A client approves the first shot in ten minutes. The character looks right: the jacket, the hairline, the way the light sits on one cheek. Then you generate shot two — a wider angle, same scene, same person — and the jaw is longer, the eyes are set differently, and the jacket has quietly become a different cut. Nobody in the room can say exactly what changed, but everyone can see it. Multiply that by fourteen shots and the film stops being about the character and starts being about the drift.
This is the single most common reason an AI video project stalls in review. Not bad models, not slow rendering. The face moves. And once a viewer's eye has locked onto a face, it registers a two-millimetre change in eye spacing faster than it registers a wrong line of dialogue.
Here is how we keep a character stable across a full sequence of AI shots, in the order we actually do it.
Why the face drifts in the first place
Every generation is a fresh guess. Even with the same seed and the same prompt, a different camera angle asks the model to invent information it has not seen: what this person's ear looks like, how the collarbone sits under the shirt, what the back of the neck does. The model fills those gaps with something plausible — and plausible is not the same as consistent.
Three things make it worse in practice. Long prompts with competing instructions, because the model averages them. Wide shots, because the face occupies fewer pixels and the model has more room to improvise. And descriptions written from memory rather than from a visible reference, which is where the real damage happens: two people describing the same character will write two different characters.
The fix is not a better model. It is treating the character as an asset with a fixed specification, the way a physical production treats a costume.
Step one: write the character sheet before generating anything
We build a character sheet as a document, not a feeling. It has four blocks, always in the same order, always with the same wording:
Subject block — age range in years, build, face shape, skin tone described plainly, eye colour and shape, hair length, texture and parting, distinguishing marks. No adjectives that need interpretation. "Late thirties, narrow jaw, deep-set brown eyes, straight dark hair parted left, small scar on the right eyebrow" survives translation into a prompt. "Handsome and charismatic" does not.
Wardrobe block — every garment named, with fabric, colour and fit. If the shirt changes in shot nine, that is a documented wardrobe change with a shot number attached, not an improvisation during generation.
Lens and light block — focal length, aperture feel, key direction, colour temperature, contrast level. This block is what stops shot two from looking like a different production even when the face is right.
Motion block — how the character moves, their resting posture, their habitual gestures. A person who stands with weight on the left leg in shot one and both feet square in shot four reads as a different person even if the face is identical.
Then, before touching the sequence, we generate a turnaround: front, three-quarter left, three-quarter right, profile, and a back view in the same wardrobe and light. If the turnaround is not consistent with itself, nothing downstream will be. This step costs a handful of generations and saves entire days.
Step two: lock the visual grammar, not just the face
Consistency is a property of the whole frame. A perfectly stable face in a shot with a different lens and a warmer grade still reads as a cut to another film.
So the sequence gets a locked technical spine: one primary focal length for dialogue, one wider length for establishing, a fixed key direction relative to the character, a fixed contrast curve, and one LUT applied at the very end rather than baked into each generation. We write these into the shot list, and the prompt for every shot inherits them rather than restating them in new words. Restating is where drift enters — the third time someone paraphrases "soft side light from camera left" as "gentle lighting from the left", the shadow moves.
We also cap how much can change per shot. One variable at a time: new angle, or new wardrobe, or new location. Never all three. In our experience a shot that changes two of the three costs more regenerations than the sequence saves.
Step three: continuity anchors between shots
In a live-action set the script supervisor watches for the same things we watch for here. Our list has six:
1. Hairline and parting, especially at the temple. 2. Collar and shoulder line — jacket cut is the fastest tell. 3. Eye line height in frame at a given shot size. 4. Light direction on the face, checked against the previous shot. 5. Hand and wrist detail near the face, which is where most generations soften. 6. Any accessory that appears twice must appear identically twice.
For each pair of adjacent shots we also carry the previous shot's final frame forward as a reference where the tools allow it. Extending from the last frame of shot four into the first frame of shot five is far more reliable than generating shot five cold from a text description. Where continuation is not available, the still frame goes in as an image reference instead.
Naming matters more than people expect. Every asset carries a code: character, sequence, shot, take. CHR-01-S07-T03. When a sequence has ninety generated clips, the only way to find the one that matches is the code, and the only way to keep the code honest is to apply it the moment the clip is made.
What still breaks
Three failure modes we have not fully solved, and say so up front:
Hands near the face. Models are much better than they were, but a hand at cheek height still produces soft knuckles and occasionally a sixth finger. We compose around it — hands lower, or out of frame — rather than fighting it with regenerations.
Fast turns and crowds. A character turning quickly through ninety degrees, or standing in a group, loses identity fastest. Practical answer: keep group shots wider and treat the principal as a silhouette, or cut away.
Reflections and screens. A face reflected in a window or a phone screen is generated as an interpretation of a face, not a copy of it. If the reflection is meaningful to the story, generate it as its own shot with the same reference and accept that it will be a near-match rather than an exact one.
Reviewing the sequence as a contact sheet
Never review shot by shot. Individual shots look fine in isolation; that is exactly the problem. We assemble all stills into one contact sheet per sequence, nine to twelve frames on a screen, and look at it from a metre away. Drift that is invisible in a viewer becomes obvious in a grid. Fixes are then ordered by cost: grade and framing first, regeneration second, reshoot of the concept last.
Every sequence ships with the character sheet, the turnaround, the shot list and the contact sheet attached. The next project that reuses the character starts from those files, not from a memory of what the character looked like.
Ask for the character sheet before you see a single frame. Ask what happens when a shot fails the contact sheet — is it regenerated, recomposed or dropped. Ask how the character is handed over so a second film can reuse it without rebuilding from scratch. A studio that answers those three questions in specifics will hold a face across a sequence. A studio that answers them with enthusiasm will not.
KAMOD is an AI video studio in Tashkent, working with brands and agencies worldwide. If you have a character that has to survive twenty shots, send us the brief and we will come back with the sheet, the turnaround and a shot-by-shot test before anyone commits to a full production.