New framework optimizes diffusion model schedules for image restoration

Researchers have developed a new analytical framework for designing schedules in Brownian Bridge Diffusion Models (BBDMs), which are used for image restoration and inverse problems. This framework utilizes a Mixture-of-Gaussians (MoG) prior to derive closed-form ideal posteriors and MMSE denoisers. The work introduces two schedule-design objectives: one based on Wasserstein distance for perceptual quality and another based on Mean Squared Error (MSE) for reconstruction fidelity, revealing an inherent trade-off between them. Experiments on controlled MoG settings and the FFHQ dataset for tasks like inpainting, deblurring, and super-resolution validate the effectiveness of these criteria. AI

IMPACT Introduces a theoretical framework and practical criteria for optimizing diffusion model performance in image restoration tasks.

RANK_REASON This is a research paper detailing a new analytical framework and methodology for diffusion models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Brownian Bridge Diffusion Models
• FFHQ dataset
• Mixture-of-Gaussians

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →