Abstract:

[Objective] Diffusion models require multi-step denoising during image generation, resulting in high inference cost. Feature caching can reduce redundant computation, but it degrades generation quality. Classifier-free guidance can improve generation quality but increases computational load. [Methods] To balance efficiency and quality, this paper improves classifier-free guidance so that it requires only a small amount of computation and can work synergistically with feature caching. In addition, we design a latent-space distillation method to train the guidance model needed for classifier-free guidance, achieving stronger guidance effects. A gradient-smoothing strategy is introduced to further improve training quality. [Results] Experiments are conducted on datasets of different resolutions using architectures such as U-ViT and DiT. Metrics such as FID, IS, precision are used to evaluate generation quality, distribution coverage, and inference efficiency. [Conclusion] The results show that the proposed method achieves a superior quality-speed trade-off across different resolutions and model architectures.

Key words: diffusion models, image generation, classifier-free guidance, inference acceleration, distillation