Alibaba Tongyi Lab Releases Z-Image-Base - Non-Distilled High-Quality Image Generation Model

On January 28, 2026, Alibaba Tongyi Lab officially released the Z-Image-Base model, the non-distilled raw checkpoint version of the Z-Image series. ComfyUI provided full support on the release day, achieving Day-0 integration.

Model Positioning

Z-Image-Base is the core foundation of the Z-Image model family. Unlike its distilled counterpart Z-Image-Turbo which prioritizes speed, the Base version preserves the full generative potential of the architecture. While it requires 30-50 sampling steps (CFG 3-5) for optimal quality and longer generation times, it produces significantly richer visual details and a higher artistic ceiling.

Key Features

Diverse Aesthetics

Z-Image-Base supports a broader range of artistic styles while maintaining exceptional photorealistic quality. Compared to the distilled version, the base model has clear advantages in artistic expressiveness and style diversity.

Fine-tuning Friendly

As a non-distilled complete model, Z-Image-Base is an ideal foundation for community fine-tuning and specialized development. Developers can train LoRAs based on this model, perform style transfer, or create customized image generation solutions.

Highly Responsive to Negative Prompts

The model is highly responsive to negative prompts, enabling precise generation control. Users can effectively avoid unwanted elements or styles through negative prompts.

Enhanced Diversity

Compared to the distilled version, the Base model has higher generation diversity, producing more creative and varied results, suitable for exploratory creative scenarios.

Performance

On NVIDIA RTX Pro 6000 Blackwell GPU, generating a 1024×1024 resolution image (30 sampling steps) takes only 13.3 seconds.

Using in ComfyUI

ComfyUI provided full support on the Z-Image-Base release day. Usage steps:

1. Update ComfyUI: Ensure you’re running the latest version of ComfyUI
2. Access Workflow Templates:
   • Click Templates in the sidebar
   • Go to Template library
   • Search for “Z-image” workflows

Official Workflow

You can download the official workflow template from:

• Z-Image Official Workflow

Model Files

Model files need to be downloaded and placed in corresponding directories:

Text Encoders (text_encoders)

• qwen_3_4b.safetensors

Diffusion Models (diffusion_models)

• z_image_base_bf16.safetensors

VAE

• ae.safetensors

File Placement

📂 ComfyUI/
├── 📂 models/
│   ├── 📂 text_encoders/
│   │      └── qwen_3_4b.safetensors
│   ├── 📂 diffusion_models/
│   │      └── z_image_base_bf16.safetensors
│   └── 📂 vae/
│          └── ae.safetensors

Recommended Parameters

• Sampling Steps: 30-50 steps
• CFG Scale: 3-5
• Resolution: 1024×1024 (recommended)

Application Scenarios

Z-Image-Base is particularly suitable for:

• Professional Photography-Grade Portraits: Fine skin textures, natural lighting effects
• Architecture & Interior Design: High-quality spatial rendering, material representation
• Artistic Creation: Diverse style exploration, creative experimentation
• Commercial Visual Design: Product photography, advertising material production
• Model Fine-tuning Base: LoRA training, style customization

Comparison with Z-Image-Turbo

Z-Image Series Ecosystem

The Z-Image series has formed a complete ecosystem:

• Z-Image-Base: Non-distilled foundation model, highest quality and flexibility
• Z-Image-Turbo: Distilled accelerated version, 8-step rapid generation
• Z-Image-Edit: Image editing specialized version (coming soon)
• ControlNet Union 2.1: Supports multiple control conditions (Canny, HED, Depth, Pose, MLSD)
• TwinFlow Accelerated Version: Experimental faster version (in development)

Related Links

• GitHub Repository: https://github.com/Tongyi-MAI/Z-Image
• HuggingFace Model: https://huggingface.co/Tongyi-MAI/Z-Image
• ModelScope Model: https://modelscope.cn/models/Tongyi-MAI/Z-Image
• ComfyUI Blog: https://blog.comfy.org/p/z-image-day-0-support-in-comfyui
• Official Website: https://z-image.ai