Qwen-Image-2.1 Image to Image: Combine and Restage Photos

Last verified: November 1, 2026

Qwen-Image-2.1 works best on edits built from reference images. Typical jobs are restaging a product, dressing a model from separate garment shots or lifting a subject onto a transparent layer. It suits this work because one model handles composition from several input images, changes to marked regions and transparent results. With Qwen-Image-2.1 Image to Image on Vidofy, you can also frame those edits as an ultra-wide 21:9 banner or a tall 9:21 vertical, as well as the usual square, landscape and portrait shapes.

Qwen describes the model's editing in four areas: working from multiple reference images, local edits, keeping people and products faithful, and a wider range of tasks. Transparency is built in, so the model generates and edits transparent images natively. For a working creator, a product shot, a lookbook frame and a cutout for a layout can all start from photos you already have, without moving between separate tools.

On the form, your base photo and any references go into Add Image, and you describe the change in plain language in Prompt. Four settings then set the frame, the resolution, the file type and whether the background stays solid or becomes transparent. The wide and tall frames are worth choosing when an edit is going into a web banner, a cinematic header, a phone story or a vertical ad. For a simple touch-up, start with a frame that matches your original.

Explore more image to image

Image To Image About Qwen Image 2.1 Use in Text To Image Seedream 4.5 Edit Seededit V3 Nano Banana 2 GPT Image 2 Qwen Image 3.0 Pro Qwen Image 3.0
Capability Snapshot

Qwen-Image-2.1 Editing Specs at a Glance

The inputs, frames and file options you get for each edit.

Input

Up to 10 images (JPG, JPEG, PNG or WEBP, up to 10 MB each) + a text prompt of up to 5,000 characters

Resolution

1K or 2K

Aspect Ratio

Auto, 1:1, 4:3, 3:4, 3:2, 2:3, 16:9, 9:16, 21:9, 9:21

Output Format

PNG, JPEG or WEBP

Background

Opaque or Transparent

Output

1 image per run

Techniques for Cleaner Qwen-Image-2.1 Edits

Mark the Edit Area on the Image Itself

Qwen-Image-2.1 reads circles and painted annotations as instructions for where an edit should happen, so a local change doesn't need a separate masking step. Draw the mark in any image editor and upload that marked copy in Add Image. Then say in Prompt what should change inside it, for example "replace the mug inside the red circle with a glass of iced coffee". Pick a mark color that stands out clearly from the photo so there's no doubt which area you mean.

Name Each Upload by Its Position

When an edit uses more than one image, say what each picture is for. Qwen's own multi-image editing example names each input by its place in the list, asking for the subject of image1 to be placed into the scene of image2. Upload the base photo first, then write prompts like "keep the woman from image 1, dress her in the jacket from image 2 and the boots from image 3". That way every reference has a clear job and the model doesn't have to guess.

Expand Rough Ideas, Keep Exact Wording Exact

Qwen releases a separate rewriting model alongside Qwen-Image-2.1. It takes a vague edit instruction plus the input images and turns them into a precise, actionable prompt for the edit. On the form, the Enhance Prompt button offers fuller versions of a loose idea for you to choose from, and the Enhance Prompt switch lets the model do that expansion itself. When your words must appear exactly as written, such as a new product name or one small fix, consider turning the switch off and writing a precise instruction yourself.

How do I edit a photo with Qwen-Image-2.1?

Five steps from upload to download.

1

Upload your images in "Add Image"

"Add Image" needs at least one picture. Upload the photo you want to change first, then any references the edit should borrow from, such as a garment, a product, a backdrop or another person. With the base image first, it's easy to say in the prompt which picture is the canvas and which ones supply the new elements.

2

Write the instruction in "Prompt"

Describe the change, not the whole picture: what to add, replace or move, and what has to stay as it is. The "Prompt" box takes up to 5,000 characters, far more than a clear edit usually needs. If you only have a rough idea, click "Enhance Prompt" under the box to see fuller versions; your text changes only if you pick one. There's also an "Enhance Prompt" switch among the settings, marked (Optional) and on by default, which lets the model expand your prompt on its own before generating.

3

Choose "Aspect Ratio", "Resolution Quality", "Output Format" and "Background"

"Aspect Ratio" starts on Auto and also offers 1:1, 4:3, 3:4, 3:2, 2:3, 16:9, 9:16, 21:9 and 9:21. Use 21:9 for a wide banner or header and 9:21 for a tall phone-screen layout. "Resolution Quality" defaults to 1K and also offers 2K; stay at 1K while you refine the instruction and switch to 2K for the version you keep. "Output Format" offers PNG, JPEG or WEBP. "Background" defaults to Opaque. Choose Transparent for a cutout and pair it with PNG or WEBP, because Transparent needs one of those two formats.

4

Click "Generate"

Each click of "Generate" returns one edited image, so treat the first run as a test of your instruction. If the change lands in the wrong place or goes too far, rewrite the prompt and run again. That works better than piling more instructions onto a single attempt.

5

Download the result

Download the image and check it at full size before you use it. Look closely at the edges of cutouts and at any lettering on products or signs, since those are the details that are easiest to miss in a preview.

