Zib distilled 4-8 steps, working good with 4 steps, but 8 steps is better.
Merged with one of my lora and 3 blocks of the turbo version.
Use euler/simple with cfg 1.
Distill loras: https://huggingface.co/alibaba-pai/Z-Image-Fun-Lora-Distill/tree/main
nvfp4 version converted with my node:
https://github.com/tritant/ComfyUI_Kitchen_nvfp4_Converter
Description
Zib-CreArt-Ultimate-distilled-nvfp4
FAQ
Comments (40)
Any chance of getting the fp32 diffusors version to allow lora training directly?
with ai-toolkit, you can train directly with the original base version, this is the best way for training lora
@jice some of my base trained loras do not look right with this model though, which is why i thought to ask about training against the model itself instead. No biggie
Anyone else get a garbled mess with the nvfp4 ?
For your information, nvfp4 is only usable with comfyui and comfyui-kitchen installed, perhaps forge neo will add support soon.
https://github.com/Comfy-Org/comfy-kitchen
@jice ooh I know I have comfyui kitchen installed even the node to make nvfp4, but neither yours or mine works as nvfp4
@TiwazM There is no special node to use nvfp4, you must use the load diffusion model node from the comfyui core. comfyui version? error message?
@jice no errors, comfyui 0.13, output looks like this https://imgur.com/a/nRu49oc
Found comfy_kitchen backend triton: {'available': True, 'disabled': True, 'unavailable_reason': None, 'capabilities': ['apply_rope', 'apply_rope1', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8']}
Found comfy_kitchen backend cuda: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['apply_rope', 'apply_rope1', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'scaled_mm_nvfp4']}
Found comfy_kitchen backend eager: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['apply_rope', 'apply_rope1', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'scaled_mm_nvfp4']}
@TiwazM I don't know why, check your graphics card drivers and Torch version; I'm using PyTorch 2.10+CU130.
@jice well its not pytorch it seem, updated from 2.9.1 to 2.10 with cu130, still nope :(
@TiwazM diffusers wheel version?
@jice I can check, what version are you using ? sure seem picky when nvfp4s that are made with comfy kitchen.
@TiwazM diffusers-0.36.0, try this too "pip install comfy-kitchen[cublas]" with cublas, otherwise, I'm out of ideas, maybe open an issue on the comfyui-kitchen GitHub repository.
I got same issue.
Oddly the Z-turbo-CreArt-UltimateV2-nvfp4 is working fine for me.
Eventually,i used dual ksamplers workflow:
ZIB-CreArt_Ultimate_distilled-nvfp4 (3 steps)
Z-turbo-CreArt-UltimateV2-nvfp4 (9 steps)
The result is fine.
@jice I got 0.36 and the cublas so I don't know I am at a loss, and tested on 2 different machines same issue
@TiwazM try the comfyui version ?
https://huggingface.co/Comfy-Org/z_image_turbo/tree/main/split_files/diffusion_models
@jice yes that works, flux 2 and flux 2 klein does too
@TiwazM Very strange that work for me and all of my friend, my turbo version nvfp4 work ?
https://civitai.com/models/2243087?modelVersionId=2614875
@jice yes fine too, its just your base or if I try to turn my zimage base into a nvfp4
@TiwazM What kind of graphics card do you have?
@jice one is a 5090 one a 5060ti works on neither
@TiwazM Okay, I have a 4060 Ti. Have you ever tested FP8 versions? The only difference between the turbo version and the base NVFP4 version is that I left some layers in FP8. This could be due to your version of Triton, if you're not handling FP8 correctly.
@TiwazM I just updated the converter, can you test it again?
@jice aah that explains it so it might just not work on blackwell cards. I will test the new converter when I get home an let you know if that helps.
@jice also blackwell has native nvfp4 but the 40 series does not that could be the big difference too. if you look at nunchaku nvfp4 works only on blackwell, the rest uses int4.
@TiwazM Yes, comfyui, check the card type. If it's a 4000, it upcasts to FP16; if it's a 5000, it uses the 5000's hardware. The 4000 works, but doesn't benefit from the acceleration. Apparently, the 5000 doesn't like the NVFP4/FP8 mix, so I left the FP8 layers in BF16.
Same on my side. Latest comfy + latest pytorch + latest 5090 drivers ... still no fp4 nowhere. Tried both your models + default comfy fp4 z image.
Any tips?
@PixelGenius I think it’s something with zimage base, I used his node to make an nvfp4 of a flux 2 klein 9B and that worked fine.
@TiwazM retry the node for convert zib normaly it work now
@PixelGenius comfyui-kitchen installed?
@jice zit works, zib not it seems
@TiwazM You upgraded the converter ?
@jice yes I checked git pull tells me I got the latest version
@TiwazM can you test this one ? if working, i copy this structure
https://huggingface.co/GuangyuanSD/Z-Image-Distilled/blob/4aab81c1d63abe5e346f63f79d83ec8c3f7566c8/RedZDX-v3-ZIB-Distilled-Lucis-NVFP4-mixed(quality).safetensors
@jice yes that one works
@TiwazM update the converter and retry, normaly its working now