More importantly, Table 3 does not isolate why component pre-training helps. The full model uses 300 epochs of component pre-training and 100 epochs of character training, while the model without pre-training uses 300 epochs of character training. Since both the data and the training schedule change, I cannot tell how much of the improvement comes specifically from component correspondences rather than the additional pre-training.
  2. I am less convinced by the results for the earlier scripts. In Appendix Table 7, Oracle Bone CosSim is only 0.0003 above the strongest baseline, and LPIPS is worse than the strongest baseline for both Oracle Bone and Bronze scripts. There are small gains for Warring States script, with clearer gains for Seal and Clerical scripts. The authors acknowledge the difficulty of earlier scripts in Appendix A.3, which is helpful, but I think this limitation should be more prominent in the paper’s claims. With only one dataset and no reported variation across independent training runs, I would be cautious about drawing broader conclusions from these results.
Questions for authors
  1. I agree that Table 3 shows a benefit from adding component pre-training, but I am not sure it shows what causes that benefit. Both the training data and the schedule change. What evidence separates the value of cross-script component correspondences from the benefit of additional pre-training? A comparison with similar data scale and training updates, for example using paired character patches rather than specifically selected components, would help me assess this claim.
  2. I am not sure what “global layout” means in the explanation of bottleneck fusion. Section 3.3 describes patch-based inputs, and Section 3.1 says that patches are generated independently. Does the model see the whole character or exchange information across patches, or does the bottleneck capture context only within each patch? Please also explain how spatially corresponding patches remain useful when components move between scripts. The whole-character examples and expert scores are useful, but I would like to understand how the conditioning and aggregation steps maintain consistency across patches.
6 comments
AuthorsOpenReview ↗

Author Response To Reviewer (1/2)

Thank you for the careful reading and constructive feedback. Your comments have helped us strengthen our work. We address your concerns below and have revised the manuscript accordingly.


Q1. On the benefit of component pre-training versus additional training.

In the original Table 3, the full model uses 300 epochs of component pre-training and 100 epochs of character training, and the variant without component pre-training uses 300 epochs of character training. These schedules are selected based on the experimental performance of each variant. We agree that the original comparison does not sufficiently explain what causes the improvement, since both the training data and the training schedule change.

We therefore include two additional variants.

  • The first removes component pre-training and uses the same 100 epochs of character training as the full model (w/o Pre. (100 epochs)).

  • The second follows the reviewer’s suggestion by replacing paired components in the first stage with paired characters (Character pre-training). Since the component and character training sets contain 337 and 427 paired samples, we use 237 epochs of character pre-training to obtain comparable training updates (300 × 337 / 427 ≈ 237 epochs). This is followed by the same 100 epochs of character training.

  • We also retain the original variant without component pre-training, which uses 300 epochs of character training (w/o Pre. (300 epochs)).

The experimental results are presented below and have been added to Section 4.6.1 and Table 3 of the revised manuscript.

TargetMethodCosSim ↑SSIM ↑LPIPS ↓
SSw/o Pre. (100 epochs)0.76320.48670.2729
SSw/o Pre. (300 epochs)0.80250.50620.2470
SSCharacter pre-training0.79990.50540.2468
SSFull model0.80940.50640.2346
CSw/o Pre. (100 epochs)0.83070.42770.2704
CSw/o Pre. (300 epochs)0.83530.44930.2556
CSCharacter pre-training0.83420.45070.2579
CSFull model0.85050.45920.2389

The comparison with w/o Pre. (100 epochs) variant demonstrates the benefit of introducing component pre-training before character training. The comparison with Character pre-training variant under approximately matched training updates further shows that this benefit cannot be explained solely by additional training updates. Pre-training with paired components is more effective than pre-training with paired character patches, supporting the value of cross-script component correspondences.


Q2. On “global layout,” spatially corresponding patches, and cross-patch consistency.

(1) Meaning of “global layout.”

“Global layout” refers to the overall spatial structure within each input patch. The bottleneck represents high-level structural features that provide an overall view of the patch. The model processes patches independently and does not see the whole character or exchange information across patches.

