DemoFusion: Democratising High-Resolution Image Generation With No $$$

High-resolution image generation with Generative Artificial Intelligence (GenAl) has immense potential but, due to the enormous capital investment required for training, it is increasimgly centralised to a few large corporations, and hidden behind paywalls. This paper aims to democratise high-resolution GenAl by advancing the frontier of high-resolution generation while remaining accessible to a broad audience. We demonstrate that existing Latent Diffusion Models (LDMs) possess untapped potential for higher-resolution image generation. Our novel DemoFusion framework seamlessly extends open-source GenAl models, employing Progressive Upscaling, Skip Residual, and Di-lated Sampling mechanisms to achieve higher-resolution image generation. The progressive nature of DemoFusion requires more passes, but the intermediate results can serve as “previews”, facilitating rapid prompt iteration.

1. Introduction

Generating high-resolution images with Generative Artificial Intelligence (GenAI) models has demonstrated remarkable potential [1], [19], [23]. However, these capabilities are increasingly centralised. Training high-resolution image generation models requires substantial capital investments in hardware, data, and energy that are beyond the reach of indi-vidual enthusiasts and academic institutions. For example, training Stable Diffusion 1.5, at a resolution of 512², en-tails over 20 days of training on 256 A100 GPUs [1]. Companies that make these investments understandably want to recoup their costs and increasingly hide the resulting models behind paywalls. This trend toward centralisation and pay-per-use access is accelerating as GenAI image synthe-sis advances in quality since the investment required to train image generators increases rapidly with image resolution. Figure 1.

Selected landscape samples of DemoFusion versus SDXL [24] (all images in the figure are presented at their actual sizes). SDXL can synthesize images up to a resolution of 10242, while DemoFusion extends SDXL to generate images at 4×, 16×, and even higher resolutions without any fine-tuning or prohibitive memory demands. All generated images are produced using a single RTX 3090 GPU. Best viewed ZOOMED-IN.