arXiv:2511.18050v1 Announce Type: new 
Abstract: Diffusion transformers have recently delivered strong text-to-image generation around 1K resolution, but we show that extending them to native 4K across diverse aspect ratios exposes a tightly coupled failure mode spanning positional encoding, VAE compression, and optimization. Tackling any of these factors in isolation leaves substantial quality on the table. We therefore take a data-model co-design view and introduce UltraFlux, a Flux-based DiT trained natively at 4K on MultiAspect-4K-1M, a 1M-image 4K corpus with controlled multi-AR coverage, bilingual captions, and rich VLM/IQA metadata for resolution- and AR-aware sampling. On the model side, UltraFlux couples (i) Resonance 2D RoPE with YaRN for training-window-, frequency-, and AR-aware positional encoding at 4K; (ii) a simple, non-adversarial VAE post-training scheme that improves 4K reconstruction fidelity; (iii) an SNR-Aware Huber Wavelet objective that rebalances gradients across timesteps and frequency bands; and (iv) a Stage-wise Aesthetic Curriculum Learning strategy that concentrates high-aesthetic supervision on high-noise steps governed by the model prior. Together, these components yield a stable, detail-preserving 4K DiT that generalizes across wide, square, and tall ARs. On the Aesthetic-Eval at 4096 benchmark and multi-AR 4K settings, UltraFlux consistently outperforms strong open-source baselines across fidelity, aesthetic, and alignment metrics, and-with a LLM prompt refiner-matches or surpasses the proprietary Seedream 4.0.

تم تقديم UltraFlux كنهج جديد لتحسين توليد الصور من النصوص، حيث يحقق جودة 4K أصلية عبر نسب أبعاد متنوعة. تتناول هذه الطريقة القيود الموجودة في المحولات الانتشارية الحالية من خلال اعتماد استراتيجية تصميم مشترك للبيانات والنماذج، باستخدام مجموعة صور تحتوي على مليون صورة تُعرف باسم MultiAspect-4K-1M، والتي تتضمن تسميات ثنائية اللغة وبيانات وصفية غنية لتحسين عملية العينة.

UltraFlux se ha presentado como un nuevo enfoque para mejorar la generación de imágenes a partir de texto, logrando calidad nativa 4K en diversos ratios de aspecto. Este método aborda las limitaciones de los transformadores de difusión existentes mediante una estrategia de co-diseño de datos y modelos, utilizando un corpus de 1 millón de imágenes conocido como MultiAspect-4K-1M, que incluye subtítulos bilingües y metadatos ricos para un mejor muestreo.

UltraFlux a été introduit comme une nouvelle approche pour améliorer la génération d'images à partir de texte, atteignant une qualité native 4K sur divers rapports d'aspect. Cette méthode s'attaque aux limitations des transformateurs de diffusion existants en adoptant une stratégie de co-conception de données et de modèles, utilisant un corpus d'images de 1 million connu sous le nom de MultiAspect-4K-1M, qui comprend des légendes bilingues et des métadonnées riches pour un meilleur échantillonnage.

UltraFlux has been introduced as a new approach to enhance text-to-image generation, achieving native 4K quality across various aspect ratios. This method addresses limitations found in existing diffusion transformers by employing a data-model co-design strategy, utilizing a 1M-image corpus known as MultiAspect-4K-1M, which includes bilingual captions and rich metadata for improved sampling.

UltraFlux: Data-Model Co-Design for High-quality Native 4K Text-to-Image Generation across Diverse Aspect Ratios

arXiv:2512.02790v1 Announce Type: new 
Abstract: With the rapid advances of powerful multimodal models such as GPT-4o, Nano Banana, and Seedream 4.0 in Image Editing, the performance gap between closed-source and open-source models is widening, primarily due to the scarcity of large-scale, high-quality training data and comprehensive benchmarks capable of diagnosing model weaknesses across diverse editing behaviors. Existing data construction methods face a scale-quality trade-off: human annotations are high-quality but not scalable, while automated pipelines suffer from error propagation and noise. To address this, we introduce a lightweight data pipeline that replaces multi-toolchains with an end-to-end model and a unified post-verification stage. For scalable quality control, we train a 7B dual-task expert model, \textbf{Qwen-Verify}, for efficient failure detection and instruction recaptioning. This pipeline yields \textbf{UnicEdit-10M}, a 10M-scale dataset spanning diverse basic and complex editing tasks. We also propose \textbf{UnicBench}, a general benchmark that extends beyond basic edits to explicitly assess spatial and knowledge-driven reasoning. To enable fine-grained diagnosis, we introduce novel metrics, including \textit{Non-edit Consistency} and \textit{Reasoning Accuracy}. Our analysis of mainstream models on UnicBench reveals their limitations and provides clear directions for future research.

