Player not loading? Watch on YouTube
This tutorial explains the creator's ComfyUI Ultimate Upscaler workflow for adding generated detail while enlarging images. It covers SDXL and FLUX variants. The speaker prefers SDXL for its smaller footprint, faster processing and lower VRAM use. On their RTX 3090, they report that FLUX consumes more memory and runs slower, though it sometimes produces better results.
The workflow divides an image into tiles, processes each area and assembles the output. The initial example uses a 1024×1024 source and a 2048×2048 target, with 512×512 tiles. A planner handles image proportions and tile sizing. Padding and masks help control visible seams; the speaker recommends increasing padding when lines remain. Higher denoise can add detail, but it also makes the result deviate further from the source.
Florence supplies captions for individual tile prompts. The walkthrough also covers checkpoint selection, sampler settings, sharpening and output previews. A later 3× example produces a 3072×3072 image. The speaker warns that large magnification values can exceed VRAM capacity and slow processing when the system uses regular RAM.
