Making a Hundred AI Images Look Like They Came From One Photographer

The failure is easy to recognise and hard to name. A site has twenty images. Each is technically fine. Together they look wrong — one is warm and shallow, the next is cool and flat, a third has that faintly plastic sheen, and the overall impression is of a page assembled from whatever was to hand.

Individually good images that do not belong together read worse than consistently mediocre ones. Coherence is a stronger signal of a real business than any single frame, and it is almost entirely a function of what you hold constant rather than how carefully you describe each shot.

The two halves of a prompt

Every image prompt is really two prompts joined together.

What it shows. The subject. A technician at a workbench. A van at a kerb. A finished installation.

How it is made. Shot type, lens character, light source and direction, colour temperature, palette, mood, depth of field.

The second half is where consistency lives, and it is the half people rewrite every time. If you describe lighting differently on each image, you get differently-lit images. Worse, anything left unstated is chosen by the model, and it will choose differently depending on the subject — which is why a set of prompts that mention no lighting at all comes back with wildly varying lighting.

The fix is structural. Write the second half once, keep it byte-identical across the batch, and vary only the first.

Building a style anchor

A style anchor is a fixed block covering five variables. Nothing exotic:

Variable	What to specify	Why it drifts without it
Framing	Shot distance and angle	Model defaults to whatever suits the subject
Lens character	Depth of field, focal feel	Background blur varies wildly by scene
Light	Source, direction, hardness	The single largest source of visible inconsistency
Colour	Temperature and a two- or three-colour palette	Warm and cool frames next to each other read as different sites
Mood	Two or three adjectives, honestly chosen	Prevents the tonal swing between clinical and cosy

A workable anchor is three or four sentences and reads like a brief to a photographer, because that is what it is. Something along the lines of: a medium shot at eye level, shallow depth of field with a soft background falloff, lit by large soft daylight from the left with gentle shadow, neutral-to-warm colour with muted blues and warm greys, calm and workmanlike, no text, no logos, no watermarks.

Then every prompt in the batch becomes anchor + subject, and only the subject changes.

The negative clauses at the end matter more than they look. Generated text in an image is almost always garbled, invented logos create a real trademark problem, and both are far easier to exclude than to remove afterwards.

The anchor also carries the realism burden for the whole batch, since the difference between a photograph and a render is decided almost entirely by the lens and lighting language. That is worth getting right once in the anchor rather than fighting per image, and the specific phrasing that produces photographs rather than renders belongs in the fixed half of the prompt for exactly this reason.

Testing it on the extremes

The mistake in testing an anchor is choosing three similar subjects, which proves only that the model is deterministic-ish.

Test on the widest spread the real set will contain. An interior close-up, an exterior wide shot, and a person mid-task. If the anchor survives all three — same tonal family, same light logic, same depth character — it will survive the rest. If the exterior comes back cold and clinical while the interior is warm and soft, the light specification is doing less work than you think and needs to be more explicit.

Keep those three test images. They are the record of what the anchor produced, and when a model updates and the look shifts six months later, they are the only way to tell whether the drift is real or misremembered.

People, and why they break everything

Faces are the hardest consistency problem, because the model generates a different person each time and human attention goes straight to the discrepancy. A set showing four different technicians who are all supposedly your one technician is worse than a set showing no people at all.

The practical resolutions, in order of how well they work:

• Hands and forearms only. Fully solves identity, keeps the human presence, and reads as documentary rather than stock.
• Backs and over-the-shoulder framing. Same benefit, works for wider shots.
• Mid-distance figures where the face is small or turned. Presence without identity.
• Deliberately different people, framed as different people. Honest, and fine when the business genuinely has a team.

What does not work is asking for the same specific person repeatedly and hoping. Hands are also where the visible artefacts cluster, which is a separate reason to frame them carefully and to look closely at the result — one of several reasons a pre-publication check of the images before you commit to them is worth doing as a defined pass rather than a glance.

Variation without incoherence

There is an opposite failure. A batch so uniform that every page carries what is visibly the same image with one object swapped looks templated, and templated is precisely the impression a small business site is trying to avoid.

The rule that resolves the tension: vary the subject, hold the treatment. Different rooms, different tasks, different angles on different objects — all shot as if by the same person on the same afternoon. That is how a real photography set looks.

This has a search dimension too. Images that are near-identical to each other tend to be grouped and collapsed rather than indexed separately, so producing twenty barely-different frames does not create twenty opportunities. It creates one, plus nineteen files. The reusing-images question is worth understanding here, because the answer changes what is worth generating at all: reuse is not penalised, but it is also not free visibility, so genuine subject variety is the only version of variation that pays.

If the difficulty is finding twenty genuinely different subjects for a service business with nothing photogenic to point a camera at, that is its own problem, and what to show when you sell a service is a longer answer than the anchor question.

When to break the anchor deliberately

An anchor applied to everything eventually produces a site where every image has the same weight, and a page with no visual hierarchy is its own kind of monotony.

Three cases justify departing from it, and they should be departures you decide on rather than drift you tolerate.

The hero. The one image at the top of a key page can afford a wider shot, more space and more atmosphere than the anchor specifies, because it is doing a different job — setting a tone rather than illustrating a point. Keep the light and colour language, relax the framing.

Genuine documentation. A real photograph of real work does not need to match a generated set, and should not be made to. Authenticity outranks consistency, and a phone photograph of an actual completed job is more persuasive than a polished frame that matches everything around it.

Diagrams and anything informational. These belong to a different visual system entirely and forcing photographic treatment on them helps nobody.

The rule of thumb is that the anchor governs the supporting cast. It does not govern every image on the site, and applying it that rigidly is how a set goes from coherent to flat.

What the coherence actually buys

Not a ranking. There is no algorithm rewarding a consistent palette.

What it buys is believability, and believability is what separates a page that reads as a real operation from one that reads as a landing page. That shows up in behaviour rather than in a metric you can point at, which makes it easy to under-invest in and easy to over-claim.

The comparison worth being clear-eyed about is against licensed stock, where the coherence problem is inverted: stock images are individually polished and almost impossible to make coherent, because they come from a dozen different photographers and the same files appear on competitors' sites. Generated imagery is the reverse — less polished per frame, fully coherent if you hold the treatment constant, and genuinely yours. The honest cost and quality comparison between the two covers the trade in full, and the coherence advantage is the part most comparisons omit.

Making it survive the team

An anchor that lives in someone's head lasts until they are on holiday.

Store it as a plain text file with the brand assets, versioned, with the three test images beside it. Include what it excludes and why. Then when a new batch is needed, the process is: open the file, paste the anchor, list the subjects, generate.

At agency scale the anchor becomes a per-client artefact, and the workflow question shifts from how do I write good prompts to how do I make sure the right anchor is used for the right client without anyone remembering to. That is an operations problem rather than a prompting one, and it is the substance of running image production across many clients at once.

The tooling side of the same problem is why brand presets exist as a product feature at all — SEOpix keeps per-brand settings so a batch inherits the right look and the right metadata without being re-specified each time. The anchor is still yours to write. The point is only that it should be written once and then applied, rather than reconstructed from memory every time somebody needs a picture.