A pasted text prompt fails on the two-person ADDICTED duet because a sentence cannot tell a video model which face belongs to which dancer, where each body should be on every beat, or what the hallway must keep looking like for 15 seconds. The result is predictable: two faces that blend into one, steps that slide away from the music, a corridor that melts, and hands with the wrong number of fingers. Jisoo Addicted AI avoids the problem by taking a photo of each person and driving the dance from a motion reference that is built into the site.
Key takeaways
- A two-person dance asks a video model to hold two separate identities, two bodies and one shared space at once. A text prompt describes none of those precisely.
- The five usual failures are identity drift, face blending, timing that slips off the music, a hallway that morphs, and malformed hands. They are different symptoms of the same missing information.
- The fix on this site is structural: a per-person photo for each face, a motion reference for the choreography, and a fixed hallway. You do not write or paste anything.
- The renders run on ByteDance Seedance models in four tiers. Standard 720p costs 10 credits and the Single pack is $6.99 for 10 credits.
- The site is independent and not affiliated with JISOO, ROSÉ, BLACKPINK, their agencies or labels, or TikTok.
To try it with two photos, open the generator. For what the trend is and how the clip is laid out, read the trend guide.
Why does everyone search for an ADDICTED prompt?
People search for a prompt because the first AI duets they saw looked effortless, and a text box feels like the obvious way in. The ADDICTED clip that started the wave shows two people dancing together in a cream-colored hallway with wooden doors and a warm light strip along the ceiling. It is short, vertical and easy to recognize, so it looks like something a single good sentence should be able to produce.
The expectation comes from single-subject AI video. Ask a model for one person walking on a beach and a short description is often enough, because there is one identity to keep, one body to animate and a loose background. A duet breaks all three assumptions at once. There are two identities, two bodies that must stay apart while moving in step, and a background that has to stay the same corridor from the first frame to the last.
What makes a two-person dance harder than a solo clip?
A two-person dance is harder because the model has to solve identity and motion twice, and then keep the two solutions from touching. In a solo clip an error in one face is an error in one face. In a duet an error in one face can leak into the other, and the leak is the first thing viewers notice.
A text prompt gives the model almost nothing to count with. "Two friends dancing in a hallway" does not say who is on the left, what either of them looks like, or which direction anyone turns on which beat. The model fills the gaps itself, and what it fills them with is an average of everyone it has seen dance.
What are the five failures a pasted prompt usually produces?
The five failures are identity drift, face blending, timing drift, hallway morphing and extra fingers. They show up together more often than alone, because they all come from the same shortage of information.
1. Identity drift
Identity drift is when a face slowly stops looking like the person it started as. A prompt can name a hairstyle or an outfit, but it cannot carry a face. Without an image, the model invents one and then changes its mind as the shot moves. By the end of 15 seconds the person on screen is a cousin of the one at the start.
2. Two faces blending
Face blending is when features from one dancer slide onto the other. It happens most when the two people share a prompt description, because a single description pulls both toward one average face. The result is two people who look like siblings, or one face that wobbles between two sets of features when the dancers overlap.
3. Timing that drifts off the music
Timing drift is when the steps no longer land on the beats. A prompt can say "in rhythm with the song", but a video model that generates silent frames has no beat to follow. It produces motion that looks musical, with a pace of its own. When the original sound is laid over the clip afterward, the two do not line up, and the mismatch gets worse as the clip runs.
4. A morphing hallway
Hallway morphing is when the corridor changes while the dancers move. Doors shift, the light strip bends, the floor changes width. Corridors are hard because they are long, repetitive and strongly perspective-driven, and small changes in one frame become obvious changes in the next.
5. Extra fingers and merged hands
Malformed hands are the oldest AI video flaw and they get worse in a duet. Hands cross, hold each other or rest on a partner, and the model may add a finger, drop one or fuse two hands into a shape that belongs to neither person.
Why can't a longer prompt fix those problems?
A longer prompt cannot fix them because the missing information is visual and temporal, and words carry very little of either. You can write three paragraphs about a face and the model still does not know the face. You can describe a dance step by step and the model still does not know where the beat is.
Adding words tends to make a duet worse for two practical reasons. First, more detail about two people gives the model more chances to blend the details together. Second, a long description competes with itself. If one sentence asks for a close embrace and another asks for clear space between the dancers, the model has to pick, and it picks differently from run to run.
This is the deeper reason the site does not offer a prompt box. A box invites a loop of guessing, rendering and paying for each guess. The information that actually removes the errors is not a better sentence. It is a photo of each person and a reference for the motion.
How do per-person photos fix identity drift and blending?
Per-person photos fix both because they give the model a real face for each dancer instead of a description. The model no longer has to invent anyone. It has to keep two faces it has already been shown, and keeping a known face is a much easier job than inventing and remembering a new one.
On Jisoo Addicted AI you can upload two solo photos, one face in each, or one photo that shows both people. The two-solo-photo route is the stronger one for keeping the pair apart. Each photo is a clean statement of one identity, so there is no shared description pulling the two toward an average. The pair stays a pair.
