Skip to main content

Qwen-Image-2.1: 7B open-source model does it all

Qwen-Image-2.1: A 7B Open-Source Model That Does It All

On September 21, the Qwen team released Qwen-Image-2.1, a new open-source image model that packs text-to-image generation, transparent image creation, and a suite of editing tools into a single 7B-parameter framework. It's a lean setup that aims to balance quality, speed, and cost.

At its core, the model uses a 32-layer Single-Stream DiT architecture. Thanks to a mixed-granularity attention structure and KV Cache reuse, it handles multi-image inputs efficiently and keeps memory use in check — a big plus for developers running it on modest hardware.

Transparent images made simple

One standout feature is transparent image generation. Based on your prompt, the model automatically decides whether to output a standard image or an RGBA image with a transparent channel. You can tweak expressions or edit text while keeping the background transparent. And if you feed it a real photo, it'll extract the subject and produce a transparent layer — perfect for design work and compositing.

Image

You can modify the subject's expression while keeping the transparent background.

Image

Text within the transparent layer can also be edited — for example, changing "BLOOM" to "Qwen-Image".

Editing that feels natural

When it comes to editing, Qwen-Image-2.1 accepts up to 10 reference images, letting you blend multiple subjects and materials into a cohesive scene. Local edits can be done via selection, brushing, or independent masks — so you can specify exactly what to change. The model also maintains impressive consistency for faces and products: after editing, facial features stay realistic, and product text and shapes remain intact.

Beyond that, the model has made strides in text rendering, lighting on people, and fine details. It handles panoramic images, infographics, and storyboards with ease. Right now, Qwen-Image-2.1 is free to use on GitHub, ModelScope, and Hugging Face.

Key Points

  • Qwen-Image-2.1 is a 7B open-source model combining text-to-image, transparent image generation, and editing.
  • It uses a 32-layer Single-Stream DiT with mixed-granularity attention and KV Cache reuse for efficient inference.
  • Transparent image generation supports expression tweaks and text edits while preserving transparency.
  • Editing accepts up to 10 reference images and offers selection, brushing, and mask-based local edits.
  • Available for free on GitHub, ModelScope, and Hugging Face.