CVPR2021
Anycost GANs for Interactive Image Synthesis and Editing
Ji Lin, Richard Zhang, Frieder Ganz, Song Han, Jun-Yan Zhu
Abstract
input projected 2x faster 4x faster 8x faster edit: no smile 8x faster edit: hair color 8x faster (a) input images (b) Anycost GANs provide consistent outputs for projected latent code (c) Consistency remains after editing less hair color Figure 1: Anycost GAN can be executed at flexible computation costs (fast preview with low cost and high-quality output with full cost), enabling interactive image editing with quick preview. The low-cost sub-generator produces consistent outputs compared to the full-cost generator during both image projection and latent code traversal, making the sub-generator an accurate proxy for various editing tasks (e.g., no smile, changing hair color). Interactive demos are available here.