Menu
Promo editors at a music desk lose the face when the chorus still starts to move. The jacket, the mic, and the eye line were approved on a photograph. The teaser then invents a second performer, or the mouth keeps singing after the hook has stopped. A Seedance API workflow does not fix that by being “cinematic.” It only helps if the route can take the still, the move, and the audio the desk already cleared.
The 2.0 card is the one that claims this job. It describes image, video, and audio inputs for motion transfer and lip-sync, and it talks about holding a character steady while the camera path gets more complicated. The prompt gallery even labels a “K-pop Girl Group Music Video Sequence” as a 2.0 example. That label is a public sample title, not a screening. The editor’s check is whether a chorus still from this campaign survives the same kind of route without a new face.
The Chorus Still That Starts To Move
A promo night is a pile of approved stills and one audio file that cannot slip. The Bollywood teaser, the K-pop cutdown, the fashion film: each one already has a face the public knows. Motion is the risky part. If the hair, the jewelry, or the jaw changes between bar one and bar two, the comments will not discuss the song. They will discuss the face. That is a reject, and it arrives faster than a color note.
Old fixes waste the night. One editor restyles the whole frame until the person looks “more alive,” which is how identity goes. Another exports a slow push on the still and calls it a video. The push is safe and also dead. The page’s reference route sits between those two failures: drive motion from a clip or a pose, and anchor style from the still, instead of asking a text prompt to remember a celebrity.
What The Old Fix Keeps Breaking
Text-only prompts are the wrong tool once a face has been approved. The prompt can name the jacket and the stage color. It cannot carry the exact person. Every extra adjective is a chance for the model to drift toward a neighboring face. On a release night that drift becomes a fan problem and a clearance problem, and the rework is a new cut, not a caption.
A New Face In The Second Bar
Watch the eyes and the jaw at the cut, not the lighting. If the second bar could be a different artist in the same outfit, discard the file. Do not “fix it in the grade.” A grade cannot put the right person back. The still has to go on the image or reference route so the face is an input, not a hope. A prompt that says the name of the artist is not a substitute for the photograph the label already approved.
A Mouth That Misses The Hook
Lip-sync is the other break. The page describes audio and voice driving, and synchronized audio, as part of reference-to-video. That is a control, not a guarantee. Play the hook and look at the mouth. If the lips keep moving through a rest, or freeze while the vocal holds a note, the take is unusable for a lyric teaser. Mute-and-ship only works when the teaser never promised a sung close-up. A chorus cut that shows the mouth has to pass the mouth.
What The Reference Route Is For
Reference-to-video on the page is three jobs: audio and voice driving, video motion transfer, and an image or style anchor. A promo desk usually needs two of them at once. The still holds the person. A short dance plate or a camera move holds the motion. The vocal holds the mouth, if the mouth is in frame. Sending only the still, and hoping the prompt invents the choreography, is how the arms become wrong.
Image Video And Audio As Inputs
Seedance 2.0 is the line written for that stack: text, image, video, and audio, with the card stressing multi-reference inputs. The create panel separates Text to Video, Image, and Reference so the route is a choice, not a footnote. Pick Reference when a plate exists. Pick Image when you only have the approved still and a simple move. Pick text only for a background plate with no recognizable person. Mixing those up is the mistake, and it shows up as a stranger in a familiar outfit.
Motion Transfer Without A New Actor
Motion transfer means the move comes from a clip you provide. The person should stay the person in the still. If the transfer pulls in the body shape of the reference dancer and drops the artist’s shoulders, the clip has failed the only test that matters on this desk. Compare the shoulders, the hands, and the jewelry against the still at two points in the shot. One match at the first frame is not enough. Identity has to survive the move.
A Reject Rule For The Teaser
Write the rule before the first generate, and keep it shorter than the prompt. The Seedance 2.0 API can come later, if a tool team wants the same route inside a CMS. The night itself is a studio decision: which input, which card, and what gets thrown away.
When The Identity On Screen Is Wrong
- If the face could be someone else, reject the take.
- If the mouth misses the hook on a sung close-up, reject the take.
- If a logo, a title card, or a lyric line turns unreadable, reject the take.
- If the camera move fights the reference plate, reject the take and fix the plate, not the adjectives.
A rejected take stays off the public cut. Do not bury it under a faster caption. The time you spend arguing that it is “close enough” is the time the correct still could have had a second preview. Close enough is how a wrong face ships.
Fast Or Quality On A Release Night
The page splits the 2.0 line into Ultra, Fast, and Mini cards, all marked Audio. Seedance 2.0 Fast is described as the quicker, lower-cost option, with the card listing 480p to 720p and 4 to 15 seconds, plus text, image, and reference. That span fits a teaser beat. It does not fit a full performance video. Ultra is the top-tier card on the same picker. Use Fast to find out whether the face and the mouth survive. Move to a quality-focused option only after the reject rule is clean. A polished wrong face is still a wrong face, and it costs the desk a longer watch.
Duration belongs in the same note. A 15-second sample title in the gallery matches the top of the 4 to 15 second span on these cards. A promo cutdown can live inside that span. A three-minute performance cannot, and no prompt will stretch the card. Cut the song to the hook, then generate. Asking for the whole track is how the job fails before the model starts.
Fashion film and a music teaser fail on different objects, so do not run them through one prompt. A lookbook clip dies when the garment changes: a neckline opens, a logo slides, a sleeve grows a seam the sample never had. Compare the cloth against the still at the turn, the same way you compare a face. A music teaser dies on the mouth and the artist. If both jobs are due the same night, give each its own still and its own reject note. A shared “cinematic stage” prompt will get both jobs wrong, and the rework will eat the only quiet hour before the post goes up.
The desk can also waste time hunting a camera word. The page does ask for camera movement in a prompt, and a dolly or a push can be named. On a release night the camera is the last argument, after identity. A beautiful orbit around the wrong person is still a reject. Write the move in one clause, then go back to the still. If the preview’s camera fights the reference plate, fix the plate or drop the clause. Do not add three more camera sentences and hope the face returns.
Where A Promo Cut Should Stop
Seedance fits a promo desk that already has an approved still and will reject a drifted face or a missed mouth. It is a poor fit when the plan is to invent the artist from a name in a prompt, or when nobody has time to compare the preview with the still. Fast can find the break. Quality cannot pardon it.
Put the still on the reference route, check the mouth against the hook, and stop when the person changes. The song can wait. The wrong face cannot.
