Skip to content

Latest commit

 

History

History
8 lines (5 loc) · 608 Bytes

201125 How to train your conditional GAN.md

File metadata and controls

8 lines (5 loc) · 608 Bytes

https://arxiv.org/abs/2011.13055

How to train your conditional GAN: An approach using geometrically structured latent manifolds (Sameera Ramasinghe, Moshiur Farazi, Salman Khan, Nick Barnes, Stephen Gould)

conditional gan에서 diversity가 떨어지는(mode collapse) 이유. latent mainfold의 큰 영역들을 singular point로 매핑하기 때문. generator가 latent manifold와 generator의 output image manifold 사이의 homeomorphism이 되도록 학습 과정을 개선하려는 아이디어. 상당히 흥미로운 개선인 듯. 리만 기하 맛 좀 볼래?

#gan #img2img #generative_model