{"id":"e14f52bb-d0fd-403f-9fc9-037e5635e41d","arxiv_id":"2505.06543","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"HDGlyph improves long-tail visual text rendering in diffusion models by disentangling text from background at pixel, noise, and latent levels, reporting accuracy gains of 5.08% in English and 11.7% in Chinese on the AnyText benchmark.","lead":"HDGlyph is a new framework that separates text drawing from background generation in diffusion models, improving accuracy for uncommon languages and small fonts. It reports clear gains over existing open-source systems, but releases no code or weights.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Multilingual-Benchmark confounds script novelty with font size: reusing English layouts for translated text changes the rendered font size, so the reported 'unseen character' accuracy drops may reflect small-text difficulty rather than cross-script generalization.","rationale":"The central claim has two legs: (1) common-text gains on AnyText-Benchmark, which are self-reported but structurally sound, and (2) long-tail gains on the Multilingual- and Complex-Benchmarks. The load-bearing weakness is in leg (2), specifically in the Multilingual-Benchmark design. Reusing the same layout boxes for translated text cannot hold font size constant; the paper both claims to eliminate font-size differences (Section 4.2.2) and later assumes font sizes are 'roughly the same' (Appendix A.1), an unsupported assumption given variable translation lengths. If Japanese/Russian stimuli are rendered at smaller sizes, the 'decline rates' in Table 2 measure size robustness, not script generalization, and the claim of 'unseen characters' is not established. This is more fundamental than the reader's Canny-edge concern because even a perfect edge extractor would not separate the confound; the test construction itself fails. The concrete test (re-render with fixed font size and compare) settles it directly. I do not see an internal inconsistency severe enough to reject; the common-text results and ablations support the framework's plausibility, so the verdict remains CONDITIONAL.","tokens_in":15690,"tokens_out":18739,"duration_ms":186269,"concrete_test":"Reconstruct the Multilingual-Benchmark renderings and measure the actual fitted font size per prompt per language in the reused bounding boxes. Then run HDGlyph and the baselines under two controlled conditions: (i) fixed font size across languages with the layout box scaled to the text, and (ii) fixed bounding box with auto-fit font size (the current setup). If the accuracy decline from English to Japanese/Russian under condition (i) is much smaller than under condition (ii), or if the measured font sizes are systematically smaller for Japanese/Russian, the unseen-character conclusion is confounded. As a secondary check, correlate per-prompt accuracy with rendered font size across all languages; a strong correlation within English would confirm that the benchmark is size-sensitive.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2.2 constructs the Multilingual-Benchmark by translating 100 English prompts from the LAION subset into Chinese, Japanese, and Russian while reusing the same text layout, claiming this 'eliminates the influence of layout variations and font size differences.' This is internally inconsistent: for a fixed bounding box, a different string length forces a different font size (or overflow). No font-size normalization is described; Appendix A.1 merely 'assumes' font sizes are roughly the same. If Japanese/Russian translations are systematically longer, their glyphs are rendered smaller, so the Table 2 accuracy drops (e.g., Ours U-Net Acc 82.00 to 59.70 for Japanese) conflate unseen-script difficulty with small-font degradation. Because the paper's novel long-tail claim is 'robust rendering even for unseen characters' (Abstract, Section 3.1.1), this confound undermines the benchmark's validity as evidence for that claim. The downstream Canny edge concern is sharpened: smaller rendered text means coarser Canny edge maps, so the model's apparent cross-script generalization may instead be small-text robustness.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HDGlyph, a framework for visual text rendering in diffusion models. It hierarchically disentangles text generation from background synthesis at three levels: at training time, a Multi-Linguistic GlyphNet and a Glyph-Aware Perceptual Loss are used to preserve glyph structures; at inference time, Noise-Disentangled Classifier-Free Guidance amplifies glyph-specific signals, and Latent-Disentangled Two-Stage Rendering with a small-text enhancement module refines backgrounds and small glyphs. The authors evaluate HDGlyph on the AnyText-Benchmark, a self-constructed Multilingual-Benchmark, and the Complex-Benchmark, reporting gains of 5.08% in English and 11.7% in Chinese text accuracy, as well as smaller accuracy drops than AnyText and AnyText2 on multilingual and small-text scenarios. The paper also provides ablations showing the contribution of each proposed component.","tokens_in":15939,"tokens_out":6212,"duration_ms":60468,"significance":"If the results hold, HDGlyph is a practical engineering contribution: it improves text accuracy over existing open-source baselines on an external benchmark, and the ablation study gives internal evidence that each proposed component contributes. The use of Canny glyph edges as a transferable structural prior and per-language LoRA experts is a reasonable design. However, the strongest claim—robust generalization to unseen scripts—rests on a self-constructed benchmark with a font-size confound, and the main comparison table mixes results from inconsistent evaluation pipelines. The AnyText-Benchmark results and the internal ablations are credible, but the manuscript needs revision before the long-tail claims are fully supported.","major_comments":[{"comment":"The Multilingual-Benchmark construction conflates script novelty with font size. The paper claims that reusing the English layout for translated text 'eliminates the influence of layout variations and font size differences,' but Appendix A.1 concedes 'we assume that the font sizes are roughly the same.' For a fixed bounding box, a translated string of different length renders at a different font size; no normalization is described. Because Table 2 reports large accuracy drops for Japanese (82.00 to 59.70) and Russian (82.00 to 62.94), these drops could be caused by smaller rendered glyphs rather than by unseen-script difficulty, which is the paper's core long-tail claim. Please measure and report per-sample font sizes (or per-character pixel heights), and rerun with size-normalized conditions—for example, matching rendered text pixel height across languages or reporting results stratified by font size.","section":"Section 4.2.2, Appendix A.1, Table 2"},{"comment":"The baseline comparison is not on a consistent evaluation pipeline. The authors adopt selected results from [15,20] (marked with †) while other rows are presumably run locally, and the resulting CLIPScore values for Glyph-ByT5 (48.02) and GlyphDraw2 (47.96) are on a completely different scale from the locally run models (roughly 80–90). The paper itself notes that different CLIPScore models produce divergent values, but it still reports these rows in the same table. This undermines the claim that HDGlyph 'maintains high image quality' relative to those methods. Please rerun all baselines under the same protocol, or clearly separate the tables and restrict claims to methods evaluated consistently.","section":"Table 1, Section 4.3.1, Appendix B"},{"comment":"The 'unseen character' claim rests on the assumption that a Canny edge map of the target glyph is a complete and transferable structural prior for scripts not seen during training. The paper provides no analysis of how Canny edge quality degrades for dense scripts (e.g., Japanese kanji, Russian Cyrillic) or for small text, and the Multilingual-Benchmark cannot resolve this because of the size confound. Please add an edge-fidelity analysis (e.g., character-level IoU between the rendered and ground-truth edge maps across languages and font sizes) or an ablation that holds font size fixed while varying script novelty.","section":"Section 3.1.1"}],"minor_comments":[{"comment":"The phrase 'Glyph-A ware' should be 'Glyph-Aware,' and the caption of Figure 2 in Appendix C says 'HMDGlyph' instead of 'HDGlyph.'","section":"Abstract, Section 3.1.2, Appendix C"},{"comment":"Appendix A and B contain unresolved citation placeholders such as '[? ]' and 'work [? ];' these need to be filled before publication.","section":"Appendix A, Appendix B"},{"comment":"The Multilingual-Benchmark is constructed from 100 randomly selected prompts, but no seed or exact list is provided; please supply the list or seed to make the benchmark reproducible.","section":"Section 4.2.2"},{"comment":"Appendix B states 'All values are expressed as percentages,' but it also reports ViT-L/14 CLIPScore values of 0.6912 and 0.6901 for AnyText and AnyText2; please specify the scaling and state which CLIPScore version applies to each