PaperlayerView exact revision on arXiv

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

arXiv:2511.22699v5cs.CV

TL;DR

Large image-generation models are costly and difficult to access, motivating Z-Image’s efficient 6B-parameter architecture and end-to-end training workflow. The resulting model delivers top-tier photorealistic synthesis and bilingual text rendering, while variants support sub-second inference and instruction-following editing on constrained hardware.

  • Problem

    High-performance image generation remains split between opaque proprietary systems and open-source models requiring massive parameter counts, limiting transparency and accessibility.

    Source: §1 · p. 8

  • Method

    Z-Image combines curated data infrastructure, a scalable single-stream diffusion transformer, streamlined training, and inference optimization in an end-to-end workflow.

    Source: §1 · p. 8, §6 · p. 53

  • Results

    314K H800 GPU hours and approximately $628K complete training, while the model achieves top-tier photorealistic synthesis and bilingual text rendering and rivals much larger models.

    Source: §1 · pp. 8–9

  • Takeaways & Limitations

    The Z-Image pipeline provides a blueprint for developing accessible, budget-friendly, state-of-the-art generative models.

    Source: §6 · p. 53

    Because Z-Image has only 6B parameters, it exhibits limitations in world knowledge and intent understanding.

    Source: §4.8 · p. 24

Abstract

from arXiv · show

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 challenges proprietary black-box models, massive open-source scaling, and synthetic-data distillation through an end-to-end, real-data workflow spanning data, architecture, training, and inference. The resulting model family delivers competitive generation and editing quality with substantially lower computational requirements, including efficient Turbo and Edit variants.

Efficient Data Infrastructure

  • Its data infrastructure combines profiling, cross-modal deduplication and retrieval, concept organization, and active curation to align data with training stages and reduce redundant or low-quality samples.The system is designed to maximize knowledge acquired per unit time while using purely real-world data rather than distilling outputs from other models.Source: §1 · p. 8

Efficiency

  • 314K H800 GPU hours, approximately $628K, suffice for the complete Z-Image training pipeline, demonstrating low-cost development versus leading models requiring orders of magnitude more resources.Table 1 bases the estimate on an H800 rental price of about $2 per GPU hour.Source: §1 · p. 8, Table 1 · p. 8

Efficient Architecture

  • Z-Image introduces S3-DiT, a single-stream multimodal diffusion transformer that enables dense cross-modal interaction and strong performance in a compact 6B-parameter model.Uniform token treatment supports text-to-image generation and image-to-image editing, while a prompt enhancer improves world-knowledge comprehension and prompt understanding.Source: §1 · pp. 8–9

Efficient Training Strategy

  • A three-phase curriculum combines low-resolution pre-training, omni-pre-training for generation and manipulation, and prompt-enhancer-aware supervised fine-tuning without separate costly stages or additional LLM training.The phases cover foundational visual-semantic alignment, arbitrary-resolution generation, text-to-image synthesis, image-to-image manipulation, and joint PE–backbone optimization.Source: §1 · p. 9

Efficient Inference and Specialized Variants

  • Z-Image-Turbo uses few-step distillation with reinforcement learning to generate high-fidelity images in 8 NFEs, with sub-second enterprise-GPU latency and support for consumer hardware below 16GB VRAM.Decoupled DMD and DMDR provide the stated quality, stability, and distribution-matching mechanisms; omni-pre-training also enables the instruction-following Z-Image-Edit variant.Source: §1 · p. 9

Results

  • Z-Image matches top-tier commercial systems in photorealistic generation and bilingual Chinese/English text rendering, while Z-Image-Edit follows image-editing instructions precisely.Figures 1–4 and Section 5.3 support competitive visual fidelity, text rendering, and editing adherence; the authors release code, weights, and an online demo.Source: §1 · p. 9

2. Data Infrastructure

Z-Image’s Efficient Data Infrastructure prioritizes data efficiency through curated, non-redundant, multilingual, and dynamically staged data. Its profiling, retrieval, knowledge-graph, active-curation, and editing-pair systems support efficient multimodal training.

