In a multi-person video edit, recognizing a face in the opening frame is only one part of the problem. The same person needs to keep the same role when the camera cuts away and returns. A performer who is central in one shot may be partly visible at the edge of another.
We build Outstanding Tiny Desk AI, a browser tool that assigns one to five uploaded photos to performers in a reference clip. This article describes our current role-mapping instructions and the checks a finished render needs. It is a developer account of the workflow, not a benchmark of identity consistency.
The format draws on Charlie Wilson's performance of “Outstanding” at NPR's Tiny Desk; his official concert announcement provides the performance context. Our template uses a 17-second reference clip, with a required lead singer and four optional band positions.
Start with the reference and the requested edit
The reference supplies the performance: where people stand, how the camera moves, and when shots change. The uploaded photos supply the intended replacement appearances.
Our current generator uses Kling 4.0. The application prepares instructions from the selected roles and submits the photos with the bundled reference clip. This simplified request shows the inputs used by our application. The URLs and prompt are placeholders; it is not a complete authenticated API request or a general specification for every Kling integration.
{
"model": "kling-4.0/reference-to-video",
"input": {
"prompt": "<editing instructions for the selected roles>",
"video_urls": ["<reference-video-url>"],
"image_urls": ["<photo-url-for-each-selected-role>"],
"resolution": "720p",
"aspect_ratio": "auto",
"duration": 17,
"keep_original_sound": true
}
}
These settings express the requested result. They do not establish that every identity, camera movement, or audio cue will survive generation unchanged.
Separate role positions from image indices
With all five photo positions filled, the mapping is:
| Photo position | Role in the reference |
|---|---|
| 1 | Lead singer in the close-up shots |
| 2 | White-cap singer at the front-center microphone |
| 3 | Back-left singer |
| 4 | Back-right singer |
| 5 | Saxophone player |
The optional positions make indexing important. The application builds a matching list of selected roles and image references; it does not send a five-image list with empty slots.
For example, when only the lead singer and saxophone player are selected, the lead maps to @image_1 and the saxophone player maps to @image_2. The saxophone remains the fifth role in the interface, but it is the second supplied image. The three unfilled roles are instructed to retain their original performers.
The useful check is whether every image reference in the instructions still points to the selected role after empty positions are removed. Confusing the interface position with the compacted image index can describe the wrong assignment before generation even starts.
Describe appearances across shots, including partial ones
The lead singer and the white-cap singer at the front-center microphone need particularly clear descriptions. Depending on the shot, either can look like the central subject. “Replace the person in front” leaves that distinction ambiguous.
A failure case worth checking is the front-center performer taking on the lead singer's appearance when the wide shot returns. An opening close-up can look correct while that later assignment is wrong.
The white-cap performer also appears partly in the lower-right corner of the saxophone close-up. That fragment belongs to the same role as the front-center performer in the wide shot. If that role has a replacement photo, both appearances should use it. If the role is unfilled, both should remain unchanged.
Role guide from the 17-second reference clip. This is reference footage, not a generated result or a before-and-after comparison.
With photos supplied for both the lead and front-center performer, the relevant instructions can be summarized as follows:
Replace the lead singer wearing the red cap and black ALABAMA hoodie
with the person in image 1.
Replace the performer wearing the white cap at the front-center
microphone with the person in image 2.
Use that same image 2 identity for the partly visible white-cap
performer in the lower-right corner of the saxophone close-up.
Keep these two identities separate across close-ups, wide shots,
cuts and partial appearances.
This is an English summary of our server's Chinese prompt. The cap and hoodie descriptions identify the source performers. They are not instructions to dress the replacements that way: the replacement's face, hair, skin tone, body shape, and clothing should follow the uploaded photo.
The instructions also preserve the original headset, microphones, and saxophone, including how they are used. Explicit descriptions clarify the intended mapping; they do not guarantee that the model follows it in every frame.
Treat input validation as application-specific
Our upload rules accept JPG or PNG, at least 341 pixels on each side, up to 10 MB, and a width-to-height ratio from 2:5 to 5:2. These are our application's rules, not a claim about every limit of the underlying model.
Use a separate photo for each person. A clear head-to-waist image gives the model information about the face, hairstyle, and clothing. Even lighting, a front-facing or slightly turned pose, and an unobstructed face make the intended appearance easier to describe. Group photos introduce another subject-selection problem before the video roles are considered.
Only use photos of yourself or adults who have agreed to appear. Our service excludes minors and celebrities. Technical acceptance of a file does not establish consent to use it.
Review identity and sound in the returned video
The workflow requests 17 seconds and asks Kling to retain the reference video's original sound. The application saves the returned video with its audio, without replacing or retiming the soundtrack afterward.
Review that returned file with sound. Audio being present does not establish accurate lip synchronization, and a correct first frame does not establish consistent identities throughout the clip.
The review checklist is:
- Does each uploaded person occupy the intended role?
- When a shot returns, does that role keep the same identity?
- Does the white-cap performer remain consistent in the wide shot and at the edge of the saxophone close-up?
- Do unfilled positions retain their intended performers?
- Are faces, hands, clothing, and instrument interactions coherent?
- Does the audio fit the visible singing and movement, particularly around cuts?
The output is a new AI render guided by the reference. Timing, framing, and movement can vary. This workflow does not provide a measured success rate or a guarantee of shot-for-shot reconstruction.
Before sharing, follow the destination platform's AI-labeling and music-use rules. Receiving a video with audio does not grant additional rights to the underlying music or footage. The tool is independent and unaffiliated with NPR, Tiny Desk Concerts, Charlie Wilson, or The Gap Band.
For this kind of edit, role mapping needs to cover the full sequence: the selected-photo index, the source performer, and every relevant appearance of that performer. The prompt describes that mapping; reviewing the returned clip establishes whether the result actually follows it.
AI assistance was used to edit and structure this article from the developer-provided workflow. It does not report an independent test or benchmark.
Top comments (0)