acceptodds
Under review as a conference paper at ICLR 2027

Blame Errant Words, Not the Canvas: Defect-Aware Spatial Credit Assignment for Visual Text Generation

Abstract

Recent reinforcement learning approaches have improved visual text generation, yet generated images still suffer from incorrect wording and malformed glyphs. We identify two coupled limitations: recognition-based rewards often overlook visible glyph defects, while whole-image scalar feedback obscures the uneven quality of individual text regions. These limitations motivate our central principle: *Blame errant words, not the canvas*, to effectively preserve local quality distinctions from evaluation to policy optimization. First, we introduce TextDART, a defect-aware region-targeted reward model that localizes individual text regions and independently evaluates their semantic correctness and glyph quality, providing fine-grained, region-precise feedback. Building on TextDART, we propose SCANNER, which turns these regional judgments into spatially differentiated policy feedback. It normalizes rewards for each text target across sampled images and casts the resulting regional advantages onto the corresponding latent tokens. Extensive experiments on existing benchmarks and our newly established PolyText-Bench demonstrate that our approach achieves state-of-the-art performance in text accuracy and glyph fidelity.

Primary area:
generative models
Keywords:
Visual Text Generation, Reward Model, GRPO, Flow Matching

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_8RjUOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

In the field of text rendering, most prior methods have focused only on improving OCR scores through GRPO with appropriately designed reward functions, while overlooking localized glyph errors. However, simple OCR detectors are insensitive to error glyphs, which makes it meaningful to develop a reward model capable of detecting localized glyph defects. To this end, the paper proposes TextDART, which can estimate defect-aware text regions. Building on this, the proposed SCANNER extends the advantage computation of vanilla GRPO to the token level, allowing correct and incorrect text regions within the same generated sample to be optimized differently. Evaluations on Chinese and English text-rendering benchmarks show that, when applied to base models such as Qwen-Image and FLUX, SCANNER outperforms other methods across multiple evaluation dimensions.

Critical strengths
  1. The paper proposes TextDART, which extends the previous TextPecker method to the region level, enabling the model to output defect-aware text regions in a single pass.

  2. The proposed SCANNER extends GRPO advantage computation to the token level, allowing different regions to be optimized differently according to their accuracy. In addition, according to the reported experimental results, it outperforms vanilla GRPO on both Chinese and English evaluation sets.

Critical weaknesses
  1. Limited Novelty. Although the paper introduces several techniques to construct a region- and glyph-sensitive GRPO-style algorithm, its overall novelty remains limited.
  • First, an algorithm targeting region-level glyph errors was previously introduced in GlyphPrinter (CVPR 2026) [2]. However, I do not see this work cited or included in the comparisons. To the best of my knowledge, GlyphPrinter proposes a region-level DPO algorithm and achieves strong performance on multilingual text-rendering benchmarks. In my opinion, SCANNER appears to adopt a similar idea and implement it within a GRPO framework.
  • Second, the idea of region-level GRPO used in SCANNER has also been explored in prior work, SpatialFlow-GRPO[1]. Concepts such as target-level policy ratios and region-level GRPO training objectives have already been implemented in a similar form in SpatialFlow-GRPO. The only apparent distinction between SCANNER and SpatialFlow-GRPO is that SCANNER introduces a global text score during the early timesteps. However, this simple modification does not establish sufficient originality. Moreover, the reward function in Eq. (5) appears similar to one previously introduced in TextPecker[4].
  • In addition, TextDART seems to be primarily a simple region-level extension of TextPecker. For example, combining a lightweight OCR detector (such as PPOCR) with TextPecker might achieve similar effectiveness and comparable inference speed.
  1. Insufficient Experiments and Coverage of Related Works.
  • The paper focuses exclusively on text-rendering models that accept text embeddings, while overlooking models conditioned on glyph images, such as EasyText[3] and GlyphPrinter[2]. Consequently, the experimental comparisons are not sufficiently comprehensive. In particular, GlyphPrinter, which is the method most closely aligned with the motivation of this paper, is not included in the comparisons. There is therefore insufficient evidence that the proposed SCANNER offers meaningful advantages over these advanced methods.
  • Furthermore, most of the qualitative cases presented in the paper focus on multi-region OCR metrics, while only a small number specifically examine glyph errors, even though glyph accuracy is the central problem the paper claims to address. I believe the model should be evaluated on more challenging benchmarks, including complex Chinese characters such as Traditional Chinese characters. To the best of my knowledge, advanced models such as Qwen-Image already achieve very high OCR and glyph accuracy in most simple scenarios. Therefore, merely outperforming other methods on relatively simple Chinese and English benchmarks is insufficient to demonstrate the practical value of the proposed SCANNER.
  1. Lack of Multilingual Text-Rendering Capabilities. The paper optimizes only for Chinese and English text rendering, without considering other languages, which limits its practical applicability. By contrast, glyph-image-conditioned methods only need to learn the mapping from text rendered in a default font to text in realistic scenes, which may allow them to generalize more effectively to other settings. According to the experiments reported in GlyphPrinter, the model is post-trained only on Chinese and English DPO datasets yet generalizes well to other languages. I therefore believe that the authors should provide additional experiments and analyses to demonstrate the main advantages of their algorithm over glyph-image-conditioned methods in improving individual glyph accuracy.

