How to Choose the Best Photos for Your Tiny Desk AI Video

Use a sharp, recent-looking photo with the person unobstructed. Avoid heavy beauty filters, stylized illustrations and extreme perspective when the aim is a realistic performance. A small face inside a busy group scene provides less useful identity detail.

Both eyes, the nose, mouth and facial outline should be readable. Soft, even light is generally easier to interpret than deep shadow or harsh backlighting. Avoid glare across glasses, a hand over the face or an angle that hides most facial features.

A portrait can show facial details clearly, but a half-body crop leaves the pelvis, legs and stance to be generated. A natural full-body photo gives the model direct structure cues, provided the face is still recognizable.

The prompts ask for adult proportions independently of the source crop and for the face to remain unchanged while completing the body. These are instructions, not a body-reconstruction guarantee. Do not treat generated hidden anatomy as a measurement of the real person.

Each photo is mapped to one role in the band. Use one person per photo and avoid collages. Putting the same person in several roles does not create a dedicated multi-angle identity workflow.

Use one adult per photo, with a clear face and even light. JPG, PNG and WebP are supported, up to 10 MiB per image. Each side must be 301–5,999 pixels, with width divided by height greater than 0.4 and less than 2.5.

Tight crops, tiny faces, blur, filters and strong selfie perspective can leave the model with incomplete or misleading cues. More participants also mean more identities and instruments to maintain. An upload passing file validation does not mean its visual content is ideal.

Choose a readable face and keep the source person distinct from the background. Compare one photo change at a time using the same output settings. Check the whole clip, because a face or body may look correct in one frame and drift during motion.