7 Creating High Resolution Images

Stable Diffusion can create striking images, but its standard output size is too small for many practical uses like printing, high-resolution displays, or detailed editing. This chapter explains why simply asking the model for bigger images often fails: larger generations demand more VRAM, and more importantly, the model was trained mostly on lower-resolution images, so it tends to produce strange repetition and distorted detail when pushed too far beyond its comfort zone.

To work around these limits, the chapter presents upscaling methods in A1111. It begins with a low-resolution image and then explores hires-fix and different upscalers, showing how scale factor and denoising strength affect the result. Smaller scaling steps and moderate denoising usually preserve the original composition better, while overly aggressive settings can blur details or reshape the image too much. The key idea is that upscaling is not just a technical enlargement; it can also add believable detail when tuned carefully.

The chapter then expands into image-to-image generation, which gives even more control over hi-res output. By sending an image into img2img, changing the seed, switching checkpoints, repeating the process, or even adjusting the prompt, the image can evolve into new styles rather than merely being enlarged. This makes hi-res creation a creative workflow as much as a scaling workflow, allowing the same base image to be refined, stylized, or transformed into something new and useful for wallpapers, prints, or other polished final outputs.