How to keep the same AI girl across a whole set

Image models do not have a person in memory. They have a cloud of faces that match a sentence. "Brunette, green eyes, hotel room" is a crowd. Each click samples a different woman from that crowd.

• You change the prompt's face words between shots — "soft jaw" on Monday, "sharp jaw" on Tuesday.
• You feed a render back in as the new "her" and repeat. Each generation adds error. By image eight she is a cousin.
• You switch tools mid-set. Different models have different average faces.
• You ask for a new hairstyle, ethnicity, or age band and expect the rest of the face to hold.
• You generate a wide shot and a tight selfie from text only. The model invents two different noses.

You will also see body drift — hands, torso length, moles that teleport. Fans notice the face first. Lock the face, then be picky about the rest.

A long prompt can bias the crowd toward one look. It cannot pick the same person twice. What actually holds a face:

A prompt is still useful. It should talk about the shot, not re-describe the person.

Treat the saved character like a model who already showed up to set. You are only moving the camera and the wardrobe.

A second set can change the room — still the same reference. If you want a new hair colour as a "new era," make it a dated event on the page ("went darker this week") and keep it for a month. Surprise redos look like a different subscription.

Clips inherit the still. A 5-second move from frame 4 is still her, because the input face is the same file. That is the photo-to-clip path: 20 credits for the still, 20 credits a second for the motion.

Do not learn this on a live wall. Generate a nine-up on purpose: three rooms, three outfits, three distances (wide, mid, tight). Same reference, same hair, no "make her hotter" in the prompt.

• Eye colour and spacing
• Nose and jaw
• Apparent age
• Teeth, if she is smiling
• The signature mark

If two of nine look like a sister instead of her, you do not have a character yet — you have a vibe. Go back, pick a different preview, or tighten the identity kit. 8 credits for another four previews is cheaper than a week of confused DMs.

If the nine-up fails, you spent $1.80 to find out — not a $24 subscription month.

Once the stills match, motion is the cheap way to make the page feel alive. Pick the still where the face is clearest and the body already suggests a move: hair over a shoulder, weight on one hip, a hand on a robe tie. Prompt the motion, not a new person.

5 seconds is 100 credits — the teaser you put on the wall and the SFW cut you post on Instagram. 10 seconds is 200, usually enough for a paid preview. Full economics in the PPV teaser guide.

Competitors advertise "100% same face." Faces still drift. Lighting changes, a smile changes the jaw, a wide lens changes the nose. Anyone who has generated more than twenty images knows this.