Z-Image Turbo
ComfyUI in the Cloud
Deploy Z-Image Turbo, from
$0.60 / hour
Z-Image Turbo ComfyUI Workflow
A ComfyUI image generation workflow powered by Z-Image Turbo, featuring a BF16 diffusion model, Qwen 3 4B text conditioning, HiDream-I1 autoencoder decoding, and efficient 8-step image generation.
Workflow Overview
This workflow utilizes the Z-Image Turbo model for AI image generation within ComfyUI. It combines a BF16 diffusion model, Qwen 3 4B text conditioning, and a HiDream-I1 autoencoder to support image generation from natural-language instructions.
Z-Image Turbo is designed for efficient image generation using a reduced number of sampling steps. The displayed workflow is configured with 8 steps, making it suitable for fast image generation workflows where inference efficiency is important.
The workflow uses a resolution of 1024 × 1024 pixels and a configurable noise seed. The selected diffusion model uses BF16 precision, while the Qwen 3 4B model provides text conditioning to guide the image generation process.
Generation Settings
Width
1024
Height
1024
Sampling Steps
8
Seed
0
Diffusion Model
Z-Image Turbo
Model Precision
BF16
Text Encoder
Qwen 3 4B
Autoencoder
HiDream-I1 AE
Model Configuration
Image Generation Model (UNet)
BF16z_image_turbo_bf16
The Z-Image Turbo diffusion model serves as the primary image generation component responsible for the denoising process. The model uses conditioning information derived from text instructions to guide the generation of images. The BF16 configuration uses 16-bit floating-point model weights, balancing numerical precision and memory requirements. Z-Image Turbo is designed for fast image generation using a relatively small number of sampling steps.
Text Conditioning Model
Qwen 3 4Bqwen_3_4b
The Qwen 3 4B model provides text conditioning for the Z-Image Turbo image generation process. It processes natural-language prompts and converts them into conditioning representations that guide the diffusion model. This component helps the workflow interpret descriptive instructions, including subjects, visual styles, lighting, composition, and other image characteristics. The text encoder plays an important role in connecting user instructions with the image generation model.
Variational Autoencoder (VAE)
Autoencoderae
The workflow selects the HiDream-I1 autoencoder model identified as ae. The autoencoder is responsible for converting latent image representations into visible images during the decoding stage. After the diffusion model completes the generation process, the autoencoder decodes the resulting latent representation into pixel-space output. The decoded image can then be used for preview, saving, or additional image processing operations within ComfyUI.
Sampling Configuration
The workflow uses a sampling step count of 8. Sampling steps control the number of iterative updates performed during the image generation process. Z-Image Turbo is designed to work efficiently with a low number of sampling steps.
| Sampling Steps | 8 |
| Diffusion Model | Z-Image Turbo |
| Model Precision | BF16 |
| Sampler | Not visible in the subgraph |
| Scheduler | Not visible in the subgraph |
The reduced step count is one of the main characteristics of the Turbo configuration. The exact sampler, scheduler, and guidance settings depend on the internal nodes connected within the complete ComfyUI workflow.
Image Dimension Configuration
The workflow includes dedicated width and height controls for configuring image resolution. Both values are set to 1024 pixels in the displayed configuration.
Image Width
1024 px
Controls the horizontal dimension of the generated image.
Image Height
1024 px
Controls the vertical dimension of the generated image.
With both dimensions configured to 1024 pixels, the workflow is set up for square image generation with a 1:1 aspect ratio. These values can be adjusted to produce images with different resolutions and aspect ratios, subject to the model and workflow requirements.
Noise Seed Configuration
The workflow includes a configurable noise seed that controls the random initialization used during image generation. The displayed seed value is set to 0.
| Seed | 0 |
| Seed Control | Configurable |
| Randomization Control | Available |
A fixed seed can help reproduce generation results when the prompt, model, dimensions, and other generation parameters remain unchanged. Changing the seed modifies the random initialization and can produce different image variations.
Additional Parameters
| Width | 1024 |
| Height | 1024 |
| Resolution | 1024 × 1024 |
| Aspect Ratio | 1:1 |
| Seed | 0 |
| Steps | 8 |
| Diffusion Model Precision | BF16 |
| Autoencoder Model | ae |
Surrounding Nodes and Workflow Integration
The Z-Image Turbo subgraph combines model selection, image dimension controls, sampling step configuration, and noise seed settings into a reusable generation component.
In a typical ComfyUI text-to-image workflow, a text prompt is supplied through an input or text conditioning node. The Qwen 3 4B text encoder processes the prompt and provides conditioning information for the Z-Image Turbo diffusion model.
The diffusion model performs the image generation process using the configured sampling steps and noise initialization. The width and height values determine the configured image dimensions. The seed controls the random initialization used by the generation process.
After the denoising process, the selected autoencoder decodes the generated latent representation into a visible image. The resulting image can then be passed to preview, image-saving, or additional image processing nodes.
The visible screenshot shows the model selections and generation parameters. Exact prompt connections, sampler configuration, internal node structure, and output connections depend on the complete ComfyUI graph.
Workflow Highlights
- ✓ Z-Image Turbo: Uses the Z-Image Turbo diffusion model for efficient AI image generation.
- ✓ BF16 Precision: Uses a BF16 diffusion model configuration for image generation.
- ✓ Qwen 3 4B Conditioning: Integrates the Qwen 3 4B text encoder for processing natural-language prompts.
- ✓ 8-Step Generation: Uses 8 sampling steps for efficient image generation.
- ✓ 1024 × 1024 Resolution: Configured for square image generation with a 1:1 aspect ratio.
- ✓ HiDream-I1 Autoencoder: Includes the selected ae model for decoding latent representations into images.
- ✓ Seed-Based Generation: Supports configurable noise initialization with the displayed seed set to 0.
- ✓ Reusable Subgraph: Consolidates model configuration, generation settings, and image dimension controls within a single ComfyUI component.
Workflow Summary
This ComfyUI workflow combines the Z-Image Turbo BF16 diffusion model with Qwen 3 4B text conditioning and the HiDream-I1 autoencoder. It is configured with 8 sampling steps, a resolution of 1024 × 1024 pixels, and a noise seed value of 0.
The Qwen 3 4B text encoder provides conditioning information, while the Z-Image Turbo diffusion model performs the image generation process. The selected autoencoder converts generated latent representations into visible images.
The workflow is designed around efficient low-step image generation. Its subgraph structure organizes diffusion model selection, text conditioning, image dimensions, sampling steps, and noise seed configuration for integration into larger ComfyUI workflows.