  2. Insufficient Theoretical Derivation. Intuitively, region-level GRPO should increase the generation probability of preferred regions while decreasing the probability of defective regions. However, the optimization objective corresponding to the GRPO formulation in Eq. (11) is not clear yet. The vanilla GRPO can be interpreted as optimizing the reward of the entire image. However, the authors do not derive the underlying optimization objective of their region-level GRPO formulation. As a result, its theoretical or physical interpretation remains unclear.

[1] Yang, Yankai, et al. “SpatialFlow-GRPO: Where Spatial Credit Drives Image Editing.”

[2] Shuai, Xincheng, et al. "GlyphPrinter: region-grouped direct preference optimization for glyph-accurate visual text rendering."

[3] Lu, Runnan, et al. "Easytext: Controllable diffusion transformer for multilingual text rendering."

[4] Zhu, Hanshen, et al. "Textpecker: Rewarding structural anomaly quantification for enhancing visual text rendering."

Questions for authors
  1. Comparison with glyph-image-conditioned models. How does SCANNER perform compared with models that take glyph images as input? Furthermore, could the authors quantify the performance gains achieved by SCANNER over competing baselines when rendering more complex glyphs, such as Traditional Chinese characters? I encourage the authors to extend the experimental evaluation to include such challenging settings. Besides, a detailed analysis of the factors underlying the improvements would further strengthen the paper.

  2. Novelty relative to prior work. Apart from several minor modifications, such as the Noise-aware Schedule, what are the key differences between the motivation and region-level GRPO algorithm proposed in this paper and those of prior works such as GlyphPrinter[2] and SpatialFlow-GRPO[1]? The answer will substantially influence my assessment of the paper’s originality and novelty. Please also cite these papers appropriately in the revised version and explicitly explain how SCANNER differs from them.

  3. Theoretical interpretation of Eq. (11). Similar to the analysis provided in GlyphPrinter, can the authors derive the underlying optimization objective corresponding to the GRPO algorithm in Eq. (11)? GlyphPrinter shows that its region-level DPO formulation essentially optimizes a region-level reward function together with a KL-divergence term. Can the authors derive an analogous result for SCANNER? I would like to better understand the theoretical or physical interpretation of Eq. (11).

  4. Role of the Noise-aware Schedule. I am curious about the practical role of the Noise-aware Schedule. Most prior works suggest that early denoising steps largely determine the overall layout and semantic content. Therefore, the rationale for assigning a larger global mixing weight during early timesteps is unclear. Even if the image content is ambiguous under extremely high noise levels, why would increasing global feedback be more beneficial than applying region-level rewards? I therefore encourage the authors to provide a more rigorous analysis of the mechanisms through which the global reward contributes to the observed performance improvements. In addition, when computing the global reward, how do the authors ensure that the concatenation order of the detected text is consistent with that of the ground-truth text?

[1] Yang, Yankai, et al. “SpatialFlow-GRPO: Where Spatial Credit Drives Image Editing.”

[2] Shuai, Xincheng, et al. "GlyphPrinter: region-grouped direct preference optimization for glyph-accurate visual text rendering."

9 comments
AuthorsOpenReview ↗

Response 1/5 — Novelty Relative to GlyphPrinter and SpatialFlow-GRPO (W1, Q2)

Thanks for your detailed and technically insightful feedback. We appreciate the recognition of TextDART's single-pass region-level defect perception and SCANNER's effectiveness on both Chinese and English generation. Below, we provide a point-by-point response to your concerns and questions.

W1 & Q2: Novelty Relative to GlyphPrinter, SpatialFlow-GRPO, and TextPecker (Part 1 of 2)

Our contribution is reliable text-target credit assignment under freely generated layouts. In prompt-only generation, requested text has a known identity but no predefined spatial support: its location and detection granularity vary across rollouts, and it may be split, merged, or omitted. Reliable optimization therefore requires establishing which regions represent the same target, how their defects should be evaluated, and when their spatial support should guide denoising updates. TextDART and SCANNER jointly address these challenges through defect-aware localization, cross-rollout target construction, and noise-dependent credit assignment.

(1) SCANNER vs. R-GDPO:

GlyphPrinter's R-GDPO learns from offline annotated regional preferences, which specify the regions to favor or suppress. However, in prompt-only online generation, establishing reliable regional supervision is itself a central challenge: text locations and detection granularity vary across rollouts, while requested content may be split, merged, or omitted. Consequently, detected regions do not automatically provide consistent units for reward comparison and credit assignment.

SCANNER addresses this challenge by constructing content-consistent, defect-aware target supervision online:

  • From input-specified regions to output-grounded targets: R-GDPO's input-conditioned regions constrain layouts; SCANNER constructs text-block identities online, preserving layout and font freedom.
  • From annotated preferences to comparable rollout feedback: R-GDPO's intra-sample pairs may differ in content and rendering difficulty, confounding glyph-quality comparisons. SCANNER compares matching targets across rollouts.
  • From detected content to explicit coverage of omissions and additions: R-GDPO contrasts correct/incorrect glyph regions; SCANNER assigns rewards/advantages to coherent text blocks for region-specific optimization.

