image2hairstyle failed to get right results

[yang1hu]

hello！when i try the image2hairstyle to get the results. i found the results is alwawys the same no matter what picture i change.when i debug ,i found the discription of the image is the following:

I think LLaVA return the bad results

[iiijasdf]

I might have a different problem when using image2hairstyle. I'm using example1.png as input but it seems like dimension mismatching.

[c-he]

image2hairstyle seems to work on my side. Though the results are reasonable, their quality is kind of poor.

[Vanessik]

@c-he Yes, the current approach is solely text-based and generation is not conditioned on any image features, that's why the quality is like that. An additional stage with personalized fine-tuning or conditioning diffusion model on image features may improve results.