2. Data Infrastructure

  • The Efficient Data Infrastructure maximizes information gain per computing unit through conceptually broad, non-redundant, multilingual, and dynamically curriculum-structured data.It comprises four synergistic modules: data profiling, cross-modal vectors, a world-knowledge topology, and active curation.Source: §2 · pp. 9–10

2.1. Data Profiling Engine

  • The Data Profiling Engine combines source-aware sampling with image metadata, technical-quality, semantic, aesthetic, safety, alignment, and multilevel-captioning analysis.It filters low-quality or weakly aligned pairs, detects AI-generated content, captures culturally specific concepts, and integrates visible text and watermarks into captions.Source: §2.1 · pp. 10–11

2.2. Cross-modal Vector Engine

  • The Cross-modal Vector Engine uses GPU-accelerated k-NN graph deduplication and multimodal retrieval to balance distributions, fill conceptual gaps, and diagnose failure-causing data.Its processing rate is approximately 8 hours per 1 billion items on 8 H800s, including index construction and 100-NN querying.Source: §2.2 · pp. 11–12

2.3. World Knowledge Topological Graph

  • The World Knowledge Topological Graph prunes and expands hierarchical concepts, then assigns semantic sampling weights using BM25 scores and parent-child relationships.Wikipedia entities, caption-derived embeddings, visual-generatability filtering, user prompts, and emerging concepts jointly maintain breadth, relevance, and balanced staged sampling.Source: §2.3 · p. 12

2.4. Active Curation Engine

  • The Active Curation Engine combines automated filtering and model-based diagnostics with human-in-the-loop verification to identify hard cases and iteratively improve annotations, captioners, and reward models.Rejected pseudo-labeled samples receive manual correction before retraining, creating a closed-loop enhancement process.Source: §2.4 · pp. 12–13, Figure 6 · p. 13

2.5. Efficient Construction of Editing Pairs with Graphical Representation

  • The editing-pair pipeline combines expert-model mixed edits, graphical pair generation, semantically filtered video-frame pairs, and controllable text rendering to scale diverse instruction-following data.Graphical construction yields 2^N+1 pairs from one input and N edited versions, while rendering controls text content, font, color, size, and position.Source: Figure 7 · p. 13, §2.5 · p. 14

3. Image Captioner

Z-Captioner is an all-in-one captioning system that combines multiple caption types, explicit OCR, world knowledge, and structured difference captions for image editing. Its design supports accurate text rendering, entity-aware generation, and clear editing instructions.

3. Image Captioner

  • Z-Captioner integrates multiple image-captioning types to describe visual content while incorporating OCR and world knowledge for text-to-image generation and editing.The pipeline explicitly includes OCR results and metadata-derived world knowledge in captions and editing instructions.Source: §3 · pp. 14–15, Figure 8 · p. 14

3.1. Detailed Caption with OCR Information

  • The captioner recognizes all image text before generating captions, preserving original languages and reducing missing or mistranslated text, especially in dense scenes.This OCR-first process follows the spirit of Chain-of-Thought (CoT) [73].Source: §3.1 · p. 15

3.2. Multi-Level Caption with World Knowledge

  • Five caption types—long, medium, short, tags, and simulated prompts—combine dense visual descriptions, world knowledge, and user-oriented prompt variation.Long captions include full OCR, subjects, objects, backgrounds, and locations; shorter forms support incomplete prompts focused on selected image regions.Source: §3.2 · pp. 15–16

3.3. Difference Caption for Image Editing

  • Difference captions convert source-target comparisons into concise editing instructions through three steps: detailed OCR-inclusive captioning, difference analysis, and instruction synthesis.The process compares visual and textual discrepancies before producing clear, useful editing instructions.Source: §3.3 · p. 16

4. Model Training

Z-Image’s training pipeline combines an efficient single-stream architecture with staged pre-training, alignment, distillation, reinforcement learning, editing continuation, and reasoning-enhanced prompting. These stages produce efficient few-step generation while improving fidelity, instruction following, editing capability, and complex-prompt understanding.

