The Qwen Image Generator is the image-creation and editing capability available through the official Qwen experience and the broader Qwen-Image model family. For most users, the safest route is to open Qwen’s official website and select its image-generation feature rather than trusting a domain simply because “Qwen” appears in its name. Qwen can create images from text, and current hosted Qwen-Image models also support reference-based editing, complex layouts, multilingual typography, and—in specific 2.1 routes—transparent output.
Product and pricing review: October 7, 2026. Commercial-use clarification: October 15, 2026. The original review covered public Qwen product pages, model announcements, Model Studio documentation and pricing. The commercial-use clarification distinguishes consumer-service restrictions from output rights and separate model or cloud agreements. No authenticated consumer image-generation task was submitted, so output-quality statements below are identified as documented capabilities rather than firsthand test results.
Planning a commercial project? Check the service and model licensing section before generating images. Free access, technical capability and rights in the result are different questions.
The official Qwen Image Generator vs lookalike websites
Search results for “Qwen Image Generator” contain several independent tools with Qwen-related domain names. Some may provide access to a Qwen model, but their subscriptions, credits, storage, moderation, privacy rules, and commercial terms belong to those operators—not automatically to Qwen or Alibaba Cloud.
A website should not be described as the official Qwen Image Generator unless it uses a Qwen- or Alibaba-controlled domain or is linked from an official Qwen property. A polished logo, an exact-match domain, or a claim that the site is “powered by Qwen” is not enough.
| Access route | Best suited to | What to verify |
|---|---|---|
| Official Qwen consumer experience | Private, non-commercial image creation or editing without writing code. | Begin at qwen.ai. Current controls, model labels, limits, and availability can depend on the account, platform, and region. |
| Qwen API Platform | Developers who want a Qwen-hosted API route. | Use the official Qwen API Platform and confirm its current model catalog, authentication, billing, and endpoint. |
| Alibaba Cloud Model Studio | Applications requiring model selection, regional deployment, API parameters, editing, or higher-volume generation. | The model, endpoint, workspace, and API key must belong to the same region. |
| Downloadable Qwen-Image model | Researchers and technical users who want local inference or framework integration. | Read the exact model card and license. The current Qwen-Image-2.1 weights are governed by a non-commercial research license. |
| Third-party Qwen-powered generator | Users who deliberately prefer an external interface or bundled service. | Check who owns the service, which model it actually uses, its privacy policy, output rights, pricing, and whether it states that it is unaffiliated with Qwen. |
Practical rule: start from qwen.ai. Treat any other generator as a separate third-party service until its relationship with Qwen is independently verified.
What “Qwen Image Generator” actually refers to
The phrase can refer to the consumer image-generation feature, a hosted API model, or downloadable model weights. These routes share Qwen branding but do not necessarily provide the same controls, prices, model versions, licenses, or reference-image limits.
Qwen’s consumer interface is the easiest no-code route, but it may not expose the precise model identifier behind every generation. Alibaba Cloud Model Studio, by contrast, lets developers explicitly choose models such as qwen-image-3.0-pro, qwen-image-3.0, and qwen-image-2.1-pro. A separate downloadable release, Qwen/Qwen-Image-2.1, can be run through supported local inference frameworks under its own license.
Which Qwen image model should you use?
The numbering can be misleading. Qwen-Image-3.0 was announced in July 2026, while Qwen-Image-2.1 was announced in September 2026. The later release date of 2.1 does not mean it replaces every 3.0 route. Current official catalogs keep both because they target different workflows.
