Qwen VLo: From "Understanding" the World to "Depicting" It

GitHub Pages · 1 year ago · 11

Alibaba's Qwen released Qwen VLo, a unified multimodal model that both understands and generates images, supporting image editing through natural language instructions like style transfer and object modification. The model uses progressive left-to-right, top-to-bottom generation and supports dynamic image resolutions with aspect ratios as extreme as 4:1, currently available as a preview in Qwen Chat. The addition of generative capabilities enables more flexible creative workflows and allows the model to verify its own understanding through intermediate outputs like segmentation and detection maps.