[PR113] D3: MISO Multimodal Image-to-Image Translation

Original Paper: MISO: Mutual Information Loss with Stochastic Style Representations for Multimodal Image-to-Image Translation

1. Image-to-image translation is fundamentally a multimodal problem. Previous methods
2. The authors assume a hierarchy between domain-invariant features(content) and domain-specific features(style) instead of a strictly dividing style and content. MISO pipeline is designed on this idea. The Mutual Information LOss(MILO) loss maximizes the mutual information between the feature and the image generated from that feature.
3. MISO was able to outperform other unpaired multimodal translation models in variety and quality.

User study results