Qwen Image 2.1 is Alibaba's compact 7B image model that generates and edits in one: native transparent PNGs, up to 10 reference images in a single edit, and mask-based local changes that leave the rest of the picture alone. Run Qwen Image 2.1 at 1K or 2K, pay per image.
Up to 5,000 characters in any language. With several references, name them by position — "the first image", "the second image". With a mask, describe only what belongs inside the white area.
Text to image defaults to 1:1. With references the default is auto, which snaps to the nearest supported ratio of the first image. For group shots and multi-product layouts a landscape ratio such as 3:2 or 16:9 usually works best.
2K has four times the pixels of 1K and costs twice as much — see the pricing table. 1K usually returns in seconds; 2K takes noticeably longer, especially with references.
For transparent output, describe only the subject — prompts that mention a scene, surface or backdrop may come back opaque. Transparent needs PNG or WebP.
PNG (default) and WebP carry an alpha channel; JPEG does not and cannot be combined with a transparent background.
On by default and it noticeably improves quality on short prompts. Turn it off when you have written an exact brief and want it followed literally.
Leave empty for a new random image each time. The same seed with the same prompt and settings reproduces the same image.
wire field: content_filter (boolean) · default true · API only
Try one of these prompts
Credit-based — 1 credit = $0.001 USD. Pay only for completed generations.
One sample at the cheapest tier (1 image).
Qwen Image 2.1 is the Qwen team's unified image model, released with open weights in September 2026. It folds text-to-image, image editing and transparency into one 7B visual generator: create a regular or transparent image from a prompt, combine up to ten reference images into one scene, or repaint a single masked region while the rest of the photo stays put. On reAPI, Qwen Image 2.1 is one model id — the request shape decides what it does.
Real-world workflows and production use cases you can build and ship with this model.
Most image models give you a picture and leave the cutout to a separate background-removal step. Qwen Image 2.1 writes a real alpha channel into the image itself, so a sticker, a product shot or a UI element comes back as a transparent PNG or WebP. Qwen Image 2.1 can also edit transparent layers and pull a subject out of an ordinary photo as an RGBA asset — the same request shape, just with a reference image attached.
Generate a transparent image
Qwen Image 2.1 accepts up to ten reference images in a single edit and composes them into one result: six portraits into a group photo, a model plus a jacket, shoes and a bag for virtual try-on, or separate pieces of furniture into a finished room. You refer to each input by position in the prompt. The more references you add the less detail each one keeps, so Qwen Image 2.1 works best with four or fewer when fidelity matters.
Send one reference image and a black-and-white mask, and Qwen Image 2.1 edits only the white area — swap a product, remove an object, restyle a sleeve — while the rest of the frame stays as it was. Without a mask, Qwen Image 2.1 still aims to preserve what defines the subject: a person's identity in a portrait, and the shape, texture and printed text of a product across variations.
Background removal is usually a separate model, a separate call and a separate failure point. Qwen Image 2.1 carries transparency inside the model, so the cutout and the image are the same output — which is exactly what sticker, e-commerce and design-asset pipelines need.
Qwen Image 2.1 is one model id on one endpoint. Leave out reference images and it generates; add them and it edits; add a mask and it edits locally. The parameters stay the same, so your integration does not branch per task.
The visual generation part of Qwen Image 2.1 is 7B parameters — 32 single-stream DiT layers with mixed-granularity attention and prefix KV-cache reuse. Alibaba's stated goal for Qwen Image 2.1 is strong image quality at low computational cost.
Both models are Alibaba Qwen image generators that also edit, and both are on reAPI at a similar per-image price. Qwen Image 2.1 is the newer release and widens almost every input: longer prompts, many references, masks and transparency. Qwen Image 2 remains the simpler single-image editor.
Comparison reflects the documented request parameters of both models on reAPI at the time of writing. It describes capabilities, not measured quality scores.
Create a key from your reAPI dashboard. New accounts start with signup credits — enough to try Qwen Image 2.1 on a transparent asset or an edit before you top up.
OpenUse the playground on this page to pick text to image or image to image, set the background, ratio and resolution, then copy the request body. The form mirrors the Qwen Image 2.1 request one to one.
OpenSend the body to the images endpoint with your key, take the task id from the response, and poll until Qwen Image 2.1 returns the image URL. Failed generations are refunded automatically.
Opencurl -X POST https://reapi.ai/api/v1/images/generations \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen-image-2.1",
"prompt": "A cute cartoon dragon sticker, bold outline",
"background": "transparent",
"resolution": "2K"
}'Rejected synchronously with the failing field named. Check enum values, ranges, and URL fields — nothing is charged.
401 means a missing or invalid Bearer key; 402 means the reserve exceeds your balance. Manage keys and credits in the console.
Reference material and outputs pass automated moderation. Rejected tasks fail with a clear error and a full refund.
A task that reaches a failed state is never charged — the reserve refunds automatically. Keep the task id and retry when ready.
Common questions about this model.
Try it in the playground or grab an API key to integrate now.
