The problem of large differences between the generated image and the original image

johnberg1 / CLIPInverter

Official Implementation of CLIP-Guided StyleGAN Inversion for Text-Driven Real Image Editing published in ACM TOG 2023

https://cyberiada.github.io/CLIPInverter/

11 stars 1 forks source link

The problem of large differences between the generated image and the original image #2

Open t1307109256 opened 10 months ago

t1307109256 commented 10 months ago

Hello, dear author, I found that when modifying a certain attribute of an image, some irrelevant areas will also be modified. How should I adjust the prompt words to avoid this situation? Can you give me some advice? Any suggestions?

johnberg1 commented 7 months ago

Hi, thanks for your interest in our work. Currently we use a 0.1 multiplier in the for the residual editing directions we add:

w_hat = w + 0.1 * encoder.forward_features(features)

In some cases, this may cause the model to apply the manipulations strongly, resulting in exaggerated outputs. You can change this multiplier to control the degree of manipulation, lower values will yield results closer to the original inversion but with less apparent manipulations.