Qwen-Image-Edit Release Notes

Qwen-Image-Edit is a specialized image editing model built upon the 20B Qwen-Image foundation. It enables high-quality semantic and appearance editing, as well as precise bilingual text modification, by simultaneously utilizing Qwen2.5-VL for visual semantic control and a VAE Encoder for visual appearance control.

Dual-Control Architecture for Semantic and Appearance Editing

Qwen-Image-Edit employs a dual-input mechanism to handle two distinct types of image modification: semantic editing and appearance editing.

Semantic Editing

Semantic editing modifies image content while preserving the original visual semantics and identity. This allows for high-level changes where overall pixels may change, but the core subject remains consistent. Key capabilities include:

  • IP Creation and Consistency: The model can generate diverse content featuring a consistent character (e.g., Qwen's Capybara mascot) across different scenarios, such as creating themed emoji packs.
  • Novel View Synthesis: The model can perform object rotations of 90 degrees or full 180-degree rotations to reveal the back side of an object.
  • Style Transfer: Portraits can be transformed into various artistic styles, such as Studio Ghibli, for use cases like virtual avatar creation.

Appearance Editing

Appearance editing focuses on low-level visual modifications where specific regions are altered while all other areas of the image remain completely unchanged. Key capabilities include:

  • Element Modification: Adding or removing specific elements, such as inserting a signboard with a corresponding reflection or removing fine hair strands.
  • Precise Color Adjustment: Modifying the color of specific small elements, such as a single letter in a word.
  • Environmental Changes: Adjusting backgrounds or changing a person's clothing.

Precise Bilingual Text Editing

Qwen-Image-Edit extends the text rendering capabilities of Qwen-Image to support precise bilingual (Chinese and English) text editing. The model can add, delete, or modify text within an image while preserving the original font, size, and style. This capability allows for the correction of complex text elements, including large headlines and small, intricate details in Chinese posters.

Chained Editing Workflow

Qwen-Image-Edit supports a chained, step-by-step editing approach to correct complex errors. By drawing bounding boxes on an image to mark regions for correction, users can iteratively refine the output. For example, in a calligraphy artwork, the model can be used to correct specific character errors one by one, fine-tuning obscure components of a character until the final result is accurate.

Performance and Availability

Evaluations on multiple public benchmarks indicate that Qwen-Image-Edit achieves state-of-the-art (SOTA) performance in image editing tasks. The model is available via Qwen Chat under the "Image Editing" feature, and weights are hosted on GitHub, Hugging Face, and ModelScope.

Sources