TextPecker: Rewarding Structural Anomaly Quantification for Enhancing Visual Text Rendering

Hanshen Zhu; Yuliang Liu; Xuecheng Wu; An-Lan Wang; Hao Feng; Dingkang Yang; Chao Feng; Can Huang

cs.CVFeb 24, 2026

TextPecker: Rewarding Structural Anomaly Quantification for Enhancing Visual Text Rendering

Authors: Hanshen Zhu, Yuliang Liu, Xuecheng Wu, An-Lan Wang, Hao Feng, Dingkang Yang, Chao Feng, Can Huang, +2 more

Organizations: Huazhong University of Science and Technology · ByteDance

Abstract

Visual Text Rendering (VTR) remains a critical challenge in text-to-image generation, where even advanced models frequently produce text with structural anomalies such as distortion, blurriness, and misalignment. However, we find that leading MLLMs and specialist OCR models largely fail to perceive these structural anomalies, creating a critical bottleneck for both VTR evaluation and RL-based optimization. As a result, even state-of-the-art generators (e.g., Seedream4.0, Qwen-Image) still struggle to render structurally faithful text. To address this, we propose TextPecker, a plug-and-play structural anomaly perceptive RL strategy that mitigates noisy reward signals and works with any textto-image generator. To enable this capability, we construct a recognition dataset with character-level structural-anomaly annotations and develop a stroke-editing synthesis engine to expand structural-error coverage. Experiments show that TextPecker consistently improves diverse text-to-image models; even on the well-optimized Qwen-Image, it significantly yields average gains of 4% in structural fidelity and 8.7% in semantic alignment for Chinese text rendering, establishing a new state-of-the-art in high-fidelity VTR. Our work fills a gap in VTR optimization, providing a foundational step towards reliable and structural faithful visual text generation.

Figures & tables

Figure 1: Existing OCR models and MLLMs struggle to perceive fine-grained structural anomalies in rendered text images, creating a key bottleneck for both VTR evaluation and RL-based optimization. Misrecognized characters are highlighted in RED .

Figure 2: Schematic illustration of the TextPecker framework . Given a generative prompt, we first sample $G$ candidate outputs $\{o_{i}\}_{i=1}^{G}$ from the reference policy model $\pi_{\theta_{\text{ref}}}$ . Each $o_{i}$ is sent to a structure-aware recognizer to extract fine-grained generated text, with markers indicating structurally anomalous text. We then compute the joint reward $\mathcal{R}_{i}$ , comprising a weighted sum of semantic alignment and structural quality scores (Sec. 3.2.1 ). Each $\mathcal{R}_{i}$ is normalized to a group relative advantage $A_{i}$ . Finally, we optimize the current policy model $\pi_{\theta}$ by maximizing $A_{i}$ while enforcing proximity to $\pi_{\text{ref}}$ via KL divergence.

Figure 3: The illustration of proposed data construction pipeline.

Data Type	Level	Samples	Proportion
Manual Annotations	Box	559.6K	39.32%
Manual Annotations	Image	131.1K	9.21%
Synthetic Anomaly Text	Box	452.5K	31.80%
Synthetic Anomaly Text	Image	100.0K	7.03%
Synthetic Normal Text	Box	150.0K	10.54%
Synthetic Normal Text	Image	30.0K	2.10%

Table 1: Statistics of our constructed text-rich image recognition dataset with structural-anomaly labels at box and image levels. Proportions are computed over all instances.

Methods	TSAP						CTR
	Image-level			Box-level			Image-level		Box-level
	P	R	F1	P	R	F1	R	NED	R	NED
English recognition
PP-OCRv5 [ 15 ]	0.000	0.000	0.000	-	-	-	0.720	0.137	-	-
GOT-OCR-2.0 [ 53 ]	0.000	0.000	0.000	-	-	-	0.610	0.186	-	-
MonkeyOCR [ 31 ]	0.000	0.000	0.000	-	-	-	0.578	0.209	-	-

Table 2: Results of Text Structural Anomaly Perception (TSAP) and Canonical Text Recognition (CTR): measuring model’s recognition ability under generated text images and TSAP-demanding prompts. Box-level results of non-supporting models are marked as ”-”.

