TextVerifier: Seeing Beyond Recognition to Verify Chinese Text Rendering

The strongest element is the IDS-conditioned text tower and its empirical justification. The ablation in Table 2 shows that replacing IDS with a single-character token drops F1 by roughly 7--10 points across stages, with Stage 3 recall collapsing from 0.858 to 0.648 because structural sensitivity is lost. The multi-prototype readout $\mathrm{score}(x,c)=1-\max_{j}\cos(f(x),\mu_{c}^{j})$ is a principled way to handle multi-modal legitimate styles without per-character threshold tuning, and the label-free Stage-3 mining of hard negatives from early versus late inpainting checkpoints is an elegant, scalable source of supervision.

A critical flaw is the circular evaluation of the downstream reward. In Section 4.5 the authors define Glyph-Acc as the fraction of target characters our TextVerifier judges correctly rendered, and then report that their full reward reaches 0.915 while a recognition-only baseline reaches 0.822. Because the judge and the reward are the same model, this is not independent evidence that human-perceived rendering errors actually decrease; it only shows the model reinforcing its own predictions. The qualitative comparisons in Figure 5 are suggestive but selective, and the comparison to TextPecker as a reward is likewise scored by TextVerifier, introducing a subtle bias toward its own texture and style preferences.

A second limitation is the modest scale and uncertain generalizability of the new benchmark. TextRenderBench-ZH is built from only 308 generated images yielding 7,358 character crops, of which just 642 are human-labeled errors. While the benchmark is the first per-character resource of its kind, coverage of rare characters, unusual fonts, and subtle stroke defects is thin, and the variance on a 642-positive test set means that small F1 differences may not be statistically robust. Finally, the global threshold $\delta$ is a single 90th percentile pooled over all training crops, which implicitly assumes that every character's legitimate manifold has the same score dispersion; if that homogeneity fails, the threshold will miscalibrate on certain characters.

The core verification claims are well-supported and the baseline comparisons are fair. The authors pit TextVerifier against OCR confidence, three general VLMs evaluated with an identical target-conditioned prompt, and TextPecker's released assessors, with TextVerifier leading by wide margins (F1 0.844 versus 0.590 for TextPecker and 0.329 for the best general VLM). The staged training ablations in Table 3 directly isolate contributions: removing Stage-1 replay drops F1 by 4.8 points, removing Stage-2 thresholding by 2.7 points, and random initialization collapses to 0.111. Where the evidence weakens is in the downstream comparisons: the Flow-GRPO reward improvements are judged by the reward model itself, so the reported gains over TextPecker in Table 4 are not independently verified and should be treated with caution.

Training hyperparameters and architecture choices are documented in unusual detail (ViT-L/16, DeBERTa-v2 large, exact temperatures and margins in Table 6), and the k-means prototype seed formula is provided. However, no code, model weights, or benchmark annotations are indicated to be publicly released, which severely limits independent reproduction. Reconstructing the 12.7M crop--IDS Stage-1 dataset requires re-crawling or licensing ~2.5M real images and running the CRAFT plus PP-OCRv6 filtering pipeline, while Stage 3 depends on training a custom JiT-H/16 inpainting model with modified 3D-RoPE to generate relative pairs. Without released checkpoints for this auxiliary model, reproducing the trajectory-mined hard negatives and the exact Stage-3 training distribution would require substantial engineering effort.