Organizations: State Key Laboratory of General Artificial Intelligence, Peking University, Beijing, China · Alibaba Group · State Key Lab of CAD&CG, Zhejiang University
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.
Figures & tables
Figure 1: Text rendering verification and its use as a reward. (A) For the character “基” in a generated image, OCR and VLM both miss the error, whereas TextVerifier correctly flags it as an error. (B) Using the per-character score as a post-training reward markedly reduces glyph errors.
Figure 2: Structure-aware dual tower and three-stage training. (a) The structure-aware CLIP-style dual encoder model architecture. (b) The three-stage training curriculum. (c) Inference procedure, and it can also serve as a reward for Flow-GRPO post-training.
Figure 3: Three-stage data construction. Left: real images pass through a data pipeline to yield stage1 data. Middle: standard glyphs are perturbed by two stroke edits into pairs. Right: early and late checkpoints of a text-region inpainting model repaint the same target character.
Figure 4: Per-character annotation examples from TextRenderBench-ZH . Images are generated by text-to-image models; each character carries a (position, identity, correctness) triple (green box: correct, red box: error), with error-versus-standard glyph pairs enlarged below.
Method
TP↑
FP↓
FN↓
TN↑
P↑
R↑
F1↑
Params
Inf.(ms)↓
PP-OCRv6 (thr =0.90 ) ( Zhang et al., 2026a )
223
140
419
6576
0.614
0.347
0.444
0.06B
17.2
PP-OCRv6 (best thr) ( Zhang et al., 2026a )
400
479
242
6237
0.455
0.623
0.526
Qwen3.8-Max ( Qwen Team, 2026 )
78
66
564
6650
0.542
0.121
0.198
—
—
GPT-5.6-sol ( OpenAI, 2026b )
37
57
605
6659
0.394
0.058
0.101
—
—
Gemini 3.7 Flash ( Google DeepMind, 2026b )
347
1122
295
5594
0.236
0.540
0.329
—
—
TextPecker (Qwen3-VL-8B) ( Zhu et al., 2026 )
491
531
151
6185
0.480
0.765
0.590
8.29B
320.4
Table 1: Main results on TextRenderBench-ZH . TextVerifier ’s three-stage F1 of 0.844 is far above all baselines; the “best thr” row tunes the PP-OCRv6 confidence ( Zhang et al., 2026a ) threshold on the test set as an upper reference only. Params is the scoring-model size; Inf. is per-character median latency (ms) on the full test set.
Table 6
Figure 5: Qualitative comparison of Flow-GRPO post-training under different rewards. Given the target text, we compare Qwen-Image ( Wu et al., 2025 ) , the NED-only ablation, TextPecker ( Zhu et al., 2026 ) , and ours; Ours yields fewer glyph errors while staying aesthetically closer to Base.
Figure 6: The positive and negative samples for stage‑2 synthesized using synthetic font files. The first column in each row corresponds to the positive sample, which represents the correct standard rendering of the font. The remaining columns are negative samples containing perturbations such as stroke addition and stroke removal.
Figure 7: The stage‑2 positive and negative samples synthesized using Chinese kaiti‑style vector data included in Makemeahanzi ( Kishore, 2016 ) . The first column of each row is the positive sample, representing the correct standard rendering of the font, and the rest are negative samples with perturbations such as stroke addition and stroke removal.
Figure 8: Architecture of the text-region inpainting model for Stage-3 relative pair construction. Three token groups, namely the full image, multiple target crops and character embeddings from Qwen3-0.6B ( Yang et al., 2025 ) , are concatenated along the sequence dimension. Custom attention mask and modified 3D-RoPE are adopted to enable multi‑crop training with absolute positional awareness.
Figure 9: The Stage‑3 positive and negative samples generated from different checkpoints of the inpainting model. In each pair, the first one is the positive sample produced by the later checkpoint, and the second one is the negative sample. For each paired positive‑negative sample, the positive sample exhibits better quality than its negative counterpart.
Figure 10: The IDS representation of Chinese characters. Left: basic decompositions, where components shared across characters are color-coded (日, 女, 木), so structurally related characters overlap at the symbol level. Middle: the twelve IDS structural operators. Right: IDS is recursive—a character expands into a tree of operators and components (想 = 相心 = 木目心)—and identical subtrees are reused across characters (相 = 木目 recurs in 想, 霜, 箱). Structural operators are shown only inside the figure.
Figure 11: Text-tower geometry versus symbolic structure (Stage-3 checkpoints). Left: text-embedding cosine for 26,136 fixed-seed character pairs grouped by normalized IDS edit distance (blue: IDS tower; red: single-character tower; box = IQR, line = median). The IDS tower’s cosine falls monotonically as structural distance grows (medians 0.65→0.05 ; Spearman ρ=−0.70 ), while the single-character tower is nearly flat ( 0.46→0.23 ; ρ=−0.34 ). Right: the same pairs stratified by shared top-level IDS operator: same vs. different medians 0.25 vs. 0.12 (IDS) and 0.28 vs. 0.24 (single-character). Pairs form a distance-stratified diagnostic sample; distances are flat token edit distances on IDS prefixes.
K
δ
P
R
F1
TP/FP/FN/TN
1
0.0408
0.775
0.804
0.789
516/150/126/6566
2
0.0346
0.803
0.799
0.801
513/126/129/6590
4
0.0284
0.836
0.815
0.825
523/103/119/6613
6
0.0253
0.813
0.827
0.820
531/122/111/6594
8
0.0231
0.827
0.847
0.837
544/114/98/6602
10
0.0214
0.830
0.858
0.844
551/113/91/6603
Appendix
Table 7: K sweep on the gold test set ( N=7,358 , 642 positives), frozen 90 th-percentile readout, Stage-3 model.
Rendering accurate Chinese text remains challenging for text-to-image models. Existing OCR-based reinforcement-learning rewards compare decoded transcripts with target strings. Such rewards overlook the compositional nature of Chinese writing: an ideograph consists of reusable components arranged through explicit spatial relations, yet OCR evaluates it as an atomic character. Consequently, visually different radical-level errors may receive equally coarse feedback, encouraging glyphs that merely resemble the target instead of faithfully reproducing its internal structure. We employ Ideographic Description Sequences (IDS), which comprise spatial operators and character components, and train an expert IDS recognizer to transcribe rendered Chinese text into this representation. Building on this recognizer, we introduce IDSpect, which deterministically decomposes the target text into IDS tokens and aligns crop-level visual IDS predictions with the target sequence. Globally unique token credit makes this comparison robust to the order of detected text regions. Combined with a whole-character semantic reward, IDSpect supplies fine-grained credit with component and spatial-relation without changing the image generator or adding inference-time cost. Experiments with GRPO post-training of Qwen-Image demonstrate that IDSpect achieves leading structural quality and semantic alignment on LongText and GenTextEval.
Yazhen Xie, Xingsong Ye, Zhineng Chen
Institute of Trustworthy Embodied AI, Fudan University · Shanghai Key Laboratory of Multimodal Embodied AI
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.
Hanshen Zhu, Yuliang Liu, Xuecheng Wu +7
Huazhong University of Science and Technology · ByteDance
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.
Mingxuan Cui, Jingpu Yang, Fengxian Ji +7
Mohamed bin Zayed University of Artificial Intelligence · Chinese Academy of Sciences Institute of Automation · Northeastern University +1