Accelerating Image Generation with Sub-path Linear Approximation Model

Chen Xu, Tianhui Song, Weixin Feng, Xubin Li, Tiezheng Ge, Bo Zheng, Limin Wang* ;

"Diffusion models have significantly advanced the state of the art in image, audio, and video generation tasks. However, their applications in practical scenarios are hindered by slow inference speed. Drawing inspiration from the consistency models, we propose the Sub-Path Linear Approximation Model (SPLAM), which can accelerate diffusion models while maintaining high-quality image generation. SPLAM treats the PF-ODE trajectory as a series of PF-ODE sub-paths divided by sampled points, and harnesses sub-path linear (SL) ODEs to form a progressive and continuous error estimation along each individual PF-ODE sub-path. The optimization on such SL-ODEs allows SPLAM to construct denoising mapping with smaller cumulative approximated error. An efficient distillation method is also developed to facilitate the incorporation of pre-trained diffusion models, such as latent diffusion models. The extensive experimental results demonstrate SPLAM achieves remarkable training efficiency, requiring only 6 A100 GPU days to produce a high-quality generative model capable of 2 to 4-step generation. Comprehensive evaluations on LAION, MS COCO 2014, and MS COCO 2017 datasets also illustrate that SPLAM surpasses the existing acceleration methods in few-step generation tasks, achieving state-of-the-art performance both on FID and the quality of the generated images."