WA-SAND: Wavelet attention diffusion with spatially adaptive noising for multi-view face generation

Area of Science:

• Computer Vision
• Generative Models
• Image Synthesis

Background:

• Generating multi-view facial images from a single input presents challenges in viewpoint control, identity preservation, and inference speed.
• Existing generative adversarial networks (GANs) and diffusion models struggle to meet these demands effectively.

Purpose of the Study:

• To propose a novel diffusion-based framework, the Wavelet-Attention with Spatially Adaptive Noising Diffusion Network (WA-SAND), for high-fidelity multi-view facial image generation.
• To enhance viewpoint control, identity preservation, and inference efficiency in single-image multi-view synthesis.

Main Methods:

• Developed a Spatially-Adaptive Diffusion Noise Schedule with Perlin noise for differentiated noise injection in facial and background regions.
• Introduced a Hybrid View Token and Angle Embedding module for accurate canonical pose representation and smooth viewpoint transitions.
• Integrated a wavelet-enhanced generator with dual attention mechanisms (band reweighting and spatial-spectral attention) for efficient high-frequency detail synthesis.

Main Results:

• WA-SAND outperforms state-of-the-art GAN and diffusion models on FFHQ and CelebA-HQ datasets.
• Achieved a Fréchet Inception Distance (FID) of 3.74 with only 24 sampling steps.
• Demonstrated significantly faster inference, requiring 0.08 seconds per image at 256x256 resolution.

Conclusions:

• WA-SAND offers a highly effective and efficient solution for single-image multi-view facial image generation.
• The proposed framework advances the state-of-the-art in controllable image synthesis and generative modeling.
• The combination of spatial noise adaptation, hybrid view conditioning, and wavelet-enhanced generation leads to superior performance.