Unsupervised content and style learning for multimodal cross-domain image translation

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Pre-trained generative models show promise for cross-domain image translation.
• Diffusion-based models often suffer from color distortion and content structure loss.
• Existing methods struggle to maintain image structure during translation.

Purpose of the Study:

• To propose an unsupervised method for cross-domain image translation that preserves content structure and accurately maps color styles.
• To address limitations of current generative models in maintaining image integrity.

Main Methods:

• Developed an unsupervised content and style learning approach.
• Utilized self-structure attention loss to preserve content structure.
• Employed color constraint loss for accurate color space mapping.

Main Results:

• The proposed method effectively separates image content structure from color style.
• Achieved high-quality multi-modal cross-domain image translation.
• Outperformed state-of-the-art methods on multiple datasets in key metrics (LPIPS, NDB, JS, IS).

Conclusions:

• The novel method successfully maintains content structure and color patterns in cross-domain image translation.
• Offers a significant advancement over existing techniques for image translation tasks.
• Demonstrates superior performance in preserving visual integrity and style fidelity.