Qwen-Image-Edit: Higher Quality and More Efficient Image Editing

Qwen-Image-Edit is an image editing model built on 20B Qwen-Image, extending Qwen-Image's strength in text rendering to editing tasks as well. It feeds the input image into Qwen2.5-VL to control visual semantics, and uses a VAE Encoder to control appearance, supporting both semantic editing and appearance editing together.

The core features can be divided into three main categories.

• Semantic editing: Editing that maintains semantic consistency even when the entire pixel set changes, such as object rotation, style transformation, and IP character generation
• Appearance editing: Editing that changes only specific regions while keeping the rest unchanged, such as background changes, clothing modifications, and adding/removing elements
• Precise text editing: Directly adding, deleting, and modifying Chinese and English text while preserving the original font, size, and style as much as possible

In the examples, various IP images are created while maintaining the personality of a capybara mascot, novel view synthesis is performed through 90-degree and 180-degree rotations, and style transfer such as Studio Ghibli style is also demonstrated. In appearance editing, fine-grained manipulations are possible, such as adding a signboard and generating its reflection, removing minute elements like hair, or changing the color of specific text only.

Text editing performance is also emphasized. English poster copy can be accurately changed, and in Chinese posters, not only large headlines but also small text can be modified. Finally, a chain editing approach is introduced where errors in calligraphy works generated by Qwen-Image are designated by region using bounding boxes and corrected step by step, showing that even difficult Chinese characters can be fixed through multiple rounds of correction.

Ultimately, Qwen-Image-Edit aims to lower the barrier to image generation, presenting a direction to establish itself as a more sophisticated and efficient foundation model for image editing.