GP-UNIT: Generative Prior for Versatile Unsupervised Image-to-Image Translation.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Unsupervised image-to-image translation models face challenges with domains exhibiting significant visual differences.
• Existing methods struggle to establish robust mappings between diverse visual domains without paired data.

Purpose of the Study:

• To introduce a versatile framework, Generative Prior-guided Unsupervised Image-to-image Translation (GP-UNIT), enhancing translation quality, applicability, and controllability.
• To address the limitations of current models in handling drastic visual discrepancies between domains.

Main Methods:

• GP-UNIT leverages generative priors from pre-trained class-conditional Generative Adversarial Networks (GANs) to establish coarse-level cross-domain correspondences.
• It applies these learned priors to adversarial translations for excavating fine-level correspondences.
• For distant domains, semi-supervised learning is incorporated to guide the discovery of accurate semantic correspondences.

Main Results:

• GP-UNIT achieves robust, high-quality, and diversified translations between both close and distant visual domains.
• The framework allows for controlled balancing of content and style consistency in translations.
• Experiments demonstrate GP-UNIT's superiority over state-of-the-art translation models.

Conclusions:

• GP-UNIT offers a significant advancement in unsupervised image-to-image translation, particularly for domains with substantial visual variations.
• The model's ability to learn multi-level content correspondences enhances its applicability and controllability.
• The integration of generative priors and semi-supervised learning provides a powerful approach for cross-domain image translation.