acceptodds
Under review as a conference paper at ICLR 2027

TextVerifier: Seeing Beyond Recognition to Verify Chinese Text Rendering

Abstract

Text‑to‑image models increasingly generate readable Chinese text that still contains local rendering defects, such as a character with an extra or missing stroke or a malformed component. Such errors are invisible to OCR and vision–language models, which are trained for recognition and are therefore invariant to the very structural deviations that define a rendering error. We formalize text rendering verification as a target‑conditioned problem: given a target character and an image crop, decide whether the crop is a correct rendering of that character. Our verifier TEXTVERIFIER feeds the text tower a character’s Ideographic Description Sequence (IDS) rather than a single‑character symbol, and is trained on 12.7M crop–IDS pairs plus label‑free “same‑character, different‑quality” hard pairs sampled from the training trajectory of a text‑region inpainting model. To evaluate this ability we build TEXTRENDERBENCH‑ZH, the first per‑character benchmark for generated Chinese text, annotating position, identity, and correctness for 7,358 characters. TEXTVERIFIER reaches an F1 of 0.844, far above OCR confidence (0.526), a general VLM (0.329), and TextPecker (0.590). Used as a per‑character reward for Flow‑GRPO post‑training, it further reduces rendering errors, showing that its score can both detect and correct errors.

Primary area:
generative models
Keywords:
Text Rendering Verification, Text-to-image generation

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_Uap7OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces TextVerifier, a validator designed to detect structural rendering errors in Chinese characters, conditioned on the target character. The method employs a dual-encoder architecture, processing both images and Ideographic Description Sequences (IDS), and is trained in three stages: using authentic character images, synthetic stroke-level perturbations, and pairs of samples with varying relative quality (collected from the training trajectories of an image inpainting model). The authors also constructed TextRenderBench-ZH, a dataset comprising 7,358 characters extracted from 308 generated images. Experimental results demonstrate that this method achieves an F1 score of 0.844, outperforming both TextPecker (0.590) and an OCR-confidence-based approach (0.526, using an optimal threshold). Furthermore, the authors integrated this validator as a reward signal into Flow-GRPO to enhance the quality of Chinese text rendering.

Critical strengths
  1. The proposed benchmark provides character-level annotations of location, identity, and correctness, enabling direct evaluation of structural errors that recognition-based metrics may overlook.
  2. Ablation experiments fully demonstrate the effectiveness of the three-stage training process.
  3. Compared to the previous related work, TextPecker, it offers significant advantages in terms of model parameters and inference speed.
Critical weaknesses

1.In Section 4.1 and Appendix E, the benchmark construction is insufficiently documented. The generating models and versions, sampling distribution, number of annotators, inter-annotator agreement, and disagreement-resolution procedure are not reported. Since distinguishing legitimate stylistic variation from rendering errors is central to the task, these omissions make it difficult to assess annotation reliability and generalization across generators and styles. 2.The comparison does not sufficiently characterize the difference in language coverage between TextVerifier and the baselines. TextVerifier is specifically designed around Chinese character structure, and its quantitative evaluation focuses on Chinese. However, TextPecker covers both English and Chinese. A Chinese-only comparison is relevant to the stated task, but provides an incomplete account of their respective applicability. More importantly, the paper does not evaluate whether Chinese-focused reward optimization preserves English or mixed Chinese–English rendering capabilities. 3.The evidence supporting TextVerifier as an optimization reward in Section 4.5 is insufficiently comprehensive. The experiments use only Qwen-Image and evaluate on oneID-ZH, leaving transferability to other generator families unverified. Although GlyphPrinter is briefly cited in Section 2, the paper lacks a substantive comparison with its region-level preference optimization approach or an experimental comparison under comparable conditions.