تم تقديم مجموعة بيانات ومعيار جديد يسمى UnicEdit-10M لمعالجة الفجوة في الأداء بين النماذج متعددة الوسائط المغلقة المصدر والمفتوحة المصدر في تحرير الصور. تتضمن هذه المجموعة 10 ملايين إدخال، وتستخدم خط أنابيب بيانات خفيف ونموذج خبير مزدوج المهام، Qwen-Verify، لتحسين مراقبة الجودة واكتشاف الفشل في مهام التحرير.

Se ha introducido un nuevo conjunto de datos y un nuevo estándar llamado UnicEdit-10M para abordar la brecha de rendimiento entre los modelos multimodales de código cerrado y de código abierto en la edición de imágenes. Este conjunto de datos, que abarca 10 millones de entradas, utiliza un pipeline de datos ligero y un modelo experto de doble tarea, Qwen-Verify, para mejorar el control de calidad y la detección de fallos en las tareas de edición.

Un nouveau jeu de données et un nouvel étalon, nommé UnicEdit-10M, ont été introduits pour combler l'écart de performance entre les modèles multimodaux propriétaires et open-source dans l'édition d'images. Ce jeu de données, comprenant 10 millions d'entrées, utilise un pipeline de données léger et un modèle expert à double tâche, Qwen-Verify, pour améliorer le contrôle de la qualité et la détection des échecs dans les tâches d'édition.

A new dataset and benchmark named UnicEdit-10M has been introduced to address the performance gap between closed-source and open-source multimodal models in image editing. This dataset, comprising 10 million entries, utilizes a lightweight data pipeline and a dual-task expert model, Qwen-Verify, to enhance quality control and failure detection in editing tasks.

UnicEdit-10M: A Dataset and Benchmark Breaking the Scale-Quality Barrier via Unified Verification for Reasoning-Enriched Edits

arXiv:2511.22699v2 Announce Type: replace 
Abstract: 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.

تم تقديم Z-Image كنموذج فعال لتوليد الصور، يستخدم بنية مكونة من 6 مليارات معلمة تعتمد على Scalable Single-Stream Diffusion Transformer (S3-DiT). يهدف هذا النموذج إلى تحدي هيمنة الأنظمة المملوكة ذات المعلمات العالية مثل Nano Banana Pro وSeedream 4.0 من خلال توفير حل أكثر عملية للاستدلال والتعديل الدقيق على الأجهزة العادية.

Z-Image se ha presentado como un modelo de generación de imágenes eficiente, utilizando una arquitectura de 6 mil millones de parámetros basada en el Scalable Single-Stream Diffusion Transformer (S3-DiT). Este modelo busca desafiar el dominio de sistemas propietarios de alto parámetro como Nano Banana Pro y Seedream 4.0 al proporcionar una solución más práctica para la inferencia y el ajuste fino en hardware de consumo.

Z-Image a été introduit comme un modèle de génération d'images efficace, utilisant une architecture de 6 milliards de paramètres basée sur le Scalable Single-Stream Diffusion Transformer (S3-DiT). Ce modèle vise à défier la domination des systèmes propriétaires à haute échelle comme Nano Banana Pro et Seedream 4.0 en offrant une solution plus pratique pour l'inférence et le fine-tuning sur du matériel grand public.

Z-Image has been introduced as an efficient image generation foundation model, utilizing a 6B-parameter architecture based on the Scalable Single-Stream Diffusion Transformer (S3-DiT). This model aims to challenge the dominance of high-parameter proprietary systems like Nano Banana Pro and Seedream 4.0 by providing a more practical solution for inference and fine-tuning on consumer-grade hardware.

UltraFlux: Data-Model Co-Design for High-quality Native 4K Text-to-Image Generation across Diverse Aspect Ratios

Was this article worth reading? Share it

VibeFrame

Chroma

4o Image Gen