Distinguish identity drift from a changed view
A profile reveals the nose differently; backlight darkens skin. Neither alone proves a new identity. Compare accepted structure: jaw shape, head-to-neck/shoulder relationships, hairline and permanent features. Expression also changes appearance; every frame need not resemble a frontal ID photo.
The asset method calls for observable recognition beyond the front. If eyebrows and eyes are the only identifiers, the back view has little support. Use already-established hair silhouette or shoulder shape without inventing scars and accessories.
Locate the first drift: inconsistent references, a wrong start frame, structural change in profile, or a different face on returning front. These need different repairs rather than one diagnosis of instability.
This page addresses identity during motion. For selecting looks across wardrobe changes, see look continuity.
Give each actual reference a clear role
| Check | What to establish |
|---|---|
| Same identity | Front, side and back references match the accepted person and current look, without mixing age stages or injuries. |
| Actual image | Open the file; an image prompt does not establish a generated picture. |
| Role | The start frame owns the opening state; identity reference owns recognition. Bind through the target’s supported method. |
| Retained facts | Clothing, hair, age and held objects follow the accepted setup rather than accidental reference differences. |
| Actual input | Check submitted media, not merely filenames listed in a document. |
Not every tool accepts front, side and back inputs together; more images are not automatically better. Use the target’s actual reference method. This article supplies no universal multi-image parameter. If a frontal reference cannot support needed side/back recognition, state the gap and prepare the relevant view.
Profile, back view and turning remain one person. Record lasting look changes separately from momentary orientation rather than repeatedly rewriting identity.
Original example: the turn delivers recognition
Original reading-room scene: Ji Zhou faces away, arranging a book cart. The entrance attendant calls “Ji Zhou.” He turns toward the entrance, ending with his face visible so viewers recognize him. The camera holds a fixed medium shot from the entrance side, looking toward the cart rather than orbiting. Clothing and injury state do not change. Accepted appearance: narrow long face, a short notch at the left eyebrow tip, naturally curly black hair tied low, and slightly forward shoulders. The cart stays put; both hands leave its handle before the turn.
Production needs an actual start frame and usable identity reference. This page provides only a paper example, with no image files or generated result. Its text is not an already-bound production task.
Ambiguous draft:
The person turns beautifully as the camera rotates. Keep a beautiful face, back to front, flowing clothes and increasingly dramatic atmosphere.
It mixes camera motion with the body’s turn and does not identify the person. Removing the turn would remove the required recognition. Return to the accepted action and camera.
A complete core prompt: who turns and where it ends
Prepare actual references and the target dialect before using this model-independent core:
Begin from the accepted rear-view reading-room start frame. Ji Zhou is beside the cart with both hands on its handle. Hold the established medium shot; the storyboard defines the entrance direction. The offscreen attendant calls “Ji Zhou.” He releases the handle, turns his body toward the entrance and brings his head around with it. He stops facing the entrance with his face visible from the current camera. Retain the same head silhouette, shoulder relationship and current low-tied curls through the turn. When the face becomes visible, match the accepted facial structure and left eyebrow notch. Keep the cart in place and clothing unchanged. The camera remains fixed, with time after the turn for recognition. The attendant speaks the name; Ji Zhou does not mouth it.
This retains trigger, release, body/head movement, face visibility and recognition. It adds no handoff, running, wardrobe change or orbit. Duration follows the accepted shot and actual action, not a universal three-second turn.
If the actual start frame does not show both hands on the cart, correct the starting description. If this camera cannot see his face toward the entrance, revise the storyboard’s camera or endpoint rather than quietly moving the camera in the motion prompt.
If drift persists, inspect the angle and submitted references before preparing a view, regenerating locally or proposing a storyboard revision. Projects permitting cuts can compare a continuous turn with cuts at actual action boundaries. A required continuous reveal retains the full turn rather than replacing it with an unrelated frontal portrait.
Review the candidate through the turn
Watch the complete action at normal speed, then locate the side/back transition and return to a visible face:
- Are hands released, body turned toward the entrance and movement stopped, rather than head and body facing opposite ways?
- Do front, side and back cues belong to the same person? Are established marks retained rather than moved or invented?
- Are normal pose/light differences mistaken for identity failure?
- Do camera, cart and clothing remain consistent, and does the attendant speak the name?
- Is the recognition endpoint actually delivered, with a usable state for the next shot?
Matching endpoints can conceal drift between them. Stills locate issues without proving natural movement. Record the failed view and changed references or instructions before comparing the candidate. A consistency instruction is not a passing result.
FAQ
Is “keep the same character” enough?
It states a goal without supplying identity evidence, real references or clear action. Locate the first drift, check submitted inputs and inspect the candidate.
Can I replace the turn with a static frontal shot?
Only if it preserves the accepted purpose. A turn delivering recognition, discovery or response cannot be silently removed during prompt revision; propose a storyboard choice.
Do it with the skill
Use $short-drama-video-prompts to diagnose this face change. Supply storyboard boundaries, accepted identity/look, actual references and the target reference method. Distinguish identity, view, binding and action errors before a local prompt revision. Retain the turn, story information and camera; list missing references instead of treating image prompts as files.
Read the method: Identity across views and look separation · View changes do not create new identities · Action conditions, direction and completion · Reference starts and state comparison
About Drama SkillsThe short-drama-video-prompts skill on GitHub