The resulting contribution is an online mechanism for establishing what to compare and where to assign credit when neither text regions nor their cross-sample correspondence are given in advance.

(2) SCANNER vs. SpatialFlow-GRPO

SpatialFlow-GRPO provides semantic-region feedback for image editing and normalizes regional rewards within instruction–label groups. For freely generated text, the critical challenge is making the comparison units consistent when detection boundaries and target coverage change across rollouts. Furthermore, spatial support inferred from the final image may not be equally suitable for assigning credit at every denoising step. SCANNER addresses these issues through four connected designs:

  • Target consistency: SpatialFlow normalizes within (instruction, label) groups tied to SFReward outputs. SCANNER builds a shared cross-rollout target partition through many-to-many matching and merging, including missing/extra text (Appendix A.1.1).
  • Local/global separation: SpatialFlow multiplies regional rewards by global perceptual quality. SCANNER excludes global quality from regional evaluation, using global text feedback as a separate advantage anchor and spatial KL (Appendix A.1.2).
  • Credit budgeting: SpatialFlow's default power weighting remains area-dependent. SCANNER assigns the same area-independent regional aggregation coefficient to each text target within an image regardless of area or token count, preventing large text from dominating small text (Appendix A.1.3).
  • Temporal attribution: SpatialFlow fixes regional advantages across denoising. SCANNER fuses global/regional credit by transition noise, emphasizing global credit early and regional credit as localization becomes reliable (Appendix A.1.4).
AuthorsOpenReview ↗

Response 2/5 — Controlled Comparisons and TextDART versus TextPecker (W1, Q2)

W1 & Q2: Novelty Relative to GlyphPrinter, SpatialFlow-GRPO, and TextPecker (Part 2)

Controlled Comparison of SCANNER and SpatialFlow-GRPO. With FLUX.1-dev and TextDART fixed, SCANNER achieves stronger overall performance (Tab. R1). The results support the benefit of SCANNER's target-level credit-assignment design beyond the reward model alone.

RL algorithmLongText-Bench / TextScore ↑OneIG-Bench / TextScore ↑OneIG-Bench / ED ↓OneIG-Bench / CR ↑OneIG-Bench / WAC ↑CVTG-2k / WordAcc ↑CVTG-2k / NED ↑GenTextEval-Bench / Qua ↑GenTextEval-Bench / Sem ↑PolyText-Bench / CS ↑PolyText-Bench / CG ↑PolyText-Bench / OG ↑
SpatialFlow-GRPO-style0.8220.84054.080.2790.5890.8090.9000.8390.6120.8650.8830.529
SCANNER0.8270.85851.060.3090.5980.8340.9140.8390.6130.8820.9050.653

Table R1. Controlled comparison of SpatialFlow-GRPO-style and SCANNER. SCANNER goes beyond applying existing region-level GRPO to text: its text-target credit assignment delivers stronger performance than SpatialFlow-GRPO-style training with the same backbone and reward model.

(3) TextDART vs. TextPecker(W1 & Q2)

Eq. (5) is motivated by TextPecker (as acknowledged in Section 4.1). TextPecker aggregates uniformly counted anomalies into an image-level reward. TextDART produces target-specific rewards from severity-weighted character-, word-, and block-level defects (Appendix A.8), with many-to-many matching, cross-rollout merging, and explicit missing/extra-text rewards (Appendix A.1.1).

Why OCR + TextPecker is not equivalent to TextDART. TextPecker supports box-conditioned evaluation, but each call accepts only one externally supplied box. This design depends on OCR localization and requires region-by-region evaluation, leading to three main limitations:

  • OCR-first evaluation propagates false boxes and confidence-filtered omissions.
  • OCR's line-oriented boxes need not match the semantic text blocks used by SCANNER.
  • N boxes require N calls, making evaluation cost scale linearly with the number of regions.

TextDART jointly localizes semantic blocks and recognizes defects in one pass, mitigating these error-propagation, granularity, and efficiency bottlenecks.

We evaluate OCRv5+TextPecker alternatives, which yield weaker overall generation quality than TextDART+SCANNER (Tab. 1). Training-side measurements (Tab. R2) show that TextDART reduces per-step time by 11.6% and GPU-hours to match the baseline's final OCR score by 41.1%.

Efficiency metricOCR+TextPeckerTextDART
Time per optimizer step (s) ↓366.84324.24
Training GPU-hours, fixed 360 steps ↓1173.891037.58
GPU-hours to match baseline's final OCR score ↓1173.89691.83

Table R2. Training-side efficiency on 32 H20 GPUs; the final row compares costs at the same OCR quality. OCR+TextPecker is not equivalent to TextDART in training efficiency: it requires substantially more GPU-hours to reach the same OCR quality and incurs higher per-step training cost.

The revised manuscript cites GlyphPrinter, clarifies distinctions in Appendix A.1.5, and adds the efficiency comparison in Appendix A.3.5, with the new content highlighted in blue.

AuthorsOpenReview ↗

Response 3/5 — Glyph-Conditioned Comparisons and Complex Traditional Chinese (W2, Q1)

W2 & Q1: Glyph-Conditioned Comparisons and Complex Glyphs

We clarify SCANNER's conditioning-agnostic scope and provide additional experiments on glyph-image-conditioned models and complex Traditional Chinese glyphs.

