GGUF conversion of the model Sick Ollie by @sickollie - All credits belong to the the original uploader
Usage recommendations:
GGUF loader nodes: https://github.com/molbal/ComfyUI-GGUF
GGUF Quant: Q8_CR
Resolution: 1440x1920
Sampler - Euler
Scheduler - Beta
CFG - 1
Steps - 9
Description
FAQ
Comments (6)
The consistency has been well maintained; hopefully, a Q6 quantitative version will be released to benefit 8GB users.
Hi, you can make a Q5_1 version yourself following the guide: https://civitai.red/articles/33482/how-to-quantize-models-to-gguf - if you have the gguf loader nodes installed, then you are all set - unfortunately I need to work to enable Q6 support, right now the formats which work well are listed in the article.
You CAN load a Q6 which you quantized somewhere else, but it will be slower to run
Tried this on an older 8GB VRAM model (RTX2070)
It actually seemed slower than the original FP8 checkpoint?
Perhaps if you try the Unet loader (Dynamic VRAM) node it would be faster. I have a 3080 laptop with also 8GB VRAM and that helps for me (Dynamic VRAM + the Q8_CR quant OR just a plain old int8 convrot checkpoint and not a GGUF also works well)
Thanks for the fast reply. I guess the node is just for ComfyUI?
I use Forge NEO.
@LurkingAI Oh, I do not use Forge NEO, so unfortunately I am not much of a help with that :(