Qwen VLo: From 'Understanding' to 'Expression'

Qwen VLo goes beyond existing multimodal models that understand image content, functioning as a unified multimodal understanding-generation model that directly generates and edits images based on that understanding. It is currently in preview version, and is available on Qwen Chat.

Its generation method uses progressive generation, which builds up images incrementally from left to right, top to bottom. This approach continuously refines predictions to enhance consistency and harmonious results, allowing users finer control over the output.

The key point is that both understanding and generation capabilities have been strengthened together. It reduces the semantic distortion or structural loss that previous models often suffered from—for example, when asked to change the color of a car photo, it naturally transforms the image while preserving the vehicle's model and shape as much as possible.

The range of supported tasks is also broad.

• Open-ended instruction editing in natural language: style transformation, scene reconstruction, detail modification
• Traditional visual tasks: generating depth map, segmentation map, detection map, edge information
• Compound editing: handling object modification, text editing, and background changes all at once

It also provides multilingual instruction support, including Chinese and English, allowing requests to be made the same way regardless of language. The article presents various demos including dogs, cartoons, photorealistic images, posters, and detection/segmentation results, showing that Qwen VLo is more than a simple generator—it is a tool that reinterprets and reconstructs based on understanding.