跳到正文
DEV Community· yina lin·· 54 分鐘前AI 評分44

Remix Motion 如何在沒有動作參考的情況下製作 Tiny Desk AI 影片模板

How we built a Tiny Desk AI video template without a motion reference

AI 導讀

Remix Motion 製作 Tiny Desk AI 影片模板時,沒有使用動作參考影片,而是以文字描述場景、人物位置和鏡頭節奏。團隊量度原片為 99 BPM,並按 1.4、6.7 和 12 秒的切點設計 15 秒影片;模板提供一、二或三張照片的版本,缺少照片的人物由 AI 虛構。生成音訊是 99 BPM 的原創靈魂樂,不使用原曲旋律或歌詞。

正文

In late October 2026, a TikTok parody put five college football coaches behind NPR's Tiny Desk, singing Charlie Wilson's performance of "Outstanding". It passed 2.8 million views, and within days a whole trend grew out of it: people put their friends, coworkers and teams behind the desk, with the same beats every time. The band is in full swing, the lead singer cracks up, and the sax player takes the solo.

We make video templates at Remix Motion: you upload a photo, pick a trend, and get your own version of it. Most of our templates work the same way: we take a real clip of the trend, turn it into a depth video, and give that to the video model as a motion reference next to the user's photo. The model copies the movement and puts the user in it.

Tiny Desk was the first template where we threw that approach out. Here is why, and what we did instead.

Why there is no motion reference

A motion reference earns its keep when the movement is the joke: a dance, a fall, someone running at the camera. Tiny Desk is the opposite. It is a fixed, front-on shot of a band at a desk. Nobody moves much. The comedy is in three moments: the band sounding great, the singer losing it, the sax taking over.

A depth video of that clip would mostly tell the model "people standing still in a cluttered room", and it would also lock the framing to the original band's positions. We would be paying for a constraint that gives us nothing.

So the scene is written out in text instead: the shelves crammed with records, who stands where, when the lead turns to the camera, who plays the closing solo. The model gets the user's photos and that description, nothing else.

Timing it against the original

Without a reference clip, timing is the part most likely to drift, and timing is what makes a trend video feel right when people post it with the trending sound. So we measured the clip that started the trend.

  • The band plays at 99 beats per minute.
  • There are three camera cuts, at 1.4, 6.7 and 12 seconds.

That gives four shots in 15 seconds:

Time Shot
0 – 1.4 s The lead singer at the mic, eyes closed, a smile starting
1.4 – 6.7 s The whole band behind the desk
6.7 – 12 s Close-up: the lead turns to the camera mid-line and cracks up
12 – 15 s The sax player takes the solo, the keyboard player grins

The description gives every shot its start and end time and states the tempo, so the cuts are meant to land on the beat, the same way they do in the original. That matters because many people post with the trending sound on TikTok, even though our own audio is not that song.

Three seats, three versions of the scene

The trend is a group joke with one star, so the template has three seats: the lead singer, the sax player and the keyboard player. Only the lead needs a photo.

One description with optional slots would leave the model guessing: if it says "the sax player from photo 2" and there is no photo 2, there is nothing to go on. So there are three separate versions of the scene: one for one photo, one for two, one for three. When a seat has no photo, its version says plainly that the model should cast a person for it. An empty seat is meant to be played by someone the AI makes up rather than left blank.

The price is the same whichever version runs, so users are never pushed to upload fewer friends.

Original music on the same beat

We did not want to ship the song. Instead, the video model generates the audio with the picture: an original soul track at 99 BPM, with the band, wordless singing and the lead's laugh where the close-up is. It shares the original's tempo and feel, not its melody or lyrics, and the office is our own set rather than NPR's footage. Anyone who wants the real song can add the trending sound when they post, and the cuts should still fall on the beat.

Which clips we show

Most of the biggest Tiny Desk videos star coaches, NBA and NFL players, or sitcom characters. We do not show those, and the template does not accept photos of public figures. The example clips on the page are from everyday creators who gave us permission, and the template is meant for you and friends who agreed to be in it.

What we would do differently

Writing the scene instead of using a reference worked here because the shot is static. For any trend where the movement is the point, we still use a motion reference. The real lesson was smaller: measure the original before writing anything. The tempo and the three cut points shaped every line of the description, and they are what make a 15-second clip feel like the trend rather than a lookalike.

If you want to try it, the template is here: Tiny Desk AI on Remix Motion. It is an independent tool, not affiliated with NPR, NPR Music, Tiny Desk or TikTok.

This post was written with help from an AI assistant and checked by our team.

來源:DEV Community · dev.to