Tobias Vontobel Seyedmorteza Sadat Farnood Salehi Romann M. Weber

Abstract

Diffusion models have emerged as the leading approach for image synthesis,demonstrating exceptional photorealism and diversity. However, trainingdiffusion models at high resolutions remains computationally prohibitive, andexisting zero-shot generation techniques for synthesizing images beyondtraining resolutions often produce artifacts, including object duplication andspatial incoherence. In this paper, we introduce HiWave, a training-free,zero-shot approach that substantially enhances visual fidelity and structuralcoherence in ultra-high-resolution image synthesis using pretrained diffusionmodels. Our method employs a two-stage pipeline: generating a base image fromthe pretrained model followed by a patch-wise DDIM inversion step and a novelwavelet-based detail enhancer module. Specifically, we first utilize inversionmethods to derive initial noise vectors that preserve global coherence from thebase image. Subsequently, during sampling, our wavelet-domain detail enhancerretains low-frequency components from the base image to ensure structuralconsistency, while selectively guiding high-frequency components to enrich finedetails and textures. Extensive evaluations using Stable Diffusion XLdemonstrate that HiWave effectively mitigates common visual artifacts seen inprior methods, achieving superior perceptual quality. A user study confirmedHiWave's performance, where it was preferred over the state-of-the-artalternative in more than 80% of comparisons, highlighting its effectiveness forhigh-quality, ultra-high-resolution image synthesis without requiringretraining or architectural modifications.