An AI image generator turns a text description into a picture, and the interesting questions are no longer about whether the picture looks good. They are about who owns it, whether you can use it commercially, whether anyone can tell it was generated, and what the tool refuses to do. This guide covers the mechanics briefly, then the parts that cost people money.
Updated October 2026.
How an AI image generator actually works
Almost every current text-to-image system is a diffusion model. The paper that made the approach practical at high resolution, Rombach and colleagues’ work on latent diffusion in December 2021, describes it in three parts: image formation is broken into a sequence of denoising steps; those steps run in the compressed latent space of a pretrained autoencoder rather than on raw pixels; and cross-attention layers let text steer the result.
In plain terms, the model learns to remove noise from images, and generation is that process run in reverse from noise, with your prompt pulling it towards an outcome. Two consequences follow. The same prompt with a different seed produces a different image, so reproducibility is a feature to look for. And the system holds no map of the world, only of what images look like, which is why hands, signage text, reflections and precise counts remain weak spots.
The prompt also does less work than people assume. OpenAI’s documentation notes that when its image tool is called from a conversation, the main model automatically revises your prompt and returns the rewritten version in a separate field. That is what actually ran. Our guide to prompt engineering covers the habits that transfer.
The main categories of AI image generator
Products blur together in marketing, but the underlying jobs are distinct:
- Text to image. A prompt in, a new picture out, judged on aesthetics alone. This is what people usually mean.
- Editing and reference-driven generation. Changing an existing picture, or using several as references. OpenAI’s Image API splits this way, with a generations endpoint and an edits endpoint.
- Masked editing, sometimes called inpainting. An image plus a mask marking the region to replace, documented within OpenAI’s edits endpoint. The workhorse for production use.
- Multi-turn refinement. Iterating on one image across a conversation rather than re-rolling a prompt, supported by OpenAI’s Responses API.
- Vectors. Adobe documents vector generation as a distinct Firefly capability, which matters if output has to scale.
Who owns the picture: copyright and licensing
This is where an AI image generator quietly creates problems, because two questions get confused. Whether the vendor lets you use the image is a contract question; whether anyone can own the copyright in it is a law question.
On the law, the US Copyright Office published Part 2 of its report on copyright and artificial intelligence on 29 January 2025. It concluded that generative AI outputs can be protected only where a human author determined sufficient expressive elements, which can include a human-authored work being perceptible in the output, or creative arrangement or modification of it, “but not the mere provision of prompts”. It also confirmed that using AI to assist creation does not bar copyright, and found no case for changing the law.
On the contracts, read the terms rather than the marketing line. Midjourney’s Terms of Service, effective 27 May 2026, say you own the assets you create to the fullest extent possible under applicable law, then list exceptions: if you or your employer earn more than $1,000,000 a year in revenue you must be on a Pro or Mega plan to own them, and upscaling someone else’s image does not transfer ownership. The same terms grant Midjourney a perpetual, irrevocable, sublicensable licence over your inputs and outputs, on an as-is basis with no warranty of non-infringement.
Training data is the other axis. Adobe states that current Firefly models were trained on licensed content such as Adobe Stock plus public domain content, and that outputs from non-beta features can be used commercially. Asked whether you can copyright Firefly output, Adobe refers users to a lawyer. If your use is a logo or a mark, that is a different legal track, covered in our piece on AI-generated logos and trademarks.
Provenance and detection: C2PA and Content Credentials
The counterpart to generation is proving where an image came from. The Coalition for Content Provenance and Authenticity publishes the standard behind Content Credentials. Its specification defines a manifest as one or more assertions about an asset plus a single claim and a claim signature, the claim being a tamper-evident structure signed with a signer’s private key. A hard binding is a cryptographic hash tying manifest to asset; a soft binding is a non-unique fingerprint or an invisible watermark.
The specification anticipates that an asset can become separated from its manifest through removal or corruption, with soft bindings used to match it back to an intact one. Screenshots and re-encodes strip credentials, and recovery depends on a lookup service existing. The standard also separates provenance from trust: it records history, while trust rests on who signed.
