How an AI Image Generator Really Works, and What You Own

You have a product page to fill, no photographer, and a generator that produces something usable in twenty seconds. The question that follows is the one every tool roundup ignores: can you sell what came out, and if a competitor lifts it, can you do anything about it?

Those are two different questions with two different answers, and most of the confusion in this area comes from treating them as one. Whether you may use an image is a matter of contract with your vendor. Whether you own it is a matter of copyright law, which your vendor does not control. A tool can grant you the first and be entirely powerless over the second.

This piece covers the mechanism first, because understanding what the machine actually does makes the rights argument obvious rather than mysterious, then the rights themselves.

How does an AI image generator actually work?

Start with the problem the researchers were solving. Diffusion models produce images by learning to reverse a corruption process: take a picture, add noise repeatedly until it is unrecognisable, then train a network to undo each step. Run the reversal from pure noise and you get a new image.

Done directly on pixels this works and is ruinously expensive. The latent diffusion paper by Rombach and colleagues puts the cost of training such models at hundreds of GPU days, with expensive inference because every generation requires many sequential steps.

Their fix is the reason you can run this on a laptop. Instead of diffusing in pixel space, they diffuse in the latent space of a pretrained autoencoder: a network that compresses an image into a much smaller numerical grid and can expand it back. The paper describes this representation as reaching, for the first time, a near optimal point between complexity reduction and detail preservation. You throw away the information the eye does not use, do the expensive work on what remains, then decode.

The second contribution is what makes a prompt matter. The authors introduce cross attention layers into the architecture, which is the mechanism that lets the denoising network condition on something outside the image, such as text or bounding boxes. Without it you have a machine that generates plausible pictures of nothing in particular. With it you have one that follows instructions.

The plain version, without the vocabulary

The model has learned what noise looks like when it sits on top of a real picture. You hand it a grid of pure noise plus a description. It predicts which part of that grid is noise, subtracts a portion, looks again, and repeats. After enough passes the noise is gone and what remains is an arrangement that matches the description. Then the autoencoder expands the small grid back into a full resolution image.

Two things follow from this that people find counterintuitive. The model is not retrieving or collaging stored images: it is predicting noise. And more steps do not mean more detail past a point, they mean a more thorough removal of noise, which is why raising the step count eventually stops changing anything worth paying for.

What actually differs between one generator and another?

This is where the tool roundups mislead, because they rank on a single axis called output quality as though it were a property of the software. Four components carry almost all of the real difference.

The autoencoder decides how much detail survives compression, which sets the ceiling on fine texture and small text. The text encoder decides how well the model understands your wording, which is why one tool follows a long compositional prompt and another quietly ignores half of it. The training data decides what the model has seen, which governs style range and also, as we will get to, the legal posture the vendor can take. The guidance strategy decides how hard the model is pushed toward the prompt, which is the trade between obedience and variety.

Once you hold those four in your head, most tool comparisons become readable. A generator that is good at typography has a better autoencoder or a text-aware pipeline. A generator that never produces a recognisable cartoon character has been trained or filtered to avoid it. Anyone building a repeatable pipeline rather than picking images one at a time will recognise this as the same reasoning behind wiring a generation workflow as an explicit graph, where each of those components becomes a node you can swap.

Why the same prompt gives you a different picture every time

Three controls explain nearly all of the variation, and knowing them turns generation from a slot machine into a process you can repeat.

The seed is the starting noise. It is a number, and the same seed with the same prompt and the same settings reproduces the same image. Most interfaces randomise it silently, which is why people believe generation is inherently unrepeatable. It is not. Record the seed and you can come back to an image months later.

The guidance scale sets how hard the model is pushed toward your text during denoising. Low values wander and produce more varied, often more natural results that ignore parts of the prompt. High values obey more literally and start to look overcooked, with saturated colour and hard edges. There is no correct setting, only a trade you are making on purpose or by accident.

The sampler and step count govern how the denoising path is walked. Different samplers reach different points from the same starting noise, which is why a workflow copied from someone else can produce a visibly different look with every other setting matched.