Models	Rewards	OneIG			LongText			CVTG-2K			GenTextEval
Models	Rewards	Avg.	Qua.	Sem.	Avg.	Qua.	Sem.	Avg.	Qua.	Sem.	Qua.	Sem.
English Rendering
SD3.5-M [ 16 ]	-	0.441	0.848	0.513	0.296	0.860	0.416	0.368	0.869	0.491	0.671	0.265
	OCR	0.572	0.941	0.627	0.295	0.944	0.498	0.513	0.943	0.671	0.940	0.462
	TextPecker	0.581	0.957	0.636	0.344	0.957	0.503	0.596	0.944	0.593	0.959	0.506
Flux.1[dev] [ 28 ]	-	0.567	0.875	0.585	0.613	0.929	0.591	0.491	0.908	0.523	0.672	0.336

Table 3: Quantitative comparisons of different RL-optimized generative models on OneIG-Bench [ 6 ] , LongText-Bench [ 20 ] , CVTG-2k [ 14 ] , and GenTextEval-Bench. Avg. : Average text score from original benchmarks; Qua. : structural Quality score. Sem. : Semantic alignment. Score measurement and reward computation are both conducted by TextPecker (Qwen3-VL).

Figure 4: Qualitative Comparisons of Text Rendering for Qwen-Image and RL-Optimized Variants. Readers are highly recommended to refer to the appendix for extensive comparisons across generative models and RL baselines.

Models	Settings		English					Chinese
	Anno.	Syn.	TSAP			CTR		TSAP			CTR
	Anno.	Syn.	P	R	F1	R	NED	P	R	F1	R	NED
InternVL3 [ 69 ]			0.206	0.165	0.183	0.759	0.102	0.190	0.128	0.153	0.927	0.069
	✓		0.795	0.970	0.874	0.938	0.042	0.583	0.966	0.727	0.849	0.148
	✓	✓	0.795	0.960	0.870	0.944	0.035	0.889	0.968	0.927	0.962	0.037
Qwen3-VL [ 60 ]			0.182	0.018	0.032	0.810	0.076	0.333	0.008	0.017	0.947	0.049

Table 4: Ablation study on effectiveness of the constructed data: Image-level recognition results across different baseline models.

Generative Model	OCR Model	Settings			GenTextEval-EN
Generative Model	OCR Model	NED	PM	SQ	Qua.	Sem.
Flux.1[dev] [ 28 ]	-				0.672	0.336
	PP-OCRv5 [ 15 ]	✓			0.976	0.602
	PP-OCRv5 [ 15 ]	✓	✓		0.976	0.644
	TextPecker	✓	✓		0.984	0.702
	TextPecker	✓	✓	✓	0.988	0.719

Table 5: Ablation study on the effectiveness of reward design: PM for Pair-wise Matching, and SQ for Structural Quality reward.

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Qualitative Comparisons of Text Rendering for Flux.1[dev] [ 28 ] and RL-Optimized Variants.

Generative Model	OCR Model	Settings			GenTextEval-EN
Generative Model	OCR Model	NED	PM	SQ	Qua.	Sem.
SD3.5-M [ 16 ]	-				0.671	0.265
	PP-OCRv5 [ 15 ]	✓			0.907	0.470
	PP-OCRv5 [ 15 ]	✓	✓		0.910	0.482
	TextPecker	✓	✓		0.956	0.498
	TextPecker	✓	✓	✓	0.959	0.506

Appendix

Table 6: Ablation study on the effectiveness of reward design: PM for Pair-wise Matching, and SQ for Structural Quality reward, measured by TextPecker (InternVL3).

Figure 6: Hard Cases of TextPecker on Gemini-Rendered [ 10 ] Visual Text with Extreme Stylization and Low Contrast Layout.

Tab. A & Methods	Normal		Extreme Stylization		Low Contrast
Tab. A & Methods	TSAP-F1	CTR-R	TSAP-F1	CTR-R	TSAP-F1	CTR-R
InternVL-3-8B [ 69 ]	0.000	0.666	0.087	0.588	0.364	0.742
TextPecker-8B	0.752	0.833	0.571	0.577	0.800	0.839

Appendix

Table 7: Robustness evaluation of TextPecker on Gemini-2.5-flash-image [ 10 ] : Performance under normal condition, extreme stylization, and low-contrast layouts.

method	rewards	weights	OneIG		LongText		GenTextEval
method	rewards	weights	Qua.	Sem.	Qua.	Sem.	Qua.	Sem.
Qwen-Image [ 55 ]	–	–	0.888	0.747	0.900	0.815	0.922	0.805
	OPA	7:1:2	0.898	0.788	0.912	0.845	0.944	0.859
	SQPA	5:2:1:2	0.943	0.828	0.941	0.889	0.970	0.893