4.1. Architecture

  • The S3-DiT architecture concatenates text, visual-semantic, and image tokens into one modality-agnostic stream for parameter-efficient cross-modal processing.Lightweight modality-specific processors feed a unified backbone, while QK-Norm, Sandwich-Norm, RMSNorm, and low-rank conditioning projections support stable, efficient training.Source: §4.1 · pp. 16–17, Figure 10 · p. 17

4.2. Training Efficiency Optimization

  • The training system combines FSDP2, gradient checkpointing, torch.compile, sequence-length-aware batching, and dynamic batch sizing to reduce memory waste and support mixed-resolution training.Frozen VAE and text-encoder components use data parallelism, while the large DiT shards optimizer states and gradients across GPUs.Source: §4.2 · p. 18, Figure 11 · p. 18

4.3. Pre-training

  • Omni-pre-training combines arbitrary-resolution, joint text-to-image and image-to-image, and bilingual multi-level caption training, enabling image-and-text conditioning up to the 1k-1.5k range.Low-resolution pre-training supplies foundational visual knowledge, including Chinese text rendering, while omni-pre-training provides initialization for both generation and editing.Source: §4.3 · p. 19

4.4. Supervised Fine-Tuning (SFT)

  • Supervised fine-tuning narrows generation toward a curated, high-fidelity distribution while tagged resampling preserves semantic diversity across rare concepts and artistic styles.Merging multiple specialized SFT variants smooths capability trade-offs and yields greater stability and robustness than any single checkpoint.Source: §4.4 · pp. 19–20

4.5.3. Results and Analysis

  • 8-step Z-Image-Turbo inference is indistinguishable from the 100-step teacher and often surpasses it in perceived quality and aesthetic appeal.Decoupled DMD restores sharp details and accurate colors, while DMDR combines distribution matching with reinforcement learning to prevent reward hacking and improve alignment.Source: §4.5.1 · p. 21, §4.5.3 · p. 22, §4.5.2 · p. 22

4.6. Reinforcement Learning with Human Feedback (RLHF)

  • RLHF uses reward modeling, offline DPO, and online GRPO to improve photorealism, aesthetics, instruction following, and semantic accuracy.The reward model evaluates instruction satisfaction, AI-content detection perception, and aesthetic quality; DPO targets objectively verifiable dimensions using VLM-generated, human-verified pairs.Source: §4.6 · p. 22, §4.6.1 · p. 22, §4.6.3 · p. 23

4.7. Continued Training for Image Editing

  • Continued editing training mixes editing pairs with text-to-image data at 512² then 1024² resolution, using a 4:1 text-to-image:image-to-image ratio to avoid degradation.A manually constructed, task-balanced SFT subset further improves editing instruction following, while synthetic text-editing data is heavily downsampled.Source: §4.7 · p. 24

4.8. Prompt Enhancer with Reasoning Chain

  • The Prompt Enhancer addresses Z-Image’s limited world knowledge, intent understanding, and reasoning by aligning a fixed pretrained VLM through structured reasoning-chain supervision.Reasoning enables location inference and step-specific content planning instead of literal coordinate rendering or generic outputs.Source: §4.8 · pp. 24–25, Figure 15 · p. 24

5. Performance Evaluation · 5.1. Human Preference Evaluation

Human-preference evaluations place Z-Image-Turbo among leading open text-to-image models, combining strong Elo rankings with low computational overhead. A separate user study finds Z-Image competitive with Flux 2 dev despite using one-fifth as many parameters.

