CycleDiff: Cycle Diffusion Models for Unpaired Image-to-image Translation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Cross-domain image translation typically requires paired training data, which is often unavailable.
• Existing diffusion-based methods struggle to align diffusion and translation processes, leading to suboptimal results.
• Generative Adversarial Networks (GANs) are common but can suffer from training instability and mode collapse.

Purpose of the Study:

• To develop a diffusion-based cross-domain image translation method that does not require paired training data.
• To address the challenge of aligning diffusion and translation processes for improved performance.
• To enhance the fidelity and structural consistency of translated images.

Main Methods:

• A novel joint learning framework is proposed to align diffusion and translation processes.
• Diffusion models are used to extract image components representing the clean signal.
• A time-dependent translation network learns complex translation mappings within an end-to-end joint learning approach.

Main Results:

• The proposed method achieves global optimization of both diffusion and translation processes.
• Experiments on various cross-modality tasks (RGB↔Edge, RGB↔Semantics, RGB↔Depth) show superior generative performance.
• The method achieved state-of-the-art Frechet Inception Distance (FID) scores on benchmark tasks, outperforming existing methods.

Conclusions:

• The joint learning framework effectively enables diffusion-based cross-domain image translation without paired data.
• The approach significantly improves image translation fidelity and structural consistency.
• This work advances the capabilities of diffusion models in complex image generation tasks.