{"id":"a64a4d15-98aa-4265-aa2f-c4c86622f8bb","arxiv_id":"2607.22101","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A unified DiT model with font-size-aware modulation and region-weighted loss outperforms existing visual text generation and editing systems on bilingual benchmarks.","lead":"InnoText is a single diffusion-transformer model that can both generate and edit visible text in images, with extra modules aimed at small text and Chinese characters. It reports better accuracy and visual quality than several previous text-generation and editing systems on English and Chinese benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Chinese generation benchmark is self-referential: InnoText-Bench is sampled from the same synthetic Seedream corpus used for training, so reported Chinese Sen. Acc gains may not generalize.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the InnoText-30K dataset, especially the Seedream-generated Chinese subset, may not be representative of real-world text images, and the self-referential benchmark can inflate apparent gains. I agree with this and refine it to focus on the Chinese generation claim, which has no external validation. The paper does have independent support: editing is evaluated on AnyText-Benchmark (external), and supplementary MARIO-Eval results give external evidence for English generation. However, the Chinese generation claim is exclusively based on InnoText-Bench, sampled from the same synthetic corpus used for training. A concrete external Chinese generation benchmark is the missing test. If the external test confirms the gains, the concern is resolved; if not, the central generation claim is weakened. The reader's verdict of CONDITIONAL is appropriate because the issue is addressable but currently unresolved. I therefore set verdict_should_be to UNCHANGED, meaning my stress test does not move the reader's conditional verdict. A secondary internal inconsistency (Eq. 10's region-weighted loss uses the size map S, while the prose says weights are on masked/modified areas) also needs clarification, but it is less load-bearing than the benchmark self-reference because it mainly affects reproducibility, not the validity of the empirical comparison.","tokens_in":14426,"tokens_out":5554,"duration_ms":61975,"concrete_test":"Construct a Chinese generation test set of 500 real-world images (e.g., storefronts, signs, posters) with held-out ground-truth text overlaid using diverse real-world fonts and perspectives, or manually verify a subset of existing OCR datasets. Evaluate the released InnoText and AnyText2 on this set and compare Sentence Accuracy. If the improvement over AnyText2 is substantially smaller than the 6.5-point gap in Table 1 (0.5907 vs 0.5262) and does not exceed statistical noise, the InnoText-Bench result is inflated by distribution overlap rather than reflecting genuine generalization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of superior visual text generation, especially for Chinese, rests on InnoText-Bench, which is sampled from InnoText-30K (§3.2). The Chinese subset of InnoText-30K is synthetic, generated by Seedream from translated and augmented Lex-10K prompts, then filtered by aesthetic scores, PPOCR re-annotation, and DeepSeek-VL2 captioning. If Seedream produces text images with systematic artifacts—limited font diversity, clean rendering, or OCR-friendly glyphs—the model trained on this distribution will score artificially high on a benchmark drawn from the same distribution. The paper provides no distributional evidence beyond NIMA aesthetics and NIQE naturalness (Fig. 4), which do not measure text realism or alignment with real-world scene text. The main editing results use the external AnyText-Benchmark, but the generation results in Table 1 use InnoText-Bench. Supplementary Table 4 reports MARIO-Eval, an external benchmark, but this is predominantly English and does not validate the Chinese generation claim. Thus the reported Chinese generation superiority (Sen. Acc 0.5907 vs 0.5262 for AnyText2) may be inflated by train-test overlap, and the model's real-world Chinese generation ability remains unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes InnoText, a Diffusion-Transformer (DiT) model that unifies visual text generation and editing in a single network by building on Flux.1 Fill dev. Three components are introduced: Font Size-Aware Modulation (FSAM), Small-Character Aware Augmentation (SCAA), and a Task-Specific Region Weighted Loss. The authors also construct InnoText-30K, a bilingual English/Chinese dataset, and InnoText-Bench for evaluation. Experiments compare with AnyText, AnyText2, TextFlux, Flux-Fill, and, in the