Skip to content
Search

Z-Image

lastmod 2026-02-13created 2025-12-084 edits

A powerful and efficient image generation model

A 6B-parameter single-stream Diffusion Transformer trained only on real data, achieving SOTA-level photorealistic image generation while drastically cutting computational cost. It runs on GPUs with 16GB or less, which makes it well suited for edge deployment.

  • Optimizes visual-semantic alignment through a data curation pipeline and a multi-stage training curriculum
  • Applies QK-Norm/Sandwich-Norm conditioning techniques
  • Supports few-step distillation based on Decoupled DMD

About

  • Alibaba's image generation model with Nano Banana Pro-level performance, open source under the Apache 2.0 license
  • An efficient 6B-parameter image generation model based on a Single-Stream Diffusion Transformer
  • Comes in three versions; Z-Image-Turbo reaches the quality of competing models with only 8 inference steps and runs in 16GB VRAM environments
  • Z-Image-Edit provides image editing based on natural-language instructions, and Z-Image-Base is scheduled for release as a base model for community fine-tuning
  • The model adopts the S3-DiT architecture, unifying text, visual, and VAE tokens into a single sequence to maximize parameter efficiency
  • Generates high-quality images even in a few steps through the Decoupled-DMD and DMDR algorithms, recording top-tier performance among open-source models

Z-Image Overview

  • Z-Image is a foundational image generation model that combines efficiency and performance, using a single-stream diffusion transformer architecture
  • Based on 6 billion parameters, with three variants: Z-Image-Turbo, Z-Image-Base, and Z-Image-Edit
    • Z-Image-Turbo: generates high-quality images with only 8 function evaluations (NFE), with sub-second inference latency on an H800 GPU
    • Z-Image-Base: a non-distilled base model supporting community fine-tuning and custom development
    • Z-Image-Edit: a version specialized for image editing, providing natural-language-based image transformation

Key Features and Performance

  • Z-Image-Turbo accurately produces photorealistic images and bilingual English/Chinese text rendering
  • The Prompt Enhancer strengthens descriptive ability based on reasoning and world knowledge
  • Z-Image-Edit supports creative image transformation and precise instruction understanding
  • Recorded top-tier performance among open-source models in Alibaba AI Arena's Elo-based human preference evaluation

Model Architecture (S3-DiT)

  • Adopts the Scalable Single-Stream DiT (S3-DiT) architecture
    • Combines text, visual semantic tokens, and image VAE tokens into a single sequence
    • Maximizes parameter efficiency compared to existing dual-stream architectures
  • This architecture enables unified text-image representation learning

Core Algorithms

  • Decoupled-DMD
    • Decoupled-DMD is the core distillation algorithm that enables Z-Image's 8-step inference
    • Splits the existing DMD (Distribution Matching Distillation) into two mechanisms: CFG Augmentation (CA) and Distribution Matching (DM)
      • CA acts as the main engine of the distillation process
      • DM acts as a regularizer to maintain output stability and quality
    • Decoupling and optimizing the two mechanisms achieves high-performance image generation even in a few steps
  • DMDR
    • DMDR (Distribution Matching Distillation with Reinforcement Learning) is a post-training technique that combines DMD with reinforcement learning (RL)
    • A complementary structure where RL maximizes DMD's performance and DMD regularizes RL
    • This improves semantic alignment, aesthetic quality, and structural coherence, and strengthens high-frequency detail

Community and Ecosystem Support

  • Cache-DiT: supports inference acceleration through DBCache, Context Parallelism, and Tensor Parallelism
  • stable-diffusion.cpp: a C++-based engine that can run Z-Image even in 4GB VRAM environments
  • LeMiCa: provides training-free timestep-level inference acceleration
  • ComfyUI ZImageLatent: provides a simple latent interface at the official resolutions

Summary

  • Z-Image is a high-performance open-source image generation model combining an efficient architecture (S3-DiT) with innovative distillation techniques (Decoupled-DMD, DMDR)
  • Z-Image-Turbo achieves both fast inference and high quality, and runs even on consumer-grade GPUs
  • Z-Image-Edit supports precise natural-language-based image editing
  • Recorded the top human preference scores among open-source models in Alibaba AI Arena
  • The Z-Image ecosystem is integrating with various community projects and expanding into a general-purpose generative model platform

See also

Favorite site

author Zer0 sitemap