The Pulse
Qwen-Image 2.1 Released on ModelScope With Transparent Image Generation and Editing
Qwen-Image 2.1 is now available as an open-source image-generation and editing model on ModelScope, with support for transparent RGBA images, localized edits and multiple reference images.
AI.info Team ·
Qwen-Image 2.1 is now available as an open-source model on ModelScope. The listing identifies it as a unified text-to-image generation and image-editing system in the Qwen family, with a focus on image quality, inference efficiency and flexible visual workflows.
The model page lists Qwen/Qwen-Image-2.1 as the model identifier and provides download and inference information. It describes a 7-billion-parameter visual generation component built with 32 Single-Stream DiT layers. The page says the architecture is designed to balance generation quality, computational efficiency and versatility.
“We are excited to open-source Qwen-Image-2.1, a unified text-to-image generation and image editing model in the Qwen family.”
— Qwen, model publisher
Generation and editing in one model
Qwen-Image 2.1 supports both text-to-image generation and image editing. The ModelScope documentation includes a text-to-image example that generates a 2,048-by-2,048 image from a written prompt. It also provides an editing example in which an input image is modified through a natural-language instruction.
The model supports localized changes through circles, painted annotations and separate masks, according to the listing. Users can also provide as many as 10 reference images. The documentation says these capabilities are intended to help preserve the identity of people and products while refining selected areas of an image.
Native transparency and RGBA output
Unlike the earlier listing information, the released model documentation specifically describes native transparent-image generation. It says Qwen-Image 2.1 can generate regular or transparent RGBA images, edit transparent layers and extract subjects from photographs.
The quick-start instructions include a transparent-image workflow that uses a prompt referring to an RGBA image, an alpha channel and a transparent background. The resulting image can be saved through the provided Python example. The page also identifies native transparent image generation among the model’s showcase capabilities.
Image quality and technical improvements
The ModelScope description highlights improved typography, portrait lighting, realistic textures and finer visual details. It presents these changes as part of an effort to produce more convincing images and better preserve visual characteristics during editing.
The model uses mixed-granularity attention and prefix key-value cache reuse, which the listing associates with a compact and efficient architecture. The page says this design delivers strong image quality at lower computational cost, though it does not include benchmark tables comparing Qwen-Image 2.1 with earlier Qwen image models or competing systems.
Supported sizes and deployment
The documentation lists preset aspect ratios from square 1:1 images to wide 16:9 and vertical 9:16 outputs. The corresponding resolutions range from 1,536 by 2,752 pixels to 2,752 by 1,536 pixels, with several intermediate portrait and landscape formats also provided.
Quick-start instructions use PyTorch, Transformers, Diffusers, Accelerate and Pillow. They show model loading through the QwenImage21Pipeline and include a CPU-offload option for memory management. The page also provides an inference API example for ModelScope, including an asynchronous image-generation request and a task-status polling loop.
Qwen-Image 2.1 is licensed under the Qwen Research License Agreement. ModelScope’s listing includes links to additional project materials, example images and inference guidance, giving users a starting point for local generation, image editing and transparent-output workflows.