| Route or model | Best fit | Important current distinction |
|---|---|---|
| Qwen consumer image generation | Fast no-code creation and account-based image editing where available. | The exact underlying model and controls may not be displayed consistently. Do not assume that every API feature is present in the consumer interface. |
qwen-image-3.0-pro | Complex compositions, dense posters, small text, multilingual layouts, interfaces, and high-detail images. | Hosted generation and editing; accepts one to three reference images through the current Model Studio API. |
qwen-image-3.0 | General generation and editing when balancing quality, throughput, and price matters. | Official documentation describes it as the quality-and-speed balance within the 3.0 series. |
qwen-image-2.1-pro | Workflows needing more reference images or prompt-controlled transparent output. | Hosted Model Studio route; supports up to ten reference images and can return an image with an alpha channel based on the prompt. |
Qwen/Qwen-Image-2.1 | Local research, framework integration, transparent asset creation, and experimental custom workflows. | Downloadable model with 7B parameters in its visual-generation component, native RGBA support, and up to ten references. Its Qwen Research License restricts the model materials to non-commercial use unless a separate commercial license is obtained. |
According to Qwen’s Qwen-Image-3.0 announcement, the 3.0 generation focuses on dense visual content, authentic details, long instructions, small text, multilingual typography, and knowledge-heavy layouts. These are vendor-documented capabilities, not a guarantee that every prompt will produce perfect text or a production-ready design.
Qwen-Image-2.1 focuses on a smaller unified generation-and-editing architecture, native transparent output, local editing, multiple reference images, and improved preservation of people and products. Use it when those capabilities are more important than selecting the numerically highest model family.
How to use the Qwen Image Generator
- Open the official Qwen website. Visit qwen.ai and enter the official Qwen experience. Avoid opening an unrelated exact-match domain from an advertisement or search result.
- Sign in if prompted. The available features, history, usage allowance, and model controls can depend on the account and platform.
- Select image generation. Use the current Image Generation or equivalent creation control shown in the interface. Labels can change, so do not rely on an old screenshot if the menu looks different.
- Describe the image precisely. State the subject, action, environment, composition, lighting, style, exact visible text, and intended aspect ratio.
- Choose an aspect ratio or size when the interface offers one. Match it to the intended use: square for many product assets, portrait for posters, and landscape for banners or article illustrations.
- Generate the image. Do not judge only the overall appearance. Check subject count, text, numbers, anatomy, crop, background, and requested objects.
- Request a focused revision. Identify what must remain unchanged and describe only the correction. For example: “Keep the bottle, lighting, camera angle, and background unchanged. Correct only the headline to ‘AUTUMN BOOK FAIR’.”
- Download the selected result. Retain the original file rather than taking a screenshot, especially when checking resolution or transparency.
- Save the prompt and settings. Record the prompt, aspect ratio, model or mode, reference images, and relevant controls when reproducibility matters.
How to write a stronger Qwen image prompt
A useful prompt describes visual relationships rather than stacking vague quality words. “Beautiful, stunning, 8K masterpiece” does not tell the model where the subject should stand, how the scene should be framed, or which text must appear.
A practical prompt can follow this structure:
Primary subject:
[Who or what must appear, including the required number of subjects.]
Action or pose:
[What the subject is doing and where it is facing.]
Environment:
[Location, background, season, time, and supporting objects.]
Composition:
[Camera angle, framing, subject position, foreground, and empty space.]
Lighting and colour:
[Light direction, softness, mood, and colour palette.]
Style and material:
[Photograph, illustration, paper cutout, 3D render, ink drawing, etc.]
Visible text:
[Write every required line inside quotation marks, with exact spelling,
capitalisation, punctuation, hierarchy, and position.]
Output:
[Aspect ratio, intended use, and any background or transparency requirement.]
Do not include:
[Unwanted objects, text, cropping, visual artifacts, or stylistic elements.]Prompt examples are not usage permissions. Before using the product or e-commerce examples for business, check the terms for your chosen service and model.
Photorealistic product-image prompt
Create a premium studio product photograph of one amber glass skincare
bottle with a matte black dropper, standing on a pale limestone pedestal.
Composition: three-quarter camera angle, bottle centred slightly to the
right, generous negative space on the left for website copy, no cropping.
Lighting: soft morning light from the upper left, subtle shadow, warm neutral
palette, realistic glass reflections, visible fine texture in the limestone.
Background: minimal cream plaster wall with a gentle tonal gradient.
Aspect ratio: 4:3.
