Prepare two separate portrait files with one adult in each. Use clear faces, even lighting and visible shoulders. A portrait supplies a face; it does not fully describe a body or a singing performance.
The photo checklist
Different backgrounds are fine. You do not need to paint an orange studio into your portrait. If you only supply a headshot, describe the intended outfit and natural body proportions so the tool does not simply inherit the reference performer's build.
- One adult per file, with permission to appear in an AI video.
- An uncovered face without heavy filters or a hand over the mouth.
- Visible shoulders or upper body, and some room around the head.
- JPG, PNG or WebP, up to 5 MB per photo.
- Separate files for the left and right people; group-photo selection is not supported.
A reference-performance prompt to adapt
This example needs two image references, a five-second performance video and its matching audio. Rename the reference labels for your chosen tool. It is an instruction example, not a guarantee of exact likeness or lip sync. The current public studio has no custom prompt or reference-video field.
Create a clearly AI-generated five-second orange-booth duet at normal speed. Use Image 1 for the adult on the left and Image 2 for the adult on the right. Preserve their recognizable faces and realistic body proportions. Use Video 1 for the scene, framing and gestures. Audio 1 is the matching five-second soundtrack. The left person sings the lead vocal from the first frame, with visible lips and jaw following the syllables. The right person dances and reacts, with brief backing responses only when the audio contains them. Keep both mouths visible. Do not swap sides, mirror people, add an intro or extend the timeline. Keep the microphone, outfits and identities stable.An example to adapt in a video tool.
Change one part at a time
Identity and body
State which portrait belongs on each side. Keep descriptions consistent with the photo. Specify a natural build and fit of clothing if the reference body is unsuitable; a headshot is not a body measurement.
Singing and movement
Say who leads and who supports. Audio determines the vocal timing. Reference video guides movement, but copying each source mouth blindly may leave the intended singer silent.
Framing and duration
Keep both faces visible and start with one short exchange. Upload only the chosen source interval and request that duration. Busy choreography, a camera change or a longer clip should be separate experiments.
Use the result to choose the next change
A provider's success status confirms delivery, not a good performance. Review the whole clip with sound and choose one specific correction. This limits confusion and avoids paying for repeated experiments without knowing what changed.
- Face drift: inspect the reference and simplify movement.
- Distorted hands: try a smaller gesture with more space between people.
- Wrong singer or closed mouth: check the lead assignment and matching audio.
- Body too broad: clarify body proportions and clothing; preserve the face.
- No sound: check mute, volume and the downloaded file before generating again.
- Unknown submission: check the existing task and charge; do not immediately retry.