1. Conditioning-Agnostic Scope and Evaluation Design

  • SCANNER operates during RL post-training and does not restrict the conditioning form of the base model. It can therefore optimize both prompt-conditioned and glyph-image-conditioned generators.
  • Direct cross-condition comparisons measure complete system performance, while same-backbone comparisons isolate the contribution of the RL algorithm. We report both below.
  • Section 5.3 already compares RL algorithms on FLUX and Qwen-Image. The new experiments extend this evaluation to EasyText and GlyphPrinter.
  • We use the official LingT2I text-rendering split [1], containing 3,000 examples across 10 languages, with paired text, scene descriptions, and glyph conditions. Gemini-4-Argon is used uniformly for text extraction and malformation evaluation.

2. Comparison with EasyText and GlyphPrinter

We evaluate the official EasyText and GlyphPrinter checkpoints and apply SCANNER to EasyText and GlyphPrinter Stage 1. In Tab. R3, SCANNER improves all six metrics on each backbone and outperforms R-GDPO on GlyphPrinter Stage 1, demonstrating benefits across both conditioning types and gains over the closest region-level preference baseline.

Base modelConditionRewardRL algorithmEn NED ↑En Accu ↑En Mal. ↓Zh NED ↑Zh Accu ↑Zh Mal. ↓
Qwen-Image-2512Prompt——0.5590.3270.1250.4740.3430.150
EasyText finalGlyph——0.9480.8000.0680.7940.6270.313
GlyphPrinter Stage 1Glyph——0.9420.8270.0840.9090.7930.152
GlyphPrinter finalGlyph—R-GDPO0.9440.8200.1400.9170.7730.222
Qwen-Image-2512PromptTextDARTSCANNER0.5890.3500.0620.4810.3600.083
EasyText finalGlyphTextDARTSCANNER0.9790.9170.0170.8930.8170.091
GlyphPrinter Stage 1GlyphTextDARTSCANNER0.9700.8900.0460.9410.8630.092

Table R3. English and Chinese results on LingT2I. Mal. denotes malformation rate. SCANNER offers meaningful advantages on glyph-image-conditioned models, improving EasyText and outperforming R-GDPO.

3. Evaluation on Complex Traditional Chinese Glyphs

We construct a 300-example generation benchmark from the public Traditional Chinese OCR test data [2]. It covers high-stroke, rare, long-text, multi-region, rotated, and vertical text, with 50 paired examples per category. Text, scenes, and layouts are matched across conditioning forms. We will release the benchmark upon acceptance.

As shown in Tab. R4, SCANNER improves both metrics in all six categories for all three backbones, directly quantifying its gains on complex glyphs.

MethodConditionHigh-strokeRareLong textMulti-regionRotatedVertical
Qwen-Image-2512Prompt0.283/0.0710.253/0.0990.935/0.0290.741/0.0760.288/0.0540.947/0.272
EasyTextGlyph0.342/0.3560.475/0.2380.560/0.6090.836/0.0510.252/0.6010.879/0.203
GlyphPrinter Stage 1Glyph0.075/0.2100.240/0.1170.157/0.1720.757/0.0160.130/0.2290.824/0.508
GlyphPrinter finalGlyph0.255/0.0830.438/0.0720.287/0.2610.643/0.0040.259/0.1220.949/0.463
Qwen-Image-2512 + SCANNERPrompt0.286/0.0310.270/0.0490.963/0.0040.857/0.0050.298/0.0370.953/0.018
EasyText + SCANNERGlyph0.350/0.1380.478/0.2170.720/0.4690.882/0.0000.350/0.3640.900/0.137
GlyphPrinter Stage 1 + SCANNERGlyph0.286/0.0450.498/0.0620.377/0.1390.938/0.0030.295/0.0920.951/0.046

Table R4. Complex Traditional Chinese evaluation. Each cell reports Character NED ↑ / malformation rate ↓. SCANNER’s benefits extend to complex Traditional Chinese glyphs, improving character accuracy and reducing glyph malformations across high-stroke, rare-character, and challenging layout settings.

4. Glyph-Specific Evidence and Factors Underlying Improvements

  • Qualitative cases and glyph metrics: Figures 4, 10, 15, and 16 in the original manuscript contain 17 cases with highlighted glyph errors, where SCANNER renders the corresponding text correctly. GenTextEval-Bench and PolyText-Bench also provide glyph-related quantitative metrics.
  • Target-consistent credit: The substantial long-text and multi-region gains are consistent with cross-rollout target alignment and explicit modeling of missing and extra text.
  • Glyph-sensitive evaluation: High-stroke and rare-character improvements, particularly the reduced malformation rates, are consistent with severity-weighted target rewards and localized credit assignment.

We have included these additional experimental results in Appendix A.3.11 and A.3.12 of the revised manuscript, with the new content highlighted in blue.

Reviewer_CTB7OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper takes a granular approach towards improving text within generated images. Current reward-based methods assigns a scalar to the entire image, leaving the image generator to dissect ambiguous reward signals. TextDART is a reward model that assigns reward per region of text, while SCANNER utilizes TextDART's local rewards to refine the generated samples. The novel contribution proves superior empirical results on the FLUX.1-dev and Qwen-Image-2512 for both English and Chinese text. The result shows consistent improvement over the baseline, Flow-GRPO, with two reward models.

