ComfyUI: basic upscaling vs Ultimate SD Upscale for 4K

This tutorial compares two ways to enlarge images in ComfyUI: a lightweight workflow with an upscale model and a tiled diffusion workflow using Ultimate SD Upscale. The presenter prefers the second method for finer detail, while describing the simpler option as useful for quick exports with less processing.

The basic workflow connects Load Image, Upscale Image by Model and Save Image, with a separate loader supplying the upscale model. Ultimate SD Upscale combines that initial enlargement with diffusion refinement. It processes smaller tiles rather than the whole image at once to reduce VRAM demands, though the presenter says it takes much longer.

The walkthrough explains the model, text conditioning and VAE connections, then covers enlargement scale, sampling settings and tile dimensions. It uses Z Image Turbo as its diffusion model. The presenter recommends leaving the positive prompt empty or describing general style, since naming an object can encourage the model to repeat it in each tile.

Denoise is the main caution. The suggested starting value is 0.2, with 0.15 to 0.25 described as a range for subtle refinement. The presenter warns that values of 0.4 and above can cause neighboring tiles to diverge and expose seams. He usually uses 1024-pixel square tiles and recommends smaller tiles when memory is limited. His experience with the seam-fix option is mixed.