PromptEnhancerImg2Img outputs Chinese even when input instruction is English

[wyhlovecpp]

Hi, thanks for releasing this model.

When testing PromptEnhancerImg2Img, I noticed that even if the edit_instruction is written in English, the enhanced prompt is often generated in Chinese.

Is this expected behavior (e.g., due to the training data being mainly Chinese), or is there a recommended way to enforce English output / preserve the input language?

Thanks!

[linqingwang]

Thank you for your feedback!

Regarding your observation: Yes, this is the expected behavior at this stage. The PromptEnhancerImg2Img model was primarily trained on a Chinese dataset specifically optimized for Yuanbao online users.

Because the training data is predominantly Chinese, the model tends to default to Chinese output even when the input instructions are in English. A dedicated bilingual version for seamless English-to-English enhancement has not yet been specifically optimized.

In the meantime, we recommend using Chinese instructions for the best results, or manually translating the output back to English if needed for your workflow. We appreciate your patience as we continue to refine the model!