Tao Han Wanghan Xu Junchao Gong Xiaoyu Yue Song Guo Luping Zhou Lei Bai

Abstract

Arbitrary resolution image generation provides a consistent visual experienceacross devices, having extensive applications for producers and consumers.Current diffusion models increase computational demand quadratically withresolution, causing 4K image generation delays over 100 seconds. To solve this,we explore the second generation upon the latent diffusion models, where thefixed latent generated by diffusion models is regarded as the contentrepresentation and we propose to decode arbitrary resolution images with acompact generated latent using a one-step generator. Thus, we present theInfGen, replacing the VAE decoder with the new generator, forgenerating images at any resolution from a fixed-size latent without retrainingthe diffusion models, which simplifies the process, reducing computationalcomplexity and can be applied to any model using the same latent space.Experiments show InfGen is capable of improving many models into the arbitraryhigh-resolution era while cutting 4K image generation time to under 10 seconds.