Use Qwen-Image-2.1 Here or Run It Yourself?

How editing through the Vidofy form compares with running the open weights in your own code.

Getting to a first edit Everything happens in the form: you upload images, write a prompt, choose the four output settings, click Generate and download the result. Running the open weights yourself means loading the model through the Diffusers pipeline on a CUDA GPU, following the quick start in Qwen's official repository. Creators who want an edited image without maintaining hardware should use the form, while developers who already run GPUs and want the model inside their own pipeline are better served by the local route.
Composing from several images You can put several images into Add Image for one run, and the prompt decides what each one contributes to the result. In code, the official pipeline takes a list of images, and the repository's example puts three separately supplied characters around a campfire in a forest. The form is the faster path for composites you direct by hand one at a time, while the local pipeline fits teams that need to generate many composites from scripted inputs.
Repeating a result The form has a Seed field that takes a number, which you can keep the same between runs while you adjust the prompt. Qwen's official editing example fixes the random generator with a manual seed so the same call can be reproduced. Anyone comparing prompt wording by eye can rely on the Seed field, while teams running automated tests of edits will set the seed in their own code.
Edit on Vidofy when you want results straight from a form. Run the open weights yourself only if you already maintain GPU hardware and need Qwen-Image-2.1 inside scripted pipelines.

Fixing Common Qwen-Image-2.1 Edit Problems

What usually goes wrong with uploads, cutouts, marked regions and variations, and how to fix it.

I picked Transparent, but the background came out solid

Transparency only survives in an output format that can store an alpha channel, and JPEG can't, which is why the form ties the Transparent background to the two formats that can. If Output Format is set to JPEG, switch it, keep Background on Transparent, and say in the prompt which subject should be kept on its own, for example the bottle without its table or the person without the street behind them. Then generate again; the source image doesn't need to change.

Add Image won't accept my file

Each upload must be in one of the accepted formats, under the per-file size cap and above the minimum pixel size, and a run can't take more than the maximum number of images. The exact limits are in the upload question in the FAQ. A phone HEIC photo or a TIFF scan isn't on the accepted list, so export it to an accepted format first. Compress or resize an oversized file, and swap a small thumbnail for the original. Then upload again and carry on; your prompt and settings don't need to change.

The area I circled lost its original detail

Circles and painted annotations cover part of the original image, so whatever sits under a thick stroke is partly hidden from the model and can come back changed. Redraw the mark as a thin outline around the area instead of painting over it, and keep strokes off faces, labels or anything that must stay exact. Name the marked area in the prompt so the instruction matches the mark, then upload the newly marked copy in Add Image and run the edit again.

I want several versions, but each run gives me one image

Each run returns a single image, so you get variations by running again. Keep the prompt and settings the same and change the number in Seed to get a different version of the same instruction. Once a version looks right, keep that Seed number while you fine-tune the wording. That way the differences you see come from your prompt changes rather than from chance.

Frequently Asked Questions

What images can I upload to Qwen-Image-2.1, and how many?

Add Image accepts JPG, JPEG, PNG or WEBP files of up to 10 MB each, at least 300×300 pixels. You need at least one image and can add up to 10 in a single run. That matches the number of reference images the model can combine in one edit, so a multi-reference composition can use its full range.

What makes one Qwen-Image-2.1 edit cost more than another?

The credit cost of a run is calculated from the options you choose and shown above the Generate button once you're signed in. Resolution Quality is the setting that changes it. Suggestions from the Enhance Prompt button don't use credits.

Will Qwen-Image-2.1 keep a person's face recognizable after an edit?

Portrait identity is one of the two areas Qwen focused on for fidelity. The model preserves facial features better, so a person stays more consistent across edits. In the prompt, say what should change, such as the outfit, the setting or the light, and say that the person's face stays the same, so your instruction backs up the model's own identity preservation.

Can Qwen-Image-2.1 change the words in an image?

Yes. Qwen shows the model replacing the word "BLOOM" with "Qwen-Image" inside a transparent layer while the transparency stays intact. Put the exact new wording in quotation marks in the prompt, name the text it replaces, and proofread the download letter by letter before it goes into a design.

Can I keep editing the same image over several rounds?

Yes. Local editing also supports sequential creation, so an image can be built up through successive edits. Download the result, upload it again in Add Image, mark the next area and describe only that change in the prompt. One change per round is easier to control than several at once.

Can Qwen-Image-2.1 turn one photo into a panorama, infographic or storyboard?

Yes. Panoramas, infographics and storyboards are among the editing tasks the model covers. Qwen's examples turn a selfie into a panorama, expand a model photograph into a detailed, information-rich infographic, and build a complete storyboard from a three-view character reference. For a panorama strip or a row of story frames, the widest landscape option in Aspect Ratio gives the composition room. A tall infographic fits better in the tallest portrait option.

References

Sources and citations used to support the content provided above.

Updated: 2026-11-01 13:38:55 2 Sources

qwen.ai

Source Link
https://qwen.ai/blog?id=qwen-image-2.1

www.alibabacloud.com

Source Link
https://www.alibabacloud.com/blog/qwen-image-2-1-compact-efficient-and-unified-image-creation_603586