Z-Image Base ComfyUI Setup: Complete Installation and Workflow Guide

ComfyUI has become the preferred interface for serious AI image generation, offering node-based workflows that provide complete control over the generation process. Setting up Z-Image Base in ComfyUI unlocks the model's full potential with customizable pipelines, LoRA integration, and advanced features. This guide walks you through every step from installation to your first generation.

ComfyUI's learning curve is steeper than simpler interfaces, but the control and flexibility it provides make it worthwhile for anyone doing serious generative work.

Prerequisites

Before starting, ensure your system meets these requirements.

Hardware Requirements

Minimum:

• NVIDIA GPU with 12GB VRAM (RTX 3060 12GB)
• 32GB system RAM
• 50GB free storage
• CUDA-capable system

Recommended:

• NVIDIA GPU with 16-24GB VRAM (RTX 4070/4090)
• 64GB system RAM
• SSD for model storage
• Latest NVIDIA drivers

Software Requirements

• Python 3.10 or 3.11
• Git
• CUDA 11.8 or 12.x
• Windows, Linux, or macOS

Installing ComfyUI

Let's get ComfyUI running on your system.

Method 1: Direct Clone (Recommended)

## Clone repository
git clone https://github.com/comfyanonymous/ComfyUI
cd ComfyUI

## Create virtual environment
python -m venv venv
source venv/bin/activate  # Linux/Mac
## or: venv\Scripts\activate  # Windows

## Install PyTorch (adjust for your CUDA version)
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

## Install dependencies
pip install -r requirements.txt

Method 2: Portable (Windows)

For Windows users, portable packages exist:

1. Download from ComfyUI GitHub releases
2. Extract to desired location
3. Run run_nvidia_gpu.bat

Verifying Installation

Start ComfyUI to verify:

python main.py

Open http://127.0.0.1:8188 in your browser. You should see the ComfyUI interface.

ComfyUI interface after successful installation

Installing ComfyUI Manager

ComfyUI Manager simplifies node installation and updates.

Installation Steps

1. Navigate to custom_nodes folder:

cd custom_nodes
git clone https://github.com/ltdrdata/ComfyUI-Manager

2. Restart ComfyUI

3. Access Manager from the interface menu

Using Manager

With Manager installed:

• Install custom nodes from the menu
• Update existing nodes
• Search for specific functionality
• Manage dependencies

Downloading Z-Image Base

Get the model files and place them correctly.

Model Download

Download Z-Image Base from HuggingFace:

• Navigate to the model page
• Download the safetensors file (~12GB)
• Note the VAE if provided separately

File Placement

Place files in ComfyUI's model directories:

ComfyUI/
├── models/
│   ├── checkpoints/
│   │   └── z-image-base.safetensors  ← Main model
│   ├── vae/
│   │   └── z-image-vae.safetensors   ← If separate
│   └── loras/
│       └── (your loras here)

Basic Generation Workflow

Let's create a simple text-to-image workflow.

Essential Nodes

A minimal workflow requires:

1. Load Checkpoint - Loads Z-Image Base
2. CLIP Text Encode - Processes your prompt
3. Empty Latent Image - Creates the generation canvas
4. KSampler - Performs the diffusion process
5. VAE Decode - Converts latent to image
6. Save Image - Saves your output

Node Connections

Connect nodes in this pattern:

[Load Checkpoint]
  ├── MODEL → [KSampler] MODEL
  ├── CLIP → [CLIP Text Encode (Positive)] CLIP
  ├── CLIP → [CLIP Text Encode (Negative)] CLIP
  └── VAE → [VAE Decode] VAE

[CLIP Text Encode (Positive)]
  └── CONDITIONING → [KSampler] positive

[CLIP Text Encode (Negative)]
  └── CONDITIONING → [KSampler] negative

[Empty Latent Image]
  └── LATENT → [KSampler] latent_image

[KSampler]
  └── LATENT → [VAE Decode] samples

[VAE Decode]
  └── IMAGE → [Save Image] images

Recommended Settings

For Z-Image Base:

Your First Generation

1. Set your prompt in the positive CLIP node
2. Add quality negatives: "blurry, low quality, distorted"
3. Configure resolution in Empty Latent Image
4. Click "Queue Prompt"

Watch the generation progress and your first Z-Image Base output appear.

