Qwen VLo: Alibaba's Unified Multimodal Model for Understanding and Generation 64

What is Qwen VLo?

Qwen VLo is a multimodal model launched by Alibaba Cloud's Tongyi Qianwen on June 27, 2025. It supports image generation and editing, style transfer, complex instruction execution, multi-image understanding and generation, image detection and annotation, as well as direct text-to-image capabilities.

Core Features

• High-quality image generation and editing: Generates and modifies images based on text or image prompts, supports progressive construction of complex scenes and real-time adjustments, and can produce multilingual images using simple instructions.

• Dynamic resolution adaptation: Uses dynamic resolution technology to generate images suited for different scenarios such as posters, illustrations, web banners, and social media covers.

• Progressive generation: Builds images gradually from left to right and top to bottom, continuously optimizing content during generation to ensure visual harmony and consistency.

• Open-ended instruction editing: Supports creative requests expressed in natural language, such as "change to Van Gogh style," "make it look like a 19th-century old photo," or "add a sunny sky." The model can respond flexibly and deliver results that match expectations.

• Multilingual instructions: Supports instructions in Chinese, English, and other languages, making it convenient for users worldwide.

Technical Features

• Deep vision-language fusion: Understands visual content and performs high-quality re-creation, maintaining semantic consistency during generation with strong detail capture capabilities.

• Strong visual editing capability: Handles complex editing instructions, completing multiple types of modifications within a single instruction while ensuring accurate subject structure and context.

Application Scenarios

• Design and creativity: Provides designers, marketers, and educators with tools that support artistic style transfer, scene reconstruction, detail optimization, and more, helping quickly bring ideas to life.

• Multilingual content creation: Meets the content production needs of users from different linguistic backgrounds.

In addition, Qwen VLo's input and output both support images of any resolution and aspect ratio, adapting to a variety of use cases. Users can view the generation process in real time and make adjustments. The model is currently still in preview stage.

FAQ

Question 1: Where does Qwen VLo surpass previous multimodal models in generation capability?

Answer: Qwen VLo's advantages lie in: more precise detail capture and semantically coherent generated content; the ability to respond to creative instructions such as artistic style transfer and complete complex tasks; and a progressive generation mechanism that delivers better visual results and more flexible creation.

Question 2: What image operations can Qwen VLo perform?

Answer: Qwen VLo can: generate images directly; modify images, such as changing backgrounds or adding elements; transfer styles, for example Ghibli style or 3D style; perform visual grounding tasks like detection and segmentation; handle complex instructions with multiple operations to complete multi-step tasks; generate images with dynamic aspect ratios (extreme aspect ratio functionality is not yet available); and re-understand its own generated content, such as identifying animal breeds.

Question 3: What limitations does Qwen VLo have in its preview stage?

Answer: During the preview stage, Qwen VLo may exhibit issues such as inaccurate generation, inconsistency with the original image, failure to follow instructions, and unstable intent understanding.

Blog: https://qwenlm.github.io/blog/qwen-vlo/