Appendix

Table 8: Additional Quantitative Comparisons of RL-Optimized Generative Models on Chinese Visual Text Benchxmarks (OneIG [ 6 ] , LongText [ 20 ] , GenTextEval) with multi reward setting. O: OCR Reward [ 34 ] , S: TextPecker Semantic Reward, Q: TextPecker Structural Quality reward, P: Pickscore Reward [ 25 ] , A: Aethetic Reward [ 45 ] . Results measurement and reward computation are both conducted by TextPecker (InternVL-3).

method	rewards	weights	OneIG		LongText		CVTG		GenTextEval		TIIF		TextAtlas		LeX
method	rewards	weights	Qua.	Sem.	Qua.	Sem.	Qua.	Sem.	Qua.	Sem.	Qua.	Sem.	Qua.	Sem.	Qua.	Sem.
SD3.5-M [ 16 ]	–	–	0.840	0.507	0.836	0.407	0.843	0.466	0.666	0.262	0.758	0.347	0.646	0.269	0.810	0.454
	OPA	7:1:2	0.908	0.588	0.913	0.508	0.895	0.621	0.896	0.461	0.886	0.483	0.916	0.436	0.894	0.563
	SQPA	5:2:1:2	0.940	0.607	0.959	0.534	0.926	0.587	0.954	0.519	0.941	0.506	0.954	0.462	0.940	0.591
Flux.1[dev] [ 28 ]	–	–	0.870	0.578	0.925	0.584	0.889	0.510	0.664	0.332	0.933	0.540	0.683	0.307	0.946	0.667
	OPA	7:1:2	0.977	0.739	0.977	0.763	0.974	0.780	0.982	0.739	0.986	0.719	0.983	0.640	0.988	0.741

Appendix

Table 9: Additional Quantitative Comparisons of RL-Optimized Generative Models on English VTR Benchmarks (OneIG [ 6 ] , LongText [ 20 ] , CVTG [ 14 ] , GenTextEval, TIIF [ 54 ] , TextAtlas [ 49 ] , LeX [ 66 ] ) with multi reward setting. O: OCR Reward [ 34 ] , S: TextPecker Semantic Reward, Q: TextPecker Structural Quality reward, P: Pickscore Reward [ 25 ] , A: Aethetic Reward [ 45 ] . Results measurement and reward computation are both conducted by TextPecker (InternVL-3).

Figure 7: Qualitative comparisons of text rendering results ( English ) among different RL baseline settings. RL-TextPecker denotes the RL setting in the main paper, and RL-SQPA refers to our enhanced RL setting as described in Sec. D .

Figure 8: Qualitative comparisons of text rendering results ( Chinese ) among different RL baseline settings. RL-TextPecker denotes the RL setting in the main paper, and RL-SQPA refers to our enhanced RL setting as described in Sec. D .

Model	Level	Samples	Proportion
AnyText [ 48 ]	Box-level	30647	7.38%
AnyText [ 48 ]	Image-level	7405	1.78%
Flux.1[dev] [ 28 ]	Box-level	91705	22.07%
Flux.1[dev] [ 28 ]	Image-level	19181	4.62%
Qwen-Image [ 55 ]	Box-level	20308	4.89%
Qwen-Image [ 55 ]	Image-level	2647	0.64%

Appendix

Table 10: Statistics of English text-rich images generated by different models, labeled at box and image levels. Proportions are computed over all instances.

Model	Level	Samples	Proportion
CogView4 [ 13 ]	Box-level	36000	11.14%
CogView4 [ 13 ]	Image-level	17005	5.26%
Kolors [ 27 ]	Box-level	35549	10.99%
Kolors [ 27 ]	Image-level	19312	5.98%
Qwen-Image [ 55 ]	Box-level	26597	8.23%
Qwen-Image [ 55 ]	Image-level	6225	1.93%

Appendix

Table 11: Statistics of Chinese text-rich images generated by different models, labeled at box and image levels. Proportions are computed over all instances.

Data Type	Level	Samples	Proportion
Manual Annotations	Box	444	41.85%
Manual Annotations	Image	417	39.29%
Synthetic Anomaly Text	Box	50	4.71%
Synthetic Anomaly Text	Image	50	4.71%
Synthetic Normal Text	Box	50	4.71%
Synthetic Normal Text	Image	50	4.71%

Appendix

Table 12: Statistics of our constructed text-rich image structural perception test dataset with structural-anomaly labels at box and image levels. Proportions are computed over all instances.