Critical strengths
  • Natural and logical approach to solve the issue of spatially local credit assignment
  • The empirical gains are significant
  • The two separate contributions merge well together in terms of the interface and practicality
Critical weaknesses
  • If I understand correctly, TextDART is part of the method being evaluated while it is also used to construct the PolyText-Bench dataset, which is clearly concerning
  • It could be possible that the Qwen model is the primary driver of empirical performance gains which would dampened the contributions of TextDART.
Questions for authors
  • Could you address the concerns I have regarding the first bullet point in weakness?
  • How is the computational cost of your approach compared to existing approaches?
  • What happens when two text regions overlap or are extremely close together, and how does that impact the spatial credit assignment?
6 comments
AuthorsOpenReview ↗

Response 1/3 — PolyText-Bench Construction and Independent Evaluation (W1, Q1)

Thanks for your constructive feedback and for recognizing the natural design of our spatial credit assignment, the significant empirical gains, and the effective integration of TextDART and SCANNER. Below, we provide a point-by-point response to your concerns and questions.

W1 & Q1: TextDART's Role in PolyText-Bench and Evaluation Reliability

(1) Benchmark Construction Is Independent of TextDART

TextDART is not involved in constructing PolyText-Bench dataset. The prompts are constructed from structured scene specifications followed by LLM-based expansion. TextDART is not used for prompt filtering, target string generation, or scene category selection. Therefore, the benchmark prompts and targets are not shaped by TextDART.

(2) Isolating TextDART's Influence and Validating Our Method

Isolating individual component contributions on PolyText-Bench.

  • Reward model effectiveness under a fixed RL algorithm. In the original manuscript, Section 5.5 compares different reward models with Flow-GRPO fixed. TextDART achieves the best overall performance on PolyText-Bench, supporting the effectiveness of its defect-aware reward signals.

  • RL algorithm effectiveness under a fixed reward model. The original manuscript further compares different RL algorithms in Section 5.5 with TextDART fixed as the reward model. SCANNER achieves the best performance on PolyText-Bench, supporting the benefit of its regional credit assignment.

Validating the combined method under independent evaluation.

  • Combined effectiveness on four independent benchmarks. The original manuscript also shows in Section 5.3 that TextDART + SCANNER achieves state-of-the-art results on LongText-Bench, OneIG-Bench, CVTG-2k, and GenTextEval-Bench, each using its own evaluator. The user study in Appendix A.3.2 also shows the superiority of our method. These results establish the effectiveness of the combined method independently of PolyText-Bench.

  • Effectiveness under an independent OCR module. TextDART is not used to construct the benchmark prompts or targets. It is, however, used in the default evaluation pipeline. We recognize that using the same model for training rewards and evaluation may favor its own preferences, and therefore additionally evaluate all methods using an independent evaluator. The default PolyText-Bench evaluation uses TextDART for text localization and malformed-glyph-aware recognition, followed by predefined metric computation. TextDART localizes and extracts text, after which all metrics are computed using predefined rules. To isolate TextDART's remaining influence on evaluation, we replace it with closed-source Gemini-4-argon for text localization and recognition. As shown in Tables R1 and R2, TextDART + SCANNER consistently outperforms all baselines, demonstrating that the combined method remains effective under an independent OCR module.

Base modelRewardRL algorithmCS ↑CG ↑OG ↑
FLUX.1-dev--0.3780.4600.259
FLUX.1-devOCRv5Flow-GRPO0.6530.7110.304
FLUX.1-devTextPeckerFlow-GRPO0.7160.7710.373
FLUX.1-devTextDARTSCANNER0.7990.8490.563

Table R1. English evaluation on PolyText-Bench using Gemini-4-argon for text localization and recognition.

Base modelRewardRL algorithmEN / CS ↑EN / CG ↑EN / OG ↑ZH / CS ↑ZH / CG ↑ZH / OG ↑
Qwen-Image-2512--0.8000.8280.2730.8180.8430.208
Qwen-Image-2512OCRv5Flow-GRPO0.7910.8200.2780.8450.8560.204
Qwen-Image-2512TextPeckerFlow-GRPO0.7930.8250.2990.8450.8580.220
Qwen-Image-2512TextDARTSCANNER0.8690.8970.4820.8550.8620.416

Table R2. English and Chinese evaluation on PolyText-Bench using Gemini-4-argon for text localization and recognition.

(3) Broader Purpose of PolyText-Bench

  • Expanding the scope of visual text evaluation. Existing benchmarks focus on prompts with ground-truth-specified text and therefore cannot assess incidental text produced in open-ended scenes. PolyText-Bench addresses this gap with an open-source paired text-critical/text-optional protocol. TextDART supports this protocol as an evaluator that localizes and assesses generated text, including text without predefined target strings.

  • Providing accessible evaluation for future research. PolyText-Bench gives researchers direct access to a systematic evaluation protocol for open-ended visual text generation. We hope it will support the development and comparison of future algorithms and encourage broader exploration of this task.

We have included these additional experimental results in Appendix A.3.9 of the revised manuscript, with the new content highlighted in blue.

