Foundations and Advanced Techniques in Diffusion Models for Efficient and Scalable Generative Systems
Noriko Hachiro, Yukiko Ayumi, Sakura Suzuki, Zoe Brian · HAL (Le Centre pour la Communication Scientifique Directe) · 2024
Diffusion models have emerged as a powerful class of generative models, achieving remarkable success in diverse applications such as image synthesis, natural language processing, and scientific simulations. These models leverage a systematic denoising process to generate data, offering high-quality and diverse outputs. However, their practical adoption is challenged by significant computational demands during training and inference. This survey provides a comprehensive overview of diffusion models, from their foundational principles to advanced techniques for enhancing efficiency. We explore strategies such as optimized noise scheduling, architecture design improvements, and accelerated sampling methods, which collectively address the computational bottlenecks. Furthermore, we highlight the transformative impact of diffusion models across domains, including healthcare, gaming,