Across a crawl of roughly 1.1 million UK careers pages, not a single employer names “diffusion model engineer” in a job title. The work is being done — by people filed under machine learning engineer or creative technologist — but the label has no market and no benchmark behind it.

The bigger problem is the public impression of this technology: somebody types a sentence and receives a pleasing picture. Almost nothing commercially valuable looks like that.

Conditioning is the craft; generating is the easy part

The requests that reach a business are constrained ones. Put our actual product, unchanged, into this scene. Keep the model’s face and change the garment. Fill this masked region and leave the rest of the frame alone.

Each is a conditioning problem, and conditioning is where the engineering lives: structural control from edge, depth or pose inputs; reference adapters that carry identity or style between outputs; masked and region-limited editing; small fine-tunes that teach a model one particular garment or person; prompt weighting that survives being handed to a non-technical user. A candidate fluent about generation and vague about control has been playing with these tools, not shipping with them.

Two questions sort that quickly. Given one photograph of our product, how would you place it in a new environment with the label still legible? A strong answer involves masking, compositing and a verification step, not a longer prompt. And what have you built that a design team used without you in the room? Handing over the controls forces the control surface to be real.

The set, not the single image, is the deliverable

One good output is a demonstration. A campaign needs forty that look as though one photographer shot them in an afternoon, and a catalogue needs four thousand. Consistency across a set is what turns a fortnight’s enthusiasm into a six-month project.

Ask what drifted. Honest answers are specific: identity slipped as the poses got more extreme, a fabric colour shifted between batches, the logo turned to mush below a certain scale. Then ask what fixed it — recorded parameters per asset, a reference set regenerated on every change, a held-back group of awkward cases checked before release, lineage stored so any published asset traces to its inputs.

The tell is whether they talk in batches. Somebody who has shipped a set will describe the rejected pile and what they changed because of it.

Provenance belongs in the contract, not in the engineer’s judgement

This is the part most likely to become somebody else’s problem after launch, and it is not an engineering decision. Which checkpoints will you build on, given that some carry terms forbidding commercial use? Do you have the rights to fine-tune on your own photography, including shoots whose contracts said nothing about training? Will you put a named person’s likeness into a fine-tune, on what consent? Do you need a supplier who indemnifies you, and will you pay the premium?

Settle that before you write the brief, with whoever advises you on contracts, and state it as a constraint the engineer works inside. Leave it open and one of two things happens. The engineer picks whichever model performs best, which is rational of them, and you inherit a licence you cannot live with. Or a careful candidate keeps asking permission, is treated as slow, and gets routed around.

What you should test a candidate on is record-keeping, not legal opinion: whether they logged what went into a fine-tune, and can still say what a weight file was built from.

Whose hardware, and what one training run costs

Have the compute conversation before the interviews, because it changes the engagement more than any technical question will. Most of this work is adaptation rather than pretraining: a fine-tune on a few dozen curated images is hours on one rented accelerator and a trivial sum, while a full fine-tune across a large library is a different order of spend — and a run that dies two thirds through has spent the money anyway. Four things to settle in writing.

  • Whose account pays, and what the ceiling is before somebody has to ask.
  • Whether your imagery may leave your estate, since renting hardware means copying customer photographs onto a machine you do not own.
  • Who owns the weights when a contractor trained them on hardware they rented themselves — cheap to resolve before it arises.
  • Whether the budget is fixed, because working inside one is a skill not every candidate has needed.

Then ask about the most expensive mistake they have made with rented hardware; anyone who has trained anything substantial has one. Candidates whose whole experience is hosted image services have never owned this problem, which is fine if you never intend to train and disqualifying if you do.

Two outputs side by side, and who is allowed to say no

Automated scores are weak here. They will report two images as broadly similar while a designer tells you one is unusable, and the designer is right. The instrument that works is a person looking at two outputs side by side and choosing.

Accept the staffing consequence. Name the person whose judgement decides, give them authority, and book their time — an hour a week of a senior designer or brand owner, indefinitely. Never let the engineer grade their own output, because they will grade on effort. Their job is to make the comparison cheap and repeatable: a fixed set of hard cases, the same conditioning inputs, two versions shown blind, the verdict recorded against the change that caused it. Measuring more systematically is its own discipline, and the model evaluation engineer guide covers what this role only borrows.

A salary range with no market to anchor it

Indicatively £75,000 to £115,000 permanent, and £130,000 upwards for somebody who has owned a production image pipeline a commercial team depended on daily. Day rates indicatively £550 to £750, at the top end where brand or likeness exposure is real.

Because no employer advertises the title, there is no market figure to point at in a negotiation, and any number you are quoted is anchored to something else — a research salary, a freelance income, or a general engineering band with nothing to do with this skill. Contract-first is often the honest answer, since the demanding phase is building the controlled pipeline and the steady state is running it. If the need is broader than imagery, the generative AI engineer and multimodal AI engineer guides cover adjacent hires with deeper pools.