Qwen has open-sourced Qwen-Image-2.1, an image model that folds text-to-image generation and image editing into a single system. Its visual generation module carries 7B parameters and natively supports the creation and editing of transparent images.

According to the team, the release focuses on four areas: model scale, transparent image handling, multi-image editing, and visual quality, with the aim of balancing output quality, inference efficiency, and cost of use.

A 7B model balancing quality and inference efficiency

The visual generation side of Qwen-Image-2.1 uses a 32-layer Single-Stream DiT architecture with 7B parameters. The Qwen team says the model retains strong image generation ability at this smaller scale, covering tasks spanning people, products, scenes, and visual design.

For inference efficiency, the model uses a mixed-granularity attention structure that treats text and images differently. Text relies on token-level causal masking, while image generation uses chunk-level masking.

The model also introduces a KV cache reuse mechanism. Input images and editing instructions are treated as relatively fixed context, precomputed and cached at the first inference step to cut repeated computation and reduce memory overhead when multiple images are fed in.

Native support for transparent image generation and editing

Transparent imagery is one of Qwen-Image-2.1's key capabilities.

The team previously released Qwen-Image-Layered, a model dedicated to transparent image generation. That capability is now folded into Qwen-Image-2.1 as a single unified model.

Through text prompts, users can ask the model for ordinary images or for RGBA images with an alpha channel. It can also generate complex transparent assets built from multiple elements, suited to e-commerce design, advertising production, game assets, and visual creation.

On the editing side, Qwen-Image-2.1 can alter subject content while preserving a transparent background: adjusting facial expressions, swapping objects, or editing text within a transparent layer.

The model can also extract a specified subject from an ordinary RGB photo and output it as an RGBA layer with an alpha channel, making downstream compositing and design easier.

Up to 10 reference images per input

Qwen-Image-2.1 accepts as many as 10 reference images, allowing multiple subjects, people, or product assets to be merged into one picture.

Users can feed in several portraits to produce a group photo, or supply images of a model, clothing, shoes, a bag, and a hat at once to generate a complete styled outfit.

In interior design, the model can draw on multiple furniture images to produce a room scheme in a consistent style. This multi-image input reduces the need to describe assets over and over, and gives users more control over complex compositions.

Selection, brushing, and mask-based editing

For local edits, Qwen-Image-2.1 supports lasso selection, brushing, and standalone masks.

Users can mark several regions for revision at once using different colors: removing a watch inside a given area, changing a subject's hair color, and swapping a clothing style.

They can also brush directly over the area to be changed, asking the model to add a person or object within that region.

Where the full original image must be preserved, the model accepts the source image and a separate mask as two inputs. It then confines edits to the masked area, limiting spillover into untouched regions.

This local editing ability also lends itself to sequential creation. Users can revise the same image step by step, building continuous visual content while keeping other areas relatively stable.

Stronger fidelity for portrait and product edits

Qwen-Image-2.1 puts particular emphasis on portrait and product consistency.

For portrait editing, the model does more to preserve facial features, so a subject keeps a reasonably stable identity after a change of clothing, background, or overall image style.

For product editing, the model works to retain original text, texture, structure, and appearance, keeping products consistent across different scenes and backdrops, useful for e-commerce imagery, ad design, and product displays.

Covering panoramas, infographics, and storyboards

Beyond standard text-to-image and local editing, Qwen-Image-2.1 handles a range of more complex visual tasks.

For panoramas, users can submit a selfie or ordinary photo and ask the model to extend the surrounding environment into a full panoramic scene.

For infographics, the model can generate complex layouts combining text, people, and assorted visual elements from a portrait photo or a thematic brief.

For storyboarding, users can input three-view character sheets and have the model produce a complete story sequence, supplying raw material for short video, animation, and visual narrative work.

Better typography and more lifelike figures

Qwen-Image-2.1 also refines text rendering, lighting on human subjects, and fine detail.

On text, the model looks beyond whether the wording is correct to how typeface, layout, and overall image style fit together, useful for posters, infographics, and product promotion images.

On people, it improves lighting relationships, facial detail, and material rendering for a more natural look.

Open-source ecosystem adapting in step

The Qwen team says Qwen-Image-2.1 is now open to developers, with community tooling being adapted alongside it. Developers can use the official code and model weights for local deployment, application development, and further optimization.

Taken together, Qwen-Image-2.1 builds on a 7B visual generation module to bring text-to-image, transparent image handling, and image editing into one model, with multi-image input, local editing, and portrait and product fidelity aimed at design, e-commerce, content creation, and visual storytelling.