Do not include hands, flowers, extra bottles, visible branding, labels,
decorative text, or a watermark.Poster prompt with exact text
Create a vertical 4:5 poster for a community book fair.
Exact visible text:
Headline at the top: "AUTUMN BOOK FAIR"
Subheading below it: "12 OCTOBER · 10:00–18:00"
Footer: "RIVERSIDE COMMUNITY HALL"
Use the text exactly as written. Do not add other words.
Visual style: elegant editorial poster, warm burnt orange and deep navy,
cream paper texture, stacked books and autumn leaves in the lower third.
Typography: bold condensed sans serif for the headline, highly legible,
large spacing, clear hierarchy.
Keep all text inside safe margins. Do not crop letters, misspell words,
change the date, or invent logos.Educational infographic prompt
Create a clean vertical infographic titled "THREE STEPS TO BETTER COFFEE".
Section 1 heading: "WEIGH"
Caption: "Measure coffee and water."
Section 2 heading: "GRIND"
Caption: "Match the grind to the method."
Section 3 heading: "BREW"
Caption: "Control time and temperature."
Use only the supplied text. Place the three sections in a clear top-to-bottom
sequence with a scale icon, grinder icon, and brewer icon.
Style: simple editorial vector illustration, high contrast, cream background,
dark brown text, muted red accents, accessible visual hierarchy.
Aspect ratio: 3:4. Avoid tiny labels, decorative paragraphs, extra facts,
brand logos, and invented measurements.Reference-image editing prompt
Use image 1 as the product reference and image 2 as the environment reference.
Preserve the exact shape, proportions, material, cap, and colour of the
product from image 1.
Place it naturally on the wooden desk shown in image 2. Match the direction
of the window light and add a physically plausible contact shadow.
Change only the environment and lighting integration. Do not alter the
product design, add a logo, change the camera angle, or introduce extra items.Transparent asset prompt
Generate one red hiking backpack as a clean e-commerce asset.
Show the full backpack at a three-quarter angle with realistic fabric,
zippers, straps, and stitching.
Output a true transparent RGBA background with clean anti-aliased edges.
Do not create a white studio background, floor, shadow rectangle, text,
logo, extra objects, or a checkerboard pattern.The last example should be used only with a route that actually supports alpha output, such as the current qwen-image-2.1-pro API or the downloadable Qwen-Image-2.1 model. Do not assume that transparency is exposed in every consumer account.
How to improve text inside Qwen images
Text rendering is a prominent Qwen-Image use case. The Qwen-Image-3.0 documentation advertises support for long prompts, text as small as 10 pixels, multiple languages, and varied fonts. Those specifications indicate model capability under selected conditions; they do not remove the need to proofread every generated word.
- Quote required text. Write
"AUTUMN BOOK FAIR"rather than loosely asking for a book-fair headline. - Preserve exact case and punctuation. Tell the model not to change capitalisation, dates, prices, separators, or symbols.
- Reduce the amount of copy. Generate the headline and short labels first. Add long body copy later in a design application.
- State the hierarchy. Separate headline, subtitle, labels, and footer instead of putting all copy in one paragraph.
- Describe placement. Specify top centre, lower left, inside a card, or another unambiguous location.
- Name the language and script. This reduces ambiguity in multilingual posters.
- Protect safe margins. Explicitly request that letters remain inside the canvas and are not cropped.
- Correct one element at a time. Ask the editor to preserve the composition while replacing only the incorrect word or number.
- Verify critical content manually. Prices, medical instructions, legal statements, formulas, dates, and factual diagrams must be checked independently.
When exact typography is essential, use Qwen to create the composition and imagery, then set the final copy manually in a design tool. A poster can look readable at a glance while still containing a wrong date, malformed number, or subtle misspelling.
Editing images and using references
The current Model Studio models support both text-to-image generation and image-to-image editing. Their reference-image limits differ:
qwen-image-3.0-proandqwen-image-3.0accept one to three reference images through the current unified API.qwen-image-2.1-proaccepts one to ten reference images.- The downloadable Qwen-Image-2.1 release also documents editing with up to ten references and supports local-edit instructions.
