Stable Diffusion HiRes Fix Guide

In this video, Silicon Thaumaturgy dives deep into the HiRes Fix feature of the Stable Diffusion Automatic1111 Web GUI, explaining its functionality, important settings, and how to use it effectively to generate high-quality images. The HiRes Fix feature is designed to address issues that arise when generating images at non-default resolutions, such as twins, triplets, or fractalization. The process involves generating an initial image at a default resolution, upscaling it, adding noise based on the denoising slider, and then performing another image generation with the upscaled noisy image. The video identifies four key variables for generating good images with HiRes Fix: denoising, the upscaler, the amount of upscaling, and subject matter. Denoising affects the complexity and detail of the final image, with higher denoising levels leading to more fractalization. The choice of upscaler impacts the image quality, with latent upscalers requiring higher denoising levels and non-latent upscalers providing better results. The amount of upscaling also influences the settings, with higher upscaling levels leading to faster fractalization. Subject matter plays a role, with single subjects like faces being less prone to issues than complex or multiple subjects. The video concludes with recommendations on settings, such as avoiding latent upscalers, using non-latent upscalers, keeping denoising within specific ranges, and limiting upscaling to 3x or below for best results. The video also provides a link to a Google spreadsheet with figures and useful prompt constructors for viewers to use.