Qwen Image 2.1
Qwen
Explore modelQwen / image model
Z-Image Turbo
Explore Z-Image Turbo, its capabilities, use cases and ComfyUI workflows for image generation on HexGrid Cloud.
Deploy Z-Image Turbo with ComfyUI
Choose a workflow to explore its setup, inputs, downloads and generation examples.
About Z-Image Turbo
Z-Image-Turbo is Alibaba's open-weight, high-speed image generation model developed by the Tongyi-MAI team, deployable on HexGrid.cloud with dedicated GPU infrastructure, configurable inference workflows, custom checkpoints and optimized model configurations.
It specializes in text-to-image generation, producing photorealistic images, illustrations, cinematic compositions, product visuals and typography-heavy graphics from natural-language prompts. The model offers strong bilingual understanding and text rendering in English and Chinese, making it suitable for creative production, advertising and rapid visual experimentation.
Z-Image-Turbo uses a 6-billion-parameter Scalable Single-Stream Diffusion Transformer (S3-DiT) architecture. It generates images using just 8 denoising evaluations through Decoupled Distribution Matching Distillation and reinforcement-learning post-training. The model supports 1024×1024 generation, configurable image dimensions and consumer-GPU deployment with approximately 16 GB of VRAM, while achieving sub-second inference on high-performance H800 GPUs.
What people use it for
Photorealistic text-to-image
Generate detailed, realistic images from natural-language prompts, including portraits, landscapes, architecture, lifestyle photography and cinematic scenes. Control lighting, camera perspective, textures, materials and visual atmosphere through descriptive instructions.
Advertising and product visuals
Create promotional images, product photography, campaign concepts, marketing banners and branded graphics. Generate realistic products, attractive compositions and visually engaging advertising materials without traditional photoshoots.
Typography and bilingual graphics
Generate posters, signage, advertisements, promotional graphics and editorial visuals containing English or Chinese text. Combine readable typography with detailed imagery and controlled visual layouts.
Cinematic and artistic imagery
Create cinematic stills, digital illustrations, concept artwork and stylized compositions using detailed prompts. Experiment with dramatic lighting, atmospheric effects, photographic techniques and artistic aesthetics.
Architecture and interior visualization
Generate architectural concepts, interior designs, realistic environments and material studies. Explore different lighting conditions, furniture arrangements, design styles and visual perspectives through text prompts.
Rapid creative prototyping
Quickly explore visual ideas, product concepts, compositions and artistic directions using the model's accelerated 8-step generation. Produce multiple iterations with configurable image dimensions and reproducible random seeds.
High-volume image generation
Generate large batches of visual content, including catalog imagery, social media assets, thumbnails, editorial illustrations and promotional creatives, using computationally efficient inference workflows.
Strengths
- High-quality text-to-image generation using a compact 6-billion-parameter model
- Scalable Single-Stream Diffusion Transformer architecture optimized for efficient image generation
- Distilled inference requiring only 8 denoising evaluations per image
- Sub-second inference demonstrated on enterprise-grade H800 GPUs
- Consumer-GPU compatibility with approximately 16 GB of VRAM
- Strong photorealistic rendering with detailed textures, realistic lighting and natural-looking subjects
- Accurate bilingual text rendering in English and Chinese
- Strong natural-language instruction following for complex visual descriptions
- Cinematic image quality with detailed control over lighting, composition, perspective and visual atmosphere
- High-quality portrait generation with realistic facial features, skin textures and photographic detail
- Support for artistic styles including illustration, painting, anime, photography and digital artwork
- Unified single-stream processing of text, visual semantic information and image latent tokens
- Qwen3-based text encoding for understanding detailed descriptions and visual relationships
- Decoupled Distribution Matching Distillation enabling high-quality few-step generation
- Reinforcement-learning post-training for improved semantic alignment, aesthetics and structural coherence
- Guidance-distilled inference without requiring conventional classifier-free guidance
- Support for square, landscape, portrait and custom image dimensions
- Configurable image width, height, random seeds and batch generation
- Optional prompt enhancement for expanding short descriptions into detailed visual instructions
- Efficient BF16 inference with support for optimized attention implementations
- Optional Flash Attention, model compilation and CPU offloading for inference optimization
- Support for quantized checkpoints and alternative optimized inference runtimes
- Open-weight architecture supporting private deployment and custom inference pipelines
- Apache 2.0 license permitting commercial use, modification and redistribution under its terms
Limitations
- Z-Image-Turbo is primarily a text-to-image generation model. Native instruction-based image editing, inpainting and multi-reference image composition are not supported by the standard Turbo checkpoint.
