7B DiT · 32 single-stream layers Native RGBA Up to 10 reference images ZeroGPU grant pending

Qwen-Image-2.1

A unified text-to-image generation and image editing model — generate regular or transparent images from text, edit transparent layers, and extract subjects from photographs, all in one model.

Try it now ↓ Model card GitHub
⏳

This Space's own ZeroGPU app is built and ready, but not yet running. Creating a Gradio Space with GPU requires a PRO subscription; a free account can still apply for a community GPU grant, which has been requested on the Community tab. Until it is approved this page runs on Static hosting (free for everyone, no compute), and the live demo below is embedded from the official Qwen Space — its GPU time is provided there.

Live demo

Real inference on Qwen-Image-2.1, embedded from the official demo Space. First generation after an idle period may take a little longer while the GPU warms up.

Make your ZeroGPU allowance last

Every visitor gets a small daily ZeroGPU allowance, charged at the requested duration rather than the actual runtime — so a few expensive images can use it up. Inside the demo below:

  • Keep the Speed (1024px) radio selected. Switching to Quality (2048px) multiplies the per-image cost several times over.
  • Under Advanced Settings, clear the Negative Prompt box — it ships containing a single space, which the app counts as "set" and charges double per step.
  • Leave Customize output size unchecked. While it is off, the Resolution preset dropdown beneath it is ignored — it displays a 2K preset by default purely as a placeholder.
  • Steps are fixed at 40 and are not exposed in the UI.
The interface defaults to English; the 语言 / Language switch at the top turns it Chinese. The allowance resets on a rolling 24-hour basis.
Embedded from Qwen/Qwen-Image-2.1 · opens in a new tab if the frame is blocked by your browser.

What this model does

Four things it is genuinely good at, straight from the model card.

🎨

Text to image

7B parameters with mixed-granularity attention and prefix KV cache reuse — strong quality at low compute cost. Supports 1:1 through 16:9 at 2K.

🔸

Native transparency

Generates real RGBA output from text and edits transparent layers directly, instead of matting a finished image.

🖌

Versatile editing

Up to 10 reference images, local edits via circles, painted annotations or masks, with identity preserved for people and products.

🔤

Text & texture

Improved typography, portrait lighting and fine detail for more visually compelling results.

Native transparent generation

Real alpha channel — shown over a checkerboard so you can see the transparency. Official examples from the model card.

Transparent RGBA generation example
RGBA from text
Transparent RGBA generation example
RGBA from text
Transparent RGBA generation example
Transparent layer editing

Text rendering

Typography is one of the headline improvements in this release.

Text rendering example

Prompt helper

Paste a prompt here to get the exact wording this model wants for a transparent background. The same formatting is what the bundled Gradio app applies when you tick its transparency checkbox.

The model follows this exact phrasing — keep the sentence order intact.