A Diffusion Model Translator for Efficient Image-to-Image Translation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Diffusion models are increasingly used for image-to-image translation (I2I).
• Existing methods are time-consuming due to iterative refinement in each denoising step.
• There is a need for more efficient I2I translation techniques.

Purpose of the Study:

• To propose an efficient Diffusion Model Translator (DMT) for image-to-image translation.
• To provide theoretical justification for domain transfer at intermediate steps in diffusion models.
• To develop a strategy for automatically selecting optimal timesteps for domain transfer.

Main Methods:

• Equipping a diffusion model with a lightweight translator (DMT).
• Theoretically justifying domain transfer at intermediate diffusion steps.
• Proposing a strategy for automatic timestep selection for domain transfer.

Main Results:

• The proposed DMT achieves efficient and high-quality image-to-image translation.
• The method demonstrates efficacy across various applications including stylization, colorization, segmentation-to-image, and sketch-to-image.
• DMT outperforms existing methods in both translation quality and computational efficiency.

Conclusions:

• The Diffusion Model Translator (DMT) offers a significant advancement in efficient image-to-image translation.
• The theoretical framework and automatic timestep selection enhance the practicality and performance of diffusion-based I2I.
• This approach is broadly applicable and improves upon current state-of-the-art methods.