We apologize for the ambiguity in our original description. We have revised the sentence as follows:

Original:

  • The injection at bottleneck allows the conditional features to capture the global layout, acting as a global guide that maintains structural stability throughout the generation.

Revised:

  • The injection at the bottleneck incorporates high-level conditional features that represent the overall spatial structure within each patch, providing structural guidance during patch generation.

(2) Spatially corresponding patches when components move between scripts.

The model learns correspondences between paired image regions (patches) without requiring exact pixel-level alignment. This mapping allows components to move between scripts within the patch. The patch width is 64, more than half the character image width of 112. Moreover, component positions remain relatively stable for most characters, so spatially corresponding patches remain useful when components move between scripts.

(3) Maintaining consistency across patches.

Conditioning provides structural guidance within each patch, but does not maintain consistency across patches. Cross-patch consistency is primarily maintained by the aggregation during the restoration stage. During restoration, we use a sliding window to extract patches. Aggregation is performed at every denoising timestep by averaging the overlapping regions and updating the image. This averaging is not performed only when combining the final patches, which helps maintain consistency across patch boundaries.

AuthorsOpenReview ↗

Author Response To Reviewer (2/2)

Having addressed the questions above, we now respond to the concerns raised in the Critical Weaknesses.

W1. On novelty and differences with OBSD, bottleneck gains.

The concern regarding the contribution of component pre-training is addressed in our response to Q1, where we present additional ablation results. Below, we explain the methodological differences from OBSD and bottleneck gains.

Our method is designed for ancient Chinese character generation, rather than simply combining existing modules. The following differences from OBSD also constitute our contributions.

  • First, component pre-training exploits cross-script component correspondences to learn component priors, whose value is supported by the additional ablation results in Q1.

  • Second, our sampling strategy combines fixed-position patches with random patches, ensuring spatial coverage while retaining sampling diversity. OBSD uses only random patch sampling, which favors central regions and may overlook boundary regions.

  • Third, we separately encode the clean source and noisy target and fuse their high-level features at the bottleneck for conditional guidance. This differs from input-level concatenation in methods such as OBSD, where jointly processing clean and noisy inputs may cause feature interference.

We acknowledge that the improvement from changing the injection position to the bottleneck alone is modest. Nevertheless, the results in Table 4 support the effectiveness of this design. The contributions of the overall method also include component pre-training and patch sampling.


W2. On early-script performance, limitations, and variation across independent training runs.

We thank you for acknowledging our discussion of the difficulties in generating earlier scripts. Below, we explain the results in Appendix Table 7. We have also added a statement about this limitation to the conclusion and provided additional results from independent training runs.

(1)Early-script performance

In the OBI and BI generation tasks in Table 7, CycleGAN achieves lower LPIPS than our method. However, CycleGAN achieves these favorable scores because of its limited ability to transform the source images, as explained below.

CycleGAN and Anc-Diff use OBI and BI interchangeably as source and target. Direct comparison of the paired original OBI and BI images gives a CosSim of 0.7531 and an LPIPS of 0.3239, which are even better than the corresponding generated results. This indicates that favorable similarity scores can be obtained by leaving the source image unchanged, without performing cross-script translation. CycleGAN performs only limited transformations, which gives it an advantage.

As discussed in Section 4.4 and Figure 4, CycleGAN largely preserves RS shapes when generating SS. Figure 12 in Appendix A.3 provides source–output–target comparisons for this task. The comparisons for translations between OBI and BI (Figure 13 and Figure 14 in Appendix A.3) show similar source-preservation patterns. Since these two scripts already share similar forms, limited changes to the source images may partly explain CycleGAN's favorable LPIPS on these tasks. On tasks with larger differences between source and target scripts, such as generating SS from RS, CycleGAN performs substantially worse than our method.

(2)Limitation in the Conclusion

For the limitation of having difficulty with earlier scripts, we have added the following sentence at the end of the conclusion to acknowledge this limitation:

  • In the future, we will focus on improving generation quality for earlier scripts where the current improvements remain limited.

(3)Variation across independent training runs

To address the concern about variation across independent training runs, we conduct 5 independent training runs for Anc-Diff, BBDM, FontDiffuser, and OBSD on the RS-to-SS and RS-to-CS tasks and report the mean and standard deviation of each metric below. These experiments assess the stability of the results.

