Qwen-Image-2.1 brings text-to-image generation, local editing, and transparent-layer handling into one open model. Qwen lists a 7B-parameter visual generation component and says the release improves inference efficiency, typography, and detail. A notable addition is native transparency: the model can generate images with transparent backgrounds or extract a subject from an ordinary photo as an RGBA layer for later layout and compositing. Read Qwen’s announcement and examples.

Qwen official mark
Qwen official mark; click to open the announcement.

What it can help create

The model accepts up to 10 reference images to combine people, clothing, products, or interior elements. For local edits, users can mark a region with circles, paint, or a separate mask. For product imagery, ask for a background change or a targeted edit while explicitly preserving the logo, texture, and product silhouette. For illustration, revise an expression, text, or color in one area instead of regenerating the full frame. Transparent layers are useful for stickers, character assets, ads, and ecommerce cutouts.

Combining several references does not guarantee perfect identity or product fidelity. Inspect masks, similar-looking subjects, reflective surfaces, and small text carefully. Do not publish a generated brand mark, label, or price without checking it; packaging text and product structure can change unintentionally.

A reproducible editing workflow

  1. Start with one source image and state what must remain and what may change—for example, preserve the bottle proportions and label while replacing the background with a plain tabletop.
  2. For a local edit, mark the region with a circle, paint, or a separate mask, and provide the original image with that mask.
  3. For multiple objects, explain what each reference contributes and how the objects relate; do not rely on a pile of images without constraints.
  4. Check transparency edges, spelling, color, and shape. Open the downloaded output over a checkerboard in your design tool to confirm the alpha channel rather than trusting a white preview background.

Open weights and deployment

Qwen says it is releasing the model weights and provides official demos and code entry points. Inference speed, VRAM use, and commercial terms depend on the specific checkpoint, runtime, hardware, and license; the announcement does not give one standard API price. Before local or product deployment, read the model card and repository license, then measure latency, memory, and quality with your real image sizes and batch patterns.