Advancements in Diffusion Models and Sampling Techniques

The field of research is witnessing significant developments in diffusion models and sampling techniques. Recent studies have focused on improving the efficiency and accuracy of these models, particularly in handling complex data distributions and inverse problems. The introduction of novel frameworks, such as Measurement-Aligned Sampling, and the extension of existing methods, like Stein Variational Gradient Descent, have shown promising results in addressing challenging tasks. Furthermore, the exploration of diffusion models in various applications, including image restoration and optimization problems, has led to innovative solutions and improved performance. Noteworthy papers include: The Effect of Stochasticity in Score-Based Diffusion Sampling, which provides valuable insights into the role of stochasticity in diffusion models. An Exact and Efficient Sampler for Dynamic Discrete Distributions, which presents a groundbreaking approach to sampling from dynamic discrete distributions. Unsupervised Imaging Inverse Problems with Diffusion Distribution Matching, which offers a versatile method for addressing inverse problems in imaging using unpaired datasets.

Sources

The Effect of Stochasticity in Score-Based Diffusion Sampling: a KL Divergence Analysis

Abstract Sound Fusion with Unconditioned Inversion Model

Measurement-aligned Flow for Inverse Problem

Branching Stein Variational Gradient Descent for sampling multimodal distributions

An Exact and Efficient Sampler for Dynamic Discrete Distributions

FRIDU: Functional Map Refinement with Guided Image Diffusion

Exploring Diffusion with Test-Time Training on Efficient Image Restoration

Unsupervised Imaging Inverse Problems with Diffusion Distribution Matching

Intrinsic Annealing in a Hybrid Memristor-Magnetic Tunnel Junction Ising Machine

Iterative Camera-LiDAR Extrinsic Optimization via Surrogate Diffusion

Two-dimensional Parallel Tempering for Constrained Optimization

Built with on top of