table.","section":"Appendix B, Section 4.2.4"},{"comment":"The explanation that a larger 𝛼2 value 'requires a faster rate' is vague; please specify how 𝛼1 and 𝛼2 change the shapes of the cosine schedules in Eqs. (6) and (7).","section":"Equation (7), Section 4.1"},{"comment":"Table 4's 'ControlNet' baseline should be clarified: it is presumably ControlNet-Canny without expert LoRA, ND-CFG, or LD-TSR, but this is not stated, and the relationship to 'SDXL+Canny' in Table 3 is unclear.","section":"Table 4, Table 3"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely to be acceptable after the authors fix the Multilingual-Benchmark confound and harmonize the baseline evaluation. The self-admitted assumption in Appendix A.1 directly contradicts a claim in Section 4.2.2, so this issue should be addressed head-on in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is a competent engineering contribution with a sensible hierarchy of disentanglement tricks, but the main long-tail evidence rests on a benchmark that doesn't isolate what it claims to isolate. The Multilingual-Benchmark reuses English text layouts for translated Chinese, Japanese, and Russian prompts. For a fixed bounding box, a different string length forces a different font size (or overflow). The paper says the setup 'eliminates the influence of layout variations and font size differences,' yet Appendix A.1 merely 'assumes' the font sizes are roughly the same. That is an internal inconsistency. So the accuracy drops in Japanese and Russian (e.g., Ours U-Net Acc falling from 82.00 to 59.70) could reflect small-text difficulty rather than unseen-script generalization. The Canny edge prior is also coarser at small sizes, so the cross-script result may partly be small-text robustness. That weakens the central 'unseen characters' claim.\n\nWhat the paper does well: the method is more than a bag of tricks. The Multi-Linguistic GlyphNet with per-language LoRA experts and the glyph-aware perceptual loss are reasonable extensions of existing ideas, and ND-CFG with a negative image condition is a neat way to amplify glyph signal without wrecking background. The ablations in Table 4 are consistent and support each component: expert LoRA helps, ND-CFG boosts accuracy but hurts background metrics, and LD-TSR recovers background quality. That internal consistency is credible. The common-text results on the external AnyText-Benchmark also look solid, though some baseline numbers are quoted from other papers with different CLIPScore versions, which the authors themselves acknowledge.\n\nOther soft spots: no code or weights released, only 100 prompts in the multilingual set, no error bars, and the Complex-Benchmark uses GPT-4o for layout, adding another variable. None of these are fatal, but they limit how much you can trust the long-tail numbers. The novelty is incremental—ControlNet-Canny, LoRA, MultiDiffusion-style blending—but the combination is coherent and the engineering is honest.\n\nThis paper is worth a serious referee, especially for a multimedia or text-rendering venue. The reviewer should push for a font-size-controlled multilingual evaluation or a clear mitigation, plus code/weights. As is, treat the long-tail claims with caution, but the common-text results and the ablation story are solid enough to engage with.","headline":"Solid engineering with a coherent stack of tricks, but the multilingual benchmark confounds unseen scripts with smaller fonts, which undermines the paper's headline long-tail claim.","tokens_in":16472,"tokens_out":2248,"would_cite":true,"duration_ms":22200,"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 diffusion framework that conditions on glyph edge maps renders unseen scripts and tiny text while preserving background quality.","keywords":["visual text rendering","diffusion models","long-tail text rendering","glyph conditioning","multilingual text generation","small text rendering","classifier-free guidance","latent disentanglement"],"falsifier":"Render a connected-stroke script such as Arabic or Devanagari at small sizes and measure per-character accuracy against the English baseline: if accuracy collapses toward the levels of the prior baselines, the edge prior is not complete for those stroke structures. A more direct check is to perturb the Canny thresholds used to build the glyph conditions and measure whether character accuracy tracks edge fidelity; if it does, the framework's ceiling is set by the edge extractor rather than the diffusion model.","tokens_in":15463,"feed_emoji":"🔤","tokens_out":11181,"duration_ms":93339,"temperature":0.7,"pith_summary":"The