TargetSourceMethodCosSim↑SSIM↑LPIPS↓
SSRSBBDM0.7617±0.00560.5110±0.01230.2831±0.0049
SSRSFontDiffuser0.7555±0.00670.4662±0.00950.3140±0.0069
SSRSOBSD0.7697±0.00290.5031±0.00270.2814±0.0022
SSRSAnc-Diff(Ours)0.8087±0.00110.5092±0.00190.2348±0.0005
CSRSBBDM0.8016±0.00970.4650±0.00330.2922±0.0114
CSRSFontDiffuser0.8155±0.01410.4501±0.00730.2758±0.0107
CSRSOBSD0.8176±0.00270.4528±0.00180.2694±0.0013
CSRSAnc-Diff(Ours)0.8511±0.00160.4590±0.00100.2394±0.0010

We also report paired significance tests against each baseline on 180 Seal Script test characters in Appendix A.5.


We hope our responses address your concerns and invite you to review the revised manuscript for the additional experiments and explanations. If any questions remain, we welcome your further feedback and will respond promptly.

Thank you again for your careful review and constructive feedback. We would greatly appreciate your feedback and would like to know if our revisions have fully addressed all your concerns. Your response will enable us to make any further adjustments as promptly as possible.


To help you easily locate where we have revised the manuscript, we list below the specific locations corresponding to each of your comments.

  • Q1 & W1: Page 9, Section 4.6.1, Table 3.

  • Q2: Page 9, Line 463-465;

  • W2: Page 15-17, Appendix A.3. Page 10, Line 522-523.

Some other experiments are presented in the responses above.

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

The paper presents a diffusion-based method to generate the form of lost ancient Chinese characters. To overcome the lack of domain-specific knowledge, a component pre-training stage is introduced to capture component priors. To enhance guidance in the generation process, a decoupled guidance control module is introduced. By decoupling the conditional feature extraction and injecting features into the U-Net bottleneck, the decoupled guidance control module avoids feature interference and strengthens conditional guidance.

Critical strengths

The paper presents a well-motivated and technically sound approach to the generation of lost ancient Chinese characters. The overall structure is clear, and the proposed methodology is described in a systematic and comprehensible manner, with sufficient details on the algorithm design and implementation. In addition, the ablation studies are sufficiently extensive to assess the contributions of individual components and validate the design choices. The paper demonstrates the effective integration of domain-specific knowledge with diffusion-based generation.

Critical weaknesses

The paper addresses an interesting and meaningful problem in the reconstruction of lost ancient Chinese characters. However, the overall algorithmic novelty appears somewhat limited, and the methodological contributions should be further clarified and strengthened. The proposed approach involves a relatively complex three-stage training process, but the current version lacks a comprehensive evaluation of training efficiency, including training time, computational cost, and resource requirements. Such an evaluation is important for assessing the practicality and scalability of the proposed method. Furthermore, the experimental comparisons are largely restricted to methods published before 2024, which may not adequately reflect the current state of the art. The authors are encouraged to conduct a more comprehensive literature review, incorporate more recent methods, and reevaluate the proposed approach against state-of-the-art alternatives.

Questions for authors
  1. The citation format for the references in the paper is incorrect and should be revised for consistency.

  2. The references corresponding to the comparison methods in Table 2 should be added.

  3. In the sentence, “Unlike conventional methods that typically concatenate conditional features with target features at the input layer, our method injects conditional features at the bottleneck Guan et al. (2024); Dhariwal & Nichol (2021),” the citation format should be corrected to ensure consistency with the paper’s referencing style.

  4. Table 1: Quantitative results. The table compares only a few closely related methods, and the comparison is limited to approaches published before 2024, making the evaluation insufficiently comprehensive. More recent methods proposed in 2025–2026 should be included. Furthermore, multimodal large language models have made substantial progress in generative capabilities over the past year, and their generation performance should also be evaluated.

  5. The proposed approach involves a relatively complex three-stage training process. How about the comprehensive evaluation of training efficiency, including training time, computational cost, and resource requirements.

