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

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Generation

Background:

• Diffusion models offer advanced virtual try-on performance but face challenges with slow inference speeds and high computational costs for quality enhancement.
• These limitations hinder real-time interactivity and practical deployment of virtual try-on systems.

Purpose of the Study:

• To develop a cost-free acceleration and quality enhancement method for diffusion-based virtual try-on systems.
• To address the trade-off between generation speed, image quality, and computational overhead in virtual try-on.

Main Methods:

• Introduced an Adaptive Caching strategy to reuse features across diffusion steps based on content similarity, adjusting caching aggressiveness.
• Implemented a Symmetric Feature Enhancement technique to amplify U-Net backbone features, improving extraction and reconstruction without significant overhead.

Main Results:

• The proposed Free-VTON method demonstrates superior speed and quality trade-offs in virtual try-on experiments.
• Adaptive Caching accelerates inference without compromising try-on quality.
• Symmetric Feature Enhancement boosts generation quality with minimal computational cost.

Conclusions:

• Free-VTON effectively overcomes the speed and quality limitations of existing diffusion-based virtual try-on methods.
• The proposed techniques enable more practical and efficient real-time virtual try-on applications.
• This work contributes to advancing the usability of diffusion models in interactive visual applications.