5.1.1. Elo-based Human Preference Evaluation

  • Z-Image-Turbo achieved 1,161 Elo points, ranking 8th overall and first among open-source models in Artificial Analysis, while placing 4th overall and first among open-source models in Alibaba AI Arena.It also had the smallest parameter count among the top ten models (6B) and the lowest inference cost ($5.0 per 1,000 images).Source: Figure 16 · p. 25, Figure 17 · p. 26, §5.1.1 · p. 26
  • The Elo evaluation compares images generated from identical prompts through hidden-identity pairwise judgments of visual coherence, detail rendering, prompt alignment, and artistic quality.Each vote dynamically updates the global Elo leaderboard based on collective human judgment.Source: §5.1.1 · p. 26
  • Z-Image-Turbo was evaluated against leading closed-source systems, including Nano Banana Pro, Imagen 4 Ultra Preview 0606, Seedream [64], GPT Image 1 [High], and FLUX.1 Kontext [Pro] [36], plus open-source HunyuanImage 3.0 [8] and Qwen-Image [76].The evaluated model is a 6B-parameter diffusion architecture using 8 NFEs.Source: §5.1.1 · p. 26
  • The results establish Z-Image-Turbo as a leading open text-to-image model whose quality, speed, fidelity, and openness support resource-constrained and interactive deployment.The findings position compact models as capable of elite-level performance without compromising usability.Source: §5.1.1 · p. 27

5.1.2. Human Preference Comparison with Flux 2 dev

  • In the Flux 2 dev comparison, Z-Image reached a 87.4% G+S Rate, comprising a 46.4% G Rate and 41.0% S Rate, with a 12.6% B Rate.The study evaluated 222 samples, each assessed by three independent annotators; Z-Image used 6B parameters versus Flux 2 dev’s 32B.Source: §5.1.2 · p. 27, Table 4 · p. 27
  • The Flux 2 dev user study used user-style prompts and three independent annotators per sample to improve robustness and reduce evaluation bias.Table 4 summarizes the G, S, B, and G+S preference rates for the comparison.Source: §5.1.2 · p. 27, Table 4 · p. 27

5.2. Quantitative Evaluation

Z-Image and its variants perform strongly across benchmarks for text rendering, instruction following, object-centric generation, prompt adherence, and image editing. Results also show competitive bilingual capability and efficiency-performance trade-offs for Z-Image-Turbo and Z-Image-Edit.

5.2.1. Text-to-Image Generation

  • Z-Image leads CVTG-2K with 0.8671 average Word Accuracy, while Z-Image-Turbo achieves the highest CLIP Score of 0.8048 and 0.8585 average Word Accuracy.Z-Image maintains accuracy as text regions increase from 2 to 5.Source: §5.2.1 · pp. 27–28, Table 5 · p. 28
  • Z-Image demonstrates robust bilingual long-text rendering, scoring 0.935 on LongText-Bench-EN and 0.936 on LongText-Bench-ZH, while Z-Image-Turbo scores 0.917 and 0.926.Z-Image ranks third in English and second in Chinese; Z-Image-Turbo preserves a strong efficiency-performance trade-off.Source: Table 6 · p. 28, §5.2.1 · p. 28
  • On OneIG, Z-Image scores 0.546 overall on English, ranks second with 0.535 on Chinese, and reaches English and Chinese Text scores of 0.987 and 0.988.It surpasses Qwen-Image and GPT Image 1 [High] on the English overall score, while Z-Image-Turbo retains strong performance with only a marginal decrease.Source: §5.2.1 · p. 28, Table 7 · p. 29, Table 8 · p. 29
  • Across general generation benchmarks, Z-Image scores 0.84 on GenEval, 88.14 on DPG-Bench, and ranks fourth on TIIF, while Z-Image-Turbo scores 0.82 on GenEval and ranks fifth on TIIF.Z-Image ties for second on GenEval, ranks third on DPG-Bench, and achieves 93.16 in DPG-Bench Attribute, exceeding Qwen-Image and Seedream 3.0.Source: §5.2.1 · pp. 29–30, Table 9 · p. 30, Table 10 · p. 30
  • PRISM-Bench places Z-Image-Turbo third on English with 77.4 and Z-Image second on Chinese with 75.3, including Chinese Text Rendering 83.4 and Composition 88.6.These results indicate strong efficiency and multilingual generation quality across reasoning and aesthetics tracks.Source: §5.2.1 · p. 30, Table 12 · p. 31, Table 13 · p. 31

