Qwen-Image-Layered: Layered Decomposition for Inherent Editability

Qwen-Image-Layered is a model capable of decomposing a single image into multiple RGBA layers, enabling inherent editability by physically isolating semantic or structural components. This approach allows users to manipulate specific layers independently, ensuring high-fidelity edits and consistency across the remaining image content.

Layered Decomposition and Image Editing

Qwen-Image-Layered transforms raster imagery into a structured representation by breaking an image down into separate RGBA layers. This physical isolation of components allows for precise editing operations where changes are applied exclusively to a target layer, leaving all other content untouched.

Key capabilities enabled by this layered representation include:

  • Recoloring: Changing the color of a specific layer without affecting the rest of the image.
  • Object Replacement: Replacing a specific layer (e.g., changing a girl to a boy) while maintaining the background and other elements.
  • Text Revision: Modifying text within a specific layer.
  • Object Removal: Cleanly deleting unwanted objects by removing their corresponding layer.
  • Resizing and Repositioning: Resizing objects without distortion or moving objects freely within the canvas.

Flexible and Iterative Decomposition

The model does not rely on a fixed number of layers, providing flexibility in how an image is structured. Qwen-Image-Layered supports variable-layer decomposition, meaning a user can decompose an image into a different number of layers (such as 3 or 8) based on the specific needs of the task.

Furthermore, the model supports recursive decomposition. Any individual layer generated from an initial decomposition can be further decomposed into additional layers, enabling a process of infinite decomposition for increasingly granular control over image elements.

Technical Implications

By bridging the gap between raster imagery and structured, editable representations, Qwen-Image-Layered moves away from traditional global image editing toward a composable framework. This ensures that elementary operations—such as resizing and repositioning—remain high-fidelity and consistent because the structural components are physically separated.

Sources