For a shop generating a catalogue, the seed is the one to care about. A consistent seed with a stable prompt template is what makes forty product shots look like they belong to the same brand, working alongside the reference photo handling that ChatGPT Images 2.5 improved for product photos, and it is the difference between a coherent page and a collage.

What training data has to do with your risk

Training data does not affect whether you hold a copyright. It affects a different exposure entirely, which is whether the output resembles somebody else's protected work closely enough to be a problem.

A model trained on licensed and public domain material, as Adobe describes for Firefly, is unlikely to reproduce a living illustrator's signature style on request because it has not been shown much of it. A model trained on a broad scrape of the open web has seen a great deal of it. Neither fact changes the copyright status of what you generate. Both change how likely you are to receive a complaint about it.

This is the honest reason vendors talk about commercial safety rather than ownership. They are describing the probability of an infringement claim arriving, which they can influence, rather than the existence of your copyright, which they cannot. Reading their marketing with that substitution in mind makes it far easier to compare.

Do you own an AI generated image?

Almost certainly not in the sense you mean, and the vendor's terms are not what decides it.

In January 2025 the US Copyright Office published the second part of its report on copyright and artificial intelligence. Its conclusion, stated in the Office's own announcement of the report, is that outputs of generative AI can be protected by copyright only where a human author has determined sufficient expressive elements, and that the mere provision of prompts does not qualify. Register Perlmutter framed the reasoning around the centrality of human creativity, warning that extending protection to material whose expressive elements are determined by a machine would undermine rather than further the constitutional goals of copyright.

The same report preserves the case that matters most to a working business: using AI to assist in the process of creation, or including AI generated material inside a larger human generated work, does not bar copyrightability. So a generated background behind a photograph you took, or a generated element inside a layout you designed, sits in a very different position from a raw output you prompted and downloaded.

If the vendor says I own the output, what does that clause do?

It transfers whatever the vendor has, and disclaims the rest. That is a real thing and it is not nothing. It is also not copyright.

Read the wording carefully and the vendors are more candid than their marketing. Midjourney's terms say you own the assets you create to the fullest extent possible under applicable law, which is a sentence that quietly hands the question back to the law rather than answering it. That same section conditions ownership for larger businesses on a plan tier: a company with more than one million dollars a year in revenue, or an employee of one, has to be on the Pro or Mega plan to own its assets. OpenAI's terms of use take the assignment route instead, assigning the user OpenAI's right, title and interest in output, while noting that outputs may not be unique and that other users can receive similar ones. Neither of those pages could be retrieved for direct citation when this article was written, so both are described here rather than linked.

Put those two facts side by side and the shape of the deal is clear. The vendor promises not to assert rights against you and passes along anything it holds. Whether a court would recognise an enforceable copyright in the result is outside the contract entirely.

The rights table

Four columns, because the roundups collapse them into one and that is where the error lives. The last column is the same for every row, which is the point of building the table.

Adobe is the row worth reading twice, because it is the only one offering something a lawyer would call protection, and it is narrower than the marketing implies. Adobe states that Firefly models were trained on licensed content such as Adobe Stock along with public domain content where copyright has expired, and that it does not train on subscribers' personal content. On the indemnity question its own enterprise page on the Firefly approach answers yes, then attaches the conditions: the organisation has to have purchased the appropriate entitlement, which requires a new contracting event, subject to the applicable terms and exclusions, and it covers select outputs rather than everything the tool produces.

That is a meaningful offer for a company with a legal department. For a one person shop on a consumer plan it is not available, and assuming otherwise is the expensive mistake in this whole area.

What does this mean if you are selling something?

Three consequences, and the first one surprises people.

A purely generated product image is close to unprotectable. If a competitor scrapes your catalogue and reuses your generated hero shot, your copyright claim is weak, because you probably never had a copyright. Other routes remain: trademark on your branding, passing off, contract terms on your site. Copyright is the one people reach for first and it is the one least likely to be there. We drew the same distinction for brand marks, where a generated logo can often be trademarked while remaining uncopyrightable, and the logic transfers directly.