Questions for authors

1.How are English letters, digits, and mixed-language text handled by the verifier and the glyph reward? Could the authors report English and mixed Chinese–English rendering performance before and after post-training to assess whether these capabilities are preserved? 2.Could the authors evaluate the reward on another compatible generator family, such as FLUX, using controlled prompts and training budgets? This would help establish whether the improvement extends beyond Qwen-Image. 3.Could the authors provide a substantive comparison with GlyphPrinter and, where feasible, evaluate the approaches under comparable conditions? An independent human evaluation of glyph correctness would also help distinguish actual rendering improvements from improvements according to the optimized verifier.

6 comments
AuthorsOpenReview ↗

Response to Reviewer Uap7 (Part I)

Comment: Chinese-specific verification and preservation of English rendering capabilities

Reviewer concern. TextVerifier is specifically designed around Chinese character structure, whereas TextPecker covers both English and Chinese. The paper does not evaluate whether Chinese-focused reward optimization preserves English rendering capabilities.

Response. We thank the reviewer for recognizing our character-level benchmark, three-stage training ablations, and model efficiency. Your suggestions regarding language coverage, cross-generator applicability, and related-method comparisons have helped us strengthen the empirical evidence and clarify the scope of our work. We have added language-capability preservation experiments, independent human evaluation, and reward-based optimization experiments on a FLUX-family model.

(1) Why Chinese rendering requires additional structural verification. We appreciate the request to clarify language coverage. Our focus on Chinese structure is motivated by a mismatch between the granularity of recognition-based evaluation and the errors we seek to detect.

For English, character-level edit distance operates on the letters within a word. If a rendering error causes a letter to be recognized as a different letter, omitted, or added, the resulting OCR string differs from the target, and NED can capture that discrepancy. A capable OCR model therefore provides a useful signal for assessing letter-level content and spelling within English words.

For Chinese, the corresponding unit is an entire Chinese character. A glyph may contain a missing or extra stroke while still being recognized as the intended character. In this case, the OCR output matches the target, and NED does not penalize the internal structural error. Recognition-based metrics can detect character substitutions, insertions, and deletions, but they cannot reveal an internal glyph defect when the recognized character identity remains unchanged. Our examples and the limited glyph-error detection performance of the OCR baseline illustrate this limitation.

Therefore, it motivates a dedicated verifier for the fine-grained structural errors within Chinese characters that recognition-based evaluation may overlook. TextVerifier complements recognition-based evaluation rather than replacing it across all languages.

(2) English rendering after Chinese-focused post-training. To examine whether Chinese-focused post-training harms English generation, we evaluated the same checkpoint obtained through TextVerifier-guided Flow-GRPO on the English Text Rendering task of OneIG, following its official text evaluation protocol.

Table 1. English rendering performance before and after Chinese-focused post-training.

ModelText score ↑Word Accuracy (WAC) ↑
Base0.92460.6886
Ours, after Chinese-focused post-training0.94370.7163
Change+0.0191+0.0277

The English text metrics improve rather than decline, indicating no observed degradation in English rendering on this evaluation. One possible explanation for the improvement is the auxiliary NED term in our training reward. Besides providing Chinese content-alignment supervision, this term encourages more complete and accurate rendering of the text requested in the prompt, potentially producing positive transfer to English generation. We present this as a possible explanation rather than an established causal mechanism.

AuthorsOpenReview ↗

Response to Reviewer Uap7 (Part II)

Comment: Independent human evaluation and the shared reward model and evaluator

Reviewer concern. An independent human evaluation of glyph correctness would help distinguish actual rendering improvements from improvements according to the optimized verifier.

Response. We agree that using TextVerifier for both reward computation and automatic Glyph-Acc evaluation requires independent human evidence to validate actual improvements.

(1) Evaluation protocol. We randomly selected 100 prompts from OneIG-ZH and compared Ours against Base and the TextPecker-reward model. Five evaluators participated in a blind GSB (Good/Same/Bad) pairwise assessment: method names were hidden, and the left–right placement of outputs was varied. Judgments were based solely on Chinese glyph correctness, without considering overall aesthetics or layout preferences. The final judgment for each image pair was determined by majority vote.

Table 1. Blind human evaluation of Chinese glyph correctness on 100 matched prompts.

