6B parameter Single-Stream Diffusion Transformer • 8 NFE sub-second inference • Bilingual text rendering • Fully open source
Copy Prompts & One-Click Design Replication.
We're proud to release Z-Image, a revolutionary 6B parameter model using Single-Stream Diffusion Transformer architecture. With only 8 NFEs, Z-Image-Turbo achieves sub-second inference on enterprise GPUs and runs smoothly on consumer devices with 16GB VRAM. Experience photorealistic quality that rivals models 10x larger.
The release of Z-Image-Base marks a pivotal moment for the open-source diffusion community. This non-distilled foundation model enables full fine-tuning, LoRA training, and custom development. More than just a model release, we're sharing our groundbreaking training methodology to empower researchers and developers worldwide.
Z-Image achieves industry-leading bilingual text rendering for English and Chinese with unprecedented accuracy. Our Single-Stream DiT architecture natively understands typography across languages, solving a challenge that has plagued image generation models for years. See how we achieve pixel-perfect text in photorealistic scenes.
Introducing Z-Image-Edit, our specialized variant for image-to-image transformations. Using natural language instructions, you can precisely edit images with creative style transfer, object manipulation, and targeted modifications. Built on Z-Image's efficient architecture for fast, high-quality edits.
Deep dive into Z-Image's revolutionary Single-Stream Diffusion Transformer architecture. Learn how we achieved an order of magnitude better parameter efficiency than competing models, enabling state-of-the-art quality with just 6B parameters and 8 inference steps. Technical insights into our distillation methodology.
The community has discovered an innovative use case: using Z-Image as a second-pass upscaler for any model. ComfyUI workflows show that Z-Image can enhance images from FLUX, SDXL, SD 1.5, and more with results comparable to commercial services. Fast enough to integrate directly into your workflow.
Z-Image now supports ControlNet, including Union 2.0, for precise control over generation. Combine Z-Image's speed with pose control, depth guidance, edge detection, and more. Perfect for professional workflows requiring both accuracy and efficiency. See examples of object removal and creative editing.
Z-Image is designed to be accessible. This comprehensive guide covers running Z-Image-Turbo on consumer GPUs with 16GB VRAM, optimization techniques for 4GB VRAM devices, and integration with ComfyUI, Diffusers, and DiffSynth-Studio. Includes performance benchmarks and best practices.
Z-Image employs a revolutionary Single-Stream DiT architecture that processes information more efficiently than traditional multi-stream approaches.
Bilingual encoder for English & Chinese
Shared embedding space
Compressed latent space representation
Efficient dimensionality reduction
Unified processing stream for efficiency
6B parameters total
Iterative refinement with only 8 NFEs
Distilled for speed
High-fidelity image reconstruction
Supports up to 2048x2048
Unlike multi-stream models, Z-Image uses a unified stream that reduces parameter redundancy while maintaining quality
Turbo variant distills the base model down to 8 NFEs without quality loss using novel training techniques
Achieves results comparable to 60B+ parameter models with only 6B parameters through architectural optimizations
Z-Image delivers exceptional performance across all key metrics, outperforming models significantly larger in size.
3-6x Faster
vs 20-50 steps for competitors
Order of Magnitude Efficient
Quality of 60B+ models
Consumer Friendly
Runs on standard GPUs
Bilingual Support
Native text rendering
Create your free account and get instant credits to start generating
Describe the image you want to generate in natural language
Click generate and watch AI create your image in seconds
Save and use your generated images for any project
Powerful 6 billion parameter model delivering exceptional image quality and detail that rivals models 10x larger, while maintaining efficiency and speed.
Revolutionary Diffusion Transformer architecture with single-stream processing that achieves unprecedented parameter efficiency and coherent generation.
Industry-leading bilingual text rendering with native support for English and Chinese, producing accurate typography where other models fail.
Blazing-fast 8-NFE turbo mode achieves sub-second inference on H800 GPUs and rapid generation on consumer hardware, 3-6x faster than competitors.
Z-Image-Edit variant enables precise image-to-image transformations using natural language instructions for creative editing and style transfer.
Complete open-source release including all three variants (Turbo, Base, Edit), weights, code, and groundbreaking training methodology for community innovation.
Runs comfortably on consumer devices with 16GB VRAM, and can be optimized for GPUs with as little as 4GB VRAM, making it accessible to all.
Excels at generating highly detailed, photorealistic images with superior detail preservation in distant elements and backgrounds compared to similar-sized models.
Full ControlNet support including Union 2.0 for precise control over pose, depth, edges, and more. Perfect for professional workflows requiring accuracy.
Z-Image-Base enables full fine-tuning, LoRA training, and distillation. Tools like DiffSynth-Studio provide comprehensive training support with low-VRAM optimization.
Use Z-Image as a second-pass upscaler/enhancer for any model (FLUX, SDXL, SD 1.5). Results rival commercial services with fast, integrated workflow processing.
Extensive community support with ComfyUI nodes, Diffusers pipelines, Replicate API, AI Runner integration, and comprehensive documentation across platforms.
