Keeping the same AI character across images — the actual workflow
The question everyone hits on day two: "I made a character I love — how do I get them again?" Text-to-image is a dice roll that invents a new face each time. The answer isn't a magic prompt — it's switching tools at the right moment.
Why re-rolling never converges
Describing the same face in the same words gets you relatives, not the same person — each generation samples a fresh face from the description's neighborhood. Consistent descriptors narrow the neighborhood; they never collapse it to a single identity.
This is the category error most people make: trying to get persistence from a tool that generates novelty. Persistence comes from a different tool entirely — the edit.
The base-image workflow
Roll until you land one strong portrait — that's your base image, and it stops being a prompt problem. Everything after is an edit on that file: outfit changes, pose changes, lighting changes, scene changes. Each edit preserves the face because that's what edits do — they change what you name and carry over what you don't.
One base, many variants: the same character in ten outfits is ten edits at a few cents each, not ten generation rolls hoping the face comes back.
Anchors that protect identity
Phrase the constraint explicitly: "keep her face and expression identical" in the edit instruction narrows what the model may repaint. The more the edit moves (extreme pose, heavy restyle), the more that anchor matters.
Chain, don't stack: change one thing per pass — outfit, then pose, then light — rather than asking one edit to move everything at once. Small anchored changes hold identity; big multi-changes drift it.
What holds and what drifts
Held nearly pixel-faithful: outfit swaps, scene relighting, weather, background changes. Held recognizably: pose changes, medium restyles (photo to anime redraws the face in the new style).
The drift case is big structural edits on photorealistic faces — a hard pose contortion can reshape features at the edges. Stylized or illustrated characters absorb structural edits far better.
Prompts that work
"Change her outfit to a rain-soaked trench coat, keep her face identical." — wardrobe pass with the anchor.
"Same woman, now leaning against the wall, arms crossed — keep face and outfit." — pose pass stacked on the previous result.
"Relight to night, neon behind her — preserve face, outfit, pose." — light pass holding everything else. The mechanics of each pass type: outfit, pose, light.