Short answer: yes, you can sell AI images — but the devil's in the details. Every model has its own Terms of Service. Every jurisdiction has its own rules. Plus, there are three universal risks (recognizable faces, trademarks, copyrighted works) that apply no matter what model you used.
This post is a practical breakdown, no legal jargon. We'll cover each of the six flagship models of 2026: what their ToS allows, which plans grant commercial rights, and where the provider companies left themselves loopholes. Plus, a precedent map and step-by-step risk management.
What Each Model Allows in 2026
GPT-Image (OpenAI)
OpenAI Usage Policies and Service Terms, current as of May 2026: users get all rights to the output, including commercial use. Quote: «You own the Output you create with our Services, to the extent permitted by applicable law». No separate licenses or upgraded plans are needed — even with free ChatGPT.
Catch: OpenAI reserves the right to use your prompts and output to improve its models. To opt out, go to Settings → Data Controls and toggle off «Improve the model for everyone». This option is off by default in the API.
FLUX (Black Forest Labs)
FLUX comes in several versions. FLUX [schnell] is Apache 2.0, completely free for any use. FLUX [dev] is a non-commercial license; using it in commercial products requires a separate license from BFL. FLUX [pro] / FLUX 2 Pro is only available via API, and its license grants commercial rights to the output on paid plans.
Quantium uses FLUX 2 Pro via the BFL API — meaning you can use the output commercially without extra steps.
Gemini Image (Google)
Google Cloud Terms of Service for the Gemini API: commercial use is allowed, and the output belongs to the user. But there's a catch: Gemini might refuse to generate if the request involves public figures, brand logos, or sensitive categories. That's a safety policy, not a legal restriction.
Midjourney
This one's the most complex. Midjourney's ToS splits users into categories:
- Free tier (if active): output belongs to Midjourney; commercial use is forbidden.
- Basic / Standard / Pro / Mega: users get rights to the output; commercial use is allowed.
- Companies with $1M+ annual revenue: require a separate Pro license ($60/month minimum).
- Stealth mode: only on Pro+ — your output isn't published in the general gallery. On other plans, your images are visible to everyone.
Stable Diffusion and Forks (SDXL, SD3, Flux-Schnell)
Stability AI's base models are released under the CreativeML Open RAIL-M or SAI License. Both licenses allow royalty-free commercial use of the output. Restrictions apply not to the results, but to how you use the model: SD is banned for generating child pornography, medical diagnoses, or disinformation.
SD3 Medium and Stable Diffusion 3.5 come with the SAI Community License: it's free up to $1M in revenue; beyond that, you'll need an Enterprise license.
Sora 2 (OpenAI) — Video
Videos from Sora 2 are covered by the same OpenAI Service Terms: commercial use is allowed, and users own the output rights. Sora 2 automatically embeds an invisible watermark and C2PA metadata. This doesn't prohibit commercial use, but any tech expert can identify the content as AI-generated.
| Model | Commercial Use | Condition | Watermark |
|---|---|---|---|
| GPT-Image | Allowed | Any plan | C2PA metadata |
| FLUX 2 Pro | Allowed | Via API / Quantium | None (optional) |
| Gemini Image | Allowed | Any plan | SynthID invisible |
| Midjourney | Paid only | Basic+ ($10/month) | None |
| Stable Diffusion | Allowed | SAI License up to $1M | None |
| Sora 2 | Allowed | Any plan | C2PA + visible |
Three Universal Risks
A model's ToS might allow commercial use. But then you run into risks that don't depend on the model itself.
Risk 1. Faces Resembling Real People
In the US, this is called the right of publicity — an individual's right to control the commercial use of their image, name, and voice. The EU has a similar concept via GDPR (biometric data) plus national laws on image inviolability. In Russia, it's Article 152.1 of the Civil Code.
If AI generates a face resembling an actor, politician, or even just a specific private individual, using it in advertising is a direct cause for a lawsuit. Solutions: drastically alter the facial features (different age, ethnicity, expressions), or get a model release from the person it resembles.
