{"id":"13251199-e4cf-4c32-9ef0-2cbd770fef2d","arxiv_id":"2501.08062","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SFGN generates Chinese characters from radical and stroke captions and transfers style via a transitive-attention mechanism, improving zero-shot generation and error-correction augmentation.","lead":"This paper introduces SFGN, a two-stage network that generates Chinese characters from text descriptions of their radicals and strokes, then copies font style from reference images using a new transitive-attention mechanism. The method targets zero-shot generation of characters not seen in training, including misspelled characters, and is shown to improve a handwritten Chinese character error-correction task when its output is used as training data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim is anchored to Tables 2–3, where SFGN's end-to-end advantage is only shown against RCN/RTN-G; its font generator alone does not beat CG-GAN or CF-Font, so SOTA superiority remains unestablished.","rationale":"I agree with the reader's conditional verdict and with the observation that the transitive-attention variance formula in Eq. 13 is unproven and load-bearing for the robustness-on-misspelled-characters claim; independent re-derivation or a numerical stability test would settle that. However, my primary concern is different and, in my reading, more decisive for the paper's headline: the quantitative support for 'outperforms current state-of-the-art font generation methods' is inconsistent. The font generator alone (the architecture that realizes the radical-level copying claim) is not better than CG-GAN or CF-Font on almost any metric in Tables 2-3; SFGN's 'highest score' refers only to the two end-to-end baselines in the bottom halves of the tables. The paper's own Section 5.2 concedes 'SFGN may have lower performance than font generation methods,' so the SOTA claim rests on (a) a task-level argument that caption-input is harder, which is plausible but unquantified, and (b) qualitative robustness figures on misspelled characters. The reader's weak-assumption pick pinpoints an internal gap (Eq. 13 derives from an i.i.d. N(0,1) assumption that is asserted, not shown), and that gap matters because the transitive-attention design is the mechanism claimed to prevent bias-driven correction of misspelled glyphs. Still, I call agreement 'partial' because the more immediate problem is the gap between the claimed superiority and the reported numbers; a check that isolates the contribution of the caption-input pipeline would decide the paper's main contribution claim, while the variance formula check would decide the internal soundness of one module. Both should be required before the central claim is accepted; neither by itself fully invalidates the method, which is why I keep CONDITIONAL rather than moving to REJECT.","tokens_in":18965,"tokens_out":2251,"duration_ms":18233,"concrete_test":"Re-run the font generation evaluation under matched input conditions. Specifically, feed the identical real content images used for the other methods (not the skeleton-builder output) into SFGN's font generator, while keeping the same style images and test splits, and report RMSE/SSIM/LPIPS/FID for seen and unseen styles. If the scores match or beat CG-GAN and CF-Font, the end-to-end claim is credible; if they remain worse, the paper should restrict its claim to 'competitive with SOTA among caption-input methods' and the verdict should be CONDITIONAL at best, since Tables 2-3 quantify a regression that the text explains qualitatively.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The abstract and conclusion claim SFGN 'outperforms current state-of-the-art font generation methods,' but the quantitative evidence in Tables 2 and 3 does not support that broad claim. The proposed font generator (rows labeled 'Font Generator') achieves RMSE 0.0233/SSIM 0.7708 on seen styles and 0.0236/0.7637 on unseen styles, which is comparable to CG-GAN (0.0223/0.7696 and 0.0237/0.7618) and CF-Font (0.0220/0.7704 and 0.0236/0.7633); on the four headline metrics it is not consistently better than either. The claim that SFGN achieves 'the highest score among these methods' is only true within the restricted end-to-end subset (RCN, RTN-G, SFGN), as the paper itself notes in Section 5.2 ('compared to other end-to-end methods'). Since all competing font generation methods take a real content image as input while SFGN synthesizes content from captions, the comparison is not controlled: the lower font-generator scores could reflect harder task conditions, but the paper does not quantify this and instead asserts 'slight performance degradation.' The load-bearing condition for the central claim is that SFGN's end-to-end pipeline is competitive with or better than image-input SOTA in the zero-shot regime; the headline evidence for this is visual (Figures 5, 6, 8) and qualitative. The reader's cited technical weakness (the unproven variance formula for transitive attention) is real but secondary; even if Eq. 13 is correct, the headline claim overreaches because the metrics only show end-to-end superiority over two methods that are not SOTA. A direct unfair-comparison control (giving SFGN's font generator the same real content images used by competitors) is the decisive missing experiment.