7 comments
AuthorsOpenReview ↗

Author Response To Reviewer (1/2)

Thank you for the constructive feedback and for recognizing the significance of our work. We have revised the manuscript and added experiments to address the concerns about novelty, recent baselines, and computational efficiency.


Q1 & Q3. On citation formatting and consistency.

We have corrected the citation formatting throughout the manuscript, including the sentence highlighted in Q3, and ensured consistency in the revised version.


Q2. On references for the compared methods in Table 2.

We have added the corresponding references in Table 2 in the revised version. These references were omitted in the original version due to space constraints, as they were already provided in Table 1.


Q4. On comparisons with recent methods and multimodal generative models.

We have expanded our comparisons to include recent methods proposed in 2025-2026: BDBM (2025), UniDB (2025), and HanziGen (2026). Their results are presented below and in Table 1 of the revised manuscript. We have also updated Figure 4, Table 2, and the Appendix (Figure 6, Table 6, Table 7, Table 9 and Table 10).

MethodSS SourceCosSim ↑SSIM ↑LPIPS ↓CS SourceCosSim ↑SSIM ↑LPIPS ↓
BDBM (2025)OBI0.47850.55030.6097OBI0.45400.50220.6176
BI0.66340.45570.3842BI0.66900.46950.4435
WSS0.67880.52690.3673WSS0.62430.48160.4998
CS0.73620.49190.2985SS0.66440.47200.4574
RS0.57740.54860.4855RS0.70660.48660.4109
UniDB (2025)OBI0.72980.47210.3317OBI0.73990.41630.3800
BI0.73950.48400.3308BI0.73040.40360.3825
WSS0.75040.47000.3148WSS0.73480.44020.3723
CS0.77510.50050.2692SS0.78460.40950.3327
RS0.74710.50730.3037RS0.81030.44020.2867
HanziGen (2026)OBI0.70880.47530.3291OBI0.73750.39640.3871
BI0.71810.47590.3219BI0.75040.36540.3819
WSS0.71960.47680.3159WSS0.74750.38950.3759
CS0.73490.46860.3109SS0.75020.38510.3732
RS0.74070.48000.2959RS0.78780.42560.3025
Anc-Diff (Ours)RS0.80940.50640.2346RS0.85050.45920.2389

These results show that our method remains competitive with recent methods.

Regarding multimodal generative models, we focus on comparisons under controlled training-data conditions to ensure fairness. Our character images are collected from publicly accessible websites, and we cannot rule out their inclusion in the pre-training data of such models. Their pre-training corpora are often not fully disclosed. We therefore cannot reliably verify whether these corpora overlap with our benchmark or include its test images.


Q5. On training time, computational cost, and resource requirements.

We agree that these measurements are important for assessing the practicality and scalability of Anc-Diff. We have revised the Computational Efficiency Analysis in Appendix A.4 of the revised manuscript and summarize the results below.

Anc-Diff consists of two training stages (component pre-training and character training) and the restoration stage. The third stage does not involve parameter updates. All measurements are obtained on a single NVIDIA GeForce RTX 4090, with a batch size of 2 for both component pre-training and character training, and 1 for restoration.

StageTimeTime/characterPeak memory
Component pre-training3.72 h—18.95 GiB
Character training1.57 h—19.64 GiB
Total training5.29 h—19.64 GiB
Restoration (180 characters)399.95 s2.22 s9.55 GiB

These results show that Anc-Diff can be trained and evaluated on a single RTX 4090, with approximately 5.29 hours of training and 2.22 seconds per generated character. This result provides practical computational estimates for assessing its practicality and scalability. We additionally report the per-character inference time of all comparison methods, including the newly added baselines, in the table below.

MethodTime (s/character)MethodTime (s/character)
CycleGAN0.02FontDiffuser0.41
Pix2pix0.02OBSD1.83
DRIT++0.23BDBM0.91
BBDM0.94UniDB1.02
ControlNet1.62HanziGen0.03
T2I-Adapter0.66Anc-Diff (Ours)2.22

Although our method takes longer than the comparison methods at inference, generation quality is more important than speed for ancient Chinese character generation. We consider an average inference time of 2.22 seconds per character an acceptable trade-off for the improvement in generation quality.