AuthorsOpenReview ↗

Response 2/3 — Gains Beyond the Qwen Backbones and Computational Cost (W2, Q2)

W2: Is the Qwen Backbone Driving the Gains?

This concern may refer to two distinct backbones: the Qwen3-VL-8B model used to initialize TextDART and the Qwen-Image generator used in RL experiments. We address both below.

(1) TextDART Provides Gains Beyond Its Qwen3-VL Backbone

  • Substantial gains over the base VLM. TextDART is fine-tuned from Qwen3-VL-8B. Appendix A.3.1 of the original manuscript shows that TextDART achieves substantial gains in text recognition and localization over this base model, demonstrating that its defect-aware capabilities do not come from the backbone alone.

  • Controlled reward-model comparison. To directly isolate the reward model's contribution, we replace TextDART with its base Qwen3-VL-8B while keeping FLUX.1-dev and SCANNER fixed. As shown in Table R3, TextDART outperforms its base model across all benchmarks and metrics, confirming that its specialized training provides gains beyond the Qwen3-VL backbone.

Base modelRewardRL algorithmLongText-Bench / TextScore ↑OneIG-Bench / TextScore ↑OneIG-Bench / ED ↓OneIG-Bench / CR ↑OneIG-Bench / WAC ↑CVTG-2k / WordAcc ↑CVTG-2k / NED ↑GenTextEval-Bench / Qua ↑GenTextEval-Bench / Sem ↑PolyText-Bench / CS ↑PolyText-Bench / CG ↑PolyText-Bench / OG ↑
FLUX.1-dev——0.6200.54385.310.1380.3790.5030.6940.6610.3520.4260.4640.281
FLUX.1-devBase Qwen3-VL-8BSCANNER0.7720.77165.500.2390.5410.7490.8700.8370.5610.7830.8080.374
FLUX.1-devTextDARTSCANNER0.8270.85851.060.3090.5980.8340.9140.8390.6130.8820.9050.653

Table R3. Reward-model comparison with FLUX.1-dev and SCANNER fixed. Base Qwen3-VL-8B denotes the backbone before TextDART's specialized fine-tuning.

(2) Gains Are Not Dependent on the Qwen-Image Generator

  • Controlled comparison on Qwen-Image. On Qwen-Image-2512, all compared methods use the same generator in Section 5.3. TextDART + SCANNER still achieves the best overall performance across rewards and RL algorithms. This within-backbone comparison rules out generator differences as an explanation for the relative gains.

  • Consistent gains on a non-Qwen generator. Section 5.3 in the original manuscript also shows consistent improvements on FLUX.1-dev. These results demonstrate that our method's effectiveness does not depend on Qwen-Image's native text rendering ability.

We have included these additional experimental results in Appendix A.3.10 of the revised manuscript, with the new content highlighted in blue.

Q2: Computational Cost

We report both the actual training cost and the cost of reaching the same text quality. All runs use 32 H20 GPUs at 1024×1024 resolution with LoRA rank 64. Training-worker GPU-hours include reward waiting time but exclude periodic evaluation. As shown in Table R4, SCANNER adds only 0.7%–2.8% to fixed-budget training cost, while reducing training-worker GPU-hours by 49.5%–74.2% at matched text quality. Under a conservative scenario that assigns the entire reward-service pool to each run, the quality-matched reductions remain 16.0%–58.8%.

BackboneCompared baselineTime / round (s), baseline → SCANNERIncreaseFull training GPU-hours, baseline → SCANNERIncreaseSCANNER GPU-hours to match baselineReduction
FLUX.1-devOCR + Flow-GRPO1931.12 → 1945.47+0.74%1029.93 → 1037.58+0.74%432.4758.0%
FLUX.1-devTextPecker + Flow-GRPO1927.12 → 1945.47+0.95%1027.80 → 1037.58+0.95%518.9749.5%
Qwen-Image-2512OCR + Flow-GRPO1196.32 → 1230.13+2.83%1701.43 → 1749.52+2.83%656.6761.4%
Qwen-Image-2512TextPecker + Flow-GRPO1197.90 → 1230.13+2.69%1703.68 → 1749.52+2.69%438.8374.2%

Table R4. Training cost and quality-matched efficiency of TextDART + SCANNER. Full training uses 360 optimizer steps for FLUX and 320 for Qwen. The last two columns use the first measured checkpoint that matches each baseline's final score under the same evaluator.

We have added a detailed analysis of this point in Appendix A.4 of the revised manuscript, with the new content highlighted in blue.

AuthorsOpenReview ↗

Response 3/3 — Spatial Credit Assignment for Overlapping and Nearby Text Regions (Q3)

Q3: Overlapping or Closely Located Text Regions

TextDART computes rewards independently for each text target. Once text targets have been detected and matched, overlap and proximity are handled during projection of target advantages onto the latent grid.

(1) Overlapping Text Regions

When multiple boxes cover the same latent position, Ours assigns it to the smallest covering region. The position therefore receives one target's advantage, avoiding direct signal mixing.

(2) Closely Located but Non-Overlapping Regions

Proximity alone does not merge targets. However, finite latent resolution can remove narrow gaps, causing one latent unit to cover multiple regions. Ours uses center-point assignment to preserve a single advantage per unit. This avoids signal mixing but can lose boundary information.