supplementary material, Flux-Text and EasyText, reporting sentence accuracy, NED, LPIPS, and human-preference results. The main claim is state-of-the-art performance on both generation and editing in English and Chinese.","tokens_in":14804,"tokens_out":5655,"duration_ms":61998,"significance":"The paper addresses a relevant gap: most strong DiT-based text models are single-task. If the claimed results hold, a single model that handles both generation and editing at good accuracy, with a new bilingual dataset, would be a useful contribution. The external editing benchmark (AnyText-Bench) and MARIO-Eval results provide some independent support, and the ablation study gives evidence for the contributions of FSAM and SCAA. However, the headline generation results, especially for Chinese, are evaluated on a benchmark sampled from the authors' own training corpus, and the main baseline comparisons are not matched in training data or compute. These issues currently leave the central generalization claim insufficiently supported.","major_comments":[{"comment":"InnoText-Bench is sampled from InnoText-30K, the same corpus used to train InnoText. The Chinese subset is generated by Seedream from translated/augmented Lex-10K prompts and filtered with PPOCR and DeepSeek-VL2; no evidence is provided that this synthetic distribution matches real Chinese scene text. The reported Chinese generation advantage over AnyText2 (Sen. Acc 0.5907 vs 0.5262) may therefore be inflated by train/test overlap. The NIMA/NIQE scores in Fig. 4 do not measure text realism or glyph distribution. Please add an externally curated Chinese generation test set or a cross-dataset evaluation, and clarify how the InnoText-Bench split avoids training overlap.","section":"§3.2, Table 1"},{"comment":"The main comparisons use baselines with their officially released pretrained weights; only AnyText and AnyText2 are fine-tuned on InnoText-30K in the ablation of Table 2. TextFlux, Flux-Fill, EasyText, and Flux-Text are not trained on the same data or with the same LoRA/compute budget in the headline tables. Since data composition and training volume strongly affect text rendering, the claimed superiority may reflect data/training disparities rather than a fundamental architectural advantage. The matched comparison in Table 6 (Ours 3M vs Flux-Text 3M) is more convincing; please extend this matched design to all headline baselines, including generation, or state explicitly where this is infeasible.","section":"§5.1, Tables 1 and 6"},{"comment":"The Task-Specific Region Weighted Loss is underspecified. The region weight S is not defined; the text says higher weights are assigned to masked or modified areas, but no formula, normalization, or dependence on the size map/mask is given. The flow-matching loss L_FM is also not explicitly defined. Without these definitions the method cannot be reproduced, and the ablation row 'w/o L_T_SRW' in Table 2 is not interpretable. Please provide the full loss expression, including how S is computed and normalized.","section":"§4.4, Eq. (10)"},{"comment":"No error bars, confidence intervals, or significance tests are reported. Several headline gaps are small: e.g., English editing Sen. Acc 0.7988 vs 0.7893 for AnyText2, Ours (3M) vs Flux-Text (3M) 0.8224 vs 0.8175, and some supplementary human-preference comparisons are near chance (e.g., 51.9% vs 48.1%). Given the evaluation set sizes (1,500 images for InnoText-Bench; 100 cases in the main human study), these differences may be within sampling noise. Please report bootstrap confidence intervals or paired significance tests for the central claims.","section":"Tables 1–6 and Fig. 10"}],"minor_comments":[{"comment":"Notation is inconsistent: §4.3 defines the small-character threshold as τ, the base resize factor as λ, and the perturbation range as ε; §5.1 says 'the size threshold λ and scaling factor ε are set to 1.5 and 0.3.' Please align the notation and specify which values correspond to τ, λ, and ε.","section":"§5.1 vs §4.3"},{"comment":"The main text refers to 'Fig. 11' and 'Fig. 12' for qualitative comparisons, but those figures appear only in the supplementary material; the corresponding main-paper figures are Fig. 6 and Fig. 7. Please fix the cross-references.","section":"§5.3"},{"comment":"Typo: 'Nagetive Effect' should be 'Negative Effect.'","section":"Supplementary §4.2"},{"comment":"Please state explicitly how the 1,500-image InnoText-Bench is split from the 30K training corpus and whether any sampled images were used during training, fine-tuning, or curation filtering. This is important for interpreting all results that use InnoText-Bench.","section":"§3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper contains a substantial number of self-citations to the authors' previous work, and the self-named InnoText-Bench is used for a key part of the evaluation. These are not, by themselves, reasons for rejection, but they increase the burden on external validation. I would encourage the editor to require, at revision, either an external Chinese generation benchmark or a clear cross-dataset analysis, and to ask the authors to release code and data to support reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, it is a serious engineering contribution: a single DiT model that does both visual text generation and editing, with a couple of well-motivated modules and a new bilingual dataset. Second, its headline Chinese-generation result rests on a benchmark sampled from the same synthetic data used for training, so the 0.5907 vs 0.5262 Sen. Acc gap should be read with caution.\n\nWhat is genuinely new: the diptych in-context input for unifying the two tasks, FSAM's size-map modulation, SCAA for small text, and the 30K bilingual dataset. The ablations are informative — removing FSAM hurts generation LPIPS most, and SCAA helps editing. The human eval, while small, favors InnoText clearly. The fairest test is Table 6, where InnoText trained on the same 3M data as Flux-Text beats it on most metrics; MARIO-Eval also gives an external check, and the numbers look good. So the core method is plausible.\n\nSoft spots. The self-referential benchmark is the big one: InnoText-Bench is drawn from InnoText-30K, which is also the training source, so the Chinese generation comparison is partially inflated. MARIO-Eval is predominantly English, so Chinese generation is not independently verified. Second, no error bars or significance tests anywhere; with 1,500 benchmark samples, some gaps could be noise. Third, Eq. 10 is confusing: it writes ||S·L_FM||2^2 for editing, but the prose describes a region-weighted loss and S is not defined there — likely a typo, but a reviewer will need it cleaned up. Minor: no code/data release yet, and the synthetic Seedream data is validated only by aesthetics/niqe, not text realism.\n\nOverall, this is a solid piece of work that deserves a serious referee. The main revisions I would push for: error bars, artifact release, and an external Chinese validation set to confirm the headline numbers. With that addressed, it becomes a strong conference paper rather than a conditional one. I would send it to peer review.","headline":"A credible unified DiT for bilingual text generation and editing, but the Chinese generation benchmark is partly self-referential and the lack of error bars makes the headline gains uncertain.","tokens_in":15258,"tokens_out":2488,"would_cite":true,"duration_ms":28755,"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":"InnoText claims that a single diffusion-transformer model can both generate and edit legible text in images across English and Chinese, outperforming task-specific systems on accuracy and visual quality.","keywords":["visual text generation","visual text editing","diffusion transformer (DiT)","font-size-aware modulation","small-character augmentation","bilingual English-Chinese dataset","scene text rendering","unified model"],"falsifier":"Hold out a set of real-world photos containing Chinese signs and labels, with varied fonts, lighting, and perspective, and run the model without any fine-tuning on that set; if OCR sentence accuracy on these real photos drops to the level of editing-only baselines, the reported edge is tied to the synthetic distribution rather than to the architecture.","tokens_in":14364,"feed_emoji":"🖼️","tokens_out":5896,"duration_ms":60742,"temperature":0.7,"pith_summary":"InnoText sets out to show that one Diffusion Transformer can handle both visual text generation (adding new text to an image from a prompt) and visual text editing (replacing text while keeping background and style) without losing to models specialized for only one task. The paper argues that existing U-Net pipelines mangle small or densely structured characters, especially in Chinese, while current transformer editors are stuck on editing alone. Its answer is a unified input format plus three ingredients: a Font Size-Aware Modulation that re-weights features by local font size, a Small-Character Aware Augmentation that magnifies tiny glyph regions during training, and a Task-Specific Region Weighted Loss that sharpens attention to edited areas. A new bilingual 30K-image dataset (English and Chinese) supports training and a two-language benchmark for evaluation. Reported results place this single model ahead of earlier systems on exact-text accuracy and perceptual