5.2.2. Instruction-based Image Editing

  • Z-Image-Edit shows competitive performance across nine ImgEdit tasks, especially object addition and extraction, and ranks third on GEdit-Bench for bilingual editing.GEdit-Bench evaluates visual naturalness and bilingual instruction following with English and Chinese instructions.Source: Table 14 · p. 31, §5.2.2 · p. 32, Table 15 · p. 32

5.3. Qualitative Evaluation … 5.3.4. Advanced Camera Controllability and Spatial Consistency

Qualitative evaluations show that Z-Image-Turbo delivers highly realistic and aesthetically strong generation, accurate bilingual text rendering, and Z-Image-Edit provides precise interactive editing with robust camera and spatial control. Across these tasks, the models demonstrate competitive performance against leading open- and closed-source systems.

5.3. Qualitative Evaluation

  • Z-Image-Turbo achieves strong visual quality across comparisons with leading open-source and closed-source image-generation models, while Z-Image-Edit is evaluated for editing, camera control, consistency, and multilingual understanding.The qualitative study covers photorealism, bilingual text rendering, complex instruction-following, localized manipulation, viewpoint control, multi-image identity consistency, prompt enhancement, and multilingual or multicultural understanding.Source: §5.3 · p. 32

5.3.1. Superior Photorealistic Generation

  • Z-Image-Turbo produces realistic close-up details, aesthetically pleasing multi-expression portraits, and authentic mobile-phone scenes, avoiding exaggerated expressions and implausible rain interactions seen in competing models.The comparisons cover Lumina-Image 2.0, Qwen-Image, HunyuanImage 3.0, FLUX 2 dev, Seedream 4.0, and other systems across facial detail, expressions, layout, posture, and environmental realism.Source: §5.3.1 · p. 32

5.3.2. Outstanding Bilingual Text Rendering

  • Z-Image-Turbo accurately renders Chinese and English text while preserving surrounding visual quality, matching Nano Banana Pro and surpassing other candidates in the presented comparisons.All results in this subsection use the Turbo version; poster examples also show correct small-text rendering and more aesthetically pleasing, realistic designs than competing systems.Source: §5.3.2 · pp. 32–33
  • Z-Image-Turbo renders posters with correct text and satisfactory design, whereas competing models produce small-character errors or repeated text in the presented examples.Qwen-Image, HunyuanImage 3.0, FLUX 2 dev, and Imagen 4 Ultra err on very small characters, while Seedream 4.0 and Nano Banana Pro repeatedly render text.Source: §5.3.2 · p. 33

5.3.3. Precise and Interactive Localized Editing

  • Z-Image-Edit supports seamless localized addition, removal, and replacement of specific objects, including changes to clothing, furniture, liquids, flowers, accessories, and background items.The examples demonstrate localized manipulation of still-life scenes, such as changing a blue scarf to orange, replacing a wooden table with marble, and inserting sunflowers or a clock.Source: §5.3.3 · p. 33
  • Bounding boxes and user scribbles guide intuitive edits with zero-seam blending, while the model harmonizes shadows, reflections, and refractions with the original illumination.The system maps edits to explicit spatial coordinates or sketch strokes and preserves physically realistic appearance, including liquid behavior inside a glass container.Source: §5.3.3 · p. 33