Invisible watermarking is the complement. Google DeepMind’s SynthID embeds watermarks into generated images, audio, text and video at creation, described as resistant to cropping, filters, frame rate changes and lossy compression, with detection through Google’s own tooling. That is the weakness of the category: a watermark only helps where the generator applies it and the checker can read it. Our guides to spotting AI-generated images and how AI text watermarks work cover what survives contact with the internet.
8 simple checks before you buy
Run these against any AI image generator subscription before committing a workflow to it. Each maps to something a vendor publishes:
- Find the ownership clause, not the headline. Search the terms for revenue thresholds, plan tiers and what happens if you cancel.
- Check what licence you grant them. A perpetual, sublicensable licence over your prompts and outputs is common and rarely advertised.
- Check the training data statement. Licensed, web-scraped or unstated are three different risk positions.
- Check for an indemnity and read its exclusions.
- Check whether outputs carry Content Credentials and whether they survive your export pipeline. A stripped credential is not a control.
- Check the editing endpoints, not just the generator. Masked editing and reference images are what production work needs.
- Check reproducibility and versioning. Midjourney’s terms say the art style and algorithms are subject to change and advise against depending on them.
- Check refusal behaviour against your real briefs. Named people, brands, medical and political content are restricted differently everywhere.
Practical limits worth knowing
Three technical limits recur whichever tool you pick. Text rendered inside an image is unreliable, so treat typography as a placeholder. Consistency across a set usually needs reference images or a custom model rather than clever prompting. And resolution claims describe the file rather than the detail in it.
The commercial limit is subtler. Because prompts alone do not establish authorship under the Copyright Office’s reading, a purely generated asset may be something you are contractually allowed to use but cannot stop anyone else copying. For a brand asset you intend to defend, that is the whole question, and it is why human editing on top is worth the time.
Common questions
How does an AI image generator work? Most are diffusion models. They learn to remove noise from images, then generate by running that process from noise while your prompt steers it, working in a compressed latent space rather than on raw pixels.
Can I use AI-generated images commercially? That depends on the vendor’s terms rather than on copyright law. Adobe says Firefly output from non-beta features can be used commercially; Midjourney says you own your assets subject to exceptions including a revenue-based plan requirement.
Can you copyright an AI-generated image? In the United States, only where a human determined sufficient expressive elements. The Copyright Office concluded in January 2025 that prompts alone are not enough, while creative modification or arrangement of the output can qualify.
How can you tell if an image was made by AI? Provenance metadata is the most reliable signal where it exists. C2PA Content Credentials record a signed history, and systems such as SynthID embed invisible watermarks. Both fail when a generator does not sign its output or a screenshot strips it.
What is the best AI image generator? There is no single answer: the tools differ most in licensing, editing features and refusal behaviour rather than in raw quality. Decide which matters for your work, then check each vendor’s documentation against it.
Sources and further reading
Where the figures and rules above come from, so you can check them:
- Part 2: Copyrightability, 29 January 2025: US Copyright Office
- Copyright and Artificial Intelligence initiative: US Copyright Office
- Specification: manifests, claims and bindings: C2PA
- Terms of Service, effective 27 May 2026: Midjourney documentation
- Firefly FAQ: training data and commercial use: Adobe
- Image generation guide: OpenAI API documentation
- SynthID watermarking: Google DeepMind
- High-Resolution Image Synthesis with Latent Diffusion Models: Rombach et al., arXiv
Photo credits: Graphic designer 5 by Ngwedi Mokgoro, CC BY-SA 4.0, via Wikimedia Commons. Penciling on Wacom Cintiq 13HD by David Revoy by David Revoy, CC BY 4.0, via Wikimedia Commons. Liljevalchs – Anders Petersen photography exhibition 9604 by Avery Jensen, CC BY-SA 4.0, via Wikimedia Commons.
The same company now sells agents that browse and act for you. See ChatGPT Dots.
Join the discussion