Source-linked AI summary
Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer
Z-Image Team, Huanqia Cai, Sihan Cao, Ruoyi Du, Peng Gao, Aiming Hao, Steven Hoi, Zhaohui Hou, Shijie Huang, Dengyang Jiang, Yuming Jiang, Xin Jin, Liangchen Li, Zhen Li, Zhong-Yu Li, David Liu, Dongyang Liu, Qilong Wu, Feng Yu, Zechao Zhan, Chi Zhang, Shifeng Zhang, Ruikai Zhou, Shilin Zhou
TL;DR
Z-Image addresses the gap between high-performing proprietary systems and parameter-heavy open-source models by introducing a 6B-parameter foundation model with an efficient single-stream architecture and optimized lifecycle. Its data infrastructure and training strategy support strong photorealistic and bilingual text-rendering results, while Z-Image-Turbo and Z-Image-Edit target fast deployment and instruction-following editing. The workflow requires 314K H800 GPU hours at approximately $628K, with Turbo providing sub-second inference and consumer-hardware compatibility.
Problem
High-performing image generators are either proprietary or so large that open-source alternatives are impractical for inference and fine-tuning on consumer-grade hardware.
Method
Z-Image combines a 6B-parameter S3-DiT architecture with curated data infrastructure, streamlined training, few-step distillation, and reward post-training.
Results
Experiments report photorealistic generation, bilingual text rendering, and performance rivaling top-tier commercial systems, with 314K H800 GPU hours and approximately $628K training cost.
Takeaways & Limitations
Z-Image-Turbo delivers sub-second H800 inference and fits consumer-grade hardware under 16GB VRAM, while omni-pre-training enables Z-Image-Edit.
Takeaways & Limitations
Z-Image’s limited 6B-parameter size constrains its world knowledge and intent understanding.
Abstract
from arXiv · showhide
The landscape of high-performance image generation models is currently dominated by proprietary systems, such as Nano Banana Pro and Seedream 4.0. Leading open-source alternatives, including Qwen-Image, Hunyuan-Image-3.0 and FLUX.2, are characterized by massive parameter counts (20B to 80B), making them impractical for inference, and fine-tuning on consumer-grade hardware. To 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). Additionally, our omni-pre-training paradigm also enables efficient training of Z-Image-Edit, an editing model with impressive instruction-following capabilities. Both qualitative and quantitative experiments demonstrate that our model achieves performance comparable to or surpassing that of leading competitors across various dimensions. Most notably, Z-Image exhibits exceptional capabilities in photorealistic image generation and bilingual text rendering, delivering results that rival top-tier commercial models, thereby demonstrating that state-of-the-art results are achievable with significantly reduced computational overhead. We publicly release our code, weights, and online demo to foster the development of accessible, budget-friendly, yet state-of-the-art generative models.
1. Introduction
Z-Image addresses the accessibility gap in image generation with a 6B-parameter model and an end-to-end workflow optimized for real-world data, architecture, training, and inference. Its variants and experiments target efficient deployment while retaining photorealism, bilingual text rendering, and editing capability.
- Motivation and approach: Z-Image challenges scale-at-all-costs development and synthetic-data distillation through a comprehensive end-to-end optimization of the model lifecycle.The workflow covers data curation, architecture design, training strategies, and inference acceleration using purely real-world data.
- Efficiency: 314K H800 GPU hours and approximately $628K complete the Z-Image training pipeline.The estimate assumes roughly $2 per GPU hour and is presented as substantially below the resources commonly required by leading models.
- Motivation and approach: 6B parameters enable Z-Image to pursue high performance with lower hardware requirements than massive open-source alternatives.The architecture uses a scalable single-stream diffusion transformer with dense cross-modal interaction.
- Specialized variants: 8 NFEs yield Z-Image-Turbo with sub-second enterprise-GPU inference and compatibility with consumer hardware under 16GB VRAM.The variant is produced through few-step distillation with reinforcement learning.