Fig. 8(b) of the original manuscript evaluates area-overlap projection. It preserves fractional area coverage but can mix target signals within shared units.

(3) Empirical Comparison and Analysis

We independently train three variants with FLUX.1-dev, TextDART, and SCANNER. Ours uses center to construct hard regional masks on the latent grid and smallest to assign each shared position to the smallest covering region. The average variant averages the advantages of all covering regions at shared positions. The pixel_area variant replaces hard latent-grid mapping with fractional pixel-area coverage within each latent unit.

As shown in Table R5, Ours outperforms both variants on every reported metric.

Base modelRewardRL algorithmRegion mappingOverlap handlingLongText-Bench / TextScore ↑OneIG-Bench / TextScore ↑OneIG-Bench / ED ↓OneIG-Bench / CR ↑OneIG-Bench / WAC ↑CVTG-2k / WordAcc ↑CVTG-2k / NED ↑GenTextEval-Bench / Qua ↑GenTextEval-Bench / Sem ↑PolyText-Bench / CS ↑PolyText-Bench / CG ↑PolyText-Bench / OG ↑
FLUX.1-devTextDARTSCANNERcenteraverage0.7920.77763.470.2470.5340.7700.8770.8320.5530.7880.8180.412
FLUX.1-devTextDARTSCANNERpixel_areasmallest0.8040.79460.930.2440.5530.7960.8910.8380.5770.8030.8380.528
FLUX.1-devTextDARTSCANNERcentersmallest0.8270.85851.060.3090.5980.8340.9140.8390.6130.8820.9050.653

Table R5. Results for different region mapping and overlap handling designs with FLUX.1-dev, TextDART, and SCANNER fixed. The last row is Ours.

Why smallest-region ownership outperforms advantage averaging. Averaging advantages at shared positions can mix conflicting signals from targets with different rendering quality. Assigning each position to the smallest covering region preserves an unambiguous, target-specific advantage. The consistently lower performance of the average variant supports this design.

Why center outperforms pixel_area. Center mapping preserves one target advantage per latent unit. pixel_area retains fractional boundary coverage but can diffuse credit across targets within shared units. Its lower scores across all metrics indicate that this spatial mixing outweighs its boundary benefit in our setting.

We have included these additional experimental results in Appendix A.3.3 of the revised manuscript, with the new content highlighted in blue.


We hope that our responses have adequately addressed your concerns. We sincerely appreciate your thoughtful comments and suggestions. If you have any further concerns, please feel free to let us know.

Reviewer_bB64OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses RL post-training for visual text generation, identifying two major limitations: recognition-based rewards overlook glyph defects, and image-level scalar rewards obscure the uneven quality of individual text regions. It proposes two complementary components. TextDART is a defect-aware, region-targeted reward model that localizes each text region, transcribes it with explicit malformed-glyph markers, and scores quality per target. SCANNER is a spatial-aware RL framework that normalizes rewards per text target across sampled images, projects the resulting advantages onto different patch tokens. This paper presents strong experimental results: TextDART outperforms other OCR baselines, especially on defect perception and localization. TextDART + SCANNER achieves state-of-the-art text accuracy and glyph fidelity across multiple benchmarks.

Critical strengths
  • The proposed methods are well-grounded. The paper identifies two common limitations of existing RL frameworks for visual text generation: recognition-based rewards overlook glyph defects, and image-level advantages average over regions of different quality. To address the former, the paper proposes TextDART, a defect-aware, region-targeted reward model that localizes each text region, detects malformed glyphs, and measures both semantic correctness and glyph quality. To address the latter, the paper designs SCANNER, an RL framework that normalizes rewards per text target and projects the resulting regional advantages onto the corresponding latent tokens. The motivation is clear, and the solution is natural and makes sense to me.

  • The proposed modules perform strongly. TextDART outperforms both baseline OCR systems and modern VLMs in text recognition, with especially large margins in defect perception and localization. TextDART + SCANNER achieves the best visual text generation quality across all benchmarks, on both backbones in different languages.

  • The experimental results are comprehensive, answering most of my questions while reading the paper. The ablation studies isolate each design choice of SCANNER, and the additional results in the Appendix further demonstrate that SCANNER generalizes to other reward models that may not be as powerful as TextDART. The qualitative results also show clear improvements over other visual text generation baselines.

Critical weaknesses
  • All reported metrics focus on evaluating text rendering quality. While this is the main focus of the paper, I believe additional metrics assessing overall image quality (e.g., aesthetic scores or other established benchmarks) should be included. Since the ultimate goal is image generation, it is important to ensure that improvements in text rendering quality do not come at the expense of overall image quality.

  • The related work discusses several region-level advantages (e.g., SpatialFlow-GRPO, Edit-GRPO), yet experiments compare only against Flow-GRPO with different rewards. Comparing with these similar region-level frameworks can better benefit the submission.

Questions for authors

I have no further questions, as they have already been addressed in the main paper and appendix.

2 comments

Thanks for your constructive feedback and for recognizing the clear motivation, well-grounded design, strong empirical performance, and comprehensive evaluation of our work. Below, we provide a point-by-point response to your concerns.

