Every long-running AI catalogue has a moment of horror: you scroll your own page and count three different women wearing your star’s name. Prompt discipline helps, but the strongest fix is a file, not a phrase — one anchor face, chosen once, that casting tools apply across stills and clips alike.
What makes a good anchor
The anchor is a production portrait, not a beauty shot: front-facing, evenly lit, face clear of hair and hands, resolution comfortably larger than any face it will be mapped onto. Produce a dozen candidates at standard prices in one sitting, judge them at full size, and pick the one that looks most like the woman your character block describes. Then freeze it. The anchor file does not get “improved” mid-season; version changes are how catalogues split in two.
Casting is geometry before it is taste
With the anchor fixed, every target gets judged on two axes: head angle within roughly thirty degrees of the source, and light arriving from the same side at similar softness. Targets that pass produce swaps you cannot see. Targets that fail produce the familiar pasted look — and no amount of re-rolling fixes a geometry problem. Choose better targets, or route the frame through generation with the block and let casting handle the next one.
One file, both formats
The same anchor casts stills at 26 credits and video through the video face swap, priced by length and resolution. That shared source is the whole system: a subscriber moving from your photos to your clips meets one continuous person. Keep the anchor with the character block in the same folder, dated, backed up. Lose it and you do not lose a picture — you lose the cast.