Qwen's new image model is downloadable, but that does not make it commercially open
Qwen released Qwen-Image-2.1 on September 20. The model combines text-to-image generation and image editing, has a seven-billion-parameter visual generation component, can work with up to ten reference images, supports local edits and can preserve identity for people and products. It also produces native transparent RGBA imagery, making it potentially useful for product imagery, design assets and compositing. The important rights change is its licence. Qwen-Image-2.1 is distributed under the Qwen Research License Agreement. The licence grants use for non-commercial research and evaluation; commercial deployment requires separate permission from Qwen.
Why visual creators should care
Open weights and commercial rights are not synonyms. A designer, studio or developer can technically download and run a model locally while still lacking permission to incorporate it into commercial production.
Practical takeaway
Record the exact model version and licence alongside generated commercial assets. Do not rely on labels such as 'open', 'open-weight' or 'downloadable' as evidence that client work is permitted.
What to watch next
Watch whether Qwen introduces a standard commercial licence for Image 2.1 and whether hosted providers offering the model clearly distinguish research and commercial usage.
Sources
Pexels is keeping generative AI outside its photography and video library
Pexels updated its contributor help guidance on September 20 and again states that content created with generative AI cannot be uploaded to the platform. The underlying Pexels Terms go slightly further: Pexels may reject material it believes was created with or materially edited or modified using generative-AI technology. The company also cautions that it cannot guarantee that every item already available through Pexels is human-produced.
Why visual creators should care
For photographers and stock contributors, this is a useful example of a marketplace drawing its boundary around both fully generated work and material that has been materially changed by generative tools. Ordinary AI-assisted editing therefore needs to be distinguished from generative modification.
Practical takeaway
If Pexels is part of your distribution workflow, preserve the camera original and keep a record of substantial generative edits. Do not assume that beginning with a photograph automatically makes a heavily generated derivative eligible.
What to watch next
Watch whether Pexels provides a more detailed definition of 'materially edited or modified' as generative retouching becomes increasingly common inside mainstream photography software.
Sources
An AI photo editor is treating uploaded photographs as processing material rather than training material
PixGrid's terms were updated September 20 for its browser-based photo editor, whose AI functions include background removal, object detection, smart inpainting and enhancement. Users retain ownership of uploaded and created imagery. PixGrid says photographs sent for AI processing are used solely to provide the requested service, are not stored after processing and are not used for AI model training. Its AI features rely on models distributed under Open RAIL licences, whose use-based restrictions still apply.
Why visual creators should care
The terms separate two permissions that are often bundled together elsewhere: permission to process a photograph for the edit the user requested and permission to use that photograph to improve or train AI.
Practical takeaway
For confidential or client-owned imagery, look specifically for processing, retention and training clauses. 'You retain ownership' does not answer any of those questions by itself.
What to watch next
Watch whether processing-only and no-training commitments become standard competitive features in AI tools aimed at professional photographers and agencies.
Sources
A multi-model video platform makes the underlying provider's rules part of every generation
VideoRouter's Content Policy took effect September 20. The platform routes image and video requests across third-party AI models and establishes its own baseline rules covering uploaded reference imagery, intellectual-property and publicity rights and synthetic depictions of real people. But it explicitly says the underlying provider's own policy still applies and may be stricter. A generation that complies with VideoRouter's general rules can therefore still be refused by the model selected to produce it.
Why visual creators should care
For filmmakers and agencies using model routers, the front-end service is only one layer of the rights chain. The model that actually processes a reference photograph or produces the clip may carry a second set of restrictions.
Practical takeaway
Record the underlying image or video model used for client assets, not simply the router through which you accessed it. If a project depends on likeness or reference imagery, check both policy layers before production.
What to watch next
Watch whether multi-model platforms begin saving the applicable provider-policy version alongside individual generations so users can reconstruct the rights position later.
Sources
A creative AI platform is separating output rights from training consent
Lino's new Terms of Service took effect September 20. Users retain the rights they hold in content they submit, create or import and give Lino a limited licence for functions such as storage, synchronization and requested AI processing. Lino says that if it plans to use user content for model training or fine-tuning it will explain the data and purpose and obtain whatever consent is legally required. For outputs generated specifically for a user, Lino assigns any rights it may hold to that user to the extent permitted by law, while warning that this does not guarantee copyright, exclusivity or non-infringement. Third-party model licences remain separately applicable.
Why visual creators should care
The terms make several commonly confused rights explicit: ownership of the source, permission to process it, permission to train on it, whatever rights may exist in the output and the separate rights of third parties.
Practical takeaway
Build those same categories into client records. 'Who owns the output?' should not replace questions about source rights, training permission, model terms and copyrightability.
What to watch next
Watch whether creative platforms increasingly require affirmative training consent rather than relying on broad service licences covering every uploaded file.
Sources
Local image processing is becoming part of the privacy promise around AI photography
Anemoia's terms dated September 20 cover AI image enhancement and upscaling, depth-map generation, pose analysis and other image-processing functions. The service says most processing occurs locally in the browser, while some functions require server processing. It says server-processed photographs are automatically deleted after processing and that user imagery is not used to train its models.
Why visual creators should care
For photographers, agencies and studios, where a photograph is processed can matter almost as much as who owns it. Client imagery, unreleased campaigns and identifiable people can create confidentiality and privacy problems even where copyright ownership never changes.
Practical takeaway
Add processing location and retention to your AI-tool checks. For sensitive projects, know whether the image stays on the workstation, enters the platform's servers or passes to another AI provider.
What to watch next
Watch whether professional imaging services begin providing verifiable local-processing modes and contractual enterprise guarantees rather than relying solely on general privacy statements.
Sources