- Empirical results: Experiments report photorealistic generation, bilingual Chinese/English text rendering, precise editing instruction adherence, and performance rivaling top-tier commercial systems.The authors publicly release code, weights, and an online demo.
2. Data Infrastructure
The data infrastructure turns large, uncurated collections into balanced, non-redundant training data through profiling, semantic retrieval, knowledge-graph organization, and active refinement. These components jointly target quality, coverage, deduplication, and model-failure remediation.
- Design goals: The infrastructure is designed to maximize information gain per computing unit rather than relying on dataset scale alone.Its target corpus is conceptually broad, non-redundant, and multilingual.
- Cross-modal Vector Engine: Cross-modal vector retrieval supports semantic deduplication, targeted sampling for conceptual gaps, and diagnosis of data clusters associated with model failures.The deduplication method uses k-nearest-neighbor search, proximity graphs, and community detection.
- Data Profiling Engine: The Data Profiling Engine computes multidimensional image-text features and supports source-specific heuristics and sampling for balanced curation.Features cover metadata, technical quality, semantic content, aesthetics, and image-text consistency.
- World Knowledge Topological Graph: The World Knowledge Topological Graph prunes low-centrality and non-visualizable concepts before guiding semantic-level balanced sampling.Sampling weights use tag relevance and hierarchical parent-child relationships.
- Active Curation Engine: An Active Curation Engine combines filtering with model feedback to refine data and address long-tail distribution deficiencies.The “Squirrel Fish” example illustrates how missing domain-specific concepts can produce erroneous compositional generations.
3. Image Captioner
Z-Captioner combines multiple caption types with OCR and world knowledge to improve image understanding, text rendering, and editing supervision. Difference captions are generated through a staged process that describes, compares, and summarizes source-target changes.
- Captioner design: Z-Captioner is an all-in-one image captioner incorporating multiple types of image captions.The captioning tasks share image understanding and depiction goals while supporting downstream synthesis and editing.
- OCR and world knowledge: World knowledge from metadata is included in captions to help identify specific entities such as public figures, landmarks, and events.The figure examples include original-language text transcription and recognition of West Lake, Hangzhou, China.
- OCR and world knowledge: Explicit OCR information is linked in the authors’ experiments to accurate text rendering in generated images.The pipeline recognizes optical characters before generating captions from the OCR results.
- Caption types: Five caption types—long, medium, short, tags, and simulated user prompts—cover detailed descriptions and incomplete real-user-style instructions.Long captions include dense image information, subjects, objects, backgrounds, locations, and full OCR results.
- Editing instructions: Difference captions specify source-to-target transformations and are produced by detailed captioning, difference analysis, and instruction synthesis.The step-by-step process moves from understanding to comparison and then concise instruction generation.
4. Model Training
Z-Image uses a parameter-efficient single-stream diffusion architecture and a staged training pipeline spanning pre-training, fine-tuning, distillation, and reinforcement learning. This workflow supports both efficient generation and editing while addressing the model’s limited world knowledge through prompt enhancement.
- Training Pipeline: The training pipeline combines low-resolution pre-training, omni-pre-training, supervised fine-tuning, few-step distillation, and reinforcement learning with human feedback.Low-resolution pre-training emphasizes cross-modal alignment and foundational visual knowledge, while omni-pre-training supports image and text conditioning at resolutions up to the 1k-1.5k range.
- Architecture: S3-DiT concatenates modality-specific representations into a unified single-stream sequence for dense cross-modal interaction and parameter reuse.The architecture uses lightweight modality-specific processors before tokens enter the shared backbone.
- Supervised Fine-Tuning: Tagged resampling balances concepts during supervised fine-tuning to preserve semantic diversity while the model converges toward a high-quality target distribution.The strategy up-weights under-represented concepts and down-weights over-represented ones.