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SFGN (Skeleton and Font Generation Network) for zero-shot Chinese character generation, combining a skeleton builder that synthesizes standard-font glyph images from radical- and stroke-level captions with a font generator that transfers target font style through a novel transitive-attention mechanism. The method is evaluated on glyph generation (Table 1), font generation for seen and unseen styles (Tables 2–3), single-font transfer (Table 4), misspelled characters, cross-lingual Korean characters, and downstream Chinese character error correction with generated data (Tables 7–8). The authors claim that SFGN outperforms current state-of-the-art font generation methods and is particularly robust on misspelled characters.","tokens_in":19498,"tokens_out":8200,"duration_ms":76323,"significance":"If the results are sound, the skeleton builder's caption-to-glyph synthesis with zero-shot generalization is a genuine advance, and the transitive-attention idea of aligning content and style at the radical level is an interesting new perspective. The quantitative results for glyph generation (Table 1) are strong, and the augmentation experiments for handwritten Chinese character error correction, while modest, support the pedagogical value of generated misspelled characters. However, the central claim that SFGN outperforms state-of-the-art font generation methods is not well supported by the reported metrics, and the theoretical justification of the transitive-attention normalization is incomplete. The paper also relies heavily on qualitative visual comparison for the misspelled-character robustness claim, which weakens the contribution.","major_comments":[{"comment":"The claims that SFGN 'outperforms current state-of-the-art font generation methods' and 'outperforms state-of-the-art generation models on both glyph and font generation tasks' are not supported by the reported numbers. In Table 2, the proposed Font Generator has RMSE 0.0233 versus 0.0220 for CF-Font and 0.0223 for CG-GAN; in Table 3, its FID is 23.12 versus 20.11 for CG-GAN and 22.08 for CF-Font. The only rows where SFGN is consistently best are the 'End-to-end' rows against RCN and RTN-G, which are text-to-glyph methods rather than font generation methods. The abstract and conclusion should be revised to claim 'comparable' performance for the font generator, or a controlled comparison that accounts for the harder caption-only input should be provided.","section":"Abstract, Section 6, Tables 2–3"},{"comment":"The variance formula V_n used to normalize the product of exponentials in the transitive-attention mechanism is stated without derivation, and the supporting assumption that queries and keys follow i.i.d. N(0,1) distributions is not justified for the projected vectors X_i U^Q_i and X_{i+1} U^K_{i+1}. Because this normalization is the core of the proposed mechanism, a derivation or a reference proving Eq. (13) is required, along with a discussion of how the N(0,1) assumption is maintained or approximated during training. Without this, the normalization is not reproducible and the theoretical foundation of the mechanism remains unverified.","section":"Section 3.2, Eq. (13)"},{"comment":"The guided loss L_g supervises the transitive-attention maps with attention maps aG generated by the skeleton builder and aR generated by the recognition model. Since the skeleton builder is trained jointly on the same data, this is a form of self-supervision that could propagate errors from the skeleton builder into the font generator. The paper should explicitly state whether aG is detached and aR is frozen, and should discuss the potential for error amplification. The ablation in Table 6 shows a large drop without L_g but does not isolate whether the guidance is transferring genuinely useful structural knowledge or merely forcing the font generator to imitate an imperfect attention pattern.","section":"Section 3.3, Eq. (21)"},{"comment":"The comparison between SFGN's end-to-end row and the font generation baselines is not controlled: SFGN receives captions as input, while the baselines receive a real content image. The text acknowledges 'slight performance degradation' for SFGN, but the abstract and conclusion do not reflect this. The paper should explicitly quantify this degradation by comparing the 'Font Generator' row (which uses a content image) with the 'SFGN' end-to-end row, and should re-frame the SOTA claim so that it is restricted to the appropriate comparison setting.","section":"Section 5.2, Tables 2–3"}],"minor_comments":[{"comment":"The citations for RCN and RTN-G are swapped: the text says 'RCN [8]' and 'RTN-G [9]', but the reference list identifies [8] as RTN-G and [9] as RCN. Please correct the in-text citations.","section":"Section 2.2"},{"comment":"The denominator in Eq. (12) is written as 'V2N −2' while the variance formula in Eq. (13) is defined generically as V_n. Please clarify that the argument is n = 2N−2 and define N in this context.","section":"Section 3.2, Eq. (12)"},{"comment":"The method name 'Strokegan' is inconsistent with the reference title 'Strokegan: Reducing mode collapse in chinese font generation via stroke encoding'; please unify the capitalization.","section":"Table 4"},{"comment":"The category 'novel characters' is used in Table 7 but is only vaguely described in the text; a concrete definition with examples would improve clarity.","section":"Section 5.4"},{"comment":"The visual comparisons in Figures 8 and 10 are difficult to read at the current size; consider enlarging the panels and adding annotations to highlight the failure cases discussed in the text.","section":"Figures 8 and 10"}],"recommendation":"major_revision","confidential_remarks":"The paper has a potentially useful two-stage architecture and a novel attention mechanism, but the headline claim of surpassing state-of-the-art font generation methods is not supported by the tabular results. The technical gap in the variance normalization needs to be addressed, and the comparison protocol for the end-to-end setting should be clarified. If these issues are resolved, the paper could be publishable in a specialized venue, but in its current form the overclaiming and missing justification are substantial."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the caption-driven skeleton builder and the misspelled-character augmentation experiments are the real contributions; the \"outperforms SOTA\" framing in the abstract is not supported by the paper's own tables. The font generator alone is at parity with CG-GAN and CF-Font, and the end-to-end comparison only beats RCN and RTN-G, which are not SOTA font generation methods. Still, the glyph generation results and the augmentation results are solid enough that this deserves a serious referee, not a desk reject.\n\nWhat is new: the skeleton builder takes radical and stroke captions and constructs glyphs without content images, which is a meaningful step beyond RCN and RTN-G. The transitive-attention mechanism, even if the formula needs derivation, is a plausible way to align style and content at the radical level. The misspelled-character generation for handwritten Chinese character error correction is a nice application, and the gains in Tables 7 and 8 look credible.\n\nWhere it is soft: Eq. 13 is stated without derivation, and the Gaussian assumption on queries and keys is not checked. That is a real gap, but secondary because the model works in practice; the paper should either derive it or treat it as a heuristic. The bigger issue is the central claim. In Tables 2 and 3, the \"Font Generator\" row is not consistently better than CG-GAN or CF-Font; on unseen styles its FID is worse than both. The end-to-end SFGN row beats only RCN and RTN-G, so \"outperforms current state-of-the-art font generation methods\" is an overstatement. The comparison is also not controlled: baselines get real content images, while SFGN gets synthetic content features. The paper mentions \"slight performance degradation\" but does not quantify it. The decisive missing experiment is to feed the font generator the same real content images used by the baselines, or at least measure the quality gap introduced by the synthetic content features.\n\nWho it is for: people working in document analysis, font generation, and handwritten character recognition. The paper is worth reading for the misspelled augmentation setup and the skeleton builder; the transitive-attention derivation needs work.\n\nRecommendation: send to peer review with major revision. Ask for a derivation or empirical validation of the variance normalization, a controlled comparison with real content images, and a toned-down abstract.","headline":"The caption-driven skeleton builder and misspelled-character augmentation results are the real contributions; the SOTA claims in the abstract overreach, but the paper deserves a serious referee.","tokens_in":19914,"tokens_out":1583,"would_cite":false,"duration_ms":16035,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A two-stage network generates unseen and misspelled Chinese characters from captions, avoiding the bias that makes older models \"correct\" them into familiar forms.","keywords":["Chinese character generation","font generation","zero-shot learning","radical analysis","transitive attention","skeleton builder","misspelled character generation","handwritten Chinese character error correction"],"falsifier":"Compute the empirical variance of the unnormalized product $\\hat\\beta$ over actual style-feature and caption pairs in the trained model and compare it with $V_n$; if the values diverge, retrain with the empirical normalization and measure whether generated misspelled characters still avoid being corrected, which would show that the stated formula, and the assumption behind it, is load-bearing. Equivalently, rescaling style features so they are provably not $N(0,1)$ should break the attention alignment if the assumption is essential.","tokens_in":18768,"feed_emoji":"✍️","tokens_out":4606,"duration_ms":46812,"temperature":0.7,"pith_summary":"The paper claims that a single two-stage network can generate Chinese characters never seen in training, including misspelled characters whose glyphs differ from correct ones by only a stroke or dot. Prior font generators tend to \"correct\" such subtle variations into familiar characters; this method avoids that by building the glyph from text captions first and then applying font style at the level of radicals, the reusable components of Chinese characters. If true, this makes font libraries and educational materials cheaper to produce and provides a new source of training data for handwriting error-correction systems. The paper supports the claim with lower distortion scores than prior methods on standard glyph and font benchmarks, and with gains in an error-correction task when generated misspelled characters are used as training data.","feed_headline":"Two-stage model generates unseen and misspelled Chinese characters from captions","feed_subtitle":"Radical-level copying beats prior font generators and turns misspelled glyphs into training data for error correction.","key_machinery":"The load-bearing object is the transitive-attention block $\\mathrm{TA}$, whose kernel $\\hat{\\beta}=\\prod_{i=1}^{N-1}\\exp(Q_i)\\exp(K_{i+1}^\\top)$ multiplies pairwise attention weights along a chain from style features to captions to content features. A variance normalization formula $V_n = D^{n-1}[\\sum_{i=0}^n (D-1)^i C^i_n e^{2n-i}] - e^n D^{2(n-1)}$ is used to rescale $\\hat\\beta$ before softmax, under the stated assumption that queries and keys are independent standard normal vectors. The skeleton builder also carries the argument: a \"blank canvas\" of position-embedded vectors is filled first with radical representations and then with stroke representations, letting the model generate glyphs from captions alone. Together the two modules turn font generation into a copy-and-place operation at radical scale.","core_discovery":"The central claim is that Chinese character font generation can be decomposed into glyph creation and font application, with both stages robust to characters that look nearly identical to training samples. The skeleton builder maps radical-level and stroke-level text captions directly to a content feature and a standard-font image, so the content channel never sees a content image and can represent unseen combinations; the font generator then aligns this content feature with style features extracted from five style images via a transitive-attention mechanism that routes attention from style image to style captions, from style captions to content captions, and from content captions back to content features. This radical-level \"copying\" path is what prevents the model from overwriting a misspelled character with the nearest familiar one. The paper additionally claims that images of misspelled characters generated this way improve a handwritten Chinese character error-correction model, and that the benefit is roughly comparable to using twenty-five times more rare-character data.","pith_inferences":["A testable extension is to apply the transitive-attention chain to other compositional writing systems or structured object generation where captions intermediate between source and target domains, since nothing in the mechanism is Chinese-specific beyond the radical captions.","The augmentation result suggests that the difficulty distribution of training data matters more than volume: data close to the decision boundary, such as misspelled characters, buys most of the error-correction gain; isolating rare-character sets at varying sizes would make this effect explicit.","Because the variance formula for transitive attention is asserted rather than derived, the method's robustness claim carries a hidden dependency; checking whether learned queries and keys are approximately $N(0,1)$, or replacing the theoretical normalization with an empirical estimate, would make the mechanism portable and safer to adopt."],"forward_implications":["Characters that are not in the training dictionary, including misspelled and internet-novel characters, can be rendered in a target font without content images.","A font style can be transferred from roughly five reference images whose radicals cover the target character, without retraining for each new font.","Generated misspelled characters work as training data for handwritten Chinese character error correction, and a small misspelled set yields gains comparable to a much larger rare-character