5.3.4. Advanced Camera Controllability and Spatial Consistency

  • Z-Image-Edit generates consistent novel views across continuous camera trajectories, supporting yaw from −135◦ to 135◦, complete 180◦ reverse views, and pitch from −75◦ to 75◦.The capability applies to both object-centric setups and complex indoor environments while maintaining spatial and geometric consistency under dramatic perspective changes.Source: §5.3.4 · p. 33
  • 3D reconstructions and camera-frustum estimates verify geometric fidelity, with generated views preserving object scale, depth cues, background parallax, shadows, reflections, and refractions without structural distortion or texture warping.The verification uses the VGGT pipeline on generated multi-view images and finds accurate camera alignments and successful sparse point-cloud reconstruction.Source: §5.3.4 · p. 33

5.3.5. Robust Multi-Image ID Consistency

Z-Image-Edit preserves character identity across stylistic and background transformations by extracting robust identity embeddings and injecting them into target layouts, supporting diverse multi-image compositions with high fidelity.

Robust Multi-Image ID Consistency

  • Z-Image-Edit preserves identity across diverse multi-image compositions, including background transplanting, style and element transfer, multi-subject composition, and subject-object interactions.Figure 29 demonstrates seamless environment blending, localized signature placement, unified multi-identity scenes, and natural interactions with referenced objects.Source: §5.3.5 · p. 33

5.3.6. Enhanced Reasoning Capacity and World Knowledge through Prompt Enhancer

The prompt enhancer uses structured reasoning to add logical reasoning, world knowledge, aesthetic refinement, and comprehensive descriptions, improving generation, editing, text rendering, and question answering. In editing, it resolves ambiguous intentions and enables physical reasoning through structured plans and realistic interactions.

Enhanced Reasoning Capacity and World Knowledge through Prompt Enhancer

  • The structured reasoning chain equips the model to solve logical puzzles, interpret user intent, render text, answer questions, and visualize scenes using world knowledge.The chain includes subject analysis, problem solving and world-knowledge injection, aesthetic enhancement, and comprehensive description.Source: §5.3.6 · p. 34
  • For image editing, the prompt enhancer transforms ambiguous instructions into structured compositions and adds physical reasoning, producing professional layouts and realistic tea-brewing interactions.Examples include redesigning a chaotic poster into an outdoor commercial scene and adding pouring water, fluid dynamics, and steam to tea-brewing edits.Source: §5.3.6 · p. 34

5.3.7. Emerging Multi-lingual and Multi-cultural Understanding Capacity

Training on bilingual data led Z-Image to develop multilingual prompt understanding and culturally aligned image generation, including local landmarks.

Emerging Multi-lingual and Multi-cultural Understanding Capacity

  • Z-Image understands prompts in multiple languages and generates images aligned with local cultures and landmarks.These capabilities emerged after training with bilingual data, as shown in Figure 34.Source: §5.3.7 · p. 34

5.3.8. Emerging Composite Instructions Understanding Capacity

After edit training, Z-Image-Edit develops emergent capacity for composite natural-language instructions, including coordinated multiattribute edits, identity-preserving transformations, and localized text rendering. Figure 35 summarizes these capabilities across bilingual prompts, source images, and edited results.

Emerging Composite Instructions Understanding Capacity

  • After edit training, the model jointly satisfies dependent edit intents, such as colorizing images while changing poses or removing elements while replacing objects and clothing colors.These results reflect faithful interpretation of the full instruction rather than isolated attribute changes.Source: §5.3.8 · p. 34
  • The model preserves facial identity during portrait transplantation with added subjects and renders stylized text on inserted screens according to the subject’s interaction and viewpoint.Examples include placing a close-up portrait in a snowy scene with a Samoyed and flowers, then adding “Z-Image” to an inserted screen.Source: §5.3.8 · p. 34
  • Z-Image-Edit handles composite instructions spanning multiattribute modifications, identity-preserving scene transformations, and localized text rendering.Figure 35 presents bilingual prompts, source images, and corresponding edited results.Source: Figure 35 · p. 52

6. Conclusion

The Z-Image series presents an end-to-end approach to efficient 6B-parameter image generation, combining S3-DiT with curated data, streamlined training, and advanced inference optimization. The resulting pipeline achieves high-quality synthesis at reduced cost while producing Turbo and Edit variants for accessible deployment.

Conclusion

  • Z-Image comprises high-performance 6B-parameter models built on S3-DiT and organized around curated data, scalable architecture, streamlined training, and advanced optimization.The optimization techniques include PE-aware supervised fine-tuning, few-step distillation, and reward post-training.Source: §6 · p. 53
  • 314K H800 GPU hours and under $630K deliver top-tier photorealistic synthesis and bilingual text rendering.These results support the paper’s challenge to the scale-at-all-costs paradigm.Source: §6 · p. 53
  • Z-Image-Turbo provides sub-second H800 inference and fits within 16G VRAM consumer hardware, while Z-Image-Edit is efficiently derived through omni-pretraining.Together, the models provide a blueprint for accessible, budget-friendly, state-of-the-art generative models.Source: §6 · p. 53

7. Authors

The paper credits core contributors and a broader contributor group, with core contributors listed alphabetically by last name and project leadership roles identified separately.

7.1. Core Contributors5

  • Core contributors include the listed authors, with Peng Gao identified as Project Leader and Aiming Hao as Image Editing Co-Leader.Core contributors are listed alphabetically by last name.Source: §7.1 · p. 53, §7.2 · p. 53

7.2. Contributors7

  • The broader contributor group comprises the additional authors listed in the contributors section.Contributors are listed alphabetically by last name.Source: §7.2 · p. 53

A. Prompts Used in the Report

This section lists the prompts used to reproduce Figures 1–3 with Z-Image-Turbo and PE disabled, spanning photorealistic scenes, text-rich designs, and editing instructions.

A. Prompts Used in the Report

  • The reported generation results can be reproduced by directly inputting the listed prompts into Z-Image-Turbo with PE disabled.The section explicitly frames these prompts as the instructions used for Figures 1–3.Source: A. Prompts Used in the Report · p. 60

A.1. Figure 1 Column #1

  • Figure 1 prompts specify six detailed photographic scenes, including selfies, rugby action, an ancient alley, a pastoral waterside, Hongcun village, and Guangzhou fireworks.The prompts define subjects, composition, lighting, environments, and photographic styles across the six cases.Source: §A.1 · pp. 60–62
  • The Figure 1 prompts also impose precise visual and textual constraints, such as jersey labels, signage, Chinese characters, reflections, motion blur, and shallow or deep focus.Examples include “Turbo,” “MAI,” “Base,” “Edit,” “Tongyi Lab,” “世德堂,” and specified photographic conditions.Source: §A.1 · pp. 60–62

A.2. Figure 2 Row #1

  • Figure 2 prompts cover a magazine cover, an inspirational desert poster, and a visual portfolio poster combining rendered Chinese and English text with graphic design.The cases specify titles, subtitles, slogans, typography, logos, illustrations, silhouettes, and coordinated color schemes.Source: §A.2 · pp. 64–65
  • The inspirational poster prompt requires five independently rendered Chinese text elements integrated into a starry desert composition.The specified text includes a top slogan, a central phrase, a main title, a subtitle, and a small signature.Source: §A.2 · pp. 64–65
  • The portfolio poster prompt combines a 3D cartoon character, sunflowers, multiple text blocks, bilingual highlights, logos, and an anthropomorphic amplifier in a high-contrast design.Its highlighted claims include “Bilingual Rendering,” “Realism & Aesthetic,” and “Complexity & Elegance.”Source: §A.2 · p. 65

A.3. Figure 3 Row #1

  • Figure 3 editing prompts request outfit and bag transfer from Figure 1 to Figure 2, style transfer from Figure 1, and a new fantasy-realistic scene with a girl, cat, and Luffy.The final case specifies autumn floral accessories, colored jewelry, a white cat, Luffy’s red straw hat and sunglasses, cherry blossoms, maple leaves, and cinematic lighting.Source: §A.3 · p. 70

Related papers

Loading 2511.22699v5…