"Z-Image has completely transformed our creative workflow. The quality and speed are unmatched in the industry."
"The bilingual text support is a game-changer for our international campaigns. Finally, AI that understands multiple languages!"
"As an artist, I appreciate how Z-Image enhances rather than replaces creativity. The instruction editing is pure magic."
"The open source nature and community support make Z-Image the best choice for developers building AI applications."
"Incredible speed without sacrificing quality. Z-Image Turbo has cut our production time by 80%."
"The 6B parameter model delivers consistently stunning results. It's become an essential tool in our design process."
Z-Image is a powerful and highly efficient image generation model with 6B parameters, using a Single-Stream Diffusion Transformer architecture. It delivers state-of-the-art photorealistic image generation that matches or exceeds models an order of magnitude larger, while running on consumer hardware with only 16GB VRAM.
Z-Image comes in three variants: (1) Z-Image-Turbo - a distilled version optimized for speed with only 8 NFEs and sub-second inference; (2) Z-Image-Base - the non-distilled foundation model for community fine-tuning and custom development; (3) Z-Image-Edit - a specialized variant for image editing tasks with natural language instruction following.
Z-Image is smaller (6B parameters), faster (8 NFEs vs 20-50), and more efficient than both FLUX and SDXL while delivering comparable or superior quality. It offers better detail preservation than SDXL and native bilingual text support. The community considers it a potential SDXL replacement for many use cases.
Beyond releasing a powerful model, Z-Image's team shared their advanced distillation training methodology with the open-source community. This enables researchers and developers to apply similar efficient training techniques to their own models, raising the bar for the entire diffusion model ecosystem.
Yes! Z-Image-Turbo fits comfortably within 16GB VRAM on consumer devices. With optimization techniques, it can even run on GPUs with only 4GB VRAM. It achieves sub-second generation on enterprise H800 GPUs and fast generation on consumer hardware.
Yes, Z-Image has excellent bilingual text rendering capabilities, supporting both English and Chinese with high accuracy. This is a significant advantage over many competing models that struggle with text generation.
Absolutely! The community has created workflows using Z-Image as a second-pass upscaler/enhancer for images from other models like FLUX, SD 1.5, SDXL, and others. Results are comparable to commercial upscaling services, and it's fast enough to integrate directly into workflows.
Z-Image-Base is the non-distilled foundation model released to unlock the full potential for community-driven fine-tuning and custom development. This release is significant because it allows the community to create specialized variants, LoRAs, and custom models based on Z-Image's powerful architecture.
Yes! Z-Image supports ControlNet for precise control over generation. There's even a ControlNet Union 2.0 version available. The community has showcased impressive results combining Z-Image-Turbo with ControlNet for tasks like object removal and pose control.
Z-Image supports multiple integration methods: ComfyUI (with community nodes and utilities), Diffusers (official pipelines), DiffSynth-Studio (for LoRA training), AI Runner, and can be accessed via Replicate API. Check the official GitHub repository for detailed setup instructions.
NFEs (Number of Function Evaluations) represent the number of inference steps needed to generate an image. Fewer NFEs mean faster generation. Z-Image-Turbo achieves excellent quality with only 8 NFEs, while competitors often require 20-50 steps, making it significantly faster.
Yes! With the release of Z-Image-Base, you can perform full fine-tuning, LoRA training, and distillation training. Tools like DiffSynth-Studio provide comprehensive support for training workflows with low-VRAM optimization.
Z-Image uses a Single-Stream Diffusion Transformer (DiT) architecture that processes information more efficiently than traditional multi-stream approaches. This contributes to the model's exceptional parameter efficiency and speed while maintaining high quality output.
Yes! Z-Image is fully open source, including the model weights, codebase, and training methodology. The team released all three variants (Turbo, Base, and Edit) to empower the community to build, fine-tune, and innovate freely.
Z-Image stands out through its unprecedented efficiency (6B params matching 60B+ models), blazing speed (8 NFEs vs 20-50), bilingual text rendering, and the groundbreaking release of both the model and training methodology to the open-source community. It's accessible on consumer hardware while delivering professional-grade results.
Tongyi-MAI (Alibaba)
Z-Image is a powerful and highly efficient image generation model with 6B parameters that employs a revolutionary Single-Stream Diffusion Transformer architecture. The model achieves state-of-the-art photorealistic image generation that matches or exceeds models an order of magnitude larger, while running on consumer hardware with only 16GB VRAM. Through advanced distillation techniques, Z-Image-Turbo generates high-quality images with only 8 NFEs, achieving sub-second inference latency.
A novel architecture that processes information through a unified stream rather than multiple parallel streams, achieving an order of magnitude better parameter efficiency.
Groundbreaking distillation techniques that reduce inference steps from 50 to 8 without quality degradation, shared openly with the research community.
Architecture-level support for bilingual text rendering in English and Chinese with unprecedented accuracy.
Three specialized variants (Turbo, Base, Edit) optimized for different use cases, all built on the same efficient foundation.