quality in both languages.","feed_headline":"Single model beats specialists at image-text generation and editing","feed_subtitle":"One transformer handles Chinese and English, big and tiny fonts, with accuracy and visual fidelity ahead of specialized systems.","key_machinery":"The central device is the Font Size-Aware Modulation (FSAM): a spatial size map encoding inverse per-character height, patchified into tokens, linearly projected, and decoded by two MLPs into per-location scale and shift parameters that multiply and add into the hidden states of each Diffusion Transformer (DiT) block. This makes the network's internal gain depend on local font scale, counteracting the degradation of tiny glyphs. Two supporting mechanisms carry the same idea further: Small-Character Aware Augmentation randomly enlarges small-character regions during training (a 'foveal' zoom), and the Task-Specific Region Weighted Loss multiplies the flow-matching loss by the size map for edi","core_discovery":"The paper's central claim is that generation and editing are the same underlying operation once the model is told what to preserve. The input stacks a glyph image (the desired text, rendered) with either a masked image for editing or a blank canvas for generation, and a mask channel tells the model which half to paint. A font-size map—a grayscale image of inverse character heights—is injected into each transformer block to recalibrate features, so small characters are not neglected. A small-character augmentation and a region-weighted loss keep tiny strokes legible and focus optimization on edited pixels. Fine-tuned on 30K curated bilingual images, the model reports the best exact-text accur","pith_inferences":["A test not run in the paper: evaluating on real photographs of street signs or product labels would reveal whether the synthetic training distribution is the reason the model wins, since the proposed benchmark samples the same corpus used for fine-tuning.","The inverse-font-height map could be reused as a general 'importance map' for any dense, small-scale visual content (charts, barcodes, fine patterns), not just text.","If the unified model also holds up under instruction prompts (e.g., 'change X to Y'), the same architecture could evolve into a text-editing assistant rather than a two-mode tool."],"forward_implications":["If the central claim holds, one trained checkpoint can switch between generation and editing by changing only the input mask, removing the need for separate pipelines and their style drift.","The font-size map is language-agnostic, so the same recipe should transfer to other complex scripts such as Japanese or Korean with minimal re-curation.","Applying the region-weighted loss beyond text—e.g., to logo or caption edits—becomes a natural extension for other precision-critical image editing tasks.","The 30K dataset, if public, gives the field a common bilingual benchmark that separates text accuracy from perceptual quality."],"fun_headline_variants":["InnoText: one model for text generation and editing","Bilingual text generation and editing with font-size awareness","Unified transformer outperforms specialized text image models","Generation and editing merged: better text images"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole advantage rests on the assumption that the synthetic Chinese images used for training and evaluation resemble real-world Chinese text scenes closely enough; if they contain hidden rendering artifacts, the measured gains could be an artifact of the benchmark itself.","fun_headline_variants_meta":{"raw":{"variants":["InnoText: one model for text generation and editing","Bilingual text generation and editing with font-size awareness","Unified transformer outperforms specialized text image models","Generation and editing merged: better text images"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000638,"raw_usage":{"total_tokens":2772,"prompt_tokens":737,"completion_tokens":2035,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":481,"completion_tokens_details":{"reasoning_tokens":1986}},"tokens_in":481,"tokens_out":2035,"duration_ms":16458,"temperature":1.0,"reasoning_tokens":1986,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T05:45:15.037434+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out a set of real-world photos containing Chinese signs and labels, with varied fonts, lighting, and perspective, and run the model without any fine-tuning on that set; if OCR sentence accuracy on these real photos drops to the level of editing-only baselines, the reported edge is tied to the synthetic distribution rather than to the architecture.","supporting_citations":[],"review_version":1}