Why colour is the thing that breaks

A studio that shoots its own on-model work carries every garment through one lighting rig, one camera profile and one retouching chain. Colour drifts there too, but it drifts the same way for every frame, and the retoucher's eye has learned the drift. The moment the work goes to a production partner, that shared memory is gone. The partner's frames are produced under different light, through a different pipeline, and judged against a recollection of the garment rather than a measurement of it.

The failure is quiet. Nobody returns a jacket because the three-quarter view faces the wrong way; they return it because the navy arrived as black. By the time the return rate on a colourway moves, the batch has been live for weeks. This is why colour, not pose and not resolution, is the first thing to specify when on-model production is outsourced, and the first thing to inspect when it comes back.

The reference set to hand over

A production partner can only be held to a reference it has. The useful reference is a photograph the brand already trusts: the packshot, flat lay or mannequin shot that sits on the product page, shot under the brand's own light and signed off by the brand's own retouching. That image is the colour authority for the garment, whatever its own small cast, because it is what the customer has already seen and bought against.

What the reference set needs, per SKU:

  • One reference per garment in the look. The hero garment and the second garment, if the look has one. The second garment is the one that drifts, because nobody thinks to send it.
  • The view the reference was shot in. A front flat lay is a reference for the front view. If the back has a different panel or print, send a back.
  • A tolerance in words. "Match the packshot" is a wish. "Within what our retouching would pass on the product page" is a tolerance a measurement can be built around. Where you have one, a numeric tolerance is better still.
  • What is not the reference. A Pantone name, a supplier's swatch scan, a lookbook image lit for mood. None of these can be measured against, and a partner that accepts them as references will produce something that looks right to someone, once.

A note on skin. Skin is not one colour and should never be corrected with the garment's gain. Lips are their own entity, cheeks and forehead are another, and a cast that reads on the whole frame usually sits on one of them. A partner that corrects the frame as a whole will fix the backdrop and shift the face.

What "measured per part" actually means

The honest version of colour control has four steps and none of them is a prompt.

  1. Generate. The frame is produced on the locked model identity from the garment's own photograph.
  2. Measure. Each part of the frame is isolated, the hero garment, the second garment, the skin, the backdrop, and compared to its own reference, band by band, not as one median. A median cannot see edges, highlights or contrast; it will pass a frame whose shadows have gone green.
  3. Correct part by part. Each part is moved towards its own reference and nothing else. One global correction that fixes a magenta backdrop will push a white shirt green; a per-part correction cannot, because the shirt is only ever compared to the shirt.
  4. Measure again, then a person. The pass is re-measured, and a retoucher looks at every frame anyway. The instrument catches the drift; the eye catches the cast that sits only on the lips, the print whose scale changed, the hardware that turned from brass to steel.

The order matters as much as the steps. Lighting is fixed before colour, because a frame with a wrong shadow on the face has to be relit, and relighting changes colour. Colour measured before a relight is colour measured twice.

The three checks a partner must run

If a partner cannot describe these three, the colour conversation is moot, because the inputs that break colour are upstream of it.

1. Preflight, before anything is spent

Every SKU folder gets a verdict before a frame is generated: will generate, will not, check first. Missing references, two garments in one folder, a character sheet that could mean two models, a styling piece filed with the plain look. At 1,000 SKUs a week nothing can be generated on a guess: in one studio's batch, 57 SKUs were held at this step with one question each, and none of those 57 became a wrong frame. The twelve rules are a one-page checklist you can run on your own folders first.

2. Measurement against your references

Per part, per frame, with a number per batch. The number we publish is the share of frames within tolerance after the pass and the share made worse, which has to be zero. On that same batch the pass lifted frames within tolerance from 80% to 88% on the 115-frame check set, with no frame worse. If a partner has no number, the partner is not measuring.

3. A person on every frame

Not a sample. Every frame, against the reference, before it is filed in your tree, with an uncorrected copy filed beside it so your own editors can see what the pass did and undo it if they disagree.

Questions to ask before a pilot

  • What do you measure colour against, and per what? The answer should name your photographs and name the parts.
  • What number do you publish per batch? Within tolerance, made worse, rejected by a person.
  • What happens to a SKU whose inputs are ambiguous? The only right answer is that it waits for one question.
  • Do we get the uncorrected frame too? If your editors cannot see the before, they cannot trust the after.
  • What does a correction round cost? Frames that fail your written rules should be redone inside the batch, not quoted as a new one.
  • Is the AI provenance written into the file? The EU transparency rule has applied since August 2026; a partner that marks every file saves you a compliance pass later.

What it looked like on a real batch

Our line is built for 1,000 to 1,500 SKUs a week, every view per SKU, in 4K. A European e-commerce studio sent one week of production: flat lays, mannequin shots, character sheets, their own lighting deck and their own product-page photographs as references. Fifty-seven SKUs were held at preflight, the rest were delivered in 4K with every view per SKU in eight days, and when the art director's notes came back the batch was reworked and redelivered within seven, with seventy-four frames regenerated and none failing the second check. The case study walks through it, and the gallery shows eighteen looks from two such batches at web size.

None of that is a promise that a model obeyed. It is a measurement, a gate, and a person, in that order, which is the only arrangement under which on-model work can leave the building and come back the right colour.

Frequently asked questions

Why does colour break when on-model photography is outsourced?

Because the supplier's output is judged against a memory of the garment rather than a measurement of it. A studio shoot carries the garment through one lighting rig and one colour pipeline; an outsourced set is produced under different light with a different pipeline, so every frame has to be measured back against the brand's own reference photographs, part by part, or the drift goes unnoticed until the customer returns the item.

What should a colour reference for on-model production look like?

A photograph of the garment that the brand already trusts: the packshot, the flat lay or the mannequin shot used for the product page, shot under the brand's own light. One reference per garment in the look, including the second garment. A swatch name or a mood image is not a reference, because nothing can be measured against it.

What does measured per part mean?

The hero garment, any second garment, the skin and the backdrop are each compared to their own reference band rather than the frame being corrected as a whole. One global correction that fixes a magenta backdrop can push a white shirt green; a per-part pass cannot, because each part is only ever moved towards its own reference.

What numbers should a supplier publish per batch?

The share of frames within tolerance after the colour pass, the share that were made worse (which should be zero), and the list of frames a person rejected and why. The line runs 1,000 to 1,500 SKUs a week, so the pass has to be measured per batch, not eyeballed: on one studio's batch it lifted frames within tolerance from 80% to 88% on the 115-frame check set with no frame made worse, and every frame still passed a human check before delivery.

Can colour accuracy be guaranteed on AI-generated images?

Not by a prompt. It can be guaranteed by measurement: generate, measure against the reference, correct part by part, measure again, and let a person reject what the instrument cannot see, such as a cast that sits only on the lips or a print that changed scale. A supplier that promises accuracy without a measurement step is promising that a model will obey, and told is not will.