UniWorld V2 - A New Generation of Image Editing Models Launched by Rabbit Show Intelligence in Association with Peking University

What is UniWorld V2

UniWorld V2 is a new generation of image editing model jointly launched by RabbitZhan Intelligence and UniWorld team of Peking University. It has significant advantages in the field of image editing, especially in Chinese comprehension and execution of complex commands. The model can accurately render artistic Chinese fonts, support fine-grained spatial control and global light and shadow fusion, and can complete difficult image editing tasks, such as moving objects out of a specified area or re-lighting, etc. UniWorld V2 is based on the UniWorld-R1 framework, and achieves efficient training and optimization through sampling, MLLM scoring, and DiffusionNFT fine-tuning, among other techniques. It has achieved excellent results in a number of industry benchmark tests, demonstrating strong generalization capabilities and high-precision editing effects. The model is applicable to many fields such as advertising, film and television, e-commerce, etc., and can significantly improve the efficiency and quality of image creation.

Features of UniWorld V2

• Accurate Chinese Font Rendering: It understands and generates complex artistic Chinese fonts, such as "Moon Full of Mid-Autumn", etc. It maintains clarity and semantic accuracy even with difficult strokes and artistic styles, and allows users to make text changes with simple commands.
• Fine-grained spatial controlSupporting the designation of editing areas by drawing frames (e.g. red rectangular boxes), the model can strictly abide by spatial constraints and accomplish difficult and delicate operations such as "moving the bird out of the red box" to ensure accurate editing.
• Global Light Blending: A deep understanding of the "re-lighting the scene" and other commands, so that the object naturally integrated into the scene, the integration of light and shadow is very high, the picture is uniform and harmonious, to avoid the problem of inconsistent light and shadow.
• multitasking adaptationIt supports a variety of task types such as text editing, red box control, object adjustment, scene relighting, etc. It covers the whole process needs from basic modification to complex creation, and meets diversified design scenarios.
• Strong Chinese comprehension: Outstanding performance in rendering complex commands and artistic Chinese fonts, far more than other similar models, can accurately execute Chinese commands, suitable for image editing needs in the Chinese environment.
• High-precision editing and generalization capabilities: High-precision editing through a reinforcement learning framework that maintains core editing capabilities even on unseen data distributions, adapts to diverse scenarios, and ensures model stability and reliability.

Core Benefits of UniWorld V2

• Strong Chinese comprehensionDeeply optimized for understanding Chinese commands, it can accurately handle complex artistic Chinese font rendering, such as "Moonlight Mid-Autumn Festival" and other difficult content, which is significantly better than other similar models, and is especially suitable for image editing needs in the Chinese environment.
• Fine-grained spatial controlSupporting the designation of editing areas through tools such as the red box, the model can strictly adhere to spatial constraints and accomplish high-precision image editing tasks, such as "moving the bird out of the red box", to ensure the accuracy and flexibility of editing.
• Global Light Blending: A deep understanding of light and shadow commands, such as "re-light the scene", can make objects naturally blend into the background to achieve a highly unified and harmonious picture effect, avoiding the problem of inconsistent light and shadow.
• Multitasking AdaptabilityIt supports a variety of task types such as text editing, object adjustment, scene relighting, etc. It covers the whole process needs from basic modification to complex creation, and meets diversified design scenarios.
• Excellent generalization properties: Maintains core editing capabilities on unseen data distributions, adapts to diverse scenarios, and ensures model stability and reliability.
• High-performance training framework: Efficient training and optimization using the UniWorld-R1 framework, combining sampling, MLLM scoring, and DiffusionNFT fine-tuning techniques to improve training efficiency and allow the use of higher-order samplers.
• Open Source and Scalability: The code and model have been made public on GitHub and Hugging Face platforms to facilitate further research and application by developers and researchers with high scalability and community support.

What is the official website for UniWorld V2

• Github repository:: https://github.com/PKU-YuanGroup/Uniworld
• arXiv Technical Paper:: https://arxiv.org/pdf/2510.16888

Who is UniWorld V2 for?

• Advertising and marketing staffUniWorld V2 can efficiently perform tasks such as text rendering, image adjustment, and other tasks to enhance work efficiency for professionals who need to quickly generate creative images for advertising design, poster production, and marketing materials.
• Film, TV and Game Production Teams: Used for character design, scene construction and special effects production, it helps artists and designers quickly realize their ideas, reduce production costs, and increase the flexibility of content production.
• e-commerce practitioner: Apply to product image optimization on e-commerce platforms, such as product display image beautification, background replacement and light and shadow adjustment, to improve product attractiveness and enhance user experience.
• Educators and researchers: It can be used as a teaching tool to help students understand the principles of image editing and multimodal techniques; its open source code can be used by researchers for further academic research and model optimization.
• Creative DesignerUniWorld V2 provides powerful technical support to graphic designers and illustrators who need to efficiently perform complex design tasks such as artistic font rendering and image compositing.
• Technology Enthusiasts & Developers: Individuals and teams interested in image editing technology, secondary development or exploring new features through open source code, UniWorld V2 provides a wealth of resources for research and applications.