When several images are supplied, identify their roles explicitly. “Combine these images” leaves the relationship ambiguous. A stronger instruction is: “Use image 1 for the person, image 2 for the coat, and image 3 for the background. Preserve the face and coat design. Change only the setting.”
Reference images improve control but do not guarantee perfect identity, brand, product, or facial consistency. Inspect small logos, packaging proportions, jewellery, clothing patterns, facial details, and object counts after every edit.
Obtain permission before uploading or transforming another person’s photograph. A technical ability to imitate or edit someone’s appearance does not establish consent or a legal right to publish the result.
Can Qwen generate a transparent PNG?
Yes, but only specific routes currently document native transparency. The hosted qwen-image-2.1-pro model can decide whether to return a standard image or an image with an alpha channel based on the prompt. The downloadable Qwen-Image-2.1 model supports generating and editing transparent RGBA images, extracting subjects, and working with transparent layers.
Do not confuse a white background, checkerboard pattern, or cutout preview with real transparency. To verify the file:
- Download the original output rather than taking a screenshot.
- Confirm that the file is PNG or another alpha-capable format.
- Open it in an editor and inspect the colour mode or alpha channel.
- Place the image over both a dark and a bright coloured background.
- Zoom in to check for white halos, jagged edges, transparent holes, and unwanted semi-transparent pixels.
A JPEG cannot preserve an alpha channel. If the downloaded file is JPEG, its background is not truly transparent.
Is the Qwen Image Generator free?
Qwen’s public homepage describes the consumer assistant as free to use, but that should not be interpreted as a guarantee of unlimited image generation. A numerical consumer allowance that applies to every account was not verified on the public product page. The current account interface is the most reliable place to check task limits, resets, queueing, and any paid options.
API access is a separate paid service. The following Model Studio rates apply to the Singapore international deployment and were verified on October 7, 2026:
| Model | Input-image price | Output price | Singapore free quota |
|---|---|---|---|
qwen-image-3.0-pro | US$0.003 per input image | 1K: US$0.04 per image 2K: US$0.075 per image | 10 images |
qwen-image-3.0 | US$0.003 per input image | 1K: US$0.03 per image 2K: US$0.03 per image | 10 images |
qwen-image-2.1-pro | See the current regional pricing table | US$0.04 per generated image | 10 images |
The official pricing page states that these free quotas are available only in Singapore and are generally valid for 90 days from Model Studio activation, model release, or application approval—whichever is later. Other regions use different prices and do not receive the same listed free quota. Model Studio charges for successfully generated images; failed requests do not consume the free quota.
Always recheck the live Model Studio pricing table before estimating production costs. Image size, region, model, reference-image count, and number of outputs can affect the bill.
For the wider billing context and image-count calculations, see our Qwen pricing guide.
Downloading model weights also does not make deployment cost-free. Local use can require a capable GPU, storage, electricity, framework setup, maintenance, and engineering time. The applicable model license may additionally restrict what the deployment can be used for.
Commercial use and Qwen image licensing
There is no single commercial-use answer that applies to the consumer product, Model Studio, a third-party generator, and every downloadable Qwen-Image model.
Images created through the Qwen consumer service
The Qwen Terms of Service include an explicit restriction: “The Services are for private and non-commercial use only.”
The same terms state that, as between the user and Qwen, Qwen assigns to the user any rights, title, and interest it has in the output. That wording does not guarantee that an output is copyrightable, unique, non-infringing, accurate, or clear of third-party rights.
The output-rights clause alone does not establish commercial permission. For advertising, paid client work or sales assets, confirm that the agreement for the exact service authorises that workflow. Do not apply the consumer-service restriction automatically to separately licensed API or downloadable-model routes.
You remain responsible for reviewing the result and for issues involving identifiable people, trademarks, protected characters, copyrighted source material, misleading content, and local disclosure requirements.
Qwen-Image-2.1 downloadable weights
The downloadable Qwen/Qwen-Image-2.1 model uses the Qwen Research License Agreement. The license defines non-commercial use as research or evaluation and grants rights to the model materials for non-commercial purposes only. Commercial use requires a separate commercial license from Qwen.