- Few-Step Distillation: Decoupled DMD separates two distillation mechanisms, enabling independent optimization of renoising schedules for content alignment and distribution matching.The resulting approach is reported to preserve sharp details and color fidelity while improving photorealism over the original multi-step teacher.
- Few-Step Distillation: Z-Image-Turbo combines Decoupled DMD with DMDR to produce 8-step inference that is frequently perceived as higher quality than the 100-step teacher.The distillation framework is presented as resolving the trade-off between inference speed and visual fidelity.
- Prompt Enhancement: A Prompt Enhancer powered by a fixed pretrained VLM supplies structured reasoning and additional world knowledge to compensate for Z-Image’s limited parameter count.The reasoning chain improves interpretation of complex prompts, including geolocation and procedural instructions.
5.1. Human Preference Evaluation
Human-preference evaluations place Z-Image-Turbo among the strongest open-source text-to-image models. It achieves competitive overall rankings while combining quality with a 6B-parameter architecture and low inference cost.
- Artificial Analysis Arena: 1,161 Elo places Z-Image-Turbo 8th overall and first among open-source models on the Artificial Analysis leaderboard.The model is reported as comparable to leading proprietary systems such as Imagen 4 and Seedream.
- Efficiency: Z-Image-Turbo has the smallest parameter count among the top ten evaluated models at 6B and the lowest inference cost at $5.0 per 1,000 images.These figures are reported alongside its 8th-place overall ranking.
- Alibaba AI Arena: 4th-place overall performance on Alibaba AI Arena again makes Z-Image-Turbo the highest-ranked open-source model.The result corroborates its leading open-source position across a second Elo-based evaluation platform.
- Flux 2 dev Comparison: 87.4% combined G+S Rate was obtained by Z-Image in a 222-sample user study against Flux 2 dev.The study reports a 46.4% Good Rate, a 41.0% Same Rate, and a 12.6% Bad Rate; Z-Image used 6B parameters versus Flux 2 dev’s 32B.
5.2. Quantitative Evaluation
Across generation, text rendering, instruction following, reasoning, and editing benchmarks, the Z-Image family reports strong and often leading performance. The results include competitive scores for the distilled model and bilingual editing capabilities for Z-Image-Edit.
- Text Rendering: 0.8671 average Word Accuracy gives Z-Image the highest CVTG-2K score, ahead of GPT-Image-1 at 0.8569 and Qwen-Image at 0.8288.Z-Image-Turbo achieves the highest CLIP Score at 0.8048 while retaining 0.8585 average Word Accuracy.
- Text Rendering: 0.935 ranks Z-Image third on LongText-Bench-EN, with strong and consistent performance reported across English and Chinese settings.The benchmark evaluates longer text rendering in both languages.
- Fine-Grained Alignment: 0.546 gives Z-Image the highest overall English OneIG score, while its English and Chinese Text scores reach 0.987 and 0.988.On the Chinese track, Z-Image ranks second overall with 0.535.
- Object-Centric Generation: 0.84 places Z-Image in a three-way tie for second on GenEval, behind Qwen-Image at 0.87.Z-Image-Turbo scores 0.82, a 2-point gap from the base model.
- Instruction Following: 88.14 ranks Z-Image third overall on DPG-Bench, while its Attribute score of 93.16 exceeds Qwen-Image’s 92.02 and Seedream 3.0’s 91.36.DPG-Bench evaluates prompt following in dense prompts.
- Instruction Following: Z-Image and Z-Image-Turbo rank fourth and fifth on TIIF, respectively, for instruction-following capability.The benchmark covers complex instructions across diverse categories.
- Reasoning and Aesthetics: Z-Image-Turbo ranks third on English PRISM-Bench with 77.4, while Z-Image ranks second on the Chinese track with 75.3.Z-Image reports Chinese Text Rendering and Composition scores of 83.4 and 88.6, respectively.
- Image Editing: Z-Image-Edit achieves competitive performance across nine ImgEdit tasks and ranks third on GEdit-Bench for bilingual editing.The model is especially competitive on object addition and extraction.