Compared againstOurs wins (G)Ties (S)Ours loses (B)Net win rate
Base (Qwen-Image)523315+37.0%
TextPecker-reward274429−2.0%

(2) Results and interpretation. Against Base, Ours wins substantially more comparisons than it loses, achieving a 77.6% win rate among non-tied comparisons. Against TextPecker-reward, the win–loss distribution is broadly balanced. These results are consistent with the trend in Table 4 of the original manuscript, where Ours substantially outperforms Base and achieves performance close to TextPecker-reward. The human judgments provide evidence independent of TextVerifier’s automatic scores that Chinese rendering correctness improves over Base.

AuthorsOpenReview ↗

Response to Reviewer Uap7 (Part III)

Comment: Applicability of the reward to another generator family

Reviewer concern. Could the authors evaluate the reward on another compatible generator family, such as FLUX?

Response. To address cross-generator applicability, we added experiments on FLUX 3 Image Klein Base, using the same TextVerifier reward formulation for Flow-GRPO post-training.

(1) Automatic evaluation. The additional results are reported below.

Table 1. Automatic Glyph-Acc before and after Flow-GRPO post-training on FLUX.

ModelAutomatic Glyph-Acc ↑
FLUX 3 Image Klein Base0.445
FLUX 3 Image Klein, post-trained with TextVerifier reward0.501

Glyph-Acc increases from 0.445 to 0.501, an improvement of 5.6 percentage points.

(2) Independent human evaluation. Since this metric is still computed by TextVerifier, we additionally conducted a human GSB comparison between the post-trained model and Base.

Table 2. Human GSB evaluation of the post-trained FLUX model against Base.

ComparisonWins (G)Ties (S)Losses (B)Net win rateNon-tie win rate
Post-trained model vs. Base344224+10.0%58.6%

The human evaluation likewise shows more wins than losses, although the advantage is modest and many comparisons are tied. Together with the automatic results, this provides additional evidence that the TextVerifier reward can improve rendering in a generator family beyond Qwen-Image.

We also observe that the FLUX base model starts with lower performance than Qwen-Image on this Chinese rendering task, and the current post-training does not close the gap between their final performance levels. This is consistent with the dependence of RL post-training outcomes on the capabilities of the starting model.

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

The paper studies Chinese text rendering verification: given a target character and an image crop, the goal is to determine whether the crop correctly renders that character.

The proposed method employs a CLIP-style dual encoder that represents target characters using Ideographic Description Sequences (IDS). The model is trained through a three-stage curriculum with stage-specific training data and loss functions.

The paper also introduces TextRenderBench-ZH, a character-level benchmark for evaluating rendering correctness.

Critical strengths
  1. The method performs well and runs efficiently. The proposed method achieves better precision and recall than the evaluated baselines while maintaining fast inference. The Flow-GRPO results also suggest that its scores provide a useful training signal, achieving performance comparable to TextPecker with a much smaller reward model.

  2. The ablations help explain the role of the training stages and losses.

Critical weaknesses
  1. It is unclear whether the gains come mainly from the method or the training data. TextVerifier receives extensive task-specific training, while the baselines are not adapted using the same data. Fine-tuning Qwen-VL or TextPecker on these data would make the comparison more convincing.

  2. The evaluation needs stronger independent checks. It is also unclear whether KK was chosen based on test results. For Flow-GRPO, TextVerifier serves as both the reward and the evaluator, so human evaluation would help confirm the gains. Figure 1 also changes text size and layout, making it difficult to separate better glyph structure from improved readability

Questions for authors

See Weaknesses and:

  1. Table 1:why does stage 2 bring only a small improvement, while stage 3 gives a much larger gain?

  2. Why does inference use a different scoring method from training?

  3. Is independent human evaluation available for the Flow-GRPO outputs? In particular, does the improvement persist when character size and resolution are controlled, and can the verifier distinguish structurally correct but blurred glyphs from clear glyphs with genuine stroke errors?