paper sets out to show that diffusion models fail at rare and tiny text because they never separate text from background, and that forcing that separation at three levels—pixel, noise, and latent—lets a single model handle common text, scripts it never saw in training, and very small fonts. The proposed framework, HDGlyph, conditions generation on the Canny edge map of the target glyph string through a Multi-Linguistic GlyphNet, trained with a glyph-aware perceptual loss that aligns predicted and ground-truth text-region edges instead of relying on OCR feedback. At inference it adds a noise-level guidance term that amplifies the glyph signal against a glyph-free negative condition, and a two-stage latent scheme that paints the background first, then refines the text, with a patch-based upscaling pass for small text. The reported payoff on a public text-rendering benchmark is 5.08% higher English accuracy and 11.7% higher Chinese accuracy than prior open-source systems, with much smaller accuracy drops on Japanese, Russian, and small-text content.","feed_headline":"Glyph edges unlock unseen scripts and tiny text in diffusion models","feed_subtitle":"Conditioning on stroke edges keeps rare-language and small-font accuracy high while backgrounds stay clean.","key_machinery":"The load-bearing mechanism is the Multi-Linguistic GlyphNet, a ControlNet-style branch initialized from a Canny-conditioned ControlNet and augmented with per-language expert LoRAs, paired with the Glyph-Aware Perceptual Loss $\\mathcal{L}_{\\text{Glyph}}$, which compares Canny edge maps of predicted and ground-truth text regions with a noise-aware timestep weighting. Together they supply a script-agnostic structural prior: the loss lives entirely on edge geometry, never on OCR recognition, which is what lets unseen characters transfer to new scripts. At inference, Noise-Disentangled Classifier-Free Guidance (ND-CFG) reweights the guidance equation with a glyph-amplifying coefficient $\\omega_{ndg}$ and a glyph-free negative image condition, while Latent-Disentangled Two-Stage Rendering (LD-TSR) re-injects background latents from a first unconditional pass and applies a cosine-scheduled guidance boost that sharpens glyph details late in denoising; for small text, the pipeline denoises overlapping latent patches at up to 2048×2048 resolution, using patch-based denoising, to recover stroke-level clarity.","core_discovery":"The paper's central claim is that long-tail text failure in diffusion models is a disentanglement problem, not a data or encoder problem. HDGlyph establishes that the Canny edge map of the target glyph string is a script-agnostic structural prior: because edges encode stroke geometry without requiring the model to have memorized the character, a network trained only on English and Chinese glyphs can render Japanese, Russian, and other scripts it never saw, legibly, from their edge maps alone. The framework separates text from non-text at the pixel level with an edge-supervised glyph branch and per-language expert adapters, at the noise level with an extra guidance weight that amplifies the glyph condition against a glyph-free negative image condition, and at the latent level with a two-stage render that first paints the background and then refines the text, plus an upscaled patch-denoising pass for small text. The experiments assert that this hierarchy is what lets one model deliver both high text accuracy and high image quality, where prior systems trade one for the other.","pith_inferences":["A testable consequence the paper does not run: if edge maps are script-agnostic, a model trained on English alone should already render Chinese and Japanese legibly; measuring that zero-shot transfer would isolate how much of the gain comes from the edge prior rather than from the English-plus-Chinese training mixture.","The appendix records that the small-text patch mechanism relies on latent priors unique to the U-Net backbone and underperforms on the DiT variant, so making the latent disentanglement architecture-neutral is the natural next step.","Because the glyph condition is just an edge map, any renderable font becomes a possible conditioning input, suggesting a route to arbitrary scripts and stylized fonts that bypasses the need for paired training data in those scripts.","The dependence on thresholded edge maps implies a ceiling: cursive or connected-stroke scripts, whose Canny edges fragment at small sizes, may need a stroke-preserving preprocessing step before the same pipeline can claim full multilingual coverage."],"forward_implications":["Text