5.3. Qualitative Evaluation
Qualitative evaluations compare Z-Image and Z-Image-Edit with leading open- and closed-source models across generation, text rendering, editing, and viewpoint control. The results highlight strong photorealism, bilingual text rendering, localized editing, and spatial consistency.
- Visual Generation: Z-Image-Turbo produces realistic close-up portraits and more natural multi-expression faces than several competing models.Comparisons include Qwen-Image, HunyuanImage3.0, FLUX 2 dev, and Seedream 4.0.
- Bilingual Text Rendering: Z-Image-Turbo accurately renders Chinese and English text while preserving surrounding image aesthetics and authenticity.Its text rendering is comparable to Nano Banana Pro and surpasses other candidates in the cited comparisons.
- Localized Editing: Z-Image-Edit supports localized object addition, removal, and replacement, including changes to colors, materials, and contents.Examples include replacing a scarf, table, or bottle liquid and removing specific objects.
- Interactive Editing: Z-Image-Edit enables bounding-box and scribble-guided edits with zero-seam blending and harmonized shadows.The examples include replacing a flower through a bounding box and adding a balloon through a scribble.
- Camera Control: Z-Image-Edit maintains spatial and geometric consistency across wide viewpoint changes, including yaw from −135◦ to 135◦ and pitch from −75◦ to 75◦.The evaluation covers object-centric setups and complex indoor environments under continuous camera trajectories.
5.3.5. Robust Multi-Image ID Consistency
Z-Image-Edit preserves character identity across stylistic, background, and multi-subject transformations. Its multi-image composition examples demonstrate consistent identity placement across varied layouts.
- Identity Preservation: Z-Image-Edit preserves character identity across stylistic and background transformations by using identity embeddings from reference images.The model supports diverse multi-image composition paradigms.
- Multi-Image Composition: The model transfers characters into new backgrounds, applies styles while retaining localized elements, and combines multiple identities in one scene.Examples include background transplanting, Chinese ink-painting conversion with a signature stamp, and multi-subject campus composition.
5.3.6. Enhanced Reasoning Capacity and World Knowledge through Prompt Enhancer
The prompt enhancer adds structured reasoning and world knowledge to improve prompt interpretation for generation and editing. It is especially useful for complex reasoning tasks and ambiguous editing instructions.
- Structured Reasoning: The prompt enhancer uses a structured chain covering subject analysis, problem solving, world knowledge injection, aesthetic enhancement, and comprehensive description.This supports tasks such as solving logical puzzles and interpreting user intent.
- Ambiguous Editing Instructions: For ambiguous editing prompts, the prompt enhancer generates a structured plan that resolves underspecified instructions into coherent scenes and typography.For “Design a poster,” it plans product placement, outdoor setting, lighting, and commercial typography.
5.3.7. Emerging Multi-lingual and Multi-cultural Understanding Capacity
After bilingual-data training, Z-Image exhibits multilingual prompt understanding and generates images aligned with local cultures and landmarks. The examples indicate emerging multilingual and multicultural capabilities.
- Multilingual Understanding: Z-Image understands prompts in multiple languages and generates images aligned with local cultures and landmarks.These capabilities emerged after training with bilingual data.
5.3.8. Emerging Composite Instructions Understanding Capacity
Z-Image-Edit demonstrates composite instruction following through multiattribute edits, identity-preserving scene transformations, localized text rendering, and controlled visual manipulation. Z-Image-Turbo shows strong portrait, scene, bilingual text-rendering, and poster-design results.
- Z-Image-Edit: Z-Image-Edit jointly satisfies multiple dependent edit intents rather than applying isolated attribute changes.Examples include colorizing a photograph while changing pose, and removing dynamic elements while replacing scene content.
