Qwen Image 2.1 is Alibaba’s image model on VdoBloom, available in two modes: Text to Image for full-scene generation, and Image Edit for editing existing photos — including mask-based inpainting that rewrites only a specific region you paint white while keeping the rest of the photo pixel-for-pixel untouched.

Disclosure: VdoBloom is our platform. Capabilities below are taken from the model’s API specification; we have not published our own test results yet.

What Qwen Image 2.1 Is On VdoBloom

Qwen Image 2.1 runs as two separate endpoints inside VdoBloom’s dashboard: qwen2-1/text-to-image at /dashboard/images/generate/ for generating a picture from a written description, and qwen2-1/image-to-image at /dashboard/images/edit/ for editing one or more existing images. Both accept prompts up to 5,000 characters in any language, which is enough room to describe lighting, composition, camera angle, and small details in one pass instead of chaining several short prompts.

Every prompt goes through an optional enhance_prompt step (on by default) that expands a short instruction into a fuller scene description before generation. You can turn it off if you want the model to follow your wording literally instead of having it filled in for you. A seed value is also returned with every result, and reusing that seed reproduces the same output — useful when you want to lock a look and then make one small, controlled change on top of it.

Because generation and editing live in two different modes, the workflow is straightforward: start a scene from scratch in Text to Image, then bring the result (or any other photo) into Image Edit whenever you need to adjust, extend, combine, or locally fix it.

Text to Image: Resolutions and Aspect Ratios

Text to Image supports two resolution tiers, 1K and 2K, with 2K noticeably slower to render. There are nine fixed aspect ratios, each with an exact pixel size at both tiers:

Aspect ratio	1K	2K
1:1	1024×1024	2048×2048
4:3	1184×896	2400×1792
3:4	896×1184	1792×2400
3:2	1248×832	2528×1696
2:3	832×1248	1696×2528
16:9	1376×768	2752×1536
9:16	768×1376	1536×2752
21:9	1568×672	3136×1344
9:21	672×1568	1344×3136

The default aspect ratio for Text to Image is 3:2. If you don’t specify one, that’s what you’ll get. Pick 2K only when you actually need the extra detail — for most drafts and iterations, 1K is faster and easier to work with.

Image Edit: Working With Reference Images

Image Edit accepts between 1 and 10 reference images per request, each a JPEG, PNG, or WebP file up to 30 MB and 25 megapixels. Order matters: the model reads references in the order you upload them, and your prompt refers to them by position — “the first image,” “the second image,” and so on. That makes it straightforward to say things like “put the person from the first image into the scene from the second image,” without having to re-describe either source in detail.

Fidelity to each individual reference drops as you add more of them. If keeping a specific face, product, or logo accurate matters, stay within about 4 references rather than using the full 10. When you are combining several people or objects into one composition, landscape ratios such as 3:2 or 16:9 tend to work better — they give the model more horizontal room to place everyone without cramming the composition.

The default aspect ratio for Image Edit is auto, meaning the output follows the aspect ratio of the first reference image unless you explicitly override it.

Mask Inpainting: How To Edit Part Of An Image With AI

Inpainting is a distinct mode inside Image Edit, triggered by supplying a mask_url alongside your reference image. The mask is a black-and-white image at the same aspect ratio as your reference photo: white marks the area you want changed, black marks the area to leave alone. Only one reference image is allowed in this mode — it’s built for precise, local edits to a single photo, not for combining multiple sources.

Step by step:

Upload the one photo you want to edit as your reference image.
Create a black-and-white mask at the same aspect ratio as that photo — paint white over exactly the region you want changed (a shirt, a background wall, an object to remove) and leave everything else black.
Upload the mask as mask_url alongside the reference image.
Write a prompt describing only what should appear inside the white area — not the whole photo. For example, “a red leather jacket” rather than “a person wearing a red leather jacket standing in a park.”
Generate. The output keeps the black area of the mask pixel-identical to your original and replaces the white area with content matching your prompt.

A few rules are specific to this mode: aspect-ratio selection and the enhance_prompt rewrite setting are both ignored — the output ratio always follows the reference image, and your prompt is used exactly as written, with no automatic expansion. Inpainting also cannot be combined with transparent background output. And because only one reference image is accepted, you can’t mix inpainting with multi-image composition in the same request.

