The face changes when the character turns
A single front-facing still cannot constrain a profile, so identity collapses on the turn. Build a multi-angle sheet, and keep the turn as a cut.
A front-facing reference is a contract about one hemisphere of a head. The moment the character turns, the model has to invent the side of the face you never showed it — jaw in profile, ear, the way the nose sits against the cheek, hair at the temple. It invents someone adjacent. You did not lose identity because the prompt drifted. You lost it because the reference never covered the angle you asked for.
This is the coverage failure, not the omission failure. Four consistency mechanisms names it: three front-facing portraits will not carry a shot from the side, and correctness tracks camera angle. The signature is precise. The character is right whenever they face you, and a different person in every profile. Padding the prompt with more eye colour does not describe a surface the model has never seen.
Two habits close it. Build a multi-angle sheet before you need the turn. And do not generate the turn inside a single clip from a front still. Cut to a new shot that is already the profile, anchored on a profile reference.
A front still is a one-sided contract
Nothing carries between generations by default. A reference image is the exception: pixels the model can copy an identity from, without that picture becoming a frame of the output. That exception has a shape. The model can re-pose, re-light, and re-frame what it has evidence for. It cannot retrieve an ear it was never shown.
A library of five near-identical hero portraits is one reference, five times. Slot count is not coverage. How many reference slots a shot actually needs ranks a new angle on the same subject as the highest-value second and third slot, specifically because it is the only way the model learns the side and the back of the head. Filling those slots with tighter crops of the same smile concentrates whatever is idiosyncratic about that one pose. Then the character can only stand one way, and the turn still breaks.
Image-to-video treats the input as frame zero, so a front still is a perfect front at t=0 that decays as soon as the head yaws. Reference-to-video is the better family for a turn you must see, because a reference is an identity the model may place at another angle — but only if you gave it more than one angle to generalise from. A single front reference asked to "turn to look off camera" is a request to hallucinate a profile.
Pause at the start, the midpoint of the turn, and the settle. If frame one is your person, the midpoint is a blend, and the settle is a neighbour, the reference ran out halfway through the yaw.
Build the turnaround before you need the profile
Do this as its own job, before any clip that contains a turn. The mascot version of the same artefact is a turnaround sheet — front, three-quarter, profile, back, one generation, one grid. For a photoreal spokesperson the grid still works, and so do four separately directed stills if you lock lighting, lens, wardrobe and expression, and only change the yaw.
Minimum set that actually covers a head turn:
| Panel | Yaw | What it pins that the others do not |
|---|---|---|
| Front | 0° | Both eyes, symmetry, the smile you will be judged on |
| Three-quarter | ~45° | How the far eye sits, cheek plane, hair at the temple |
| Profile | 90° | Nose-to-lip line, ear, jaw hinge, the silhouette |
| Back three-quarter | ~135° | Hair volume, nape, earring, the collar from behind |
Front plus profile is the pair that does most of the work. Three-quarter is the panel you will use most often in finished shots, because it is how you photograph a person talking to someone slightly off camera. Back three-quarter is optional until you write a shot that needs it, and then it is not optional.
Generate the grid in one pass if the character is stylised. Consistency inside a single generation is stronger than consistency across four calls. For photoreal, one pass can fuse four people into a grid; if that happens, generate four stills from the same master and freeze everything except yaw.
The same woman as the reference, studio key from
camera left, 85mm, grey seamless, neutral expression,
closed mouth, hair in a low bun. Camera at eye height.
Three-quarter view, head turned 45 degrees to her left,
shoulders still square to camera.
Change only the last two lines for each panel. Do not improve the wardrobe description between panels. Do not change the smile. A sheet whose expression varies with yaw is four identities.