Methods	Chinese				English
	Image		Box		Image		Box
	TSAP-F1	CTR-R	TSAP-F1	CTR-R	TSAP-F1	CTR-R	TSAP-F1	CTR-R
InternVL3-8B (Baseline)	0.106	0.955	0.244	0.791	0.183	0.759	0.304	0.570
TextPecker-8B (Anno)	0.866	0.849	0.906	0.815	0.874	0.938	0.809	0.918
TextPecker-8B (Anno + Syn)	0.901	0.917	0.955	0.995	0.850	0.931	0.840	0.944

Appendix

Table 13: Performance of TextPecker on Real-only Test Splits

Figure 9: Statistics of the RL prompt set for RL-based VTR optimization (English: word-level; Chinese: character-level).

Figure 10: Comparison among CVTG-2K [ 14 ] , OneIG-Bench [ 6 ] , LongText-Bench [ 20 ] , and Our Proposed GenTextEval-Bench with Respect to the Length of Rendered Texts in English (Left) and Chinese (Right).

Figure 11: The prompting template used for our TextPecker.

Parameter Category	Canonical Chinese Text	Canonical English Text	Structural Anomaly Construction
Basic Text Configuration
Vertical text probability	10%	10%	10%
Number of elements per sample	3–10	3–10	3–10
Text length range	1–25 characters	3–25 characters	1–25 characters
Font Settings
Number of font types	976	976	976

Appendix

Table 14: Key parameters for canonical text rendering and structural anomaly construction.

Figure 12: This figure shows manually annotated structural anomalies (box-level and image-level) from various generative models, alongside synthetic structural anomalies generated by our rendering engine for data augmentation.

Figure 13: This figure presents evaluation samples of TextPecker, showcasing its performance in detecting structural anomalies across diverse text rendering scenarios. $\omega=1$ for structural quality score visualization.

Explore similar work

TextAlign: Preference Alignment for Text Rendering with Hierarchical Rewards
May 19, 2026Mingxuan Cui, Jingpu Yang, Fengxian Ji +7Diffusion Model Alignment Reward Modeling
Mohamed bin Zayed University of Artificial Intelligence · Chinese Academy of Sciences Institute of Automation · Northeastern University +1
TextVerifier: Seeing Beyond Recognition to Verify Chinese Text Rendering
Oct 14, 2026Yiyang Chen, Guo-Hua Wang, Yaoli Liu +4T2I Generation Evaluation T2I Diffusion Models
State Key Laboratory of General Artificial Intelligence, Peking University, Beijing, China · Alibaba Group · State Key Lab of CAD&CG, Zhejiang University
Residual Decoder Adapter: ID-Preserving Tokenizer Adaption for Autoregressive Text Rendering
Jun 1, 2026Dongxing Mao, Jinpeng Wang, Jiahao Tang +6Image Tokenization Visual Tokenization
Central South University · University of Oxford · Microsoft Research

May 19, 2026J/K move · Enter open · S save

TextAlign: Preference Alignment for Text Rendering with Hierarchical Rewards

Mingxuan Cui, Jingpu Yang, Fengxian Ji, Qian Jiang+6

Mohamed bin Zayed University of Artificial Intelligence · Chinese Academy of Sciences Institute of Automation · Northeastern University+1

Diffusion Model Alignment Reward Modeling T2I Generation T2I Diffusion Models

Open paper AskPDF arXiv

Faithful text rendering remains a persistent weakness of large text-to-image generative models, as it requires both semantic instruction following and fine-grained glyph-level structure. Prior methods often improve this ability through architecture-specific modules or encoder modifications, which complicate deployment across foundation models. We study text rendering as a post-training preference-alignment problem and propose TextAlign, a non-invasive framework that keeps the generator architecture unchanged. The key component is a hierarchical vision-language model (VLM)-based reward that decomposes rendering errors into global, word, and glyph levels, then converts binary defect judgments into a scalar preference signal. The resulting signal supports both Group Relative Policy Optimization (GRPO) and Direct Preference Optimization (DPO). Experiments on FLUX.1-dev and Z-Image-Turbo show consistent gains in OCR-based text accuracy without degrading general generation quality. Compared with strong foundation and text-rendering baselines, including SD3.5, Qwen-Image, AnyText, and TextDiffuser, these results indicate that reward design offers a scalable alternative to model redesign for improving text rendering.

Abstract

Explore similar work

TextAlign: Preference Alignment for Text Rendering with Hierarchical Rewards

TextVerifier: Seeing Beyond Recognition to Verify Chinese Text Rendering

Residual Decoder Adapter: ID-Preserving Tokenizer Adaption for Autoregressive Text Rendering