Tiny Desk AI Faces, Outfits and Lip-Sync: What to Expect

A reference is guidance, not a frame-perfect edit

This application uses portraits to guide five band roles and reference clips to guide a four-scene performance. The generator makes new frames. It does not simply replace every source pixel while preserving the original body, camera timing and mouth shape exactly.

That difference matters if your requirement is precise identity replacement or exact lip-sync. The recognizable scene structure can be present while a face, hairstyle, outfit or gesture changes. A successful task status means the video completed, not that it passed a perfect-likeness test.

Why multi-person shots need special attention

In a group shot, several identities, outfits and positions compete for the model's attention. A person can remain unchanged, be replaced only after the beginning of a shot, or resemble the wrong reference. A person who looks convincing in one scene may look different in another.

Check the group shots as well as the lead close-up. Looking only at one flattering frame misses the very errors that matter most in a five-person video. Clothing and body shape also need review; replacing a face is not the same thing as reproducing the entire person.

Original soundtrack is not the same as exact lip-sync

The completed video is combined with the reference soundtrack after the scenes are generated. This preserves the chosen audio through assembly, but it does not force the newly generated mouth shapes to sing each syllable at the right instant.

A shift in a smile, breath, head turn or phrase length can make a close-up feel out of sync even if the total duration matches. The saxophone shot and backing vocals introduce different visible performance cues. Matching the final running time is necessary for the format, but it is not proof that every performer is synchronized.

What can improve the input

Use one clear portrait per role, make the face large enough to recognize, show useful hair and outfit details, and avoid occlusions. Do not rely on fine lettering on a jersey as a stable identifier. Outfit guidance can express a preference, but it does not enforce exact reproduction.

Read the photo requirements. Both 480p and 720p use the same model workflow; higher resolution is not a face-consistency or lip-sync guarantee. See the format comparison for the actual differences.

Review the result in passes

1. Watch the entire video with sound for overall pacing and mouth timing.
2. Check the start of each scene for an original face that only changes later.
3. Count the visible people in group shots and match each to the intended role.
4. Compare hair, clothing and body shape across the recurring appearances.
5. Inspect the lead close-up and saxophone section separately.

Do not purchase based on an expectation of exact one-to-one motion transfer. If you need an exact performance for professional delivery, this generative workflow may not meet that requirement.

Failure and quality concerns are different

A task that fails returns the credits it used. A completed file may still have a missing replacement or a visual defect. Contact support@gettinydesk.com with the task ID and identify the scene and time where the issue occurs. Reporting “the backing singer is unchanged at the start of the ending” is more useful than sending only “bad quality.”

A separate silent MP4 is not included. You can mute or replace audio in an editor, but changing the audio alone does not repair generated mouth timing. Review the workflow guide before starting another task.