Presentation / Installation
Parsing-conditioned Anime Translation: A New Dataset and Method
SessionImage and Video Editing
DescriptionThe abstract art of anime presents a challenging misaligned image translation problem. We propose a new anime translation framework that leverages prior knowledge of a pre-trained StyleGAN model. The proposed framework incorporates disentangled encoders, four tailored losses, and a FaceBank aggregation method to generate in-domain animes. We further introduce the Danbooru-Parsing dataset, which connects face semantics with appearances, enabling constrained translation settings. Experiments demonstrate editability and extend the method to manga images. We produce the first feasible solution to anime translation.
Presenter
Event Type
Technical Paper
TimeMonday, 7 August 20232:11pm - 2:22pm PDT
LocationPetree Hall D
ACM Digital Library
Technical Papers pdfs
Session Time & Location
Research & Education
Livestreamed
Recorded
FC
FCS
V
VS
EFC
