Advancements in Diffusion Models and Multi-Objective Optimization

The field is currently witnessing significant developments in diffusion models and multi-objective optimization. Researchers are focusing on improving the efficiency and stability of diffusion models, particularly in offline reinforcement learning settings. Novel methods, such as variational adaptive weighting and frequency-decoupled guidance, are being proposed to enhance the performance of these models. Meanwhile, advancements in multi-objective optimization are leading to the development of new algorithms with improved convergence rates and sample complexity. Noteworthy papers include: Fast and Stable Diffusion Planning through Variational Adaptive Weighting, which achieves competitive performance with up to 10 times fewer training steps. Guidance in the Frequency Domain Enables High-Fidelity Sampling at Low CFG Scales, which proposes frequency-decoupled guidance to improve image quality at low guidance scales. STIMULUS: Achieving Fast Convergence and Low Sample Complexity in Stochastic Multi-Objective Learning, which introduces a new algorithm with state-of-the-art convergence rates and sample complexities.

Sources

Fast and Stable Diffusion Planning through Variational Adaptive Weighting

Online Algorithms for Recovery of Low-Rank Parameter Matrix in Non-stationary Stochastic Systems

Guidance in the Frequency Domain Enables High-Fidelity Sampling at Low CFG Scales

STIMULUS: Achieving Fast Convergence and Low Sample Complexity in Stochastic Multi-Objective Learning

On sharp stable recovery from clipped and folded measurements

Identifiability and Maximum Likelihood Estimation for System Identification of Networks of Dynamical Systems

Rethinking Oversaturation in Classifier-Free Guidance via Low Frequency

Built with on top of