- Image editing requires a separate model such as Z-Image-Edit or an additional image-editing workflow.
- The distilled checkpoint is optimized for 8 denoising evaluations. Increasing the number of steps does not necessarily improve quality and may produce unexpected results.
- The standard Turbo configuration uses guidance_scale 0.0. Conventional classifier-free guidance is not supported in the same way as the undistilled Z-Image foundation model.
- Negative prompting is not effective as a standard control in the guidance-distilled Turbo pipeline. Strong negative-prompt control is better supported by the foundation Z-Image model.
- The Turbo checkpoint prioritizes image quality and inference speed over output diversity and fine-tuning flexibility. Repeated prompts may produce less variation than the foundation model.
- The official model comparison does not recommend Turbo as the primary checkpoint for fine-tuning. Custom training is better suited to the undistilled Z-Image foundation model.
- The model generates static images only. Native video generation, animation, audio synthesis and synchronized audiovisual output are not supported.
- The standard inference example uses 1024×1024 resolution. Larger images increase memory consumption and inference time, and quality may vary outside recommended generation configurations.
- Although bilingual text rendering is strong, long paragraphs, small typography, mathematical notation and complex multi-column layouts may still contain errors.
- Complex compositions involving multiple interacting subjects, precise spatial relationships or exact object counts may produce inconsistent visual results.
- Consistent reproduction of specific characters, products or visual identities across generations requires additional conditioning, specialized fine-tuning or compatible custom workflows.
- Transparent RGBA generation is not a dedicated native capability of the standard Turbo checkpoint. Transparent assets require additional processing or compatible background-removal workflows.
- Sub-second generation depends on high-performance hardware and optimized inference configurations. Consumer GPUs may require substantially longer inference times.
- Full-precision inference requires significant GPU memory, particularly when generating high-resolution images or processing multiple requests simultaneously.
- Prompt enhancement requires a separate language model or preprocessing component and introduces additional computational overhead.
Specifications
Generation, input and serving details.
ProviderQwenFamilyZ-ImageCategoryimageWorkflows1
Generation
- Modes
- Text-to-image
- Model
- Z-Image-Turbo
- Model family
- Z-Image
- Architecture
- Scalable Single-Stream Diffusion Transformer (S3-DiT)
- Model parameters
- 6 billion
- Text encoder
- Qwen3-based 4B language model
- Resolution
- 1024×1024 standard generation; configurable dimensions for larger or smaller outputs
- Default resolution
- 1024×1024 in the official reference example
- Aspect ratios
- Square, landscape, portrait and custom aspect ratios
- Image dimensions
- Configurable width and height, subject to memory and model constraints
- Output formats
- PNG; JPEG and WebP through compatible output processing
- Inference steps
- 8 denoising evaluations
- Diffusers step configuration
- num_inference_steps 9 in the official example, corresponding to 8 transformer forward evaluations
- Guidance scale
- 0.0 recommended for Turbo
- Generation framework
- Flow-matching diffusion with few-step distillation
- Distillation
- Decoupled Distribution Matching Distillation (Decoupled-DMD)
- Post-training
- Distribution Matching Distillation combined with reinforcement learning (DMDR)
- Attention mechanism
- Unified single-stream attention across text conditioning and image latent tokens
- Image decoder
- Variational autoencoder with latent image reconstruction
- Batch generation
- Configurable through compatible inference workflows
- Model precision
- BF16 recommended for standard inference
- Inference performance
- Sub-second inference demonstrated on H800 GPUs; actual performance depends on hardware and resolution
- Bilingual support
- English and Chinese prompt understanding and text rendering
Inputs and Controls
- Text prompt
- Required natural-language description of the target image
- Input image
- Not supported as a native conditioning input by the standard Turbo text-to-image pipeline
- Reference images
