DiffI2I: Efficient Diffusion Model for Image-to-Image Translation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Diffusion Models (DM) are state-of-the-art for image synthesis but show limitations in image-to-image translation (I2I) tasks.
• Existing DMs for I2I often require extensive iterations and large models, leading to artifacts and inefficiency, especially for tasks like super-resolution that demand fidelity to ground truth (GT) images.

Purpose of the Study:

• To propose a novel Diffusion Model (DM) framework, DiffI2I, designed for efficient and high-performance image-to-image translation (I2I).
• To address the limitations of traditional DMs in I2I tasks by introducing a method that generates accurate results with reduced computational cost.

Main Methods:

• DiffI2I integrates three components: a compact I2I prior extraction network (CPEN), a dynamic I2I transformer (DI2Iformer), and a denoising network.
• A two-stage training process is employed: pretraining to capture a compact I2I prior representation (IPR) using CPEN, and Diffusion Model (DM) training where the DM estimates the IPR from input images.

Main Results:

• The compact IPR in DiffI2I enables more accurate outcomes compared to traditional DMs.
• DiffI2I utilizes a lighter denoising network and requires fewer iterations, significantly reducing computational burdens.
• Extensive experiments across various I2I tasks demonstrate state-of-the-art performance.

Conclusions:

• DiffI2I offers a simple, efficient, and powerful solution for image-to-image translation (I2I) tasks.
• The proposed framework achieves superior performance while substantially lowering computational requirements, making advanced I2I accessible.
• DiffI2I represents a significant advancement in applying Diffusion Models (DM) to practical image-to-image translation challenges.