arXiv 2407.03297
Improved Noise Schedule for Diffusion Training
By Tiankai Hang, Shuyang Gu, et al.
Published 2024-07-03
Mindmap
Browse the paper's core ideas, clusters, and relationships in a structured outline.
Diffusion models have emerged as the de facto choice for generating high-quality visual signals across various domains. However, training a single model to predict noise across various levels poses significant challenges, necessitating numerous iterations and incurring significant computational costs. Various approaches, such as loss weighting strategy design and architectural refinements, have been introduced to ex…