The short answer
A 58th Floor AI video is made from two photos, not from footage. You add one photo of each person, keep the 58th Floor set, and press Create: in about two to five minutes you get a 12-second vertical clip of both of you trading a hook and a verse at one vintage mic hanging in a seamless teal studio.
Three things decide whether yours looks real or looks like AI slop: the sharpness of the two photos, which face you put in the right slot, and whether you fix a bad take instead of posting it.
What you need before you start
- Two photos, one person in each. A normal phone selfie in daylight is enough.
- About five minutes of waiting per render. You can close the tab while it works.
- 10 credits for the default clip (12 seconds, 720p, 9:16). The price is on the Create button before you press it.
The two photos do not need to be taken together, in the same city, or in the same year. If you already have one picture with both of you in it, you can skip the second photo entirely and use the shared-shot mode instead.
Step by step
Step 1: Pick who leads the hook and who takes the verse
This is the one decision people get wrong, and it is not cosmetic. The 58th Floor set is built around a fixed four-shot edit:
- Left slot finishes the hook in the opening two-shot, then bounces along beside the mic.
- Right slot takes the verse, the close-up, the head-to-toe wide and the final face-off.
So the right slot gets roughly three quarters of the screen time. Put the bigger personality, or the person whose photo is sharper, on the right. If you are making the clip as a surprise for someone, put the person receiving it on the right.
Step 2: Take the two photos
Both photos need to pass the same test: the model has to be able to read one face clearly and carry the outfit through the whole clip. The rules below are the difference between a clip that looks like you and a clip that looks like a cousin of you.
- One person per photo. No group shots, no one cropped at the edge of the frame, no arm around a shoulder.
- Face forward, both eyes visible. A slight angle is fine; a profile is not. The mouth must be unobstructed.
- Natural light, no hard shadows. Stand facing a window. Overhead bulbs in a dark room flatten the face and the render inherits that flatness.
- Sunglasses off. Also off: masks, scarves over the chin, hands near the face, hair across the eyes. Eyes and mouth are what the lip-sync is built on, and the model cannot invent them convincingly.
- Skip heavy beauty filters. The smoother the photo, the less face there is left to keep.
- Chest-up or waist-up. The outfit from the photo is the outfit in the clip, so at least show the shoulders and the top half of what you are wearing.
- Same lighting on both photos. One person indoors at night and one in bright sun reads as two different rooms in the same shot.
Formats: JPG, PNG or WEBP, up to 20 MB per image. A 1080p phone selfie is more than enough; do not upscale a blurry screenshot.
Step 3: Open the generator and add both photos
Go to the 58th Floor AI video generator and scroll to the studio. The 58th Floor set is already selected, so there is nothing to configure before you upload.
Two ways to add the photos:
- Two portraits (default). Drag each photo into its slot, or press Ctrl+V (⌘V) to paste a screenshot. Left is the lead, right raps as well.
- One shared shot. If your photo already has both of you in it, switch the portrait mode to "1 shared shot" and upload just that one. The person on the left of the picture leads and the person on the right backs them up.
Step 4: Keep the set and check the take settings
The default is 720p, 12 seconds, 9:16 - which is what you want for TikTok, Reels and Shorts, and it costs 10 credits. Change it only if you have a reason:
- 480p is the cheaper option if you are testing whether the photos work at all.
- 8 seconds is the cheapest way to find out whether the lip-sync lands before you spend on the full clip.
- 1080p is worth it only if you are posting somewhere that keeps the detail.
- Every model, quality and length shows its own credit price as you pick it, so you never press Create blind.
You can also switch the room here: Hotel Lobby, Luxury Lobby, Studio Booth and Street Cypher are one tap away and your photos stay in place. The 58th Floor set runs on Seedance only, so Wan 3.0 is off on this page.
Step 5: Tell it what to rap about (optional)
Type a topic and the bars get written around it. "Sam's 30th birthday", "eight years of putting up with each other", "our shop opens at seven" - a specific topic beats a generic one every time, because a specific topic gives the verses something to land on.
Leave it blank and you get a generic duo track instead. If you would rather use your own song, upload an MP3 or WAV (2 to 15 seconds) and pick the section you want - but only audio you own the rights to.
Step 6: Press Create and download the MP4
The button shows the price before you press it. Most renders land in two to five minutes; you can close the page and the finished clip will be waiting in your works.
The download is a clean MP4 with no watermark at every quality level. The song is baked into the file, so post it as-is rather than adding a separate sound.
Photo requirements at a glance
If you only remember one thing from this guide, remember this list. Almost every disappointing 58th Floor clip traces back to one of these.
- One person per photo, face clearly in view
- Front-facing, both eyes visible, mouth unobstructed
- Sunglasses, masks and hats pulled low: off
- Natural daylight, no hard shadows across the face
- Chest-up or waist-up so the outfit carries over
- JPG, PNG or WEBP, up to 20 MB
- Matching light on both photos
- No heavy filters, no upscaled screenshots
When it goes wrong: the five common failures
AI video is not deterministic, so the same two photos can produce a great take and a bad one. These are the failures people actually hit, and what to do about each.
