UPDATE YOUR COMFYUI AND INSTALL MISSING CUSTOM NODES
(V2 remaps nodes to a cleaner setup, removed prompt cycler as its going to be a paid feature on my discord from now on so if you got it already, congrats.)
(V1 is the fp8 version of the NEW Z-Image-Turbo image model! (really a quant but the model is so small it doesnt matter))
Base workflow designed for EXTREMELY FAST, high quality generations.
Generation Times are 30 seconds or less.
MY PC SPECS:
8GB VRAM (RTX 3070)
16GB RAM
i-7
LINKS TO FILES:
Text Encoder
https://huggingface.co/Comfy-Org/z_image_turbo/tree/main/split_files/text_encoders
VAE
https://huggingface.co/Comfy-Org/z_image_turbo/tree/main/split_files/vae
GGUF
https://huggingface.co/unsloth/Z-Image-Turbo-GGUF/tree/main
fp8 Model
https://huggingface.co/T5B/Z-Image-Turbo-FP8/tree/main
bf16 model (v2)
Yes, its uncensored as well.
Description
LOW VRAM WORKFLOW<3
FAQ
Comments (14)
Fast and light and so GREAT. This is groundbreaking
mindblowing isnt it!?
Doing God works here
great work, any idea how we can use it on apple silicon macs?
im sorry but i know absolutely nothing about macs. :(
Get the workflow, eliminate the prompt part on the left and leave the groups with the loaders, clip text encoders and sampler, vae decode and latent image and save image. Mac MPS does not like fp8, so convert the fp8 to fp16 (Copilot can help you with the python instructions to do it in 5 minutes) and in the load diffusion model use your new fp16 model and change weight_dtype to default. Make sure to use the vae and clip models kindly included in the links of this post article. It works under 30 seconds in M3 Max pro with the example in the workflow. Option B: use the bf16 model also linked in the article without having to convert from fp8 to fp16 yourself although it uses more memory.. Option 3: use a quantized model from huggingface (Q5 k-m is the one I am using) with a gguf loader. In quality bf16 will give you the best from what I have seen. Good luck
@javir dude your actually a legend for this. Ill get to work when i get home and get right on it 💪 am i able to create a workflow with the framework to provide it FOR people or is this a mac only functionality thing?
@realrebelai I don't see why it should not work on a PC as well as in a Mac. I will try it on an Nvidia 4090 once I am done digesting the turkey. Thank you for your fantastic post and all the links, it saved a lot of time.
@javir i had meant for me to maybe get a workflow for the mac users who struggle with conversions. i could do it for them if thats possible to do! your very welcome tho, happy i could help! thats what im here for! :)
@realrebelai got it, here it is https://civitai.com/models/2172300?modelVersionId=2446238 with credit to your useful information and workflow. Once there is a new comfyui desktop update for mac, I am sure other workflows can be used that have more power. Thank you
@javir nah the credits all yours bro! great job, this is what its all about. just helping the community because you can!
Best workflow. unfortunately no upscaler..
use my seedvr workflow!
USE MY SEEDVR2 GGUF WORKFLOW TO UPSCALE THE IMAGES AFTER GENERATING!