W1: Overall Image Quality beyond Text Rendering

We have conducted additional automatic and human evaluations of image aesthetics and image-prompt consistency.

(1) Automatic Evaluation on DrawBench

We conduct automatic evaluation on DrawBench [1], a comprehensive text-to-image benchmark covering diverse prompt categories. We use PickScore to assess image aesthetics and CLIP cosine similarity to measure image-prompt consistency.

As shown in Table R1, TextDART with SCANNER matches the best PickScore and CLIP cosine similarity. These results indicate that our improvements in text rendering do not degrade overall image aesthetics or image-prompt consistency.

BasemodelRewardRLalgorithmAesthetics/PickScore↑Consistency/CLIPcosine↑
Qwen-Image-2512--22.980.278
Qwen-Image-2512OCRv5Flow-GRPO22.980.279
Qwen-Image-2512TextPeckerFlow-GRPO23.000.276
Qwen-Image-2512TextDARTSCANNER23.000.280

Table R1. Automatic evaluation on DrawBench. Image aesthetics and image-prompt consistency are measured by PickScore and CLIP cosine similarity.

(2) Human Evaluation of Overall Image Quality

To complement the automatic metrics, we conduct a human evaluation following the same protocol as the human evaluation in our appendix. Evaluators assess the generated images in terms of overall image aesthetics and image-prompt consistency.

As shown in Table R2, TextDART with SCANNER receives the highest human preference on both criteria. These results provide direct human evidence that our method improves overall image quality and prompt alignment alongside text rendering.

BasemodelRewardRLalgorithmImageAesthetics↑Image-PromptConsistency↑
Qwen-Image-2512--21.0%14.3%
Qwen-Image-2512OCRv5Flow-GRPO17.6%16.9%
Qwen-Image-2512TextPeckerFlow-GRPO28.1%24.6%
Qwen-Image-2512TextDARTSCANNER33.3%44.2%

Table R2. Human preference shares for image aesthetics and image-prompt consistency using Qwen-Image-2512.

Overall, TextDART with SCANNER improves text rendering while achieving the best image aesthetics and image-prompt consistency among all compared methods.

We have included these additional experimental results in Appendix A.3.7 of the revised manuscript, with the new content highlighted in blue.

W2: Comparison with Existing Region-Level RL Frameworks

These region-level RL frameworks are relevant baselines, but both are designed for image editing and are not directly compatible with visual text generation. SpatialFlow-GRPO uses editing-specific rewards and semantic regions, and has no public implementation. Edit-GRPO assigns separate rewards to edit and non-edit regions for locality preservation. For a controlled comparison, we fine-tune FLUX using TextDART as the shared reward and adapt the core RL design of each method. We reproduce SpatialFlow-GRPO-style training following its paper. For Edit-GRPO-style training, we compute advantages for target text regions using Critic-TextDART and for non-target text regions using Free-TextDART.

The controlled comparison in Table R3 demonstrates the effectiveness of SCANNER's advantage design over the two region-level RL frameworks.

RLalgorithmLongText-Bench/TextScore↑OneIG-Bench/TextScore↑OneIG-Bench/ED↓OneIG-Bench/CR↑OneIG-Bench/WAC↑CVTG-2k/WordAcc↑CVTG-2k/NED↑GenTextEval-Bench/Qua↑GenTextEval-Bench/Sem↑PolyText-Bench/CS↑PolyText-Bench/CG↑PolyText-Bench/OG↑
Edit-GRPO-style0.6470.74761.260.2010.4840.5640.6980.7500.4180.4960.5350.593
SpatialFlow-GRPO-style0.8220.84054.080.2790.5890.8090.9000.8390.6120.8650.8830.529
SCANNER0.8270.85851.060.3090.5980.8340.9140.8390.6130.8820.9050.653

Table R3. SCANNER versus adapted SpatialFlow-GRPO and Edit-GRPO on five visual text generation benchmarks.

We have included these additional experimental results in Appendix A.3.8 of the revised manuscript, with the new content highlighted in blue.

Reference: [1] Saharia, Chitwan, et al. “Photorealistic text-to-image diffusion models with deep language understanding.”


We hope that our responses have adequately addressed your concerns. We sincerely appreciate your thoughtful comments and suggestions. If you have any further concerns, please feel free to let us know.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal and Revised Manuscript

Dear Reviewer bB64,

Thank you again for your constructive feedback. As the discussion period approaches its end, we would like to briefly follow up on our responses and revisions.

We have made substantial updates to address your concerns, including automatic and human evaluations of overall image quality (PickScore and CLIP consistency on DrawBench, plus a human preference study, showing that our text-rendering gains do not come at the expense of image aesthetics or prompt consistency; Appendix A.3.7), and a controlled comparison with region-level RL frameworks (adapting SpatialFlow-GRPO-style and Edit-GRPO-style training with TextDART as the shared reward, where SCANNER achieves stronger performance across five benchmarks; Appendix A.3.8). These additions further strengthen the empirical evaluation of TextDART and SCANNER.

We hope our responses and the revised manuscript have addressed your main concerns. We would greatly appreciate it if you could revisit the new evidence and reconsider your assessment accordingly. We would also be happy to clarify any remaining questions.

Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.