Adding a photo of a specific garment (in the demo, a coat) to the reference images so the model dresses the subject in it while keeping her character consistent.
The host lists a changed outfit among the flaws he has to fix. When the garment drifts, he re-prompts with the right references rather than accepting the result. See generate-then-correct-hybrid-workflow.
A fashion sketch can play the same role as a garment photo. See Compact and non-verbal inputs instead of long descriptions.
The demo combines a garment reference with a location reference and a period cue in one prompt: 'Make her wear the coat and stand straight in this location in the 1970s'. The prompt names the garment, the pose and the era, so the references supply the coat and the place.