LinkGAN: Linking GAN Latents to Pixels for Controllable Image Synthesis

Jiapeng Zhu, Ceyuan Yang, Yujun Shen, Zifan Shi, Deli Zhao, Qifeng Chen

Abstract

This work presents an easy-to-use regularizer for GAN training, which helps explicitly link some axes of the latent space to an image region or a semantic category (e.g., sky) in the synthesis. Establishing such a connection facilitates a more convenient local control of GAN generation, where users can alter image content only within a spatial area simply by partially resampling the latent codes. Experimental results confirm four appealing properties of our regularizer, which we call LinkGAN. (1) Any image region can be linked to the latent space, even if the region is pre-selected before training and fixed for all instances. (2) Two or multiple regions can be independently linked to different latent axes, surprisingly allowing tokenized control of synthesized images. (3) Our regularizer can improve the spatial controllability of both 2D and 3D GAN models, barely sacrificing the synthesis performance. (4) The models trained with our regularizer are compatible with GAN inversion techniques and maintain editability on real images
BibTeX
@inproceedings{zhu2023linkgan,
  author = {Jiapeng Zhu and Ceyuan Yang and Yujun Shen and Zifan Shi and Deli Zhao and Qifeng Chen},
  title = {{LinkGAN: Linking GAN Latents to Pixels for Controllable Image Synthesis}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2023},
}
Figure