Fudan's FlowTranslator Does Image Translation Without Training

FlowTranslator, a training-free in-image machine translation system from Fudan University's Institute of Trustworthy Embodied AI, reaches 40.1% exact-match accuracy translating English text in 1,280 ScenePair photographs into Chinese. Against UniTranslator, a system fine-tuned on 7M in-domain pairs, it posts average BLEU 20.6 vs 12.8 and COMET 57.2 vs 50.7 on the PRIM benchmark, the paper reports.

The pipeline trains nothing. PP-OCRv5 extracts text boxes, a call to Doubao-Seed-2.0-lite returns the translations plus paired source and target captions, and a frozen 6B flow-matching text-to-image model called Z-Image-Turbo does the rendering. The trick is making the editor actually edit. The authors start from FlowEdit, an inversion-free flow-editing procedure, and observe it won't work out of the box for IIMT: the text region is such a small minority of the frame that "the two branches barely differ at the outset and the edit never starts."

Two mask-confined trajectory controls fix this. Region Coupling Release partially decorrelates target-branch noise inside the text region early in the trajectory; Coupling Expansion amplifies the resulting displacement mid-trajectory. On PRIM the two controls together lift average BLEU from 16.88 to 20.57, with Coupling Expansion alone recovering roughly 90% of the gain.

Script coverage inherits whatever the frozen backbone knows. English-to-Russian sits at 4.6 BLEU because Z-Image-Turbo "renders Cyrillic unreliably, producing near-Latin glyph substitutions that OCR cannot read." The whole system runs on a single NVIDIA RTX 4090, and arrived among a cluster of Chinese multimodal diffusion papers posted the same day.