I have been playing with Z Image Turbo inside ComfyUI for a while now.
What I like is simple. You get realistic images in less than a second, the graph is light, and it still works on low VRAM cards.
Z Image itself is a 6B parameter image model that was built for speed. The ComfyUI example page explains that it is a distilled diffusion transformer made for sub second inference.
On top of that we now have Z Image Turbo, a faster distilled version that plugs straight into ComfyUI with a small set of files from the official Comfy Org repository on Hugging Face.
So I built one workflow around it.
It does four things.
- Text to image with two samplers
- Auto prompt from Qwen 3 4B
- Image to prompt for product style images
- Optional upscale at the end
You do not need to write long prompts every time. You can give one short idea and let the graph write a better Z Image Turbo prompt for you, or you can drop in a reference image and ask it to follow that style.
File You Need To Download
Let us start with the files. If you already run other ComfyUI workflows on your PC this part will feel familiar.
All paths are inside your main ComfyUI folder.
1. Z Image Turbo diffusion models
You have three main options right now.
BF16 model
z_image_turbo_bf16.safetensors- Size is around 12 GB on disk
- Best if you have 16 GB or more VRAM
Save it here
ComfyUI/models/diffusion_models/
FP8 models
The community also uses FP8 builds when they want a lighter load. In most tests BF16 gives slightly higher quality, FP8 gives more headroom on VRAM.
If you see a file name with fp8 in the Z Image Turbo tree, save it in the same diffusion models folder.
GGUF model for very low VRAM
If your GPU is not strong, you can use the Q4 K M GGUF model.
- Name is similar to
Z-Image-Turbo-Q4_K_M.gguf - Works on 2 GB to 3 GB VRAM, especially if you also use CPU offload
Save it in the folder where your GGUF loader node expects models, for example
ComfyUI/models/diffusion_models/
In this workflow I treat BF16 as the main option and GGUF as the fallback when users write on Reddit that they only have a laptop GPU or a small cloud instance.
2. Text encoder, Qwen 3 4B
Z Image Turbo expects a Qwen text encoder.
Save it here
ComfyUI/models/text_encoders/
In my graph this file does double duty.
It runs the normal text encode for prompts, and it also works inside the auto prompt system where I use a smaller Qwen model to expand short text into a full Z Image Turbo prompt, similar to how I already use Qwen in my image edit workflows.
3. VAE file
For the VAE I am using the same file that I use in my Flux workflows.
Save it here
ComfyUI/models/vae/
If you already have Flux set up, you probably already have this file in place.
4. Color Z Image Turbo film LoRA
For style I use one LoRA that was trained for Z Image Turbo.
Save the LoRA here
ComfyUI/models/loras/
This LoRA does not replace your subject. It mainly pushes the color grading, contrast, and film grain. It gives a movie frame feeling without changing identity.
Workflow Download