5 comments
AuthorsOpenReview ↗

Response to Reviewer hNec (Part I)

Comment: Contributions of the method versus the training data

Reviewer concern. It is unclear whether the gains come mainly from the method or the training data. TextVerifier receives extensive task-specific training, while the baselines are not adapted using the same data. Fine-tuning Qwen-VL or TextPecker on these data would make the comparison more convincing.

Response. We thank the reviewer for recognizing our detection performance, inference efficiency, and training ablations. Your suggestions have helped us strengthen the experiments. We have added a task-adapted VLM baseline, blind human evaluation, and readout comparisons using a frozen model. This part addresses the contributions of task-specific data and method design.

(1) Task-adapted VLM baseline. We agree that comparing off-the-shelf models with a task-trained TextVerifier does not fully separate the contributions of training data and method design. Following your suggestion, we fine-tuned Qwen3.5-9B using our Stage-2 and Stage-3 training data. We used the same target-conditioned evaluation prompt as in Appendix D and evaluated the model before and after SFT on TextRenderBench-ZH.

Table 1. Glyph-error detection before and after task-specific SFT.

MethodPrecision ↑Recall ↑F1F_1 ↑
Qwen3.5-9B, without task-specific fine-tuning0.7140.1230.210
Qwen3.5-9B, SFT on our data0.3380.4800.397
TextVerifier0.8300.8580.844

Task-specific data benefit the VLM: SFT increases F1F_1 from 0.210 to 0.397. Recall improves from 0.123 to 0.480, while precision decreases, indicating that the adapted model detects more errors but also produces more false positives. Under the current SFT configuration, its performance remains substantially below TextVerifier.

(2) Different supervision semantics across training stages. An important consideration when adapting our data to other models is that the three stages provide different forms of supervision, as described in §3.3.

Table 2. Data characteristics and supervision across the three stages.

StageData characteristicsAvailable supervision
Stage 1Predominantly correctly rendered real-character crops and standard-font renderingsCharacter identity and normal glyph representations
Stage 2Correct glyphs and synthetically constructed erroneous glyphsAbsolute correct/incorrect labels
Stage 3Paired renderings of the same target character from early and late inpainting-model checkpointsNoisy relative quality relationships, rather than definitive positive/negative labels

In particular, an early-checkpoint rendering is not necessarily incorrect, and a late-checkpoint rendering is not necessarily correct. Consequently, converting these samples into a uniform absolute-correctness SFT objective may not be the most suitable way to exploit all available supervision. Our curriculum uses identity alignment, absolute discrimination, and relative ranking to accommodate these distinct data characteristics.

We view the proposed VLM adaptation as an informative comparison. The results confirm the value of the training data, while the supervision differences motivate our stage-specific objectives.

(3) Additional supporting observations. We would also like to highlight two points:

  • TextPecker also uses similar synthetic negative samples. When constructing its SFT data, TextPecker synthesizes negative samples using an approach conceptually similar to our Stage 2. Nevertheless, its final F1F_1 on TextRenderBench-ZH is 0.590, substantially below TextVerifier’s 0.844.
  • Our ablations support the model and training design. The ablation study examines changes to both the training stages and the model design, including replacing IDS with single-character inputs. None of these ablated configurations matches the full model’s performance, providing complementary evidence for the effectiveness of our chosen model and training approach.
AuthorsOpenReview ↗

Response to Reviewer hNec (Part II)

Comment: Shared reward model and evaluator, and independent human evaluation

Reviewer concern. For Flow-GRPO, TextVerifier serves as both the reward and the evaluator, so human evaluation would help confirm the gains.

Response. We agree that when TextVerifier serves as both the reward model and the automatic Glyph-Acc evaluator, an increase in the automatic metric alone is insufficient to confirm a reduction in actual glyph errors. We have therefore added an independent human evaluation.

