AI fashion model

Where the figure comes from

There are two routes, and they behave differently in production. A library model is a pre-generated figure the tool offers you — pick a body type, age and skin tone and render against it. A reference model starts from an image you supply, so the same face and body carry across everything you produce.

The distinction matters more than it first appears, because it determines whether a catalogue holds together. Picking from a library per SKU tends to produce a category page where every product is on a different person — technically fine images, incoherent as a set.

• Library — fastest start, no input needed, weakest at catalogue coherence
• Reference — one identity you supply, held across every render
• Brand model — a reference figure treated as a fixed brand asset, the way a real house model would be

Identity consistency is the hard problem

Generating one convincing figure is largely solved. Generating the same figure two hundred times, across different garments, poses and lighting, is the part that separates usable production tools from demos — and it is what brands mean when they complain that image two hundred does not match image one.

It is worth auditing deliberately rather than assuming. Put twenty renders of the same model side by side and look for the drift that a single image hides: face, hair, and the small proportions of build.

• Face — the most noticeable drift, and the one shoppers register first
• Skin tone under different scene lighting, where warm and cool sets pull it in opposite directions
• Hair length and style, which quietly changes between renders
• Height and build relative to the garment, which changes how the fit reads

Why brands use them

The first reason is coverage: figures cost nothing to schedule, so SKUs that never justified a booking can have on-model imagery at all. The second is representation — showing a garment across body types is a repeat render rather than a repeat casting, which is why extended size ranges are the clearest use case.

The third is control. A generated figure has no availability, no ageing out of a contract, and no usage window expiring — the practical frictions that make a real house model an ongoing commitment rather than a one-off cost.

Limitations and disclosure

The technical limits are the ordinary ones for generated imagery: hands, text and logos on the garment, fine lace and mesh, and complex layering. All of them are why a human should look at output before it is published.

The non-technical limits matter more. Disclosure requirements for synthetic depictions of people are real, differ by market and channel, and are still moving — several advertising regimes and platforms now require it, so the rule is to check each channel rather than assume one answer covers all of them. Separately, a generated figure must not resemble an identifiable real person you have no rights to; that is a publicity-rights problem no amount of generation solves.

Finally, there is a representation critique worth taking seriously rather than dismissing: using synthetic bodies to depict diversity, while not hiring the people being depicted, is a position a brand should decide on deliberately rather than by default.

• Disclosure of AI-generated people is required in some markets and on some platforms — verify per channel
• A generated figure must not be an identifiable real person you lack rights to
• Hands, garment text and sheer fabric remain the routine defects
• Depicting diversity synthetically instead of hiring for it is an editorial choice, not a neutral default

Common questions

See also

Doing this with TinTin