DiffEM: Learning from Corrupted Data with Diffusion Models via Expectation Maximization

📝 Original Info

• Title: DiffEM: Learning from Corrupted Data with Diffusion Models via Expectation Maximization
• ArXiv ID: 2510.12691
• Date: 2025-10-14
• Authors: ** 논문에 명시된 저자 정보가 제공되지 않았습니다. **

📝 Abstract

Diffusion models have emerged as powerful generative priors for high-dimensional inverse problems, yet learning them when only corrupted or noisy observations are available remains challenging. In this work, we propose a new method for training diffusion models with Expectation-Maximization (EM) from corrupted data. Our proposed method, DiffEM, utilizes conditional diffusion models to reconstruct clean data from observations in the E-step, and then uses the reconstructed data to refine the conditional diffusion model in the M-step. Theoretically, we provide monotonic convergence guarantees for the DiffEM iteration, assuming appropriate statistical conditions. We demonstrate the effectiveness of our approach through experiments on various image reconstruction tasks.

💡 Deep Analysis

📄 Full Content

Reference