Find the first frame where identity changes
Watch the result once at normal speed, then pause near the first moment a face stops matching the portrait. Drift often begins during a run, head turn, partial occlusion, embrace, or the hard cut into the memory. Identifying that moment tells you whether the source image, motion, or transition needs attention.
Replace the weaker portrait first
Use an original image with one large, evenly lit face, both eyes visible, and minimal compression. A tiny social-media crop, beauty filter, sunglasses, dramatic side profile, or another person in frame gives the model fewer stable identity cues. Do not replace both photos at once unless both are clearly weak.
Keep Photo 1 and Photo 2 visually compatible
The two people do not need matching backgrounds, but similar crop size and lighting can make the pair easier to distinguish. Avoid one extreme close-up beside one distant full-body image. Confirm that Photo 1 is the person who turns and Photo 2 is the partner who recognizes them.
Remove optional direction before adding more
Each scene already includes its action order, camera, lighting, and sound design. If you added a fast spin, costume change, extra location, or dramatic camera move, retry with the built-in prompt alone. More words do not necessarily create more control in a 15-second generation.
Know what a retry can and cannot fix
AI video output is probabilistic, so another run can differ even with the same inputs. A clearer portrait and simpler motion improve the conditions but cannot guarantee exact identity in every frame. Review hands, faces, clothing, and audio throughout the full result before sharing it.