Z-Image: A 6B-Parameter Open-Source Image Generation Model Challenging the Scale-At-All-Costs Paradigm
By
[Submitted on 27 Nov 2025 (v1), last revised 22 Jun 2026 (this version, v4)]
Summary
The Z-Image team introduces an efficient 6B-parameter image generation foundation model built on a Scalable Single-Stream Diffusion Transformer (S3-DiT) architecture. Unlike dominant proprietary systems (e.g., Nano Banana Pro, Seedream 4.0) and massive open-source alternatives (20B-80B parameters), Z-Image achieves competitive performance with significantly reduced computational overhead — completing full training in 314K H800 GPU hours (~$630K). The model supports few-step distillation (Z-Image-Turbo) for sub-second inference on enterprise GPUs and compatibility with consumer hardware (<16GB VRAM), plus an editing variant (Z-Image-Edit). It excels at photorealistic image generation and bilingual text rendering, rivaling top-tier commercial models while being open-source.
Source
bskyZ-Image: A 6B-Parameter Open-Source Image Generation Model Challenging the Scale-At-All-Costs Paradigmarxiv.orgKey quotes
· 5 pulledTo address this gap, we propose Z-Image, an efficient 6B-parameter foundation generative model built upon a Scalable Single-Stream Diffusion Transformer (S3-DiT) architecture that challenges the 'scale-at-all-costs' paradigm.
By systematically optimizing the entire model lifecycle -- from a curated data infrastructure to a streamlined training curriculum -- we complete the full training workflow in just 314K H800 GPU hours (approx. $630K).
Our few-step distillation scheme with reward post-training further yields Z-Image-Turbo, offering both sub-second inference latency on an enterprise-grade H800 GPU and compatibility with consumer-grade hardware (<16GB VRAM).
Z-Image exhibits exceptional capabilities in photorealistic image generation and bilingual text rendering, delivering results that rival top-tier commercial models.
We publicly release our code, weights, and online demo to foster the development of accessible, budget-friendly, yet state-of-the-art generative models.
You might also wanna read
Training-Free Single-Image Diffusion Model Achieves Fast, High-Quality Generation
This paper presents a training-free approach to single-image diffusion models. Instead of training a neural network on a single image (which
arxiv.org·16d ago
Iris: Pure C Implementation of Flux 2 Image Generation Model Inference Pipeline
Iris is a pure C implementation of an inference pipeline for image generation models, specifically designed for Flux 2 models. The project p
github.com·5mo ago
ZSE: Ultra Memory-Efficient LLM Inference Engine for Running Large Models on Consumer GPUs
ZSE (Z Server Engine) is an ultra memory-efficient LLM inference engine with native INT4 CUDA kernels that enables running large language mo
github.com·3mo ago
Helios: A 14B Parameter Real-Time Video Generation Model for Minute-Scale Content
Helios is a 14B parameter video generation model that achieves real-time performance at 19.5 FPS on a single NVIDIA H100 GPU while supportin
alphaxiv.org·3mo ago
FLUX.2 [klein] Fast Image Generation Models Released with Under-Second Inference
FLUX.2 [klein] is a new family of fast image generation models that unify image creation and editing in a single compact architecture. The m
bfl.ai·5mo ago
Moebius: A 0.2B Parameter Lightweight Image Inpainting Framework Matching 10B-Level Performance
Moebius is a lightweight 0.2B parameter image inpainting framework that achieves performance comparable to 10B-level industrial foundation m
hustvl.github.io·1d agoComments
Sign in to join the conversation.
No comments yet. Be the first.