- Not supported natively by the standard Turbo checkpoint
- Image editing
- Requires a separate image-editing model or additional workflow
- Negative prompt
- Not supported as an effective standard control in guidance-distilled Turbo inference
- Style control
- Natural-language descriptions of photographic style, illustration techniques, lighting, composition and visual aesthetics
- Camera and composition
- Prompt-based control over perspective, framing, focal length, depth of field, lighting and photographic appearance
- Text rendering
- Supports English and Chinese text within generated images
- Language support
- Strongest documented support for English and Chinese
- Inference steps
- 8 effective denoising evaluations; 9 scheduler steps in the official Diffusers example
- Guidance scale
- 0.0 recommended
- Sampling scheduler
- FlowMatchEulerDiscreteScheduler
- Random seed
- Optional integer seed for reproducible image generation
- Image width
- Configurable; 1024 pixels in the official reference example
- Image height
- Configurable; 1024 pixels in the official reference example
- Number of images
- Configurable through batch generation workflows
- Prompt expansion
- Optional through a separate prompt-enhancement model or preprocessing workflow
- Prompt reasoning
- Optional reasoning-based prompt enhancement for expanding semantic context and visual descriptions
- Image consistency
- Controlled through detailed prompts and random seeds; no native reference-image identity locking
- Checkpoint selection
- Z-Image-Turbo for accelerated inference; undistilled Z-Image for greater sampling flexibility
- LoRA training
- Not recommended on the distilled Turbo checkpoint; foundation Z-Image is better suited for training
- LoRA inference
- Supported through compatible community adapters and custom inference workflows
- Model precision
- BF16 recommended for supported GPUs
- Attention backend
- PyTorch SDPA by default; optional Flash Attention 2 and Flash Attention 3
- Model compilation
- Optional transformer compilation to improve repeated inference performance
- CPU offloading
- Supported for reducing GPU memory requirements
- Quantization
- FP8 and other reduced-precision configurations available through compatible optimized runtimes
- Image post-processing
- Optional resizing, upscaling, format conversion, background removal and additional image processing
Serving
- Provider
- Alibaba / Tongyi-MAI
- API platform
- HexGrid.cloud
- Deployment
- Private GPU deployment with configurable image generation workflows
- Generation workflows
- Text-to-image generation, high-resolution image synthesis, batch generation and prompt-enhanced image creation
- Model checkpoint
- Tongyi-MAI/Z-Image-Turbo
- Checkpoint formats
- Safetensors
- Output format
- Static RGB images, typically PNG
- Model availability
- Open-weight checkpoint available for private commercial deployment
- Model license
- Apache 2.0
- Commercial use
- Permitted under Apache 2.0, subject to applicable license and trademark requirements
- Fine-tuning
- The foundation Z-Image model is recommended for custom fine-tuning; Turbo supports compatible inference adapters
- Model architecture
- 6B Scalable Single-Stream Diffusion Transformer with Qwen3-based text conditioning
- Inference frameworks
- Official Z-Image inference implementation, Diffusers, ComfyUI and compatible optimized inference runtimes
- Infrastructure
- Dedicated GPU infrastructure with private model execution and configurable inference pipelines
- GPU requirements
- Approximately 16 GB VRAM for optimized consumer-GPU inference; additional memory may be needed depending on precision and resolution
- Memory optimization
- CPU model offloading, reduced-precision inference, optimized attention and model compilation
- Parallel inference
- Batch processing and multi-GPU execution through compatible inference frameworks
- Customization
- Custom checkpoints, inference parameters, sampling configurations, prompt enhancement and image post-processing workflows
- Optimization
- 8-step distilled inference, optional Flash Attention, model compilation and quantized model execution
- Performance
- 8 effective denoising evaluations per image, with sub-second generation possible on H800-class hardware. Actual throughput depends on GPU hardware, resolution, batching and optimization settings
- Content safety
- Configurable content filtering and image safety checks through compatible deployment workflows