.png){kind=icon}
Return
Robust GAN inversion
.png){kind=icon}
PRE
AI
DOI：10.1080/10556788.2025.2581586.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
GAN inversion aims to faithfully reconstruct a real target image while preserving the editability offered by the latent space of a pretrained generator. We introduce WRanGAN, a weight-randomized variant of StyleGAN2 that learns adaptive per-parameter regularization coefficients. During inversion we optimize a latent code together with a low-dimensional noise vector controlling the randomized weights. This yields lower distortion than prior tuning approaches (e.g. PTI) while remaining in the original W space, thus retaining editing quality. Extensive experiments on FFHQ and LSUN Church demonstrate consistent improvements in reconstruction, editing, and computational cost.
Keywords:
GAN inversion
regularization strategy
optimization
Journal
O
IF:
1.4
Papers:
31
Citations:
0
Organization
Cited Papers
.png){kind=icon}
ShareSaveShareSave
View more