Do we actually need one?

Usually not on the first attempt, and the test is cheap. Try a detailed standing instruction for tone and conventions, then retrieval over your own material for knowledge, and see what remains unsolved. What survives both is the real case for adaptation, and it is a much smaller problem than the one you started with.

What does it cost to keep?

The recurring cost is redoing it when the base underneath is retired or superseded, which happens on the provider's schedule rather than yours. That means fresh examples, another training run and another evaluation pass, periodically, for as long as you rely on it. Ask any vendor quoting for one what the renewal looks like.

Will it stop the model inventing things?

No, and expecting that is the commonest disappointment. Adaptation changes how it writes rather than what it knows, so a confident wrong answer comes back in your house style instead of a generic one. Grounding it in your own material is what addresses the accuracy question.

Glossary

Custom model

A custom model is one adapted to a particular organisation's material or voice, and because the general model underneath it keeps changing, the adaptation is something maintained rather than something owned.

In plain terms

A version shaped to your own material, your own house style or your own kind of work. It is sold as something that becomes yours, and the thing it was built on top of carries on changing underneath it. That is the part worth understanding before you commission one, because it turns a purchase into a commitment.

Why it matters

Because the word custom implies a possession and the reality is closer to a subscription. The adaptation is anchored to a particular general model, and when that model moves on the adaptation has to be redone rather than carried across, so the cost is recurring in a way the original quote rarely reflects.

How it works

The adaptation sits on top of something general rather than replacing it. A capable base is shaped with your examples so it follows your conventions, uses your vocabulary and produces your shapes, and none of that changes what the base actually knows. It changes how it behaves, which is usually what people wanted and not always what they asked for.

The upkeep is the buried cost, and it arrives when the base moves. Providers retire and replace models on their own schedule, and an adaptation built on a retired one does not transfer, so keeping a custom model current means periodically redoing the work with fresh examples. Budgeting once for a recurring obligation is the commonest mistake here.

Much of what people commission a custom model for is available far more cheaply. A standing instruction sets tone and conventions at no ongoing cost, and retrieval puts your own material in front of a general model without altering it at all, so the honest sequence is to exhaust both before adapting anything. Genuine cases exist and they are rarer than the sales conversation suggests.

It appears in products as a premium tier rather than as a project you run. Several tools let you train on your own imagery or define a house style once and generate against it repeatedly, which is the same idea packaged so the upkeep is the vendor's problem instead of yours. That packaging is usually the better trade for anyone who is not building a product on top of it.

Try these in order

Laying these out as an order rather than as options is the whole of the advice, because the conversation that leads to a custom model almost never starts at the left. It starts with a genuine frustration, usually that output does not sound like the organisation or does not know something it should, and a custom model is the answer that sounds proportionate to a serious problem. The two cheaper arrangements solve most of it. A carefully written standing instruction fixes voice and format permanently and is edited in a text box when the house style changes. Retrieval fixes the knowledge half and stays current because updating a document is the whole of the maintenance. What remains after both is a narrow and real category, and a team that has worked through the order arrives at it able to say precisely what the adaptation is for, which is also what makes the eventual evaluation possible.

Seen in the wild

A generation platform built around control, where training on your own imagery holds a character or brand steady across a whole production run.
Leonardo AI
A design tool where a brand look is defined once and generated against repeatedly, turning one-off generation into asset production.
Recraft
A product-photo specialist running its own purpose-built model by default, with training on your own catalogue sold above it as an enterprise tier.
Photoroom

Common misconceptions

People assume

It teaches the model our knowledge.

In fact

It teaches behaviour far more than facts. Adaptation reliably shifts tone, format and conventions, and it is a poor and expensive way to install information that changes, which is what retrieval does better and cheaper. Teams wanting the model to know their material usually needed the other thing.

People assume

Once built, it is ours.

In fact

It is anchored to a base that will be retired on somebody else's schedule, and the adaptation does not move across on its own. Treat it as something with a renewal date rather than an asset on the books, and the budgeting follows correctly from there.

Telling them apart

Custom model vs Fine-tuning

Custom model

The thing you end up with, and its upkeep.

Fine-tuning

One technique for producing it.

One is the outcome and the other is a method, so a custom model may be produced by fine-tuning or by several cheaper arrangements sold under the same word.

Questions

Do we actually need one?: Usually not on the first attempt, and the test is cheap. Try a detailed standing instruction for tone and conventions, then retrieval over your own material for knowledge, and see what remains unsolved. What survives both is the real case for adaptation, and it is a much smaller problem than the one you started with.
What does it cost to keep?: The recurring cost is redoing it when the base underneath is retired or superseded, which happens on the provider's schedule rather than yours. That means fresh examples, another training run and another evaluation pass, periodically, for as long as you rely on it. Ask any vendor quoting for one what the renewal looks like.
Will it stop the model inventing things?: No, and expecting that is the commonest disappointment. Adaptation changes how it writes rather than what it knows, so a confident wrong answer comes back in your house style instead of a generic one. Grounding it in your own material is what addresses the accuracy question.

Key takeaways

It shifts behaviour reliably and installs knowledge poorly.
The base moves on the provider's schedule, so the adaptation has a renewal date.
Exhaust a standing instruction and retrieval first; what survives is the real case.
Bought as a product tier, the upkeep becomes the vendor's problem instead of yours.

Tools that use this

Leonardo AI
Training on your own imagery to hold a character or brand steady.
Recraft
A brand look defined once and generated against repeatedly.
Photoroom
Its own model by default, per-catalogue training sold as an enterprise tier.