arXiv 2409.03550
DKDM: Data-Free Knowledge Distillation for Diffusion Models with Any Architecture
By Qianlong Xiang, Miao Zhang, et al.
Published 2024-09-05
Mindmap
Browse the paper's core ideas, clusters, and relationships in a structured outline.
Diffusion models (DMs) have demonstrated exceptional generative capabilities across various domains, including image, video, and so on. A key factor contributing to their effectiveness is the high quantity and quality of data used during training. However, mainstream DMs now consume increasingly large amounts of data. For example, training a Stable Diffusion model requires billions of image-text pairs. This enormous…