Risk 2. Trademarks in the Frame
If a generated image clearly shows a Coca-Cola bottle, a Nike sneaker, or an Apple logo, that's a trademark infringement. Even if you didn't prompt for those brands. The model might have 'remembered' the brand from its training dataset.
Solution: Before commercial use, visually check the frame for identifiable logos. If a brand is critical to the story, license it or use a generic substitute.
Risk 3. Recognizable Works and Artist Styles
This is the most debated point. The prompt "in the style of Greg Rutkowski" isn't formally banned, but Greg Rutkowski filed a class-action lawsuit against Stability AI in 2023 and is still fighting it in court. By 2026, many platforms (Adobe Firefly included) banned living artists' names from prompts.
The safe bet: use artists whose work is in the public domain (they died over 70 years ago in most places). For example, "in the style of Caravaggio" or "in the style of Hokusai" carries no legal risk. "In the style of Greg Rutkowski" is a gray area.
Court Cases 2024-2026
Andersen v. Stability AI (USA, 2023-2026). A class-action lawsuit by artists against Stability AI, Midjourney, and DeviantArt for using protected works in their datasets. In August 2024, the court partially dismissed the suit, but kept the induced copyright infringement claims against Midjourney. The case is ongoing, with a final decision expected in 2026.
Getty Images v. Stability AI (UK, 2023-2026). Getty claims Stable Diffusion was trained on 12 million images from its database. In December 2025, the High Court of England issued an interim ruling favoring Getty on a key point. A full decision is expected in Fall 2026.
New York Times v. OpenAI (USA, 2023-2026). This lawsuit is about GPT, but the court's reasoning applies directly to DALL-E / GPT-Image. The NYT claims these models "memorize" training data and can reproduce it verbatim. The case is active.
Thaler v. Perlmutter (USCO, 2023-2025). Stephen Thaler tried to register an AI image as an autonomous AI work without a human author. The USCO denied it, and an appeals court upheld that denial in March 2025. This solidified the position: in the US, copyright requires a human author.
What About Russia?
Russia doesn't have a specific law for AI content. General Civil Code rules apply:
- Article 1257 of the Civil Code: The author is a citizen whose creative work produced the piece. AI can't be an author.
- Article 1259 of the Civil Code: For copyright to arise, the work must be the result of creative effort. An AI image without significant human input is a questionable object of protection.
- Article 152.1 of the Civil Code: Protection of a citizen's image. Applies to faces resembling real people.
Rospatent's position (letter from October 2024): Registering AI images as copyrighted works isn't forbidden, but registration doesn't mean a court will recognize rights if there's a dispute. In practice, they use a hybrid model: "a human assembled the prompt, picked from 50 options, and refined it in Photoshop — so it's the work of a human author."
Commercial use (sales, advertising, merch) in Russia is not formally prohibited, as long as the model's ToS are met and no third-party rights are violated (per Article 152.1 of the Civil Code and trademark law).
By Industry: Where Risks Are Higher
The risk level from commercial use of AI images varies a lot by industry and application. Here's a practical breakdown.
High Risk
- Ads with Faces. Any campaign using an AI-generated person as a 'brand face' risks a lawsuit if features match a real person. We've already seen cases in the US from 2024-2025.
- Book Covers, Movie Posters. High resolution, wide distribution, long-term use – these factors combined really grab the attention of rights holders.
- Merch (T-shirts, Mugs, Posters). If a product features an image resembling a real person or a protected character, a lawsuit is almost inevitable. Famous case: an Etsy designer got a cease-and-desist for AI images of Harry Potter characters.
- Political Ads. Deepfake laws in many places ban using AI images of politicians without clear labeling.
Medium Risk
- Stock Photos for Marketing Materials. If the face is abstract, with no recognizable features, the risk is low. If you suspect it's a real person, that risk climbs.
- Illustrations for Articles and Blogs. Standard commercial use means low risk, as long as there are no recognizable elements.
