A simple upscale workflow for Qwen-Image. The idea behind it is basic: upscale the image to twice its size with UltraSharpV2, add some noise, do a second pass with Qwen. A second more adapted model could be used for the upscale, but I didn't bother as I prefer to use a single model for my workflows.
Description
Replaced the group node by the fancy subgraph feature
FAQ
Comments (4)
What do you mean exactly with "add some noise" in the second step? Newbie here.
Usually, you run your sampler with 1.00 denoise, meaning you take pure noise to convert it to an image. Here, I use the un upscaled image as the input latent and I run it with 0.35 denoise, meaning Comfy makes the image a little noisy, but most of the structure of the image is preserved.
@LususNaturae Thanks for the reply. So i run your workflow once, and i add the generated image to the workflow again setting the noise higher, like 0.75 and run again?
@tahin47757646 the more noise you add, the more the images changes from the original input.
Images are created from noise, so if you want to just upscale and image you want to add little noise so the original image persists.