Advanced Workflows

Once basics work, explore advanced configurations.

Adding LoRA Support

Insert a LoRA loader between checkpoint and samplers:

[Load Checkpoint]
  └── MODEL → [Load LoRA] model

[Load LoRA]
  ├── MODEL → [KSampler] MODEL
  └── CLIP → [CLIP Text Encode] CLIP

Configure LoRA strength (typically 0.7-1.0 for full effect).

Image-to-Image

Replace Empty Latent Image with a loaded image:

[Load Image]
  └── IMAGE → [VAE Encode] pixels

[VAE Encode]
  └── LATENT → [KSampler] latent_image

Adjust denoise (0.3-0.7) to control how much changes.

Batch Generation

Generate multiple images efficiently:

[Empty Latent Image]
  batch_size: 4

This generates 4 images per queue, useful for exploring variations.

Complex workflows enable sophisticated generation pipelines

Optimization Tips

Get better performance from your setup.

Memory Management

For limited VRAM:

Enable attention optimizations:

• Install xformers: pip install xformers
• Add --use-pytorch-cross-attention to launch args
• Use --lowvram flag if needed

Quantization:

• fp16 models use half the VRAM
• Some quality trade-off

Speed Optimization

Faster generation:

• Reduce steps (20 instead of 30 for previews)
• Use smaller batch sizes if VRAM limited
• Enable CUDA graphs: --cuda-malloc

Caching:

• Model weights are cached after first load
• Subsequent generations start faster

Quality Optimization

Better outputs:

• Use full 30+ steps for final renders
• CFG 7-8 for balanced quality
• Higher resolution when VRAM allows
• Appropriate negative prompts

Troubleshooting

Common issues and their solutions.

Model Won't Load

Symptoms: Error messages when loading checkpoint

Solutions:

• Verify file placement in correct folder
• Check file isn't corrupted (re-download)
• Ensure sufficient VRAM
• Check ComfyUI console for specific errors

Out of Memory

Symptoms: CUDA out of memory errors

Solutions:

• Enable --lowvram mode
• Reduce resolution
• Close other GPU applications
• Use fp16 model if available

Slow Generation

Symptoms: Generation takes very long

Solutions:

• Check CUDA is being used (not CPU)
• Update GPU drivers
• Verify xformers installation
• Monitor GPU usage to diagnose

Black/Blank Outputs

Symptoms: Generated images are black or empty

Solutions:

• Check VAE is loaded correctly
• Verify node connections
• Try different seed
• Reduce CFG scale

Frequently Asked Questions

Does Z-Image Base work with A1111?

This guide covers ComfyUI. A1111 may have limited support depending on extension availability.

Which Python version should I use?

Python 3.10 or 3.11 are recommended. 3.12 may have compatibility issues.

Can I run ComfyUI without a GPU?

Technically yes with CPU mode, but generation will be extremely slow (minutes to hours per image).

How do I update ComfyUI?

Run git pull in the ComfyUI directory, then update dependencies if needed.

Where do I find community workflows?

Check ComfyUI subreddit, Civitai workflows section, and GitHub repositories.

Can I use multiple LoRAs?

Yes, chain multiple Load LoRA nodes together.

How do I save my workflow?

Use the Save button in ComfyUI to export as JSON. Load button imports saved workflows.

Why is my first generation slow?

Model loading takes time on first generation. Subsequent generations are faster.

How do I use different aspect ratios?

Change width and height in Empty Latent Image node.

Can I run multiple workflows at once?

ComfyUI queues workflows. Multiple can be queued but execute sequentially.

ComfyUI with Z-Image Base provides a powerful local generation setup that gives you complete control over your creative process. The initial setup effort pays off in flexibility and capability.

For users who prefer simpler interfaces without local setup, Apatero offers instant access to Z-Image models alongside 50+ other options, with LoRA training available on Pro plans.