My first attempt at adding more creativity to Z-image-turbo
Added fp8 version
If you want to convert your masterpieces to nvfp4, you can use my custom node.
https://github.com/tritant/ComfyUI_Kitchen_nvfp4_Converter
Description
Nvfp4 version, 40% faster on rtx 5000 with cu130
FAQ
Comments (21)
Thanks for your great work, REALLY GREAT MODEL!!!! BTW, how fp4 format made, which util/framework you have used? WOuld you provide simple explanation?
I used comfyui kitchen, this "library"
https://github.com/Comfy-Org/comfy-kitchen/blob/main/comfy_kitchen/tensor/nvfp4.py
@ding214 My node for convert https://github.com/tritant/ComfyUI_ZimageTurbo_nvfp4_Converter
i have a node keep missing in the workflow: CacheDitControl
JWIntegerin subgraph 'Z Image Turbo Inference'
JWStringin subgraph 'Z Image Turbo Inference'
any tip?
Just remove it or use manager for install missing node
how i can use nvfp4 model?
In workflow from images i got error "nvfp4"
You need the last comfyui and torch2.9.0+cu130, it working with the core load diffusion model of comfyui
this is the first nvfp4 model for z i see. Very nice. it is so fast. I tested it with 20 steps. The speed increase is more than double with NVFP4. I hope there coming more and more models with nvfp4.
@Popelt this is a turbo version, it work better with 9 steps
lol 8 steps enough
Hey please in your workflow, the cacheditcontrol doesnt work and i installed the missing nodes? what is that ?
The cache claims to speed up inference by 2x, but there is a lot of quality loss, i use it for my tests, you can remove it.
Title says FP16, but the model name under it says BF16. Which of the two is it?
bf16 i change the title thx
Is there any chance to user that with Forgeneo ?
If you're talking about nvfp4, I think for now only comfyui can use it (maybe swarmui), you'll have to ask the Forge Neo developers to implement it. Otherwise, z-image is supported by Forge Neo.