The two faces are on the wrong sides
Symptom: the person you wanted as the lead finishes the hook, and the verse and close-up went to the other one.
Cause: you uploaded the photos in the opposite order. The slots are not interchangeable - left slot leads, right slot takes the verse, the close-up and the wide.
Fix: swap the two uploads. Do not re-roll the render and hope; the ordering is the input, so fix the input. With a shared shot, the person on the left of the picture leads, so either switch to two separate portraits and order them yourself, or use a photo where the right person is already on the left.
The two of you blend into one face
Symptom: somewhere in the clip, especially in the wide and the face-off, the two faces drift toward each other and end up looking like one person with two hairstyles.
Cause: this is identity bleed. It is most likely when both photos are framed the same way, lit the same way, and the two people look alike - same hair length, same colour, same age, similar build.
Fix: make the two photos look less alike. Different framing (one chest-up, one waist-up), different backgrounds, clearly different hair. If the two of you genuinely do look alike, pick photos where the outfit and hair make the difference obvious, and regenerate - it usually resolves on the second or third take.
The mouth does not match the words
Symptom: the lips move, but not on the syllables. Sometimes the mouth opens on a pause and freezes mid-word.
Cause: usually one of three things - a photo where the mouth is already open or caught mid-word, an uploaded track with layered or heavily processed vocals, or a clip long enough for the sync to drift by the final shot.
Fix, in order of how often it works:
- Regenerate. The sync is generated per take, so a re-roll is the first and cheapest thing to try.
- Use a closed-mouth or relaxed-smile photo. A photo taken mid-laugh or mid-sentence hands the model a mouth position it then has to animate around.
- Switch to AI rap instead of an uploaded track. Bars written for the clip follow the motion the set already produces; an external track with fast delivery, overlaps or heavy autotune is much harder to match.
- Drop to 8 seconds. Fewer syllables per shot, less room to drift.
- Trim your track to one clean voice. If you uploaded your own song, pick a section with a single vocal line and no ad-libs stacked on top.
The mouth barely moves at all
Symptom: the face sings, but the lip movement is small or stops for a stretch.
Cause: the photo is too small in frame, the face is turned away, or the jaw is hidden by a scarf or a hand.
Fix: a tighter, front-facing photo where the jawline and the mouth are unambiguous. This failure and the previous one share the same root cause, so a better photo fixes both.
The render fails or the outfit changes mid-clip
Symptom: the task errors out, or the outfit from the photo morphs into something else halfway through.
Cause: a failed render is a normal platform event, not a charge; a morphing outfit usually means the outfit was only partly visible in the photo, so the model had to invent the rest.
Fix: a failed render returns its credits automatically - a bad take costs nothing, so just run it again. For outfit drift, use a photo that shows the whole top half of what you are wearing, and avoid photos where the outfit is half cut off by the frame edge.
What it costs
The default clip is 10 credits. Credit packs start at $6.99 for 10 credits (about one video), $29 for 50 (about five) and $99 for 350 (about 35); pack credits never expire and there is no subscription required. Check the price on the Create button before you commit, because 1080p and longer lengths cost several times more than the 720p default.
Two cheap habits: test the photos at 480p or 8 seconds first, and never pay for a resolution you cannot post.
Before you post
- Label it as AI. Both TikTok and Instagram ask you to disclose realistic AI-generated media, and it is the honest thing to do either way.
- Use photos of people who agreed to be in it. No celebrities, no public figures, no photos of someone else pulled off the internet.
- Keep the vertical crop. The clip is already 9:16 - do not reframe it into 1:1 or you lose the hanging mic.
- Do not claim it is real footage. The whole point of the format is that it is obviously a bit much.
- No affiliation. This is an independent tool and is not connected to any artist, label or show the trend borrows its look from.
Frequently asked questions
Do I need a video or just photos?
Just photos. Two of them, one person in each, or one photo with both of you in it. Nothing is filmed and nothing is edited.
How long does it take?
Usually two to five minutes per clip. You can close the tab while it renders and collect the finished video from your works.
What if my photos get rejected?
A failed upload is usually a file problem, not a content problem: check the format (JPG, PNG or WEBP) and the size (under 20 MB). If a render fails, the credits come back automatically.
Can I use the same two photos on a different set?
Yes. Hotel Lobby, Luxury Lobby, Studio Booth and Street Cypher are all one tap away, and your photos stay uploaded when you switch rooms.
Can I make it with just one person?
No - the 58th Floor format is a duet. The one-photo route is the effects tab instead, where a single picture becomes a five-second clip.
Is this a face swap or a filter?
Neither. A face swap pastes a face onto existing footage; here a new performance is generated from your photos, which is why both of you move, react and get your own shot.
Do I have to show my face straight on?
For this set, yes. Front-facing photos with the eyes visible are what makes the lip-sync and the identity hold up.
Make yours next
Two photos in, a 12-second duet out, everything else already built into the room. Open the 58th Floor AI video generator or read the step-by-step guide to the Hotel Lobby trend if you want the orange booth instead, and check the pricing before you start.