AI image generation for business: models, uses and limits
Current image models are dependable for placing real products in new settings, exploring marketing concepts, producing mock-ups and doing editing work such as background replacement, provided a person reviews every image before it is published. They remain unreliable for exact product detail, long or small text and strict layout. The law on training data and ownership of outputs is unsettled, so choose tools and licences deliberately.
What can current image models do reliably?
The change that matters for business is not prettier pictures. It is control. The leading models now follow instructions about an existing image (“replace the background with a pale oak floor and leave the chair untouched”), take several reference images at once, and hold a subject steady across a series of edits.
In practice these jobs work well enough to plan around:
- generating a realistic scene from a written description
- editing by instruction: swap a background, relight, remove an object, extend the canvas to a new aspect ratio
- carrying a product, a person or a visual style from reference images into a new picture
- short, legible text, such as a three-word headline
- transparent backgrounds, and output large enough for print
The limits are just as consistent, and the vendors state them. OpenAI’s image generation guide says its models can still struggle with precise text placement, with keeping recurring characters or brand elements visually consistent, and with exact positioning in layout-sensitive compositions. That list describes the whole field fairly.
Expect a share of every batch to be unusable, and budget review time for it.
Which model families matter now?
Names and version numbers change every few months, so treat this as a snapshot from the time of writing and choose by capability and licence.
| Family | How you get it | Suits | Check before you commit |
|---|---|---|---|
| OpenAI GPT Image | API and ChatGPT | Editing by instruction, masks, reference images, transparent backgrounds | The vendor’s own notes on text placement and consistency |
| Google Gemini image models (“Nano Banana”) | Gemini API and Google’s apps | Combining many references, text in images, conversational edits, output up to 4K | Every output carries an invisible SynthID watermark |
| Black Forest Labs FLUX | Hosted API, with some models released as open weights | Photorealism, multi-reference editing, self-hosting | Open-weight licences differ: FLUX.2 [dev] is non-commercial unless you buy a licence |
| Stability AI Stable Diffusion 3.5 | Open weights you host yourself | Custom styles, full control, private data | The free community licence applies only below US$1 million in annual revenue |
Google’s image generation documentation says its current models accept up to 14 reference images, and it now describes the older Imagen line as legacy.
Design teams also use Midjourney for concept work and Adobe Firefly inside Creative Cloud. Adobe says Firefly is trained on licensed and public-domain content and offers intellectual property indemnity on qualifying plans. Read which plan and which model that covers, because partner models offered in the same app are treated differently.
Whatever you pick, keep the model behind a thin interface in your own system. Expect to swap it.
Which business uses hold up?
Consider a furniture retailer with 1,200 products that wants each one shown in four room settings. Studio shoots for all of them are out of reach. Image generation can do the job, and the way it fails is instructive.
Give a model a photo of a sofa and ask for “this sofa in a bright Scandinavian living room”, and many results will look superb and be wrong. The legs are a different shape. The fabric has gained a weave. The proportions have shifted by a few centimetres. A customer who orders from that picture receives a different sofa, which is a returns problem and can amount to misleading advertising.
The dependable method is to keep the product’s pixels from a real photograph, cut out cleanly, and generate only the room, floor, light and shadow around it. Where exactness matters less, generation from references is fine.
Uses we would back:
- Product imagery variations. New backgrounds, seasons and settings around a true product image.
- Marketing concepts. Thirty directions for a campaign before anyone books a photographer. It is internal and low risk, and it changes how briefs get written.
- Mock-ups. Packaging, signage, exhibition stands and placeholder art during product design.
- Editing and background work. Background removal, extension for new formats, retouching, upscaling. It is the least exciting use and the one that pays back fastest.
Uses we would avoid: pictures of real people without their consent, anything presented as a record of an event, and any image that is itself a factual claim, such as before-and-after results, portion sizes or property photographs.
How do you keep images on brand?
Three things do most of the work.
A reference library. Collect approved product shots and style frames and supply them as references every time. Current hosted models take ten or more references in one request, which is enough to fix a product, a setting and a visual style together.
A written image style guide. Lighting, palette, camera angle, what never appears. Turn it into reusable prompt blocks so that two designers get the same look.
