All-in-One Multilingual Scene Text Recognition with Script-aware Mixture-of-Experts
ScriptMoE is presented as a single multilingual text recognizer that beats both per-language STR baselines and heavyweight VLMs on the paper’s benchmarks.
The authors build TextMuSS-10M, a synthetic dataset covering 10 scripts and 229 languages. Their model uses one shared visual encoder, then routes each image to top-2 script-aligned experts plus a shared expert. On TextMuSS-Bench, it reaches 82.06% accuracy, 1.31 points above the strongest STR baseline cited. In a CC-OCR setup, swapping ScriptMoE into PP-OCRv5 raises F1 from 65.71% to 80.89%, just above the best VLM result reported. Source: HF Daily Papers' note.
The authors build TextMuSS-10M, a synthetic dataset covering 10 scripts and 229 languages. Their model uses one shared visual encoder, then routes each image to top-2 script-aligned experts plus a shared expert. On TextMuSS-Bench, it reaches 82.06% accuracy, 1.31 points above the strongest STR baseline cited. In a CC-OCR setup, swapping ScriptMoE into PP-OCRv5 raises F1 from 65.71% to 80.89%, just above the best VLM result reported. Source: HF Daily Papers' note.
score 4