A single photo of both people works too, and it is the convenient choice when you only have one picture together. The tradeoff is that the model must separate the two faces itself, so a sharp, evenly lit photo with both faces clearly visible matters more. If you have the choice, two solo photos are the safer pick for a duet.
What counts as a good photo is simple: the face is sharp, front-facing or close to it, and not hidden by sunglasses, a hand or heavy shadow. The photo does not need to match the outfit in the final clip. In the Hallway Duet look, the dancers wear their own clothes.
Two rules apply to every upload. Photos of celebrities cannot be uploaded, and that includes JISOO and ROSÉ. Photos of minors cannot be uploaded either. The site is built for you and a friend, partner or family member, not for putting a famous face into the clip.
How does a built-in motion reference fix timing and the hallway?
A built-in motion reference fixes timing and the hallway because it tells the model what the dance and the space should do, frame by frame, so the model does not have to imagine either. The choreography is carried by the reference instead of by a sentence, and the corridor is carried by a fixed setup instead of by a description.
For timing, the output is the right length for the trend: 15 seconds, matching the two-person stretch of the original clip. The video is generated silent and the trend's original sound is mixed in afterward. Because the motion follows a reference built around that same stretch of the song, the steps land where the beats land when the sound goes on. You do not have to line anything up in an editor.
For the hallway, the one look on the site is called Hallway Duet and it keeps the original setting: the cream corridor with wooden doors and the warm light strip, with the dancers in their own clothes. A fixed setup means the space is part of the setup and not a thing the model has to recall from a sentence. The result is a corridor that is far less likely to wander.
The site does not change the scene: the AI simply replaces the two dancers in the reference video with the two people in your photos, and the clothes follow your photos. If your priority is a clean, recognizable result, start with two sharp photos.
Which model tier should I pick for a duet?
Pick the tier by how much detail you need to hold, not by habit. The site offers four ByteDance Seedance tiers: Standard on Seedance 2.0 Mini, Fast on Seedance 2.0 Fast, Pro on Seedance 2.0 and Ultra on Seedance 2.5. Higher tiers generally hold faces, hands and fast motion more closely, and they cost more credits.
At 720p, the default for the 9:16 vertical format, the prices in credits are:
| Tier | Model | 720p credits |
|---|---|---|
| Standard | Seedance 2.0 Mini | 10 |
| Fast | Seedance 2.0 Fast | 17 |
| Pro | Seedance 2.0 | 20 |
| Ultra | Seedance 2.5 | 37 |
Pro and Ultra also offer 1080p. Standard is a sensible first render when the photos are clean and you want to see how the pair looks in the hallway. If the faces drift a little or hands look unreliable, moving up a tier is a cleaner fix than re-rendering the same tier and hoping.
The credit packs never expire, and there is no free tier. The Single pack is $6.99 for 10 credits, which is one Standard 720p video. Packs are non-refundable. If a render fails or is blocked, the credits are returned automatically. That is credits and not money, so a render that finishes but looks wrong still counts as a completed render.
What should I do when the clip still has a flaw?
When a clip still has a flaw, change an input and not a sentence. The inputs you control are the photo, the number of photos and the tier. Almost every visible problem maps to one of those three changes.
If a face drifts, swap in a sharper, more front-facing photo. If the two faces blend, use two solo photos in place of a single shared photo. If hands look wrong or fast motion smears, move up a tier. If the corridor feels unstable, render again or try a higher tier. If the result is fine but you want a different outfit, upload photos where you wear it.
For a longer walk-through of the usual problems, one by one, read 8 ADDICTED duet AI fails and how to fix each. When the clip is ready, the next question is how to post it, which the TikTok, Reels and Shorts guide covers.
Frequently asked questions
Why do the two dancers sometimes look like relatives?
Two faces can pull toward a shared average when the model has only one picture or one description to work from. Using two solo photos, one face in each, gives each dancer a clear identity of their own, which reduces the sibling effect. A sharp, front-facing photo of each person helps most.
Does the clip follow the beat of the song?
The motion is driven by a reference built around the two-person stretch of the original dance, and the trend sound is mixed in after the video is generated. That keeps the steps close to the beats without any editing on your side. Because models vary, a small offset on a single render is possible.
Is a one-photo upload worse than two photos for a duet?
A single photo of both people works, but the model must separate the two faces itself, so it depends more on a sharp, evenly lit picture. Two solo photos are the safer choice when you want both faces kept distinct. If you upload only one person, the second dancer is a generated passerby.
Can I upload a photo of JISOO or ROSÉ to appear in the duet?
No. Photos of celebrities, including JISOO and ROSÉ, cannot be uploaded, and neither can photos of minors. The tool is for you and someone you know. The site is independent and not affiliated with JISOO, ROSÉ, BLACKPINK, their agencies or labels, or TikTok.
What should I wear in the photos for a first render?
Wear what you want to see in the duet, because the clothes follow your photos. Plain, solid colours in clear light read best against the cream corridor. Keep your face sharp and your shoulders visible, then check how the pair behaves in the hallway before changing anything else.