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

Abstract

Generating high-fidelity multi-view facial images from a single input poses significant challenges in computer vision, requiring precise viewpoint control, robust identity preservation, and efficient inference. To address these challenges, we propose the Wavelet-Attention with Spatially Adaptive Noising Diffusion Network (WA-SAND), a diffusion-based framework that incorporates a spatially adaptive noise scheduling mechanism, a hybrid view conditioning strategy, and a frequency-aware generative process. The Spatially-Adaptive Diffusion Noise Schedule divides the diffusion process into two distinct phases guided by Perlin noise, enabling differentiated noise injection for facial and background regions. The Spatial Noise Embedding module encodes per-pixel signal-to-noise ratios, thereby enhancing spatial awareness during denoising. For controllable view synthesis, we introduce a Hybrid View Token and Angle Embedding module that combines discrete learnable view tokens with continuous Fourier-based angle embeddings, ensuring accurate canonical pose representation and smooth viewpoint transitions. The diffusion process is embedded within a wavelet-enhanced generator equipped with dual attention mechanisms-band reweighting and spatial-spectral attention-which improve high-frequency detail synthesis with minimal computational cost. Extensive evaluations on FFHQ and CelebA-HQ demonstrate that our framework outperforms state-of-the-art GAN and diffusion models. In particular, WA-SAND achieves an FID of 3.74 with only 24 sampling steps, requiring just 0.08 s per image at 256 × 256 resolution, significantly faster than others.