Qwen-Image-2.1 release notes / what's new

Qwen-Image-2.1 unifies image generation and editing in a 7B parameter model

Qwen-Image-2.1 is an open-source image model that integrates text-to-image generation and image editing into a single architecture. It features a visual generation component with 7B parameters and 32 Single-Stream DiT layers, balancing high generation quality with inference efficiency and reduced computational cost.

Native transparency and unified creation

Qwen-Image-2.1 provides native support for generating and editing transparent images (RGBA), unifying capabilities previously found in dedicated models like Qwen-Image-Layered.

  • Text-to-Transparency: The model uses prompts to determine whether to output a standard RGB image or an image with a transparency channel.
  • Transparent Layer Editing: Users can edit elements within a transparent layer—such as changing a subject's expression or modifying text (e.g., replacing "BLOOM" with "Qwen-Image")—while maintaining the transparent background.
  • Subject Extraction: The model can process an RGB photograph and extract a specific subject as an RGBA layer with transparency for use in design and composition.

Advanced image editing capabilities

Qwen-Image-2.1 introduces several enhancements to image editing, focusing on reference capacity, local control, and fidelity.

Multiple reference images

The model supports up to 10 reference images, allowing for the combination of multiple distinct assets into a single coherent composition. Examples include:

  • Group Portraits: Combining six individual portraits into one group photograph.
  • Virtual Try-on: Merging five inputs (model, clothing, shoes, bag, and hat) into a complete outfit.
  • Interior Design: Using 10 images of furnishings to generate a complete room arrangement.

Local editing and region selection

Qwen-Image-2.1 offers three methods for specifying edit regions:

  1. Circles: Using different colored circles to identify multiple regions for simultaneous edits (e.g., removing a watch, changing hair color, and replacing clothing).
  2. Painted Annotations: Marking specific areas in white to add new elements, such as adding a diver to a specific part of an image.
  3. Separate Masks: Accepting the original image and a separate mask as two distinct inputs to preserve the original content while editing only the masked region.

Sequential local edits can also be used to create simple animations by making successive changes while preserving the rest of the scene.

Fidelity and task coverage

The model improves identity preservation for portraits (facial features) and consistency for products (text, textures, and shape). It also expands its task coverage to include:

  • Panoramas: Generating a panorama from a single selfie.
  • Infographics: Expanding a model photograph into a detailed composition.
  • Storyboards: Turning a three-view character reference into a complete storyboard.

Inference efficiency and architecture

To optimize inference, particularly when using multiple reference images, Qwen-Image-2.1 employs a mixed-granularity attention architecture.

  • Attention Masking: Text, system prefixes, and editing instructions use a token-level causal mask, while image generation uses a chunk-level mask.
  • KV Cache Reuse: Input images and editing instructions are treated as static context, computed and cached during the first step to reduce memory usage and increase speed.

Visual aesthetics and typography

Qwen-Image-2.1 features refined aesthetics with a focus on typography and portraiture. Text rendering now accounts for type styles, layout, and the relationship between text and the overall composition. Portrait generation is further improved through enhanced lighting and fine details for a more realistic appearance.

Sources