How AI Image Generation Works: A 2026 Business Guide
How AI image generation works in 2026: diffusion models, not GANs. Plus the EU AI Act marking rules that start on 2 August 2026. A guide for UK business teams.
By Ivan Pylypchuk, CEO of SoftBlues. Has led Claude and automation projects for finance, legal and healthcare teams across the UK and Ireland. Last updated: 29 July 2026.
From 2 August 2026, Article 50 of the EU AI Act requires providers of generative AI systems to mark synthetic image, audio, video and text output in a machine-readable format so it can be detected as artificially generated (European Commission). If your marketing team has been quietly generating product visuals for the last two years, that date matters more than any model release.
SoftBlues is an AI consultancy at softblues.io, working with regulated mid-market companies across the UK and Ireland. We are a registered Anthropic Partner Network member and a Google Cloud Partner. We generate images in our own production work, including most of the covers on this blog, so this guide is written from doing it rather than from reading about it.
AI image generation is the process of producing a new image from a text description using a model trained on large numbers of image and caption pairs. Most systems in use in 2026 are diffusion models: they start from random noise and remove it step by step, guided by the text prompt, until an image emerges. A smaller group, including OpenAI's newer image models, generate images the way a language model generates text, one piece at a time. The older approach, generative adversarial networks, is largely historical for this task. What has changed for businesses is not the picture quality but the rules and the rights attached to the output.
Key facts
How does AI image generation actually work?
A diffusion model is trained by taking real images and adding noise to them in small steps until nothing recognisable is left. The model learns to undo each of those steps. At generation time you run that process backwards from pure noise, and a text encoder steers each step towards your prompt.
That is why generation takes a few seconds rather than being instant, and why the same prompt gives you a different image each time. You are watching a guided cleanup, not a lookup.
Two details matter in practice. The model works in a compressed representation of the image rather than on raw pixels, which is what makes it fast enough to be useful. And the text encoder, not the image part, sets the ceiling on how well the system follows a detailed brief. When a tool ignores half your prompt, that is usually the text side, not the drawing side.
What changed since the GAN era?
Generative adversarial networks pit a generator against a discriminator: one makes images, the other judges them, and both improve. It was the dominant approach around 2017 to 2021, and it produced convincing faces. It was also unstable to train and hard to steer with text.
Diffusion replaced it for text-to-image because it trains more reliably and takes direction better. Here is the practical difference:
| Approach | How it generates | Text control | Where you see it in 2026 |
|---|---|---|---|
| Diffusion | Removes noise over many steps | Strong | Most mainstream image tools |
| Autoregressive | Predicts the image piece by piece | Very strong on detailed briefs | Newer multimodal models |
| GAN | Generator competes with a discriminator | Weak | Niche and legacy uses |
If you read an explainer that stops at GANs, it was written before the shift and will not describe the tool you are actually using.
What can a UK business use this for?
The uses that hold up are the ones where the image is disposable, internal, or heavily reviewed. Marketing teams produce concept visuals and variants for testing. Product teams mock up interfaces before committing design time. Training teams illustrate internal material that would never have justified a photoshoot.
The pattern that works is generate, then review, then publish, with a named person accountable for what goes out. That is also the pattern the EU AI Act carve-out is built around. We run our own blog covers this way, and the wider story of running a company on AI is in our Claude operating system case study. It is also why we know how much of the effort sits after the model finishes.
If you are weighing up whether generated imagery belongs in your workflow at all, the same questions apply as with any other AI purchase, and we set them out in 10 questions to ask before signing a generative AI consulting contract.
What are the rules for AI-generated images in the UK and EU?
Three things to hold in view.
The EU AI Act transparency obligations bite from 2 August 2026 and reach any provider putting a generative system on the EU market, which includes UK companies serving EU customers. We cover the wider picture in what the EU AI Act means for UK companies.
UK copyright law has not settled. The Getty ruling narrowed one route of attack, but it turned largely on where the training happened and on what model weights contain. It did not give anyone a general licence, and the trade mark findings went partly against Stability.
Provenance standards such as C2PA content credentials are becoming the practical answer to "prove this was reviewed". Whichever tool you use, being able to show who approved an image and when is the part that survives a regulator's question. That is a governance job, and we set out how mid-market teams handle it in AI governance for UK mid-market companies.
Where AI image generation is not the right fit
It is a poor choice for anything that has to be factually accurate: technical diagrams, medical illustration, real products a customer will receive, or anything a person might rely on. Models produce plausible images, not correct ones.
It is also the wrong tool where you need the same subject to appear identically across a long series. Consistency has improved and it is still the thing that quietly eats the time you thought you were saving.
And if the honest reason for using it is to publish more content faster with no one checking it, the EU AI Act carve-out does not cover you, and neither would we.
Frequently asked questions
Do AI image generators still use GANs?
Mostly no. Generative adversarial networks led text-to-image work until around 2021 and have since been replaced by diffusion models, with some newer systems using an autoregressive approach instead. GANs still appear in narrower tasks such as some upscaling and face work.
What is the difference between diffusion and a GAN?
A GAN has two networks competing, one generating and one judging. A diffusion model learns to reverse a noise process and generates by removing noise step by step, guided by your prompt. Diffusion trains more reliably and follows text instructions far better, which is why it took over.
Do I have to label AI-generated images under the EU AI Act?
From 2 August 2026, providers of generative AI systems must mark output in a machine-readable format under Article 50. Deployers publishing AI-generated content also have disclosure duties, with an exemption where the content has been through human review or editorial control and a named person holds editorial responsibility. Check your own position with a lawyer; this is a summary, not legal advice.
Does the EU AI Act apply to a UK company?
It can. The Act reaches providers and deployers placing systems or output on the EU market, so a UK company serving EU customers is often in scope even without an EU entity. SoftBlues works with UK and Ireland teams on exactly this question during discovery, and it is usually narrower than people fear.
Who owns an image generated by AI?
Ownership depends on the tool's terms and on your jurisdiction, and UK law is unsettled. The November 2025 Getty judgment addressed whether training and model weights infringe, not who owns the output. Read the licence of the tool you use before you put an image on a product.
Is it cheaper than hiring a designer?
Per image, yes, by a wide margin. Across a real workflow the saving is smaller than the demo suggests, because review, brand consistency and rights checking still take a person's time. SoftBlues treats generated imagery as a drafting step that a human finishes, which is where the honest saving sits.
Can we run image generation on our own infrastructure?
Yes. Open-weight diffusion models run on your own hardware or in your own cloud tenancy, which suits teams with data or residency constraints. You take on the running costs and the maintenance in exchange for keeping everything inside your boundary. SoftBlues builds this kind of self-hosted setup for UK and Ireland clients whose data cannot leave their tenancy.
SoftBlues is a registered Anthropic Partner Network member and a Google Cloud Partner. We are practitioners, not slide-deck consultants: we use these tools in our own production work before we recommend them, and we put working AI into production in 90 days at a fixed price, with a money-back guarantee if it fails. You can see the range of that work on our generative AI consulting page. If you want a straight answer on where generative AI earns its place in your business, and where it does not, book a discovery call.
See it in production
Systems we have built and run for clients, with the numbers that came out of them.
