A better Tiny Desk AI prompt
Separate identity from performance: one portrait controls the lead, one reference controls the session.
Give each reference a clear job
Use the portrait for the lead vocalist’s identity and the video for motion, the room, band placement and camera timing. A text prompt alone cannot reliably recreate a particular performance, its exact edit or its music.
Example prompt structure
Edit the supplied performance reference. Use the portrait person only for the lead vocalist. Recast the supporting musicians as a fresh ensemble of fictional adults for this video, with new faces, hairstyles and outfits. Apply the new identity from the first visible frame. Keep it across every camera angle, profile, close-up, occlusion and reappearance; do not blend faces or restore the original face after a cut. Preserve the supporting roles, positions, actions, room, lighting, camera moves, shot changes and exact source timing. Preserve the lead’s gesture and mouth-opening timing. Do not transfer the portrait identity to another band member. Keep the portrait’s clothing consistent. Preserve performance headphones, microphones and instruments even if absent from the portrait. Keep each newly created supporting identity consistent through all shots and occlusions, and keep the number of people unchanged. Do not revert to the reference faces or copy the lead identity onto the band. Do not introduce new shots, text, logos, duplicate people or new audio.
Add output constraints after the edit scope
Specify the intended timeline, aspect ratio and audio workflow only when the model supports them. In this studio, the source controls the 19-second vertical timeline, visual generation is silent and a matching configured soundtrack is attached afterward.
Common problems to look for
A vague instruction such as make me sing at a tiny desk can create a new band or room rather than edit the supplied clip. Competing identity descriptions can mix faces. A different audio track can break mouth timing. Review the result after every camera cut and near the end of the clip.
Use the preset without writing a prompt
The studio sends a fixed server-side prompt and template. You choose the portrait and review the fixed 19-second, 480p session. The client cannot replace the model, reference video or credit price, which keeps the workflow controlled and its cost clear. Supporting identities vary by generation, and model consistency is not guaranteed.