Qwen Image 2.1 Pro

Qwen Image 2.1 Pro — Text-to-Image Generation and Editing

What is Qwen Image 2.1 Pro?

Qwen Image 2.1 Pro is the Alibaba Qwen team's unified image model that generates images from text and edits existing images in a single checkpoint. Its compact 7B visual generator keeps iteration fast while delivering a standout capability: native transparency. The model writes true RGBA PNGs with an alpha channel straight from a prompt, so stickers, icons, and product cutouts arrive ready to use without a separate background-removal pass. It also edits with up to 10 reference images, which makes multi-subject composition, virtual try-on, and group photos straightforward.

Key Features

• •Native transparent (alpha) PNG output directly from a prompt
• •Unified text-to-image generation and image editing in one model
• •Up to 10 reference images for multi-subject composition and identity preservation
• •Subject extraction from ordinary photos into reusable RGBA assets
• •Native 2K output (up to 2048x2048) across many aspect ratios
• •Strong complex text rendering and typography, bilingual and multilingual

Best Use Cases

Designers and developers use Qwen Image 2.1 Pro to generate die-cut stickers, app and game UI art, and ecommerce hero cutouts without manual masking. Its multi-reference editing suits fashion try-on (assemble an outfit from separate garment shots), product-in-scene composites, and moodboard-consistent ad sets, because the model holds identity and product details across references. The unified generate-and-edit design replaces separate diffusion, matting, and inpainting steps: recolor an object, swap a background, add or remove elements, or extract a subject, all through one API. Strong text rendering extends it to posters, infographics, and storyboards.

Prompt Tips and Output Quality

For transparent output, ask explicitly for a transparent background and a real alpha channel, and keep prompt_extend off — with it on, the model can paint a checkerboard instead of producing true transparency. For edits, pass reference image URLs in order and name exactly what to keep unchanged so edits stay local. In testing, transparency, multi-reference composites, appearance edits, and photo-to-RGBA extraction all returned clean results on the first attempt. Typography is strong but can carry occasional character errors, so proof the final copy.

How to Use Qwen Image 2.1 Pro

Qwen Image 2.1 Pro works in two modes from one endpoint: send a prompt alone to generate an image, or add reference images to edit and compose. The parameter that most shapes your result is prompt_extend, so set it deliberately per task.

Recommended Settings

• •Transparent assets (text-to-image): describe a die-cut subject and add "transparent background, real alpha channel"; set prompt_extend to false; size 1328*1328. With prompt_extend on, the model paints a checkerboard instead of real transparency.
• •Editing and composition: pass 1 to 10 image URLs in images, keep prompt_extend false and prompt_extend_mode direct (agent is text-to-image only), and name what to preserve so edits stay local.
• •Size: every size bills the same and does not change quality, so choose by aspect ratio — 13281328 square, 1664928 wide, 9281664 tall, up to 20482048.
• •seed fixes reproducibility; watermark stays false for clean assets.

Examples to try

Each example below was run end to end at the settings shown, and the parameter that carried the result is noted.

• •Native transparent sticker (text-to-image). Prompt: "A die-cut glossy vinyl sticker of a potted monstera deliciosa houseplant in a terracotta pot, bold clean white sticker border, flat illustration, on a fully transparent background with a real alpha channel, no background fill." The output was a true RGBA PNG with see-through split-leaf holes. Tip: keep prompt_extend off or the model paints a grey checkerboard.
• •Single-image appearance edit (needs one source image). Prompt: "Repaint the scooter's body from red to brushed matte navy blue and add one thin white racing stripe along the side panel. Keep the chrome, wheels, kickstand, background street and lighting exactly the same." The body changed colour while the chrome, seat, and street stayed identical. Tip: list every element to keep unchanged.
• •Subject extraction to RGBA (needs one source image). Prompt: "Extract just the glazed bundt cake (not the plate) from this photo and place it on a fully transparent background as a clean RGBA cutout. Remove everything else - the plate, counter, whisk, cutting board, lemons and towel. Output a PNG with a real alpha channel, no background fill, no checkerboard pattern." The cake was lifted onto transparency with the clutter removed. Tip: ask for a real alpha channel and keep prompt_extend off.

Reference-guided composition (up to 10 images)

Qwen Image 2.1 Pro accepts up to 10 reference images in the images array, addressed by their order in the list.

• •Multi-reference outfit (two references: first a person, second a jacket). Prompt: "Dress the person in the first image in the olive-green bomber jacket shown in the second image. Keep her face, hairstyle and pose exactly the same. Plain light-grey studio background." The exact jacket was placed on the subject with face and pose preserved. Tip: refer to each image by position and name what to preserve; 2 to 3 clean references are more reliable than many.

Further ideas the model is built for: "assemble a full outfit from separate shoe, bag and hat references", "combine six portraits into one group photo", or "turn this product photo into a transparent PNG".

How do I get a transparent background?

Ask for a transparent PNG and a real alpha channel in the prompt, and keep prompt_extend off.

How many reference images can it take?

Up to 10, applied in the order you supply them.

Can it both generate and edit?

Yes — one model handles text-to-image generation and image editing.

What resolution does it output?

Up to native 2K, 2048x2048, in square, portrait, and landscape ratios.

Does it render text?

Yes, with strong bilingual and multilingual text rendering for posters and layouts.