Qwen Image 2.1 Turbo

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

What is Qwen Image 2.1 Turbo?

Qwen Image 2.1 Turbo is the fast, cost-efficient tier of Alibaba Qwen's Qwen-Image-2.1, an open-weight model that unifies text-to-image generation and image editing in a single checkpoint. It turns a prompt into a finished image and, just as easily, edits an existing one, so you do not switch between a generator and a separate edit model. Its standout trait is native transparency: ask for a transparent background and the model returns a real RGBA PNG with a true alpha channel, no background-removal pass required. The Turbo tier keeps the same 7B visual-generation architecture and feature set while running a shorter sampling schedule for speed and lower cost.

Key Features

• •Native transparent (alpha) PNG output straight from a text prompt, plus transparent-layer editing.
• •Multi-reference editing with up to 10 reference images to combine people, garments and props while preserving identity.
• •Unified generation and editing: local recolor, object edits, and subject extraction into clean RGBA cutouts.
• •Refined typography for English and Chinese, with attention to type style, layout and composition.
• •Native 2K output up to 2048 x 2048 across square, portrait and landscape aspect ratios.
• •Reproducible seeds and prompt_extend auto-rewriting (direct for editing, agent for text-to-image).

Best Use Cases

Designers use Qwen Image 2.1 Turbo to generate stickers, logos, UI assets and product cutouts without a separate background-removal tool. E-commerce and marketing teams compose virtual try-ons and product scenes from several reference photos, swap outfits, and recolor SKUs while shape, texture and branding stay intact. In testing, a single reference photo recolored a scooter from red to navy while leaving chrome, wheels and background untouched, and a cluttered food photo yielded a clean transparent cutout of just the cake. The typography strength suits posters, packaging, social graphics and infographics where readable text must sit naturally in the layout.

Prompt Tips and Output Quality

Name concrete subjects and, for transparency, spell out transparent background, real alpha channel, no background fill, and turn prompt_extend off so the instruction survives. For edits, pass references in order and address them as the first image and the second image. Output is fast and crisp; verify long or tiny text and inspect the alpha channel, since prompts that name a background can return an opaque result.

How to Use Qwen Image 2.1 Turbo

Qwen Image 2.1 Turbo handles two jobs from one endpoint: generating images from a prompt and editing existing ones. Leave the images field empty for text-to-image; add one or more reference images to edit, combine or extract. The prompt is your main lever, so describe concrete subjects, and for any transparent result spell out transparent background, real alpha channel, no background fill.

Recommended Settings

• •Transparent assets (stickers, cutouts, UI): prompt_extend false so the alpha instruction is preserved, size 1328*1328, save the output as PNG.
• •Editing and composition: place references in the images array (up to 10), keep prompt_extend_mode direct (agent is text-to-image only), and name what must stay the same so the edit stays local.
• •Resolution: every size bills the same; use 20482048 for 2K, 1664928 for 16:9, 928*1664 for 9:16.
• •Reproducibility: fix seed (12345 in these examples) to repeat a result; change it for variations.

Example edits and prompts to try

Each example below ran end to end at the settings shown, with identity, framing and background carried over exactly as described.

• •Native transparent generation (default). 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 result is a crisp die-cut sticker on real transparency. Tip: keep prompt_extend off and save as PNG so the alpha channel survives.
• •Local recolor edit. 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." Only the body changed; everything else held. Tip: list what must stay the same to keep the edit local.
• •Subject extraction to a transparent cutout. Prompt: "Extract just the glazed lemon 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." The cake lifts off the cluttered scene with clean edges. Tip: explicitly name what to remove.

Reference-guided edits

The model takes up to 10 reference images; address each by position in the prompt.

• •Multi-reference outfit swap (two references). 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 first reference is the person, the second is the garment; face, hair and pose carried over. Tip: pass references in the exact order your prompt names them.

Further prompts the model is built for: "combine these product photos into one studio scene", "change the sign text to read OPEN", and "extract the subject onto a transparent background".

Does it really output transparency?

Yes, a genuine alpha channel from the model's RGBA pipeline, not a PNG trick.

How many reference images can I use?

Up to 10 in a single editing or composition request.

Is it the same as Qwen Image 3.0?

No. 3.0 is a hosted, closed line focused on information-dense layouts; 2.1 is the open-weight, unified model.

Can it edit transparent layers and extract subjects?

Yes, it edits RGBA layers and lifts subjects from photos into transparent cutouts.

What languages does the typography handle?

English and Chinese, with layout-aware text placement.