Z-Image logoZ-Image
Developer Resources

Run Z-Image Locally

The definitive resource for GGUF/FP8 models, ComfyUI workflows, and LoRA training insights.

Model Repository

Z-Image Turbo
ReleasedInference

Z-Image Turbo

8-step distilled model. The fastest inference speed suitable for real-time generation.

Original (Safetensors)VRAM: 16GB+ • Best Quality
FP8 (T5B)VRAM: • Balanced
FP8 (drbaph)VRAM: • Alternative
GGUF (Q4/Q8)VRAM: • Low VRAM
Z-Image Base
Coming SoonTraining

Z-Image Base

Non-distilled foundation model. Requires more steps (20+) but ideal for Fine-tuning.

Links will appear here upon release
Z-Image Edit
Coming SoonEditing

Z-Image Edit

A variant fine-tuned on Z-Image specifically for image editing tasks. Allowing for instruction-based editing.

Links will appear here upon release

ComfyUI Workflow

Step-by-step setup to prevent errors.

Step 1: Environment Prep

If you're a new user, install ComfyUI. If you're already using ComfyUI, update it via the Manager by selecting Update All.

Install ComfyUI Update All Nodes (Manager)

Step 2: Download Essentials

Place these files in the correct folders. Turbo needs a specific text encoder.

qwen_3_4b.safetensors
Download
z_image_turbo_bf16.safetensors
Download
ae.safetensors
Download

Step 3: Load Workflow

Don't build from scratch. Use the community verified JSONs.

Download Turbo JSON

LoRA Ecosystem & Training

Distillation Warning

Z-Image Turbo is a distilled model. Distillation compresses the latent space, making it significantly harder to train LoRAs compared to the Base model.

Recommended Tools

Ostris AI-Toolkit
@ostrisai

The bleeding-edge tool for training flux/z-image architecture. Updated recently to support Z-Image.

View Repository

Strategy Roadmap

Now
Inference Only
Use Turbo for generating images. Experiment with Ostris tool for style transfer only.
Next
Wait for Base
For character/object training, the Z-Image Base model is required for stability.

Community Tutorials

Official ComfyUI DocsLow Vram Workflow(GGUF)Deploy on runpod