What it is
An all-in-one multilingual scene text recognizer uses script-aware MoE decoding plus a 10M synthetic dataset to cover 10 scripts and 229 languages. Instead of one recognizer per language or a huge VLM, it uses one shared encoder with a sparse script-aware decoder: an image-level router sends each crop to top-2 script experts plus a shared expert.
Key results
- 82.06% on TextMuSS-Bench, beating the strongest STR baseline by 1.31%
- CC-OCR F1 improves 65.71%→80.89% by swapping in ScriptMoE
- 80.89% CC-OCR F1 slightly beats the best VLM at 80.73%
Shows script-level sparse specialization can outperform both per-language OCR pipelines and much larger VLM recognizers in multilingual STR.