Enhanced Multi-Scale Cross-Attention for Person Image Generation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Generative Adversarial Networks (GANs) are widely used for image generation.
• Person image generation presents challenges in synthesizing realistic appearance and shape.
• Existing GANs often struggle with effectively integrating multi-modal features for complex generation tasks.

Purpose of the Study:

• To propose a novel cross-attention-based GAN, named XingGAN, for improved person image generation.
• To enhance the fusion of appearance and shape information for more accurate synthesis.
• To develop a computationally efficient method that rivals diffusion-based model performance.

Main Methods:

• Developed XingGAN with two generation branches for appearance and shape.
• Introduced novel cross-attention blocks for feature transfer and embedding updates.
• Implemented multi-scale cross-attention blocks for long-range correlation learning.
• Proposed an enhanced attention (EA) module to refine attention weights.
• Integrated a densely connected co-attention module for feature fusion.

Main Results:

• XingGAN outperforms existing GAN-based methods in person image generation.
• The proposed method achieves performance comparable to diffusion-based models.
• XingGAN demonstrates significantly faster training and inference speeds compared to diffusion models.

Conclusions:

• The novel cross-attention mechanisms in XingGAN effectively improve person image synthesis.
• XingGAN offers a compelling alternative to diffusion models, balancing performance and efficiency.
• This work advances GAN-based approaches for complex image generation tasks.