(1) Evaluation protocol. We randomly selected 100 prompts from OneIG-ZH and compared Ours against Base and the TextPecker-reward model, using the same prompt set for both comparisons. Given approximately 30 target characters per prompt on average, we adopted a GSB (Good/Same/Bad) pairwise evaluation focused specifically on Chinese glyph correctness. Five evaluators assessed the image pairs with method names hidden and left–right positions swapped across comparisons. Judgments were based solely on glyph errors involving strokes and components, without considering overall aesthetics or layout preferences. The final judgment for each image pair was determined by majority vote.

Table 1. Blind human evaluation of Chinese glyph correctness on 100 matched prompts.

Compared againstOurs wins (G)Ties (S)Ours loses (B)Net win rate
Base (Qwen-Image)523315+37.0%
TextPecker-reward274429−2.0%

(2) Results and interpretation. Against Base, Ours wins substantially more comparisons than it loses, achieving a 77.6% win rate among non-tied comparisons. Against TextPecker-reward, the win–loss distribution is broadly balanced. These judgments are consistent with the automatic Glyph-Acc results in Table 4: Ours substantially outperforms Base while remaining close to and slightly below TextPecker-reward. The human evaluation provides evidence independent of TextVerifier’s automatic scores that Chinese glyph correctness improves over Base.

AuthorsOpenReview ↗

Response to Reviewer hNec (Part III)

Comment: Smaller immediate gains from Stage 2 and larger gains from Stage 3

Reviewer concern. Why does Stage 2 bring only a small improvement, while Stage 3 gives a much larger gain?

Response. We thank the reviewer for carefully examining our paper. This is an important observation. We believe it is necessary to distinguish the immediate performance gain at the end of a stage from that stage’s contribution to subsequent training.

(1) Complementary roles of Stages 2 and 3. Stage 2 provides absolute correctness supervision through synthetic errors. However, a distribution gap remains between manually constructed errors, such as stroke additions and deletions, and actual generation errors. Its immediate improvement on a benchmark of generated images is therefore limited. Stage 3 instead uses paired outputs from a generative model’s training trajectory, whose error patterns more closely resemble the target setting, thereby further improving sensitivity to actual generation defects.

However, relative supervision in Stage 3 cannot independently establish an absolute boundary between correct and incorrect renderings: late-checkpoint outputs may still contain errors, while early-checkpoint outputs may already be correct. The absolute-discrimination supervision provided by Stage 2 therefore plays a complementary role.

(2) Evidence from the ablations. The results in Table 3 of the original manuscript support this complementary relationship.

Table 3 (original manuscript). Ablations of Stage 2 and its absolute-discrimination loss.

ConfigurationPrecisionRecallF1F_1
Full three-stage training0.8300.8580.844
Skip Stage 2 and proceed directly from Stage 1 to Stage 30.7670.8610.811
Remove the Stage-2 absolute-discrimination loss during Stage 30.7740.8640.817

Skipping Stage 2 reduces final F1F_1 by 3.3 percentage points; removing its absolute-discrimination loss retained during Stage 3 reduces final F1F_1 by 2.7 percentage points. Both ablations achieve recall close to that of the full model, but with substantially lower precision. This supports our interpretation: Stage 2 and its absolute-discrimination supervision help control false positives, and their value cannot be assessed solely by the immediate F1F_1 gain at the end of Stage 2.

(3) Precision–recall trade-off. We do not claim that Stage 3 outperforms Stage 2 on every metric. Precision decreases from 0.864 to 0.830, while recall increases from 0.754 to 0.858, reflecting a trade-off between greater error sensitivity and an increase in false positives. Overall, synthetic absolute supervision helps constrain the decision boundary, while generator-derived pairs improve sensitivity to the target error distribution; together, they produce the final performance gain.

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

TextVerifier detects incorrectly rendered Chinese characters, including errors that OCR can still read correctly. It uses IDS-based training and automatically constructed character examples. The paper introduces a human-labeled benchmark and uses the verifier as a Flow-GRPO reward.

Critical strengths
  1. This addresses a real gap in OCR-based evaluation. The detection results look strong against the reported baselines and the character-level benchmark is useful.