Then crop the panels and register them as one character, not four. Set up a reusable character and attach front, three-quarter and profile at minimum. Downstream shots should point at that asset by key, then name the panel the shot actually needs: "use the profile reference." Hygiene still applies. Same light, same crop convention, no mixing a clean studio profile with a frame-grab from a previous clip. Mixed-condition sets are how you get a third person who is the average of the two.
Don't generate the turn. Cut it.
A head turn inside one generation is a continuous invention of angles you may not have covered. Even with a three-panel sheet, a five-second yaw asks the model to interpolate surfaces between the panels, on every frame, while also doing motion. That is a hard ask. A cut is a cheap ask.
Write the sequence as two shots.
Shot A. Front or three-quarter, held. The line, the look, the hold before the turn. Generate it from the matching panel. Prompt no yaw.
Shot B. Profile, already settled, looking at the new target. Generate it from the profile panel. Prompt no yaw.
Cut on the intention to turn — a blink, a breath, a look off — the way coverage always has. The audience reads a turn. The model never had to draw one.
When you truly need to see the motion, stop using a single still as an image-to-video start and hoping. Pin both ends.
- Take the front panel as frame one and the profile panel as the last frame.
- Generate first-last-frame so the clip is a bounded interpolation between two identities you already approved.
- Keep the two stills in the same location, lighting and lens family, or the interpolation reads as a morph rather than a yaw.
- Keep the clip short. A two-second turn between matching stills is a camera-and-neck problem. A six-second turn is an identity problem with extra time to drift.
Do not also prompt "she slowly turns her head then smiles then looks back." One motion. The stills already contain the faces; the prompt only has to describe the yaw and the camera.
If the model you are on will not take two stills, generate shot A and shot B as separate clips from the matching references and cut.
Shot language that keeps identity in frame
Once the sheet exists, the prompt should stop re-describing the face. Re-description fights the reference. Spend the prompt on the angle, the eye-line, and what must not happen.
| Instead of | Write |
|---|---|
| "She turns her head to the left" on a front-only ref | Two shots, or first-last-frame between front and profile stills |
| "A beautiful woman with green eyes and a sharp jaw" | "From the profile reference, looking camera-left, hold" |
| "360 orbit around her face" | A locked three-quarter; orbits invent the back of the head |
| "Looks over her shoulder at the camera" | Back three-quarter panel first, then the look |
Eye-line is the small version of the same bug. A front reference asked for a profile gaze will often yaw the whole head to fake the look, and that unplanned yaw is where identity slips. If you need eyes off-camera while the face stays front, say "eyes only, head still." If you need the head to yaw, you are back in the shot plan above.
For a sequence of many angles, fill reference slots with categories (front, profile, three-quarter), not with duplicates. A nine-slot model fed nine frontals is still a one-sided contract. Review every clip that contains a yaw at three timestamps, not at the poster frame. The poster is almost always the front, which is the angle your reference already could do.
FAQ
I only have one good photo of the spokesperson. Can I synthesise the profile?
You can generate a profile candidate from the front still, then you have to approve it as a new master, not treat it as recovered truth. Generate several, pick the one that still reads as the same person, and only then put it in the sheet. If none of them do, do not ship turns. A guessed profile that you then chain from is how the neighbour becomes the lead.
Why does a three-quarter still survive a small turn when a front still does not?
Three-quarter already contains most of a profile: the far cheek, the nose against it, one ear. A small additional yaw stays inside evidence. A front still contains none of a profile. The first twenty degrees of yaw are already invention.
Should the sheet be one grid or four files?
One grid for stylised characters, because the model holds all panels at once. Four files for photoreal if the grid starts fusing identities. Downstream generations should be fed the cropped panel that matches the shot. A full grid in a single reference slot often teaches the model "this character is a comic page."
Can motion control replace the sheet?
No. Motion control copies a performance onto a subject. It still needs a subject that is correct at the angles the performance visits. A driving clip of a head turn mapped onto a front-only character is a well-timed collapse. Build the sheet, then transfer motion if the timing is the thing you cannot prompt.