Text-to-image models can now produce Chinese text that looks readable at a glance, but the characters often contain local rendering defects—an extra stroke, a missing component, or a malformed part. These errors are invisible to OCR systems and vision-language models, which are trained for recognition and therefore ignore the very structural deviations that define a rendering error.

The paper introduces TextVerifier, which formalizes 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. Instead of feeding the text tower a single-character symbol, TextVerifier uses the character's Ideographic Description Sequence (IDS), a structural decomposition. The model was trained on 12.7 million 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, the authors built TEXTRENDERBENCH-ZH, the first per-character benchmark for generated Chinese text, with position, identity, and correctness annotations for 7,358 characters.

On that benchmark, TextVerifier reaches an F1 of 0.844, far above OCR confidence (0.526), a general VLM (0.329), and the existing TextPecker (0.590). The paper also shows that when TextVerifier's score is used as a per-character reward for Flow-GRPO post-training, it further reduces rendering errors—indicating the same score can both detect and correct defects.