- Z-Image-Edit: Z-Image-Edit preserves facial identity while inserting subjects, backgrounds, and interactive text elements.The model transplants portraits into snowy scenes, adds a Samoyed and flowers, and renders text onto an inserted screen from the subject’s viewpoint.
- Z-Image-Turbo: Z-Image-Turbo renders close-up portraits with character expressions, skin textures, aesthetic composition, and realistic people-background relationships.The comparisons emphasize subtle expressions, skin, clothing, hair, layout, and posture details.
- Z-Image-Turbo: Z-Image-Turbo and Nano Banana Pro accurately generate the expected Chinese and English couplets in comparative text-rendering examples.Z-Image-Turbo also produces correct, aesthetically pleasing, and realistic Chinese and English poster designs.
- Z-Image-Edit: Z-Image-Edit supports localized addition, removal, and replacement using bounding boxes or user scribbles while maintaining lighting, shadows, and glass refractions.The figure presents ten diverse edits radiating from a central still-life reference image.
6. Conclusion
The Z-Image series is a family of high-performance 6B-parameter models built on S3-DiT and designed through an end-to-end efficiency strategy. The workflow achieves strong image synthesis, bilingual text rendering, fast Turbo inference, and image editing within a comparatively modest resource budget.
- Conclusion: The Z-Image series comprises high-performance 6B-parameter models built upon a Scalable Single-Stream Diffusion Transformer.Its strategy combines curated data, scalable architecture, streamlined training, and inference optimization.
- Conclusion: 314K H800 GPU hours and under $630K complete the workflow while delivering top-tier photorealistic synthesis and bilingual text rendering.The pipeline includes PE-aware supervised fine-tuning, few-step distillation, and reward post-training.
- Conclusion: Z-Image-Turbo provides sub-second inference on an H800 GPU and fits within 16G VRAM consumer-grade hardware.Z-Image-Edit is derived efficiently through the omni-pretraining paradigm.
7. Authors
The supplied author section lists the paper’s contributors and identifies the project’s leadership and contributor-order conventions.
- Authors: Huanqia Cai through Shilin Zhou are listed in the first author block.The names are presented in sequence in the supplied author listing.
- Authors: The author section contains two supplied blocks of names followed by explanatory contribution notes.The contribution notes define the ordering and leadership symbols.
- Authors: Chenglin Cai through Jia are listed in the second author block.The supplied passage ends mid-list at “Jia.”
- Authors: Core contributors are listed alphabetically by last name.The notation identifies this as convention 5.
- Authors: Project leadership is marked with † for Project Leader and ‡ for Image Editing Co-Leader.The notation identifies these roles in the author list.
- Authors: Contributors are listed alphabetically by last name.The passage separately labels this convention as item 7.
A. Prompts Used in the Report
The report appendix provides prompts and instructions used to reproduce showcased Z-Image-Turbo and Z-Image results. The examples span photographic scenes, poster designs, style and clothing transfer, and composite image editing.
- A.1. Figure 1 Column #1: Figure 1 prompts specify detailed photographic scenes including an elevator selfie, a dynamic stadium scene, and an ancient village landscape.The prompts describe subjects, environments, lighting, composition, and material details.
- A.1. Figure 1 Column #1: The elevator-selfie prompt describes a young East Asian woman, reflective stainless-steel walls, fluorescent lighting, and a dark marble floor.It is supplied in Chinese with an English translation.
- A.1. Figure 1 Column #1: The appendix includes bilingual translations for several prompts, including the elevator selfie, village landscape, poster designs, and fantasy scene.These translations make the requested visual content explicit in English.
- A.2. Figure 2 Row #1: The Figure 2 examples include magazine-cover and motivational-poster designs with specified typography, silhouettes, color transitions, and rendered Chinese text.The prompts define both visual composition and textual content.
- A.3. Figure 3 Row #1: The Figure 3 instructions request clothing and bag transfer, style matching, and a detailed fantasy-realistic composition.The examples combine reference-image transformations with a multi-subject scene.