  2. The IDS and training-stage ablations support the design. Mining examples from the inpainting model's checkpoints is a practical idea. The compact verifier is also attractive as a reward model.

Critical weaknesses
  1. The RL experiment uses TextVerifier as both the reward and the evaluator. A higher Glyph-Acc can reflect better agreement with this judge. I would like an independent check of whether the post-trained model actually makes fewer glyph errors.
Questions for authors

Is there a human evaluation of glyph errors after Flow-GRPO, using the same prompts for the base model and the different reward variants?

1 comment
AuthorsOpenReview ↗

Response to Reviewer sCjg

Comment: Independent human evaluation of glyph correctness

Reviewer concern. The RL experiment uses TextVerifier as both the reward and the evaluator. A higher Glyph-Acc can reflect better agreement with this judge. Is there a human evaluation of glyph errors after Flow-GRPO, using the same prompts for the base model and the different reward variants?

Response. We thank the reviewer for recognizing the relevance of our research question, the detection performance, the IDS and training-stage ablations, and the practicality of a compact reward model. We also appreciate the concern about independent evaluation.

We agree that using TextVerifier as both the reward model and the automatic Glyph-Acc evaluator limits the independence of the original evaluation. Given the limited availability of reliable tools for fine-grained Chinese glyph verification, we used TextVerifier because it substantially outperformed the evaluated alternatives on our benchmark. Nevertheless, this does not exclude the possibility that post-training improves agreement with the optimized evaluator. We have therefore added a blind human evaluation to independently assess rendering correctness.

(1) Human evaluation protocol. We randomly selected 100 prompts from OneIG-ZH and compared Ours against Base (Qwen-Image) and the TextPecker-reward model, using the same prompt set for both comparisons.

Each prompt contains approximately 30 target characters on average, making exhaustive character-by-character, stroke-by-stroke assessment of every output highly labor-intensive. We therefore adopted a GSB (Good/Same/Bad) pairwise evaluation. For each pair of images generated from the same prompt, evaluators judged whether Ours was better, comparable, or worse solely in Chinese glyph correctness, considering missing or extra strokes, malformed components, and related structural errors. Five evaluators participated in the blind assessment. Method names were hidden, and the left–right placement of outputs was varied to reduce identity and position biases. The final judgment for each image pair was determined by majority vote, yielding 100 final judgments per comparison.

(2) Human evaluation results. All wins, ties, and losses below are reported from the perspective of Ours.

Table 1. Blind human evaluation of Chinese glyph correctness on 100 matched prompts.

Compared againstWins (G)Ties (S)Losses (B)Net win rateNon-tie win rate
Base (Qwen-Image)523315+37.0%77.6%
TextPecker-reward274429−2.0%48.2%

The rates are defined as

\qquad \text{Non-tie win rate}=\frac{G}{G+B}.$$ Against Base, Ours wins 52 comparisons, loses 15, and ties 33, achieving a 77.6% win rate among non-tied comparisons. These judgments provide evidence independent of TextVerifier’s automatic scores that the glyph reward improves Chinese rendering correctness. Against TextPecker-reward, the judgments are broadly balanced: Ours wins 27 comparisons, loses 29, and ties 44. This supports our observation that the two glyph rewards achieve comparable rendering performance, while TextVerifier uses a substantially smaller reward model (**0.36B vs. 8.29B**). **(3) Consistency with automatic evaluation.** The human judgments follow the same comparison trend as Table 4: Ours achieves substantially higher automatic Glyph-Acc than Base (**0.915 vs. 0.812**), while remaining close to and slightly below TextPecker-reward (**0.915 vs. 0.919**). Thus, the improvement over Base is supported by both automatic evaluation and independent human assessment. We will include the human evaluation protocol and results in the final manuscript. The GSB study adds independent pairwise evidence of improved glyph correctness, addressing the concern that the reported gains might only reflect greater agreement with TextVerifier. We thank the reviewer for prompting this additional validation.

