The field of generative modeling is witnessing significant advancements with the integration of diffusion-based models in molecular generation and image reconstruction. Recent developments indicate a shift towards designing more efficient and adaptive models that can handle complex tasks such as molecular optimization, image analysis, and conditional generation. Researchers are exploring novel approaches to improve the performance of diffusion models, including the incorporation of graph neural operators, Bayesian experimental design, and progressive inference-time annealing. These innovations have led to state-of-the-art results in various applications, including molecular design, MRI acquisition, and image reconstruction. Noteworthy papers in this area include the introduction of GrIDDD, a graph diffusion model that can dynamically grow or shrink chemical graphs during generation, and Diffusion Tree Sampling, a method that enables scalable inference-time alignment of diffusion models. Additionally, the development of Antibody Design and Optimization with Multi-scale Equivariant Graph Diffusion Models has shown promising results in therapeutic and diagnostic development. Overall, the field is moving towards more sophisticated and efficient models that can tackle complex tasks and provide high-quality results.
Diffusion-Based Models in Molecular Generation and Image Reconstruction
Sources
Graph Diffusion that can Insert and Delete
Active MRI Acquisition with Diffusion Guided Bayesian Experimental Design
Progressive Inference-Time Annealing of Diffusion Models for Sampling from Boltzmann Densities
Mesh-Informed Neural Operator : A Transformer Generative Approach
Uncertainty Quantification for Linear Inverse Problems with Besov Prior: A Randomize-Then-Optimize Method
Consistent Sampling and Simulation: Molecular Dynamics with Energy-Based Diffusion Models
Geometric-Aware Variational Inference: Robust and Adaptive Regularization with Directional Weight Uncertainty
Noise Consistency Training: A Native Approach for One-Step Generator in Learning Additional Controls
Ctrl-Z Sampling: Diffusion Sampling with Controlled Random Zigzag Explorations
Diffusion Tree Sampling: Scalable inference-time alignment of diffusion models
Antibody Design and Optimization with Multi-scale Equivariant Graph Diffusion Models for Accurate Complex Antigen Binding