Adaptative Noise Scheduling for Equivariant Diffusion Models

The Equivariant Diffusion Model (EDM) has emerged as a powerful generative framework for producing high-quality 3D data. EDM operates by progressively denoising a sampled input through a graph neural network trained to reverse a predefined noise process, all while maintaining E(3) equivariance. A critical component of this noising process is the noise scheduler, which determines how noise is injected during the forward process and removed during generation. The choice of noise schedule significantly influences model behavior, affecting training stability, convergence, and the quality of the generated samples. Despite its importance, noise scheduling strategies have not been extensively studied in the context of EDM. My master thesis aims to explore how adaptive noise scheduling approaches impact the performance and generative quality of EDM across diverse tasks.