Qwen Image 2.1

Qwen Image 2.1 generates images with a real alpha channel straight from your prompt. Create transparent stickers, logos and lettered posters in 2K.

First Frame *

Upload First Frame

Last Frame *

Upload Last Frame

Prompt: 0 / 2048

Generate

Create Alpha-Channel Assets and Photo Edits with Qwen Image 2.1

Last verified: November 3, 2026

Qwen Image 2.1 is an open-weight image model from Alibaba's Qwen team. It was released on September 20, 2026 . It can return either a regular picture or an image with a real transparency channel, and the prompt decides which one you get . Text-to-image generation and image editing share one set of weights. The visual generation component has just 7B parameters and uses mixed-granularity attention with prefix KV cache reuse . A Qwen3-VL 8B vision-language encoder reads both your instructions and any condition images . In practice, stickers, logos and layered assets come out ready to composite, with no separate background-removal pass.

Native transparency is the big step up from the previous generation. Qwen launched Qwen-Image-Layered in December 2025 as a separate model for transparent images, and this release builds that ability into the unified model . It can generate RGBA images from text, edit existing transparent layers and extract subjects from photographs. For editing, it takes up to 10 reference images and understands local edit regions marked with circles, painted annotations or separate masks . Compared with its predecessor, the biggest visual gains are in the look of text and portraits, and inference with multi-image inputs is much faster .

The model generates natively at 2K, and its default size is 2048 × 2048 . On this page it runs in Text to image and Image to image. Both modes have a Background setting that switches the output between opaque and transparent.

Explore Qwen AI's Models

Qwen AI Qwen Image 2 Use in Text To Image Use in Image To Image

Capability Snapshot

Specs That Shape Every Render

The frames, inputs and alpha options you can set here, plus the architecture behind them.

Resolution Quality

1K or 2K (1K by default)

Aspect Ratio presets

1:1, 4:3, 3:4, 3:2, 2:3, 16:9, 9:16, 21:9 and 9:21, plus Auto in Image to image

Transparent output

Background set to Opaque or Transparent; Transparent works with PNG or WEBP

Reference images (Image to image)

1 to 10 per run

Prompt length

Up to 5,000 characters

Model core

7B-parameter visual generator with a 64-channel RGBA VAE

Check These Before You Generate

Six quick checks that prevent lost transparency, cramped compositions and rejected uploads.

Pair Transparent with PNG or WEBP

If Background is set to Transparent, Output Format has to be PNG or WEBP, because JPEG has no alpha channel to hold the cutout. Set both before you generate so the result keeps its transparency.

Write cutout prompts around the subject alone

This model reads the prompt to decide whether to output an RGBA image. Its model card suggests opening with wording like 'This is an RGBA image with transparency' and saying the background is transparent . Describe the isolated subject, its edges and its lighting, not a room behind it, so the prompt and the Transparent setting ask for the same thing.

Choose the frame before you compose

Aspect Ratio goes from square to ultra-wide 21:9 and tall 9:21. Write the scene for the shape you pick: spread subjects across the width for 21:9 and stack them top to bottom for 9:21. In Image to image the default is Auto, so switch to a fixed ratio when the image has to fit a set placement.

Prepare uploads within the limits

Images smaller than 300×300 pixels, files over 10 MB, and formats other than JPG, JPEG, PNG and WEBP won't upload. Export each reference inside those limits, and save transparent layers you want to edit as PNG or WEBP, since JPEG throws away the alpha channel.

Mark local edits on the image itself

The model understands edit regions drawn as circles or painted annotations, and it also accepts a separate mask. Draw the mark on your picture before you upload it, then say in the prompt exactly what should change inside that area.

Choose 2K when fine detail matters

The model works natively at 2K , so set Resolution Quality to 2K for small lettering, strands of hair and product textures. 1K is fine for quick concept passes where speed matters more than fine detail.

Choose Your Generator

Qwen Image 2.1 or Seedream 5.0 Lite: Match the Generator to the Job

Most rows compare the two generators as you'll find them on Vidofy: modes, prompt length, frame shapes, resolution tiers, uploads and transparent output. The last row gives each creator's own description of its model's design, so you can pick Seedream 5.0 Lite or Qwen Image 2.1 based on what you need to deliver.

8 Criteria 2 Options

Feature/Spec	Qwen Image 2.1	Seedream 5.0 Lite
Modes you can run	Text to image and Image to image	Text to image and Image to image
Prompt box limit	5,000 characters	2,048 characters
Aspect Ratio choices	1:1, 4:3, 3:4, 3:2, 2:3, 16:9, 9:16, 21:9, 9:21 (Image to image adds Auto)	1:1, 2:3, 3:2, 3:4, 4:3, 9:16, 16:9
Resolution Quality choices	1K, 2K (default 1K)	2K, 3K (default 2K)
Add Image uploads (Image to image)	1 to 10 images; JPG, JPEG, PNG, WEBP; up to 10 MB each; at least 300×300 pixels	1 to 8 images; JPG, JPEG, PNG, WEBP; up to 10 MB each; at least 300×300 pixels
Transparent background output	Background setting (Opaque or Transparent) in Qwen Image 2.1 Text to image and Image to image; Transparent needs PNG or WEBP	No Background setting in either mode of this generator
Images per run	1	1
Model design	Text-to-image generation and image editing unified in a single model	Unified multimodal image generation model