Comments on the paper

2 comments
Area_Chair_miYmOpenReview ↗

Reviewer Discussion

The authors have provided their rebuttal and additional experiments. Given the differing review scores, please take another look at the responses and discuss any remaining concerns. Feel free to update your scores if your concerns have been addressed.

AuthorsOpenReview ↗

Overall Comment: Rebuttal Summary and Key Contributions

Dear Area Chair and Reviewers,

We sincerely thank the reviewers for their constructive feedback and the Area Chair for facilitating the discussion. We have conducted additional experiments to address the main concerns and would like to briefly summarize the new evidence and the broader contributions of our work.

1. Core Contributions

Our work highlights a gap between character recognition and glyph correctness: a Chinese character may be correctly recognized by OCR or VLMs despite missing strokes, extra strokes, or malformed components. We formulate target-conditioned glyph verification, introduce TextVerifier with IDS-based representations and three-stage training, and construct TextRenderBench-ZH with character-level annotations.

TextVerifier achieves 0.844 F1 in glyph-error detection, compared with 0.526 for OCR confidence and 0.590 for TextPecker, while using a compact and efficient model. As a Flow-GRPO reward, it also improves Qwen-Image's Glyph-Acc from 0.812 to 0.915, connecting fine-grained verification to generation improvement.

2. Rebuttal Summary: Additional Evidence

(a) Contributions of data and method design. To disentangle these factors, we trained a task-adapted Qwen3.5-9B baseline using our Stage 2/3 data, improving its F1 from 0.210 to 0.397. This confirms the value of our task-specific data, although the adapted VLM remains below TextVerifier (0.844). Our ablations show complementary roles for absolute correctness supervision and generator-derived relative quality supervision. Taken together, data construction and method design are complementary rather than interchangeable. Our three-stage data pipeline provides rich supervision signals with different semantics, while our design, e.g., the IDS-based representations, stage-specific training objectives and so on, enables the model to exploit these signals more effectively. Their joint design is an important factor underlying TextVerifier's strong performance.

(b) Independent evaluation of reward optimization. We recognize the concern that using TextVerifier as both reward and evaluator could inflate reported gains. We therefore added a blinded, five-rater GSB study on 100 matched prompts. Against Qwen-Image Base, our method obtains 52 wins / 33 ties / 15 losses (77.6% win rate excluding ties), independently supporting improved glyph correctness. Against TextPecker-reward, results are comparable (27 / 44 / 29), while our reward model is much smaller (0.36B vs. 8.29B). Thus, the observed improvement over Base is not supported solely by our own automatic metric.

(c) Applicability, language preservation, and scope. On a FLUX-family generator, TextVerifier-guided Flow-GRPO increases automatic Glyph-Acc from 0.445 to 0.501; independent human comparisons also favor the optimized model (34 wins / 42 ties / 24 losses), albeit modestly. English rendering does not degrade on OneIG: Text Score rises from 0.9246 to 0.9437 and Word Accuracy from 0.6886 to 0.7163. We also expanded the comparison with GlyphPrinter, clarifying differences in objectives and conditioning. Our scope remains Chinese glyph-structure verification; since RL does not fix layout or text size, we do not claim improvements under strictly controlled typography. We will clarify these limits and benchmark annotation details in the revision.

3. Key Strengths and Broader Significance

We believe the paper's value extends beyond an improvement on one metric. First, it identifies an underexplored failure mode and provides a precise task formulation and character-level benchmark for studying it. Second, it offers a coherent solution combining complementary supervision, structure-aware representations, and efficient inference, now supported by stronger baselines, ablations, and independent human evidence. Third, it shows that an efficient verifier can also serve as a practical reward for improving text-to-image generation, with initial evidence across two generator families.

These contributions form a connected path from problem identification and evaluation to model design and downstream optimization. We believe the central claims are well supported within the stated scope.

We appreciate Reviewer Uap7's acknowledgement that our responses addressed their concerns and their stated intention to raise the score. We thank all reviewers and the Area Chair again for their time and thoughtful consideration!

Sincerely,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.