Z-Image
Open-source 6B-parameter image generation model family from Tongyi-MAI using a single-stream diffusion transformer.
At a Glance
Self-hosted Z-Image model family and code under the Apache License 2.0, with a perpetual, no-charge, royalty-free license grant.
Engagement
Available On
Listed Nov 2026
About Z-Image
Z-Image is an image generation foundation model family from Tongyi-MAI, published as an open-source repository under the Apache License 2.0. It has 6 billion parameters and is built on a Scalable Single-Stream Diffusion Transformer (S3-DiT). Checkpoints are distributed through Hugging Face and ModelScope, with online demos on both.
What It Is
Z-Image is a text-to-image model family with variants for different needs. Z-Image-Turbo is a distilled version that runs in 8 function evaluations and, per the README, offers sub-second latency on H800 GPUs and fits in 16G VRAM consumer devices. Z-Image is the foundation model aimed at high-quality generation, aesthetics, diversity, negative prompting, and fine-tuning. Z-Image-Omni-Base and Z-Image-Edit cover generation plus editing and instruction-based image editing; the model zoo lists them as to be released.
How It Works
The S3-DiT architecture concatenates text, visual semantic tokens, and image VAE tokens at the sequence level into one unified input stream. Turbo's speed comes from Decoupled-DMD distillation, and the DMDR method combines distribution matching distillation with reinforcement learning. Both are documented in accompanying arXiv papers.
Running It
Users can run inference with PyTorch natively (pip install -e . then python inference.py) or through the Hugging Face diffusers ZImagePipeline. Turbo uses guidance scale 0 and about 9 steps. The base Z-Image model recommends guidance 3.0-5.0, 28-50 steps, and negative prompts, at resolutions from 512x512 to 2048x2048. Optional attention backends, compilation, and CPU offloading are documented.
Community Ecosystem
The README lists community projects supporting Z-Image, including Cache-DiT, stable-diffusion.cpp (documented for GPUs with as little as 4GB VRAM), DiffSynth-Studio for LoRA and full training, vllm-omni, SGLang-Diffusion, Candle, LeMiCa, and MeanCache. The README states Z-Image-Turbo ranked 8th overall on the Artificial Analysis text-to-image leaderboard and first among open-source models.
Community Discussions
Be the first to start a conversation about Z-Image
Share your experience with Z-Image, ask questions, or help others learn from your insights.
Pricing
Open Source (Apache-2.0)
Self-hosted Z-Image model family and code under the Apache License 2.0, with a perpetual, no-charge, royalty-free license grant.
- Apache License 2.0 (OSI-approved)
- Z-Image-Turbo, Z-Image, Z-Image-Omni-Base and Z-Image-Edit variants
- 6B parameter image generation model family
- Free to use, modify and distribute under Apache-2.0 terms
- Self-hosted; own hardware or inference costs apply
Capabilities
Key Features
- 6B-parameter image generation model family
- Z-Image-Turbo 8-step distilled inference
- Scalable Single-Stream DiT (S3-DiT) architecture
- Bilingual English and Chinese text rendering
- Negative prompting and CFG support in base model
- Image editing variants (Z-Image-Edit, Omni-Base)
- Decoupled-DMD and DMDR distillation methods
- PyTorch native and diffusers inference
- Checkpoints on Hugging Face and ModelScope
- Apache 2.0 licensed
