search

bytedance / DEADiff

[CVPR 2024] Official implementation of "DEADiff: An Efficient Stylization Diffusion Model with Disentangled Representations"
Apache License 2.0
190 stars 4 forks source link

seems unable to accurately reproduce the style of the reference image？ #4

Open akk-123 opened 2 months ago

akk-123 commented 2 months ago

can you support deadiff with diffusers?

akk-123 commented 2 months ago

what's more, deadiff seems unable to accurately reproduce the style of the reference image, I test with app.py

Input image: 4 generate image: (prompt: a man, portrait) image

input image: 3 generate image: image

longmans commented 2 months ago

what's more, deadiff seems unable to accurately reproduce the style of the reference image, I test with app.py

Input image: 4 generate image: (prompt: a man, portrait) image

input image: 3 generate image: image

The same, still can't reproduce the result.

zhang2002forwin commented 2 months ago

Hello, is there a problem in ldm/modules/new_attention.py？In picture, self.atten2 function doesn't have the 'img_weight'. How do you solve this problem? Thanks a lot!!! I have tried to remove the parm 'img_weight', and then there is another problem in CrossAttention's forward function. Snipaste

Tianhao-Qi commented 2 months ago

Hello, is there a problem in ldm/modules/new_attention.py？In picture, self.atten2 function doesn't have the 'img_weight'. How do you solve this problem? Thanks a lot!!! I have tried to remove the parm 'img_weight', and then there is another problem in CrossAttention's forward function. Snipaste

This issue may stem from that you've not correctly installed xformers, so the computation of joint text-image cross attention layer will run into trouble. Therefore, please carefully check the package xformers.

zhang2002forwin commented 2 months ago

Thanks!! Your work is so fantastic!