SPDiff: Enhancing Diffusion Models for Scalable Periodic Image Generation

Periodic images have broad applications including texture synthesis, pattern design, and panoramic scene generation. Recent progress in diffusion-based image generation enables high-quality image synthesis. However, existing models exhibit inherent spatial biases that prevent them from producing periodic output simply through prompts like ``periodic''. Moreover, creating high-resolution images or images with unusual aspect ratios often leads to loss of periodic structure and visual degradation. This paper proposes a novel SPDiff framework that enhances diffusion models for scalable periodic image generation. To enable periodic image synthesis at different high resolutions, we introduce Cross-Subimage Splice-and-Focus, a training-free approach that combines overlapped tiling with a refined attention strategy. We further incorporate a RandomShift mechanism to explicitly break spatial biases, thus improving image naturalness. Additionally, we develop a Periodic Expansion Pipeline (PEP) to produce large periodic units from a reference image. Extensive experiments demonstrate that our framework enables scalable periodic generation with high visual quality.