FlowTranslator: Training-Free In-Image Machine Translation with Flow-Based Editing
FlowTranslator reports 40.1% exact-match accuracy on photographs translated into Chinese, without task-specific training.
The framework uses OCR and an LLM to build translation prompts for an existing text-to-image diffusion model. Two controls adjust the editing trajectory to improve translated text while preserving the surrounding image. Under an EasyOCR read-back evaluation on PRIM, it scores 20.6 BLEU and 57.2 COMET, outperforming the training-based UniTranslator baseline. The authors also report stronger results than existing editing systems on ScenePair photographs, while noting that some comparisons cover text rendering rather than the full translation task.
HF Daily Papers' note
The framework uses OCR and an LLM to build translation prompts for an existing text-to-image diffusion model. Two controls adjust the editing trajectory to improve translated text while preserving the surrounding image. Under an EasyOCR read-back evaluation on PRIM, it scores 20.6 BLEU and 57.2 COMET, outperforming the training-based UniTranslator baseline. The authors also report stronger results than existing editing systems on ScenePair photographs, while noting that some comparisons cover text rendering rather than the full translation task.
HF Daily Papers' note
score 4