A fine-tuned style. If references are not enough, train an adapter on an open-weight model using a few dozen images you own. The technique is the same LoRA method used for language models, described in our fine-tuning guide. Some cloud platforms offer managed tuning of image models as well. Hosting a model yourself brings GPU operations with it, a cloud and MLOps cost to count before you choose that route.
Two rules save a lot of rework. Never let the model draw your logo; place the real vector file in the design tool. And check brand colours by value, because a model’s “navy” is whatever it likes.
Can AI put text in images?
Short text, mostly yes. Google describes its models as able to produce legible text for infographics, menus and marketing assets. OpenAI says text rendering has improved and still lists placement and clarity as weak points.
Our working rule: let the model set a few words while you are exploring a concept, and add all published text in the design tool. You need the licensed brand typeface, exact spacing, legal lines and translations, and none of those should depend on chance.
If generated text does stay in an image, proofread every character, in every language. The usual failure is a small misspelling in the second line that nobody reads until it is printed. Never rely on generated text for prices, ingredients, dosage or terms.
What is the legal position?
This is a summary for planning, not legal advice. There are four separate questions.
Training data
Whether training a model on copyrighted images without permission is lawful is contested, and the answer differs by country. In Getty Images v Stability AI, decided by the High Court in London in November 2025, Getty abandoned its claim about training because there was no evidence the training took place in the UK. The court then held that a model which does not store or reproduce the works is not an “infringing copy”, and it found limited trade mark infringement where watermarks appeared in outputs. Getty has been given permission to appeal on the copyright point, and cases brought by artists and film studios in the United States are still running.
The developer carries most of this risk, but you can be drawn in through what you publish. Do not prompt with artists’ names, characters or other companies’ brands, and screen outputs for lookalikes.
Copyright in outputs
The US Copyright Office concluded in January 2025 that prompts alone do not make the user the author of an output. Protection can cover human-authored elements within it, and creative arrangement or modification of AI material, and using AI as an assisting tool does not rule protection out. Other countries take different positions.
The practical reading: a purely generated image may be free for a competitor to copy, so keep layered files and a record of the human work.
Licence terms
Read four things. Is commercial use allowed? Who owns outputs as between you and the vendor? Are your uploads used for training? Does any indemnity apply, and on which plan? Open weights do not mean free for business use, as the table above shows.
Disclosure and provenance
The EU AI Act’s transparency rules, which the European Commission lists as applicable from 2 August 2026, require providers to mark generated content in a machine-readable way and require those who publish deepfakes to disclose them. Systems that were already on the market before that date have until 2 December 2026 to meet the marking duty. A voluntary code of practice published in June 2026 sets out how. Our article on the EU AI Act for product teams has the wider picture.
The main technical standard is C2PA Content Credentials: signed metadata that records where a file came from and how it was edited, backed by a group that includes Adobe, Google, Microsoft and OpenAI. Metadata can be lost when a file is re-saved or uploaded, so keep your own record as well.
Some channels add their own rules. Google’s merchant listings, for instance, require AI-generated product images to carry a specific IPTC metadata value.
A workflow with human review
- Write a brief: purpose, channel, and what must be accurate (the product, its colour, any claim).
- Use only inputs you own or have licensed. No real people without consent.
- Generate in batches from the approved references and prompt blocks.
- A designer makes the first cut for artefacts, anatomy, physics and brand fit.
- Someone who knows the product checks it against the real thing.
- Check rights: logos, characters, recognisable faces, stray watermarks.
- Finish in the design tool with the real logo, real type and corrected colour.
- Keep provenance metadata, add any required labels, and store the prompt, model version, references and approver.
- Publish, and test the workflow again whenever the model changes.
Steps 5 and 6 are the ones teams skip under deadline. They are also the ones that prevent the expensive mistakes.
How to start
Choose one high-volume, low-risk job for the first week. Background replacement on existing product photographs is the usual candidate. Take 20 products, run them through the workflow above, and record two numbers: the share of generated images that pass review, and the time per approved image compared with how you do it today.
Then make the first real decision: a hosted API or an open-weight model you run yourself. Sensitive images, very high volume or a need for a custom-trained style point towards open weights. Everything else points towards hosted. It is the same reasoning as any build or buy choice.
Content teams should read this alongside our guide to generative AI for content and marketing teams. If you need the pipeline built into your catalogue or design tooling, that is part of our generative AI work.