Multilingual scene text recognition (STR) has long been split between two imperfect options: deploying a separate recognizer for every language, which raises serving costs and compounds errors, or relying on large vision-language models (VLMs) that are expensive and still unreliable across many scripts. A new Hugging Face paper proposes a third path.
The work introduces TextMuSS-10M, a large-scale synthetic scene text dataset spanning 10 scripts and 229 languages. The authors say it supplies balanced supervision for languages where real training data is scarce or unavailable.
On top of that data, the team proposes ScriptMoE, a script-aware Mixture-of-Experts architecture. It keeps a single shared visual encoder but replaces the dense decoder with a sparse MoE block: an image-level router sends each image to the top-2 script-aligned experts, while a shared expert captures cross-script knowledge.
On the authors' assembled TextMuSS-Bench (10 scripts, 10,899 images), ScriptMoE reaches 82.06% accuracy, edging out the strongest STR baseline by 1.31%. In the end-to-end CC-OCR multilingual task, swapping only the recognizer inside PP-OCRv5 for ScriptMoE raises F1 from 65.71% to 80.89% — slightly above the best VLM (80.73%) while using a fraction of the parameters.
The authors frame the result as an all-in-one recognizer that is simpler than per-language experts, lighter than VLMs, and more accurate than both.