How to Make the Viral Hotel Lobby AI Video (Migos COLORS + Higgsfield Genjutsu)

Short answer: it's Quavo and Takeoff's 2022 COLORS performance of 'Hotel Lobby,' with the two rappers AI-swapped for any pair you like. The orange set, the hanging mic, the timing, the gestures, and the audio all stay; only the people change. The tool most tutorials name is Higgsfield Genjutsu's Motion Transfer; the model behind the cleanest renders is ByteDance's Seedance 2.5; and one-click templates on Summrs, Mitte, and CapCut's Pippit will do it from two photos with no prompt. It's the Yung Lean STORM format with two people instead of a crowd — and a cleaner tutorial, because the hard part is one thing: don't let it swap left and right.

The clip, and why it took off again

'Hotel Lobby' came out May 20, 2022 on Quality Control and Motown, Quavo and Takeoff recording as Unc & Phew. COLORS posted their performance on June 17, 2022 — 'the first in our series celebrating Black Music Month created in collaboration with Motown Records.' Four and a half months later, on November 1, Takeoff was shot and killed in Houston at 28. That's worth saying plainly before the tutorial, because every version of this meme edits him out of his own performance — and because the person who gets to decide how that feels has weighed in.

Quavo posted the real clip on September 23 — 9.2 million views — and on Facebook, per Complex: 'Aye Take we viral again The original!!!' The trend started a week earlier: the earliest version Know Your Meme traces is a September 16 TikTok that put Suits' Harvey and Mike in the booth; two cats followed on September 18 (7.8 million views), Napoleon and Mbappé the next day. XXL put it on hip-hop timelines on September 23 — 'famous pairs rapping along to the track' — and Complex followed with the Obamas, Walter White and Jesse, Haaland and Bellingham, and the pushback: fans objecting to Takeoff's likeness being used, and to the tools 'binge collecting our images.' The Source reports a 540% jump in US streams of the song, though it names no data source.

Check you have the right clip

• Solid orange COLORS set. If there's no orange wall, it's a different video.
• Two performers, left and right, never three. One hanging mic on a cable between them.
• Tight cuts between the two, handheld feel, ~15–30 seconds — not the full song.
• The audio is always 'Hotel Lobby.' The joke is the pairing: rivals, a couple, two mascots, two cats.

The tools, in the order people actually use them

• Higgsfield Genjutsu — Motion Transfer. Launched September 1: 'take the motion from any video and build a new scene around it.' Reference video 4–30 seconds, up to 30 reference images on the web app (the API takes eight), output up to 1080p. Higgsfield's own example: a 15-second clip at 1080p is 144 credits, about $7.20; API rates run from $0.32 a second at 480p to $1.63 at 1080p. Plans start around $19 a month with a free tier — check in-app, the numbers move. Higgsfield doesn't say what model Genjutsu runs on, so neither will we. There's no 'Hotel Lobby' preset; the community's 4–30 second reference clips are just Genjutsu's input limit.
• Seedance 2.5, direct. ByteDance's July 31 release takes 'up to 30 images, 10 video clips, and 10 audio clips as reference materials in a single pass,' tagged with @ syntax — which is where the @Video1 / @Image1 convention in every tutorial comes from. Reachable through Dreamina (CapCut), Kapwing, or API resellers; the cleanest 1080p renders we've seen ran about $16 per 30-second generation via one developer's API write-up. Seedance's Draft Mode is unavailable in the US per provider restrictions, so some routes are region-dependent.
• One-click templates. Summrs has a 'Migos Hotel Lobby' face-swap page (50 free credits on signup, from $15/month or $5 for 50 credits); Mitte does a left-photo/right-photo swap over 30 seconds of the performance; CapCut's Pippit lists a template (misspelled 'Quack And Takeoff'). Starrd is cited by XXL at about $5 per 15-second video but we couldn't find its template page. Templates are the beginner path: two photos, no prompt, less control, more copyright refusals.

The method

• Files. A 15–30 second cut of the COLORS performance. Photo A for the person replacing the rapper on the left; Photo B for the right. Front-facing, even light, face and shoulders — full body if your cut shows bodies. No sunglasses, no group shots.
• In Higgsfield: Video → Genjutsu → Motion Transfer. Upload the video first, then Image 1, then Image 2 — that order is how the tool maps them, and getting it wrong is the number-one failure. Add a short description (a preset alone can work, but for two people you want the prompt).
• Generate three to five times and keep the one where nobody changed sides.

Replace both performers in @Video1 with the two people from the reference images. @Image1 replaces the performer on the LEFT. @Image2 replaces the performer on the RIGHT. Take only the person from each image: face, hair, body, outfit. Ignore their pose and background. Keep everything else from @Video1: camera, cuts, timing, gestures, lip-sync, orange backdrop, hanging microphone, lighting, audio. Do not add, remove, or re-time anything. Never swap sides or blend their faces.

The short version circulating in X how-tos — 'Keep everything the same in the video. Just replace the faces with the two people in the 2 images provided.' — works about as often, with more side-swaps. One creator's YouTube Short lays out the whole thing in four steps: open Genjutsu, pick Motion Transfer, upload the clip, add one image per performer, paste the prompt, generate; the first generation is free.

The three ways it breaks

Swapped sides. The tool puts Photo A on the right rapper. Cause: image order or an ambiguous prompt. Fix: video first, Image 1 = left, Image 2 = right, and say LEFT and RIGHT in the prompt. Jewelry and clothes that won't leave. Rings, the watch, and the chains from the original persist on your new person, hair color drifts, pockets vanish — the developer who ran it through Seedance documented all four. Name the outfit you want and say what to remove. Copyright refusals. Some tools recognize the clip and decline — one Starrd user: 'spend $7 and it says copyright issue cant make it.' That's the tool telling you something true; see below.

What you're actually copying — and how to credit it

The thing that makes these clips land isn't the AI. It's a real performance — two rappers who'd been trading bars since Gwinnett County, staged by COLORS in the one-mic, one-color format the Berlin studio built its name on — and the audio is a Quality Control/Motown recording. A template makes the clone easy; it doesn't make it licensed. Higgsfield's own terms require that you 'have obtained all necessary rights, releases, and consents' for any identifiable person you upload, bar content that 'depicts them without permission,' and require you to 'disclose that Output is artificially generated.' YouTube requires disclosure when a video 'makes a real person appear to say or do something they didn't do,' with removal or Partner Program suspension for non-disclosers; TikTok requires labeling for realistic AI depictions of people.

• A personal meme of you and a friend: fine, in practice. Label it AI and credit the source.
• Two celebrities: you're using two likenesses without consent. Platforms tolerate parody; the people depicted, and the NO FAKES Act now moving through Congress, may not.
• A brand, a logo, or an ad: not without the music, the performance, and the likenesses licensed. Same warning as STORM.
• The credit line: 'Re-creation of Quavo & Takeoff – Hotel Lobby, A COLORS SHOW (2022). Made with [tool].' Link the original.

How this maps to the Yung Lean STORM trend

Our read

This is the friendlier of the two trends, and it's still not free. Quavo reposting the original is as close to a blessing as these things get, and the format is forgiving — two faces, tight shots, no crowd to dissolve. But the pushback in Complex's comments is the honest note: the second performer in every one of these clips is a man who was killed months after he filmed it, and the tools that make the swap trivial are also, as one fan put it, 'binge collecting our images.' Make yours, label it, credit it, and if the tool refuses on copyright grounds, that's the answer, not a bug. Our AI video ranking covers where Seedance 2.5 and the rest sit for everything that isn't a meme.