set.","The three transmitted attention maps are interpretable as locating components in the style image, deciding which components are needed, and placing them in the content image.","The method extends to at least one other compositional script, Korean, after fine-tuning on that script's training set."],"supporting_citations":[{"why":"Supplies the RCN baseline for zero-shot glyph generation that the skeleton builder extends and improves on.","marker":"[9]"},{"why":"Supplies the RTN-G transformer-based glyph baseline and the joint optimization idea the skeleton builder builds on.","marker":"[8]"},{"why":"Provides the LF-Font component-wise style baseline for few-shot font generation.","marker":"[3]"},{"why":"Provides the CG-GAN component-level supervised baseline and the motivating failure cases of biased generation.","marker":"[4]"},{"why":"Provides the FsFont copying-based few-shot font baseline that most closely resembles the proposed radical-level transfer.","marker":"[24]"},{"why":"Provides the DG-Font unsupervised deformation baseline used in comparisons.","marker":"[21]"},{"why":"Provides the EMD style-content disentanglement baseline used in comparisons and augmentation experiments.","marker":"[17]"},{"why":"Supplies the TAN handwritten Chinese character error-correction model used in the augmentation experiments.","marker":"[25]"},{"why":"Supplies the RTN-R radical recognition model that provides the content loss and the attention guidance.","marker":"[41]"},{"why":"Supplies the standard attention mechanism from which the transitive-attention kernel is extended.","marker":"[35]"}],"fun_headline_variants":["Radical-level font generation handles unseen glyphs and misspellings","Two-stage model turns captions into unseen and misspelled characters","Zero-shot font generation preserving subtle character variations","From captions to fonts: robust generation of novel and misspelled glyphs","Skeleton+font network generates rare glyphs and boosts error correction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The transitive-attention block assumes that query and key vectors behave as independent standard-normal variables, because its variance normalization formula is derived from that assumption and is stated without proof.","fun_headline_variants_meta":{"raw":{"variants":["Radical-level font generation handles unseen glyphs and misspellings","Two-stage model turns captions into unseen and misspelled characters","Zero-shot font generation preserving subtle character variations","From captions to fonts: robust generation of novel and misspelled glyphs","Skeleton+font network generates rare glyphs and boosts error correction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000178,"raw_usage":{"total_tokens":1331,"prompt_tokens":1015,"completion_tokens":316,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":631,"completion_tokens_details":{"reasoning_tokens":227}},"tokens_in":631,"tokens_out":316,"duration_ms":3344,"temperature":1.0,"reasoning_tokens":227,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:29:15.554160+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the empirical variance of the unnormalized product $\\hat\\beta$ over actual style-feature and caption pairs in the trained model and compare it with $V_n$; if the values diverge, retrain with the empirical normalization and measure whether generated misspelled characters still avoid being corrected, which would show that the stated formula, and the assumption behind it, is load-bearing. Equivalently, rescaling style features so they are provably not $N(0,1)$ should break the attention alignment if the assumption is essential.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the RCN baseline for zero-shot glyph generation that the skeleton builder extends and improves on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the RTN-G transformer-based glyph baseline and the joint optimization idea the skeleton builder builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the LF-Font component-wise style baseline for few-shot font generation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the CG-GAN component-level supervised baseline and the motivating failure cases of biased generation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the FsFont copying-based few-shot font baseline that most closely resembles the proposed radical-level transfer."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the DG-Font unsupervised deformation baseline used in comparisons."},{"cited_title":"Zhang, Y","cited_arxiv_id":null,"evidence_quote":"Provides the EMD style-content disentanglement baseline used in comparisons and augmentation experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the TAN handwritten Chinese character error-correction model used in the augmentation experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the RTN-R radical recognition model that provides the content loss and the attention guidance."}],"review_version":1}