accuracy no longer trades against background quality: the two-stage latent scheme recovers most of the image-quality loss that aggressive glyph guidance costs.","Unseen characters become renderable without retraining: the edge prior transfers across scripts, so the English- and Chinese-trained model produces legible Japanese and Russian with far smaller accuracy drops than the prior baselines.","Small text is addressed by upscaled patchwise denoising, cutting the measured accuracy drop on the small-text benchmark to 8.07%, versus 13.43% and 10.51% for the closest open-source systems.","The design is portable across diffusion architectures: the paper implements it on both a U-Net backbone and a DiT backbone and reports consistent gains on both.","Supervising glyphs through edge alignment removes the dependence on OCR models during training, eliminating a source of bias that recognition-based losses introduce."],"supporting_citations":[{"why":"Supplies the training corpus and the primary public evaluation benchmark on which the headline accuracy gains are measured.","marker":"[21]"},{"why":"The closest prior open-source system whose long-tail accuracy drops on multilingual and small text serve as the comparison baseline.","marker":"[20]"},{"why":"Provides the Canny-conditioned ControlNet whose structural priors initialize the Multi-Linguistic GlyphNet.","marker":"[26]"},{"why":"The DiT backbone on which the second implementation of the framework is built and tested.","marker":"[7]"},{"why":"The SDXL backbone on which the main U-Net implementation of the framework is built and tested.","marker":"[19]"},{"why":"Contributes the Complex-Benchmark of small-sized mixed-script posters used for the small-text evaluation.","marker":"[15]"},{"why":"Defines classifier-free guidance, the mechanism that ND-CFG extends with glyph-amplifying and negative-image weights.","marker":"[10]"},{"why":"Supplies the patch-based denoising strategy adapted for the upscaled small-text rendering pass.","marker":"[1]"}],"fun_headline_variants":["Glyph edges render unseen scripts and tiny text in diffusion models","Disentangling glyphs from scenes improves long-tail text rendering","Edge maps as script-agnostic keys for stable text in images","HDGlyph: stroke-conditioned diffusion handles rare words at any size","Two-stage rendering with edge priors fixes small and unseen text"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the Canny edge map of the target text is a complete and transferable structural prior, so a model trained only on English and Chinese glyphs can render unseen scripts and very small fonts from those edge maps alone; if edge extraction loses stroke-level detail for dense or tiny glyphs, the unseen-character and small-text claims weaken.","fun_headline_variants_meta":{"raw":{"variants":["Glyph edges render unseen scripts and tiny text in diffusion models","Disentangling glyphs from scenes improves long-tail text rendering","Edge maps as script-agnostic keys for stable text in images","HDGlyph: stroke-conditioned diffusion handles rare words at any size","Two-stage rendering with edge priors fixes small and unseen text"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00025,"raw_usage":{"total_tokens":1562,"prompt_tokens":961,"completion_tokens":601,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":577,"completion_tokens_details":{"reasoning_tokens":512}},"tokens_in":577,"tokens_out":601,"duration_ms":5510,"temperature":1.0,"reasoning_tokens":512,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:39:09.020005+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render a connected-stroke script such as Arabic or Devanagari at small sizes and measure per-character accuracy against the English baseline: if accuracy collapses toward the levels of the prior baselines, the edge prior is not complete for those stroke structures. A more direct check is to perturb the Canny thresholds used to build the glyph conditions and measure whether character accuracy tracks edge fidelity; if it does, the framework's ceiling is set by the edge extractor rather than the diffusion model.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The DiT backbone on which the second implementation of the framework is built and tested."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The SDXL backbone on which the main U-Net implementation of the framework is built and tested."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the patch-based denoising strategy adapted for the upscaled small-text rendering pass."}],"review_version":1}