- Icons and Pictograms. Simple shapes are almost risk-free.
Low Risk
- Prototypes and Mockups for Presentations. Internal use without publishing means minimal risk.
- Sketches and Moodboards. If a human creates the final work, AI acts as a reference.
- Abstract Illustrations, Patterns, Textures. No people or recognizable objects makes this a virtually risk-free zone.
AI Images on Stock Sites: Platform Policies
Selling AI images on stock photo sites is a whole other set of questions. Platform policies vary:
- Adobe Stock: accepts AI content marked 'Generative AI.' It requires the author to own rights to the output (meaning using commercial model subscriptions).
- Shutterstock: only accepts submissions through its own AI generator. Uploading images from other models is forbidden.
- Getty Images, iStock: AI images have been strictly forbidden since September 2022 due to a lawsuit against Stability AI.
- Vecteezy, Freepik: accept AI content with mandatory labeling.
- Etsy: allows the sale of AI digital goods with mandatory labeling in the listing description.
Profits from AI stock are modest (cents per download), but with 1000+ pieces, it can become a steady passive income. The main catch is moderation: stock sites actively reject weak work, and a year after the AI boom, quality standards have jumped significantly.
Watermarks and C2PA: What They Are and Why They Matter
By 2026, all major AI providers embed invisible tags into their output. These are two different technologies.
C2PA (Coalition for Content Provenance and Authenticity) — an open standard. It records origin info in file metadata: what model generated it, when, and by whom. OpenAI, Adobe, Microsoft, Sony, and BBC support it. You can delete this metadata, but that violates many providers' ToS, and it's against the law in some places.
SynthID (Google DeepMind) — an invisible watermark embedded into image pixels or video frames. You can't remove it without serious quality degradation. It's used in Gemini Image, Imagen 3, Veo 2, and Veo 3.
OpenAI invisible watermark — OpenAI's own solution for DALL-E 3, GPT-Image, and Sora 2. It works similarly to SynthID, but it's proprietary.
What does this mean in practice? Any tech-savvy expert can tell if an image or video was AI-generated. You can't hide its AI origin. For commercial use, this isn't a ban. But if you try to pass off AI content as "shot on camera," you run a serious risk of exposure.
Risk Management: Your Checklist Before Commercial Use
Who's Liable If an AI Image Violates Rights?
Common misconception: "I just typed a prompt — the AI service is responsible for the outcome." In practice, liability shakes out differently.
OpenAI ToS, Indemnification section (current for 2026): The company reimburses corporate API clients for damages from copyright infringement lawsuits, but only if they meet certain conditions (like using standard safety filters and not trying to bypass policies). Free and Plus users don't get this protection.
Microsoft Copilot Copyright Commitment: A similar program for corporate users. Microsoft covers fines and legal costs in third-party IP infringement disputes.
Adobe Firefly: Adobe officially guarantees Firefly trains only on licensed Adobe Stock data and public domain content. That gives you maximum legal clarity, but the model's quality is noticeably behind FLUX and Sora 2.
Midjourney, FLUX, Sora 2, Kling: By default, all responsibility falls on the user. If your image infringes someone's rights, you're on the hook in court.
What this means for business: For big commercial campaigns, it makes sense to use models with indemnification (OpenAI Enterprise, Adobe Firefly, Copilot for Business). For smaller tasks, use models without indemnification, but with user-side risk management.
What's in Quantium
Quantium's Terms of Service state users get all output rights the model provider offers. That means for FLUX 2 Pro, Gemini Image, GPT-Image, and Sora 2 in Quantium, commercial use is allowed, no extra steps needed.
Quantium doesn't save your prompts for training or share them with providers (other than for the generation itself). C2PA metadata is saved in image and video files for transparency.
Related articles: AI Content Copyright, Midjourney Alternatives, Chat Model Comparison, All Image Models.
Try Quantium Free
You get 20 credits a month on the free plan. Plus, 30+ AI models in one Telegram bot.
Open Bot →Read Also
