even though the download section show fp8. it is int8 since civitai site doesnt have int8 option to choose. regardless the file is int8.
this model is tuned by https://huggingface.co/AX1Y2JP/Krea-2-Turbo-INT8-ConvRot/blob/main/krea-2-turbo-int8-convrot-aggressive.safetensors all credit goes to him https://huggingface.co/AX1Y2JP. if he wants me to remove it. please do tell i will remove it. i just wanted the int8 version of krea2t model to be uploaded here
Description
use this custom node https://github.com/BobJohnson24/ComfyUI-INT8-Fast
FAQ
Comments (11)
What are all these differences? So much getting released recently and I'm getting confused
From what ive read on convrot stuff, the quality is higher and its faster.
70% faster than fp8 on old hardware, use comfyui int8 fast node
INT8 is the new future for image gen i guess. it's higher quality plus faster gen speed on old to new gpu. use less vram. for now we have to use comfyui-int8-fast custom node. comfyui did get support for it natively but its not optimized as of now. use the custom node and test you'll see the speed and quality diff
I update comfyui to 0.27. it still doesn't work.
# ComfyUI Error Report ## Error Details - Node ID: 91 - Node Type: UNETLoader - Exception Type: KeyError - Exception Message: KeyError: 'int8_rowwise'
you need this custom node https://github.com/BobJohnson24/ComfyUI-INT8-Fast
@AI_Aurora unet_name krea2TurboINT8Convrot_v10.safetensors
weight_dtype default
model_type krea2
on_the_fly_quantization false
enable_convrot true
lora_mode Stochastic
it stops at Ksamlpler
@ironman5 lora mode default
@AI_Aurora sorry , I've tried all 3 type :None,Stochasitc,Dynamic, all stop at ksampler.
@ironman5 as of new comfyui update default load unet node also works nicely. or just copy the workflow from the image posted here
He clearly states that you must use a custom node to put it in the node folder, then search for int8 and a dedicated loader will appear, provided it must be 0.27.0