Qwen 开源 Qwen-Image-2.1 图像模型

Qwen-Image-2.1: Compact, Efficient, and Unified Image Creation | Qwen Blog Publication About Try Qwen Chat Qwen-Image-2.1: Compact, Efficient, and Unified Image Creation September 20, 2026 · 6 min · 1117 words · Qwen Team | Translations: 简体中文 GITHUB HUGGING FACE MODELSCOPE We are excited to open-source Qwen-Image-2.1 , an image model in the Qwen family that balances generation quality, inference efficiency, and cost. Qwen-Image-2.1 unifies text-to-image generation and image editing in a single model, with just 7B parameters in its visual generation component and native support for generating and editing transparent images. This update introduces four key improvements: Compact and efficient : A lightweight architecture and inference optimizations balance image quality and computational cost. Native transparency, unified creation and editing : Generate regular or transparent images from prompts, edit transparent layers, and extract subjects from photos. Versatile editing : Use up to 10 reference images , specify local edits, preserve people and products, and handle a wide range of tasks. Realistic textures and refined aesthetics : Improved typography, portrait lighting, and fine details make generated images more visually compelling. Compact and Efficient # Qwen-Image-2.1 uses a lightweight architecture. Its visual generation component contains 32 Single-Stream DiT layers and 7B parameters . Despite its compact size, the model delivers strong image generation quality. The Qwen-Image-Bench comparison below shows its performance alongside other open-source and closed-source models. Qwen-Image-Bench evaluation comparison Inference efficiency is another focus, particularly when editing with multiple input images. This optimization comes from a mixed-granularity attention architecture . Text, including the system prefix and editing instructions, uses a token-level causal mask, while image generation uses a chunk-level mask. KV cache reuse allows input images and editing instructions to serve as static context, computed and cached in the first step to improve inference efficiency and reduce memory usage. Qwen-Image-2.1 mixed-granularity attention architecture Native Transparency, Unified Creation and Editing # Native transparency is a major addition in this release. In December 2025, we introduced Qwen-Image-Layered as a dedicated model supporting transparent image generation. Qwen-Image-2.1 now integrates this capability into a unified model, using the prompt to determine whether to output a regular image or an image with a transparency channel. Here are examples of transparent images generated directly from text:

Beyond individual subjects, the model can generate more complex transparent images composed of multiple elements, opening up further possibilities for design and asset creation. By unifying generation and editing, Qwen-Image-2.1 also supports editing transparent images directly. For example, it can change a subject’s expression while retaining the transparent background. The input is on the left and the edited result on the right. Text within a transparent layer can also be edited. In the following example, “BLOOM” is replaced with “Qwen-Image”. This capability extends to real photographs. Given an RGB image, the model can extract the desired subject as an RGBA layer with transparency, making it easier to reuse elements in subsequent design and composition work. Versatile Editing # Qwen-Image-2.1 improves editing in four areas: multiple reference images, local editing, fidelity preservation, and task coverage. Multiple References: Up to 10 Input Images # Qwen-Image-2.1 supports up to 10 reference images , combining multiple subjects and assets into a coherent composition. In the group portrait example below, six individual portraits on the left are brought together in one photograph. A group photograph generated from six portrait references For virtual try-on, five inputs—a model, clothing, shoes, a bag, and a hat—can be combined into a complete outfit. A complete outfit generated from five reference images For interior design, the model can use 10 images of furnishings to generate a complete room arrangement. An interior generated from ten furnishing references Local Editing: Flexible Region Selection # Qwen-Image-2.1 supports circles, painted annotations, and separate masks to specify where an edit should take place. The first example uses circles of different colors to identify three regions at once: “Remove the metal watch in the blue circle, change the hair in the red circle to black, and replace the area in the green circle with gray short-sleeved linen pajamas.” The annotated input is on the left, and the result is on the right. Painted annotations provide another way to identify a region. In this example, the model adds a diver to the area marked in white on the right side of the input image. Circles and painted annotations obscure some of the original content. To preserve the complete input, Qwen-Image-2.1 also accepts the original image and a separate mask as two inputs . The following example asks the model to generate a cowboy on horseback, using these two images: The model edits the region specified by the mask, producing the following result: Mask-guided editing result: a cowboy on horseback Local editing can also support sequential creation. By making successive changes while preserving the rest of the scene, edited images can be assembled into simple animations, as illustrated by the capybara Calabash Brothers example below. Capybara Calabash Brothers: animation through successive local edits Fidelity: Preserving People and Products # Fidelity improvements focus on portrait identity and product consistency . For portraits, the model better preserves facial features so that a person’s identity remains more consistent across edits. Each pair below shows the input on the left and the edited result on the right.

For product editing, the model aims to preserve text, textures, and shape, keeping the product’s defining features consistent in a new image.