Visual Text Rendering on OCR
0.95OCR AccuracyVGPO
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| VGPOKL Regularization=w/o KL2025.12 | 0.95 | 5.33 | 3.98 | 0.9 | 22.17 | 3.07 | |
| VGPOKL Regularization=w/ KL2025.12 | 0.94 | 5.34 | 4.08 | 0.98 | 22.44 | 3.14 | |
| Flow-OPDCold-start=Model Merging2026.05 | 0.94 | — | — | — | — | — | |
| Flow-GRPOKL Regularization=w/o KL2025.12 | 0.93 | 5.13 | 3.66 | 0.58 | 21.79 | 2.82 | |
| Flow-GRPOKL Regularization=w/ KL2025.12 | 0.92 | 5.32 | 4.06 | 0.95 | 22.44 | 3.14 | |
| GRPO-OCRTeacher Model=True2026.05 | 0.92 | — | — | — | — | — | |
| Flow-OPDCold-start=Supervised Fine-Tuning (SFT)2026.05 | 0.92 | — | — | — | — | — | |
| SFT+GRPO-MixInitialization=SFT2026.05 | 0.86 | — | — | — | — | — | |
| Merge+GRPO-MixInitialization=Model Merging2026.05 | 0.86 | — | — | — | — | — | |
| GRPO-Mix2026.05 | 0.83 | — | — | — | — | — | |
| GRPO-PickscoreTeacher Model=True2026.05 | 0.69 | — | — | — | — | — | |
| GRPO-deqaTeacher Model=True2026.05 | 0.66 | — | — | — | — | — | |
| GRPO-GenevalTeacher Model=True2026.05 | 0.65 | — | — | — | — | — | |
| SD3.5-M2025.12 | 0.59 | 5.39 | 4.07 | 0.87 | 22.34 | 3.07 | |
| SD-3.5-M2026.05 | 0.59 | — | — | — | — | — |