Second, the marketplace you sell on has its own rules that operate independently of copyright. Platforms increasingly require disclosure of synthetic imagery or restrict it in particular categories, which is a separate compliance surface from ownership. We went through those in detail when covering the marketplace rules that apply to AI generated products, and none of them care whether you hold a copyright.

Third, the way you use the image changes its legal status. A raw output is the weakest position. The same output composited into a designed layout, retouched, arranged with your own photography and typography, becomes part of a work with human authorship in it. That is not a trick, it follows directly from the Copyright Office's carve out for assistive use, and it happens to describe what a competent product page looks like anyway. The practical guidance we set out on what a generated product image can honestly claim pushes in the same direction for a different reason.

Where do Content Credentials fit into this?

They answer a question next to the rights question, and the two get conflated constantly.

C2PA Content Credentials attach a signed record of how a file was made. The C2PA explainer describes a manifest of assertions covering when and where an asset was created, what modifications happened using what tools, and use of AI, bound to the file with cryptographic hashes so that any alteration to the asset or its provenance breaks the match.

What it does not do is stated just as plainly in the same document: provenance information alone cannot tell you whether the content is true, accurate or factual, and it makes no value judgement about whether a set of provenance data is itself true. Credentials can also be stripped, which is why the specification pairs them with soft bindings such as watermarking so a stripped file can still be matched back.

For a seller the useful reading is narrow. Credentials are evidence about process, not a licence and not a title. They help you prove what you did, which matters if a platform asks, and they help a buyer trust an image. They do not make an unprotectable image protectable. If you are thinking about durable marking of published material, the trade offs we covered for watermarking text on published pages apply here in the same shape: a mark that survives is a mark that constrains what you can do to the file.

Does any of this change if you run the model yourself?

It changes who you can blame, and little else.

Running open weights locally removes the vendor from the picture. There is no terms of service granting you output rights because there is no service. The model licence governs your use of the weights, sometimes with conditions on redistribution or acceptable use, and it typically says nothing about the pictures. The copyright analysis on the output is identical, because it was never about the tool.

What you lose is any indemnity, any content filtering the hosted services apply, and any audit trail. What you gain is control over the four components described earlier, plus the ability to keep prompts and outputs off a third party's servers. For a shop generating a large catalogue that last point is often the real driver, and it is a reasonable one.

A working checklist before you publish an image

Answer four questions and you have covered the ground that actually causes losses.

Which plan am I on, and does its terms grant commercial use at this revenue level? The Midjourney revenue condition is the clearest example of a rule that catches growing businesses precisely when they stop being small.

Is there an indemnity, and does it apply to me? For nearly every reader the answer is no, and the correct response is to stop treating the vendor as a backstop.

Did a human determine expressive elements here? If the honest answer is that a prompt did all the work, treat the image as free to be copied and plan accordingly rather than budgeting for enforcement you cannot do.

Do I have a record of how this was made? Credentials, source files or an ordinary archive of prompts and edits. Platform disclosure requests are becoming routine, and reconstructing provenance after the fact is far harder than keeping it.

Sellers who are assembling a storefront and its imagery at the same time have an advantage here, because the record keeping can be part of the build rather than an archaeology project later. That is one of the reasons we bake product content decisions into the flow when you build an online store with AI instead of generating assets in one place and publishing them in another.

What to watch next

Two things will move this. The first is litigation over training data, which affects the vendors' ability to make commercial safety promises rather than the copyright status of your output. A vendor that trained on licensed material is insulated from that in a way a vendor that scraped is not, which is why training data belongs in a rights table at all.

The second is the boundary of assistive use. The Copyright Office left a door open for human contribution, without drawing a line for how much is enough. Registrations and any disputes that follow will draw it, and where it lands decides whether a lightly edited generation counts as yours or as nobody's.

Until then the safe operating assumption is simple. Treat the licence as permission to use, treat copyright as absent unless a human genuinely shaped the expression, and keep the receipts.