This restriction concerns the licensed model materials. Rights in a particular output can involve additional terms and laws, so do not infer unrestricted commercial permission from the ability to download the weights.
Model Studio and third-party platforms
Hosted Model Studio use has a separate agreement. In Section 4.48 of the Alibaba Cloud International Website Product Terms of Service, Alibaba Cloud states that Model Studio does not claim intellectual-property ownership of outputs and permits use of inputs and outputs subject to applicable law, the agreement and its rules. Additional model or component licences can also apply. This is not an unrestricted commercial-use guarantee.
A third-party image generator applies its own terms, even when it calls a Qwen model behind the scenes. Check the operator’s conditions and any applicable model licence; paying for a subscription does not establish all necessary rights.
Review the terms that apply to the exact product, model, provider, account, and intended use. “No watermark” and “downloadable weights” do not automatically mean unrestricted commercial use. This is a practical summary, not legal advice; obtain provider clarification or qualified advice for a high-stakes project.
Privacy when uploading photos or reference images
The consumer interface, Model Studio API, third-party generators, and local deployment do not share one universal privacy policy. Review the current Qwen Privacy Policy for consumer use and the relevant Alibaba Cloud or third-party agreement for another route.
Do not assume that an uploaded image is deleted immediately, never logged, never reviewed, or never processed outside the device unless the exact applicable policy says so.
- Do not upload passwords, API keys, access codes, or screenshots containing active credentials.
- Avoid identity documents, medical images, confidential contracts, unreleased products, and private client assets.
- Obtain appropriate permission before uploading another person’s private photograph.
- Remove unnecessary personal details and hidden metadata where appropriate.
- Check whether a share link or gallery entry becomes public before distributing it.
- Use an approved business account or deployment route when organisational data-handling requirements apply.
A local model can reduce reliance on a remote inference provider, but “local” is not always synonymous with offline. Model downloads, prompt enhancers, extensions, analytics, storage integrations, or other workflow components may still communicate over a network.
Qwen Image API: the current developer route
Alibaba Cloud Model Studio currently provides a unified Qwen Image Generation and Editing API. It supports both text-to-image and image-to-image workflows through OpenAI-compatible and native DashScope protocols.
| API detail | Current documented behaviour |
|---|---|
| Available unified models | qwen-image-3.0-pro, qwen-image-3.0, and qwen-image-2.1-pro |
| Protocols | OpenAI-compatible synchronous calls plus DashScope synchronous and asynchronous routes |
| Generation endpoint | The OpenAI-compatible route uses /images/generations for both generation and editing. Editing supplies the image extension field; it does not use the official OpenAI /images/edits multipart format. |
| References | One to three images for the Qwen-Image-3.0 series; one to ten for qwen-image-2.1-pro |
| Output count | One to six images per documented request |
| Output format | PNG |
| Size range | Total pixel area from 512×512 to 2048×2048, with aspect ratios from 1:8 to 8:1 for the current unified models |
| Prompt length | Recommended maximum of 4,500 tokens |
| Negative prompts | Supported by the Qwen-Image-3.0 series |
| Prompt enhancement | Optional direct or agent rewriting, subject to model and generation/editing mode |
| Seed | Supported for repeatability within the documented integer range, although identical seeds do not guarantee portability across model changes |
| Watermark parameter | Available in the API and currently defaults to false |
| Result URLs | Generated image URLs are valid for 24 hours and should be downloaded promptly |
The model, workspace-specific endpoint, and API key must belong to the same region. A valid key can fail when paired with a model or domain from another region.
The API currently accepts reference images through public URLs or Base64 data. Its documented input formats include JPEG, PNG, BMP, TIFF, WebP, and GIF, with a maximum of 10 MB per input image and recommended dimensions between 384 and 2048 pixels.
Read the current Qwen Image Generation and Editing API reference before implementation. Parameters such as image, negative_prompt, seed, prompt_extend, enable_thinking, and watermark are Model Studio extensions and must be passed according to the selected protocol.
