Improved Residual Learning in Diffusion Models.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Diffusion models (DMs) excel at generative modeling, producing high-quality images.
• Residuals in generated images, caused by accumulated modeling errors in DMs, present a persistent challenge.

Purpose of the Study:

• To propose a novel residual learning framework to mitigate errors in diffusion model-generated images.
• To enhance the performance of pre-trained diffusion models through effective residual correction.

Main Methods:

• Developed a parameterized correction function within a residual learning framework.
• Introduced an advanced training approach using an ODE sampler bank for refined residual simulation.
• Incorporated a score consistency maintenance technique to improve model convergence.

Main Results:

• The proposed framework demonstrated significant improvements in residual correction for diffusion models.
• The correction function exhibited remarkable transferable residual correction capabilities across different DMs and datasets.
• Extensive experiments validated the framework's effectiveness and superiority on multiple datasets.

Conclusions:

• The novel residual learning framework effectively addresses and corrects residuals in diffusion model-generated images.
• The framework's transferable correction capability offers a versatile solution for enhancing various pre-trained diffusion models.
• The advanced training approach and techniques contribute to improved model convergence and performance.