
I tried GGUF before, but it messed with the skin a lot. FP8 didn’t work for me either because it added those tiling grids to the image.
Since the models get downloaded directly through the workflow, you don’t have to grab anything manually. Just be aware that the first image will take a bit longer.
I'm just using the standard SeedVR2 workflow here, nothing fancy. I only added an extra node so I can upscale multiple images in a row.
The base image was generated with Z-Image, and I'm running this on a 5090, so I can’t say how well it performs on other GPUs. For me, it takes about 38 seconds to upscale an image.
Here’s the workflow:
Test image:
https://imgur.com/a/test-image-JZxyeGd
Custom nodes:
for the vram cache nodes (It doesn't need to be installed, but I would recommend it, especially if you work in batches)
https://github.com/yolain/ComfyUI-Easy-Use.git
Seedvr2 Nodes
https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler.git
For the "imagelist_from_dir" node
when an upscaler is so good it feels illegal
byu/Ok-Page5607 inStableDiffusion
