Free-VTON: Cost-free acceleration and quality enhancement for diffusion-based virtual try-on

With the breakthrough of diffusion model in the field of image generation, the virtual try-on based on diffusion models has significant advantages in try-on performance. However, diffusion-based inference suffers from slow generation speed. Additionally, improving image quality generally relies on more and more network parameters, which raises computational cost. These issues hinder the real-time interactivity and practical deployment of virtual try-on systems. In this work, we propose Cost-Free Acceleration and Quality Enhancement for Diffusion-based Virtual Try-On called Free-VTON. Specifically, we introduce an Adaptive Caching acceleration strategy that adaptively caches and reuses features according to the similarity between features in adjacent diffusion steps. Aggressive caching is used when the similarity is high, and conservative caching is used when the similarity is low. Different caching trajectories are applied to different samples, and the caching rhythm is adaptively adjusted based on the content. This strategy accelerates the inference process without affecting the quality of try on. In addition, we introduce a Symmetric Feature Enhancement technique, which symmetrically amplifies the backbone features on both sides of the U-Net during inference to enhance feature extraction and reconstruction generation capabilities. Similarly, this technique improves generation quality with almost no additional computational overhead. Experiments demonstrate the superiority of our method for speed and quality trade-offs. Code will be avaliable at https://github.com/PERSIST10/freevton.