noise image

[hehehahi4]

Thanks for your work. I have a question.
Why the images are noisy? It seems not be explaned in paper. Which type of images did you use in experiment, noise_gs, noise_line or origin?

[EchoDreamer]

Thanks for your interest. For the evaluation and steering experiments of modality preference, we use the "origin" images, which are sourced from COCO2014. The code in line 14-17 of the "steer_qwen_search_scale.py" file, we point the image path of coco2014. We only test the shifting of modality preference using noisy images. Perhaps the Hugging Face file misled you, and I will make modifications to the repository.