Explore a comprehensive video explanation of the TUNIT paper, which introduces a novel approach to truly unsupervised image-to-image translation. Delve into the innovative method that eliminates the need for paired images or domain labels, allowing for fully unsupervised image translation from a source image to the style of one or multiple reference images. Learn about the joint training of a guiding network that provides style information and pseudo-labels. Follow the detailed breakdown of the architecture, including pseudo-label loss, encoder style contrastive loss, adversarial loss, generator style contrastive loss, and image reconstruction loss. Gain insights into the experimental results demonstrating the model's ability to separate domains and translate images effectively, even outperforming set-level supervised methods in semi-supervised settings.
Rethinking the Truly Unsupervised Image-to-Image Translation - Paper Explained
Overview
Syllabus
- Intro & Overview
- Unsupervised Image-to-Image Translation
- Architecture Overview
- Pseudo-Label Loss
- Encoder Style Contrastive Loss
- Adversarial Loss
- Generator Style Contrastive Loss
- Image Reconstruction Loss
- Architecture Recap
- Full Loss
- Experiments
Taught by
Yannic Kilcher
Related Courses
-
Big Self-Supervised Models Are Strong Semi-Supervised Learners
-
AttnGAN: Fine-Grained Text to Image Generation with Attentional Generative Adversarial Networks
-
Generative Adversarial Networks for Image Synthesis and Translation - Dr. Jan Kautz
-
CMU Multilingual NLP - Unsupervised Translation
-
NeurIPS 2023 Research Highlights - Circuits, Collaborative Filtering, Clustering, and Privacy
-
CMU Multilingual NLP 2022 - Unsupervised Machine Translation