Can Qwen Image run locally?
Yes. The current downloadable route is Qwen-Image-2.1, with model files also published through the official Hugging Face repository.
The release documents support for frameworks and tools including Diffusers, ComfyUI, vLLM-Omni, and SGLang. It supports text-to-image generation, image editing, transparent RGBA output, local edits, and multiple reference images.
Local suitability depends on the exact framework, precision, output resolution, offloading strategy, and available hardware. Do not publish or rely on one universal GPU-memory figure without matching it to those conditions.
- Confirm that the repository and model card are published by Qwen.
- Read the Qwen Research License before downloading or deploying the model.
- Use the package versions currently required by the official repository.
- Begin at a documented resolution before increasing memory demand.
- Do not assume that community quantisations, ComfyUI nodes, or repackaged checkpoints have identical behaviour or licensing.
- Treat downloaded custom nodes and extensions as executable software that must be reviewed before installation.
How to inspect a generated image before publishing
A visually attractive result can still fail the actual task. Review the original downloaded file at full size.
- Are all requested subjects present, with the correct count?
- Is every visible word, date, price, number, and symbol correct?
- Are hands, eyes, teeth, limbs, reflections, and shadows coherent?
- Are logos, product proportions, packaging, and reference-image details preserved?
- Does the composition match the requested camera angle and subject placement?
- Are important elements cropped or placed too close to the edge?
- Does the file have the required aspect ratio and sufficient resolution?
- Is transparency genuine, with a usable alpha channel and clean edges?
- Is an unexpected watermark, signature, brand, or protected character visible?
- Does a factual infographic contain verified information rather than plausible-looking invented data?
- Does the chosen service and model licence permit the intended use, and have relevant input and third-party rights been checked?
Important information should also be available as accessible text outside the image. Do not rely on an AI-generated poster or infographic as the only source of legal, medical, safety, financial, or instructional content.
Common Qwen Image Generator problems
| Problem | Likely cause | Practical fix |
|---|---|---|
| The site looks like Qwen but asks for unrelated credits | You may be using a third-party generator rather than Qwen’s official service. | Return to qwen.ai. Review the other site’s About, Terms, Privacy, and affiliation disclosure before paying. |
| Image generation is missing | The account, platform, region, application version, or rollout can differ. | Use the official web experience, sign in, check the current tools menu, and update the official app where applicable. |
| Text is misspelled | The prompt contains too much copy, ambiguous hierarchy, or insufficient exact-text instructions. | Quote the exact text, reduce the amount, specify placement, and request a targeted correction. Add critical copy manually if accuracy is mandatory. |
| The wrong number of people or objects appears | The scene is overloaded or the count is not prominent. | Put the exact count at the beginning, simplify the scene, and repeat the restriction in the “Do not include” line. |
| Important content is cropped | The requested framing conflicts with the ratio or the prompt does not reserve safe margins. | Specify full-body, uncropped, generous margins, subject position, and the intended aspect ratio. |
| A reference image is ignored | The role of each image is unclear, the model limit was exceeded, or the selected route does not support the requested number of references. | Number each reference, assign one role to each, and confirm the selected model’s current reference limit. |
| The subject changes during editing | The instruction asks for broad regeneration rather than a constrained edit. | List the exact attributes to preserve and state which single region or property may change. |
| The “transparent” image has a white background | The route lacks alpha support, the prompt requested only visual isolation, or the downloaded file has no alpha channel. | Use a documented 2.1 transparency route, request true RGBA output, download PNG, and inspect its alpha channel. |
| The API rejects a valid key | The key, workspace, endpoint, and model may belong to different regions. | Rebuild the configuration from one regional Model Studio page and use the matching workspace-specific domain. |
| An API result URL no longer works | Model Studio result URLs are retained for 24 hours. | Download and store successful outputs immediately rather than using the temporary URL as permanent hosting. |
| Local generation runs out of memory | The model, resolution, precision, or workflow exceeds available resources. | Use the official setup, reduce resolution or batch size, apply a documented offloading method, and avoid assuming that every community workflow has the same requirements. |
| The prompt is refused | The request may conflict with current usage or safety rules. | Remove unlawful, exploitative, deceptive, or privacy-invasive content. Do not attempt to bypass moderation safeguards. |
Frequently asked questions
What is the official Qwen Image Generator website?