These results have been added to Tables 8 and 9 in the revised version.

AuthorsOpenReview ↗

Author Response To Reviewer (2/2)

Most concerns raised in the Critical Weaknesses have been addressed in our responses to Q1–Q5, particularly those regarding baseline coverage and computational efficiency. Below, we focus on the remaining concerns about methodological novelty and contributions.

Weakness. On methodological novelty and contributions.

  • The novelty of Anc-Diff lies in incorporating domain-specific priors of strokes and radicals into cross-script character generation. Component pre-training learns reusable transformation patterns from paired components, and the learned parameters are transferred to whole-character training. This design exploits the compositional structure of Chinese characters when whole-character training data are limited.

  • Anc-Diff also introduces a decoupled conditioning design: a separate encoder extracts features from the conditioning image and injects them into the U-Net bottleneck, separating clean-condition encoding from noisy-target encoding.

  • Furthermore, a shared patch sampling strategy connects component pre-training and character training, maintaining a consistent local learning scheme across both stages. Random patches provide sampling diversity, while fixed-position patches ensure full spatial coverage, including boundary regions.

Together, these designs provide a task-specific approach that combines stroke and radical priors, conditional guidance, and component-to-character transfer for ancient character form generation.


We hope our responses address your concerns and invite you to review the revised manuscript for the additional experiments and explanations. If any questions remain, we welcome your further feedback and will respond promptly.

Thank you again for your careful review and constructive feedback. We would greatly appreciate your feedback and would like to know if our revisions have fully addressed all your concerns. Your response will enable us to make any further adjustments as promptly as possible.


To help you easily locate where we have revised the manuscript, we list below the specific locations corresponding to each of your comments.

  • Q1 & Q3: Page5, Line 231.

  • Q2: Page 8, Table 2.

  • Q4: In Page7, Table1, we add some recent methods proposed in 2025-2026: BDBM (2025), UniDB (2025), and HanziGen (2026). After adding several new baseline methods, we have also made corresponding revisions to the following sections: Figure 4, Table 2, and the Appendix (Figure 6, Table 6, Table 7, Table 9 and Table 10)

  • Q5: Page17, Appendix A.4, Table8, Table9.

Comments on the paper

2 comments

Thank you for your response. We would like to confirm whether our revisions have addressed all your concerns. From our perspective, we have carefully responded to every point you raised and provided additional experimental results. If anything remains unresolved or requires further clarification, please let us know. We will be glad to make further revisions. This work is very important to us.

AuthorsOpenReview ↗

Summary of rebuttal

We sincerely thank all reviewers, Area Chairs (AC), Senior Area Chairs (SAC) and Program Chairs (PC) for the time, effort, and recognition of our work. We deeply appreciate the invaluable and constructive suggestions provided during the review process, which have significantly strengthened the quality of our paper. We have meticulously addressed every concern and made every effort to resolve the reviewers' questions.


For Reviewer TiMk:

Thank you for acknowledging that most of your initial concerns (citation formatting, missing references, recent baseline comparisons, the comprehensive evaluation of training efficiency, and methodological novelty and contributions) have been addressed.

Regarding your new concern about reorganizing and strengthening the experimental comparisons, we have provided a detailed response in our updated rebuttal and revised the manuscript accordingly in Sections 4.3, 4.4, and Appendix A.1. We have carefully revised the manuscript according to your suggestions and hope these updates resolve your remaining concerns.


For Reviewer NqFM:

Thank you for recognizing our work and responses. In our revised manuscript, we have thoroughly addressed all of your concerns regarding component pre-training, cross-patch consistency, differences with OBSD, early-script performance, variation across independent training runs, and so on.

We particularly appreciate your increasing the score from 2 to 3.


For Reviewer atE6:

Thank you for recognizing our work and responses. In our revised manuscript, we have thoroughly addressed all of your concerns regarding component pre-training, early-script performance, and evaluation of genuinely missing historical forms, and so on.

We particularly appreciate your increasing the score from 2 to 3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.