Feature Deep Dive

Where Each Generator Fits Your Workflow

Frame shape or pixel count

Qwen Image 2.1's ratio list includes cinema-wide and full-height phone frames, and Image to image adds an Auto option. That means a website hero or a vertical story can be composed for its final shape from the start, not cropped out of a 16:9 frame. Seedream 5.0 Lite gives up those extreme shapes but has a higher top Resolution Quality tier. That higher tier helps for large prints and detailed key visuals that fit one of the standard ratios both generators offer.

References, prompt room and alpha output

For multi-image edits, Qwen Image 2.1 accepts two more references per run and more than twice as much prompt text. That helps when your brief explains what each upload contributes and how the result should look. It is also the generator here with a Background setting, so you can make stickers, logos and packaging overlays without leaving the page. If your edit needs eight references or fewer and a short prompt, the two generators accept the same file formats and sizes.

Choose by Deliverable: Cutouts and Extreme Frames, or the 3K Tier

Use this quick guidance to pick the best option for your workflow.

How to choose: Pick Qwen Image 2.1 if you need a transparent background, an ultra-wide 21:9 or tall 9:21 frame, nine or ten reference images in one edit, or a prompt longer than 2,048 characters. Its generator here offers all four. Pick Seedream 5.0 Lite if you need the 3K Resolution Quality tier. Both run Text to image and Image to image, return one image per run, and accept the same upload formats and file sizes.

Go From Idea to Finished Image

Six steps from picking a mode to your generated picture.

Pick Text to image or Image to image

In the model list, choose Text to image to start from words alone, or Image to image to work from your own pictures.

Upload references with Add Image

In Image to image, use Add Image to upload one to ten pictures within the size and format limits. Text to image skips this step and goes straight to the prompt.

Write in the Prompt box or tap Enhance Prompt

Describe the scene, subject or edit in up to 5,000 characters. If you only have a rough idea, the Enhance Prompt button under the box offers fuller versions, and your text only changes when you pick one.

Set Aspect Ratio, Resolution Quality, Output Format and Background

Choose the frame (Auto is available in Image to image), 1K or 2K, a file type, and Opaque or Transparent. Transparent needs PNG or WEBP.

Adjust the optional Enhance Prompt switch and Seed

The Enhance Prompt switch is on by default and lets the model expand your prompt before generating. Turn it off to send your words exactly as written. Type a number into Seed if you want to set one yourself.

Click Generate

Press Generate to create one image. For another take, adjust the prompt or settings and run it again.

Frequently Asked Questions

What is Qwen Image 2.1 best used for?

It's built for assets that need a real alpha channel. Qwen Image 2.1 decides from your prompt whether to output a regular image or an RGBA one, so stickers, logos and product layers come out transparent with no separate cutout step . The same 7B-parameter visual generator also edits with up to 10 reference images , so one model can both create transparent pieces and revise the photos you use them with.

Can Qwen Image 2.1 remove the background from a photo?

Yes. The model can extract subjects from photographs as part of its transparency features . Open Image to image, upload the photo with Add Image, set Background to Transparent with PNG or WEBP output, and describe only the subject you want to keep, for example: "Extract the woman in the yellow coat as a clean cutout; the background is transparent."

Can I edit an image that already has transparency?

Yes. The model edits transparent layers natively, and Qwen points to editing text inside transparent images for compositing work . Upload the PNG or WEBP layer in Image to image, keep Background on Transparent with an alpha-capable Output Format, and describe the change, such as recoloring a logo, swapping a word or adding an outline.

How do I get clean, readable text in my images?

Text is one of the areas where this model improves most on its predecessor . Put the exact words in double quotes, as Qwen's own sample prompt does for a neon shop sign , and say where the text goes and what style it should be: headline, small caps or script. Keep each line short, and if you want your wording sent exactly as typed, turn the Enhance Prompt switch off.

Is Qwen Image 2.1 open source, and what license does it use?

Qwen released the model's weights openly under the Qwen Research License Agreement . You can read the agreement on the model's Hugging Face page. On this page you run it directly in the Text to image and Image to image modes.

Is this the same as Qwen-Image-2.1-Turbo?

No. Turbo is a separate checkpoint that uses its own saved 8-step schedule, while the original Qwen-Image-2.1 checkpoint has its own defaults. Both default to the 2048 resolution level . The generator on this page runs Qwen Image 2.1 in Text to image and Image to image.

How is the cost of a Qwen Image 2.1 run shown?

Every version in the model list shows the credits a run starts from, and any visitor can see this. Once you're signed in, the cost of the options you've picked is calculated and shown above the Generate button. In both Text to image and Image to image, Resolution Quality is the setting that changes the cost. Suggestions from the Enhance Prompt button don't use credits.