Start from qwen.ai, the official Qwen website, and select its image-generation experience. Do not assume that an exact-match Qwen Image domain is official.
Is Qwen Image Generator free?
Qwen describes its consumer assistant as free to use, but a universal numerical image limit was not verified on the public page. Hosted API calls are separately billed, with limited free quota available for eligible Singapore Model Studio accounts. Third-party generators set their own prices.
Which model does the Qwen consumer interface use?
The consumer interface may not always display a precise model identifier. Do not assume that it automatically exposes the newest API or downloadable model. For explicit model selection, use an official developer route such as Model Studio.
Is Qwen-Image-2.1 newer than Qwen-Image-3.0?
Qwen-Image-2.1 was announced after Qwen-Image-3.0, but the current catalog keeps both for different workflows. Qwen-Image-3.0 targets rich layouts, text, and hosted generation, while 2.1 emphasises a compact unified architecture, multiple references, local editing, and native transparency. It is not a simple replacement sequence.
Can Qwen generate readable text inside an image?
Current Qwen-Image documentation specifically targets text-heavy layouts and multilingual typography, but no image model should be trusted to reproduce every word perfectly. Quote required text, limit the copy, define its hierarchy, and proofread the result at full resolution.
Can Qwen edit an existing photograph?
Yes. Current Model Studio Qwen-Image-3.0 and 2.1 Pro routes support image-to-image editing. Editing controls in the consumer product can vary by account and platform.
Can Qwen create a transparent PNG?
The hosted qwen-image-2.1-pro route and the downloadable Qwen-Image-2.1 model document native alpha-channel output. Confirm availability in the route you are using and inspect the downloaded PNG rather than relying on its preview.
Can I use Qwen-generated images commercially?
Not automatically. Qwen’s consumer terms restrict the service to private, non-commercial use; assigning output rights does not by itself establish commercial permission. Model Studio and third-party generators have separate agreements. Using the downloadable Qwen/Qwen-Image-2.1 model materials commercially requires a separate commercial licence beyond the research licence. Check the terms for the exact service and model before generating images for a commercial project.
Is there a Qwen Image API?
Yes. Alibaba Cloud Model Studio currently offers a unified generation-and-editing API through OpenAI-compatible and DashScope protocols. Its current listed models include qwen-image-3.0-pro, qwen-image-3.0, and qwen-image-2.1-pro.
Can I run Qwen Image locally?
Yes. Qwen-Image-2.1 is available as downloadable model weights with support for frameworks such as Diffusers and ComfyUI. Local use still requires suitable infrastructure, and the current Qwen Research License limits the model materials to non-commercial use without a separate commercial licence.
Sources checked
The product references below are retained from the October 7, 2026 review. For the October 15, 2026 commercial-use clarification, the consumer-service restriction and output-rights wording were checked in indexed text from Qwen’s official Terms of Service; the Qwen-Image-2.1 licence and Alibaba Cloud’s Model Studio terms were read directly. This was not a clause-by-clause review of the complete consumer agreement or a new verification of all prices and features.
- Official Qwen website and consumer-product overview.
- Official Qwen-Image-3.0 announcement.
- Official Qwen-Image-2.0 announcement.
- Official Qwen-Image-2.1 announcement.
- Alibaba Cloud Model Studio Qwen Image Generation and Editing API reference.
- Alibaba Cloud Model Studio pricing table.
- Alibaba Cloud Model Studio rate-limit documentation.
- Official Qwen-Image-2.1 repository.
- Official Qwen-Image-2.1 model card and Qwen Research License Agreement.
- Qwen Terms of Service and Qwen Privacy Policy.
- Alibaba Cloud International Website Product Terms of Service, Section 4.48: Model Studio.