This is the feature to reach for when you want to swap a background, remove an object, change an item of clothing, or fix one detail without regenerating the whole photo and risking every other part of it changing along with it — which is the usual problem with a plain text-prompt edit that has no mask to constrain it.

Transparent Background Output

Qwen Image 2.1 can output a real alpha-channel PNG or WebP with a transparent background instead of a solid one. This only works with PNG or WebP output — not JPEG, which has no alpha channel to store transparency in. When you use it, describe only the subject in your prompt (a product, a character, a logo) rather than a background, since anything you describe as a background will conflict with the transparency setting. It’s useful for product cutouts, sticker-style assets, and anything you plan to place over your own background layer later. Note that transparent background and mask inpainting are mutually exclusive within the same request — you can use one or the other, not both.

Qwen Image 2.1 vs GPT Image 2.5 vs Flux 2 Pro

VdoBloom also offers GPT Image 2.5 (listed on the platform as Flare/Sunburst) and Flux 2 Pro for editing work. Here’s how the three compare on the specifics that matter for editing:

Feature	Qwen Image 2.1	GPT Image 2.5	Flux 2 Pro
Max reference images	10	16	8
Resolutions	1K, 2K	1K, 2K, 4K	1K, 2K
Mask inpainting	Yes (dedicated `mask_url` mode)	Not available	Not available
Transparent background output	Yes (real alpha PNG/WebP)	Not exposed on VdoBloom	No
Max prompt length	5,000 characters	20,000 characters	5,000 characters

The short version: Qwen Image 2.1 is the one with a dedicated mask-based local-edit mode and native transparent-background output on VdoBloom. GPT Image 2.5 goes further on raw capacity, supporting up to 16 reference images and rendering up to 4K. Flux 2 Pro sits in between on reference count, at up to 8. None of this is a claim about which model produces better-looking results on any given photo — that depends on your specific image and prompt, and we haven’t published comparative quality testing between them.

How To Use Qwen Image 2.1 On VdoBloom

For a new image from a text description, go to /dashboard/images/generate/ and choose Qwen Image 2.1.
For editing an existing photo — including mask inpainting — go to /dashboard/images/edit/ and choose Qwen Image 2.1.
Write your prompt (up to 5,000 characters, any language), pick a resolution and aspect ratio, and decide whether to leave enhance_prompt on or write the exact wording you want followed.
For inpainting specifically, upload your one reference photo and its black-and-white mask, then describe only what belongs inside the white area.
Reuse the returned seed value if you want to reproduce a result or make a small follow-up change without the rest of the image drifting.
Cost is shown before you generate: 3 credits at 1K, 5 at 2K. New VdoBloom accounts start with 10 free credits to try it.

Frequently Asked Questions

Can I edit just part of an image with Qwen Image 2.1?

Yes — that’s what mask inpainting does. Paint the area you want changed white on a mask image, upload it together with your one reference photo, and describe only what should replace the white area. Everything under the black part of the mask stays untouched.

Does Qwen Image 2.1 support transparent backgrounds?

Yes, as a real alpha-channel PNG or WebP. It isn’t available with JPEG output, and it can’t be combined with mask inpainting in the same request.

How many reference images can I use?

Up to 10 in standard Image Edit mode, but fidelity to each one drops as you add more — stay within about 4 if accuracy to a specific face, product, or logo matters. Mask inpainting only accepts 1 reference image.

What resolutions does Qwen Image 2.1 support?

1K and 2K, across nine fixed aspect ratios with set pixel dimensions. 2K takes noticeably longer to render than 1K, so it’s worth reserving for final output rather than every draft.

How is this different from GPT Image 2.5 on VdoBloom?

Qwen Image 2.1 is the only one of the two with a dedicated mask-based inpainting mode and transparent-background output. GPT Image 2.5 supports more reference images (up to 16) and higher resolution (up to 4K). Which one to use depends on whether you need a precise local edit or a large multi-reference composition at higher resolution.

How much does Qwen Image 2.1 cost on VdoBloom?

3 credits per 1K image and 5 credits per 2K image, the same on Text to Image and Image Edit (mask inpainting included). The cost is shown before you generate. New accounts start with 10 free credits, so you can try both tabs before spending anything.

Qwen Image 2.1: Mask Inpainting & AI Photo Editing Guide