search

Jamie-Cheung / ArtBank

ArtBank: Artistic Style Transfer with Pre-trained Diffusion Model and Implicit Style Prompt Bank (AAAI2024)
31 stars 2 forks source link

Why does the test come out badly？ #8

Open BBABM opened 1 month ago

BBABM commented 1 month ago

I used the author's trained model and the default code to run it, and I feel that the results are much worse, is there still a need to change the parameters? scenery_resized fstylized “Vangoh” car_resized carstylized “picasso“ It feels like the content is badly distorted

BBABM commented 1 month ago

@Jamie-Cheung I feel there should be something wrong with the parameters, do I still need to change the parameters in the code, thank you very much!

Jamie-Cheung commented 1 month ago

@Jamie-Cheung I feel there should be something wrong with the parameters, do I still need to change the parameters in the code, thank you very much!

You can adjust the parameter "strength" during testing. When this method is used, it will get better results on some content images, but not on some content images. This is because the weight is only trained on the style images. In inference, if the content of the content images does not appear in the trained style images , and the result will have a large degradation.