A performance music video with AI holds together when the artist is a reference sheet, not a description. Approve a set of stills of the artist first (front, three-quarter, profile, the outfit), then feed those stills into every shot: a reference-to-video model for the wide performance shots, a start-frame model for the close ones, and a lip sync model for the lines where the mouth has to match the vocal. If the artist is a real person, their written consent comes before the first image. The surprise for most people: the lip sync tools on 8frame make a new clip from a still and the audio. None of them re-syncs the mouth on footage you already shot.
TL;DR
| Decision | What to use on 8frame | Price in credits | What it needs |
|---|---|---|---|
| The artist sheet | CHARACTER POSES template (on /workflows) with an image model | Nano Banana 2 11 at 1K; Seedream 5 Lite 6 | Photos of the artist (with consent) or an approved design |
| Wide shots with the artist in a new place | Happy Horse reference-to-video | 95 for 5 s at 720p, 189 at 1080p | 1 to 9 character references plus a prompt |
| Same, alternative | Kling O1 Reference | 73 for 5 s | Reference images; check the node for its limit |
| A close shot from an approved still | Kling v3 Standard | 86 for 5 s with sound, 57 silent | A start image |
| Singing lines | Pruna Avatar / OmniHuman v2 / Hailuo H3 Max Lip Sync | 3.375 per s at 720p, 6.075 at 1080p / 14.175 per second of audio (5 s = 71) / 34, 54, 108, 216 per 5 s | A portrait still plus the vocal |
| Edit to the song | Video Editor timeline | no generation | All clips plus the master |
We quote price and settings, not results.
Consent before the first image
A performance video sells the artist's face. Two situations:
- A real artist. Get written permission that names what you will generate (a music video, its teasers, the thumbnail) and where it will run. Permission that is specific protects both sides when the clips are cut into teasers later. If you are a director working for an artist, the label or management usually signs, but the artist should know their face will be generated.
- A virtual artist. A character you designed has no consent problem, but it still needs to stay one character. Treat its approved sheet like a logo.
Never use a well-known musician's face or voice "in the style of" without a licence. YouTube asks creators to disclose content that "Makes a real person appear to say or do something they didn't do", and a generated performance by a real singer is that.
Build the artist sheet
Open the CHARACTER POSES template from /workflows or start a blank board. The goal is six to ten stills you would sign off as "this is the artist":
- Front, three-quarter and profile, neutral light.
- The stage outfit, full length, front and back.
- Two expressions you will need: eyes closed singing, looking into the lens.
- One still per location you plan (the artist on the rooftop, in the club, in the rain), made from the sheet.
Seedream 5 Lite takes the most reference images on the canvas (up to 14), so it can see the whole sheet at once. Nano Banana 2 and GPT Image 2 take up to 10. Put the approved stills in one frame on the board labelled "canon", and pull references only from there. Old drafts in the reference set are how a face drifts.
Shots that need the artist, and shots that do not
Most music videos are not wall-to-wall close-ups. Split the shot list before you generate:
| Shot | Needs the artist's face? | Route |
|---|---|---|
| Crowd, city, light, abstract inserts | No | Text or image to video, any model |
| The artist small in a wide frame | Yes, but small | Happy Horse or Kling O1 Reference with the sheet |
| Hands on the guitar, the mic, the shoes | Outfit and hands only | Start frame from a sheet still |
| The artist mid-shot, moving | Yes | Kling v3 Standard from an approved location still |
| The artist singing a line | Yes, with the mouth | Lip sync from a portrait still plus that line of vocal |
The fewer shots carry the face, the fewer chances for drift, and the cheaper retakes are.
The performance shots, step by step
Wide and mid shots. Attach the sheet to Happy Horse (1 to 9 character references) or Kling O1 Reference and describe the action and the place, not the face. "The singer from the references walks across the wet rooftop at night, neon behind, camera tracks at chest height" works better than a list of facial features. Happy Horse opens at 1080p (189 for 5 s); switch to 720p (95) for first takes.
Close shots from a still. When the shot starts on an approved still, a start-frame model keeps the most of it. Kling v3 Standard needs a start image and returns 720p; 86 credits for 5 seconds with sound, 57 silent. Turn sound off: your master is the soundtrack.
Singing lines. Cut the vocal into the lines you want on screen. Each line plus one portrait still goes into a lip sync node:
- Pruna Avatar: portrait plus audio, billed per second of output (the audio length, rounded up), 3.375 credits per second at 720p and 6.075 at 1080p, up to 35 seconds; the output keeps the portrait's shape.
- OmniHuman v2: image plus audio, 14.175 credits per second of audio (71 for 5 s), 4 to 35 seconds.
- Hailuo H3 Max Lip Sync: 5-second floor, audio clipped at about 14.8 seconds; 34, 54, 108 or 216 credits per 5 seconds at 480p, 768p, 1080p or 2K.
Use the isolated vocal stem if you have one, not the full mix: a lip sync model listens for a voice.
Edit it to the song
Open the Video Editor, put the master on a sound layer and cut the clips to it. Mute every clip's own audio. Cut on the beat; the clip lengths you generated (5 or 10 seconds) give you room to trim to the bar. Filter presets apply per clip, so pick one grade (say Cinematic) for all performance shots so they match.
Teasers, Shorts and the thumbnail from the same sheet
The release needs more than the video: vertical teasers, a Shorts cut, a thumbnail, sometimes a still for the single's page. Make them from the same canon frame, not from new prompts. The Kling tool opens at 9:16, so a teaser can be a new take of a scene you already approved, reframed for vertical, rather than a crop that loses the artist's head. For the thumbnail, render a still from the sheet at 2K or 4K with an image model (Nano Banana 2 is 17 credits at 2K and 22 at 4K) and put the title on in your design tool. One sheet, many formats: that is what keeps the artist recognisable across a campaign, as our guide to keeping characters consistent across a campaign describes for brand work.
Which model holds the artist best?
We have not measured models against each other on identity, so choose by input:
| You have | Use | Limit |
|---|---|---|
| A sheet, no matching still for this shot | Happy Horse (1 to 9 refs) or Kling O1 Reference | Check the face in every take |
| An approved still of exactly this shot | Kling v3 Standard | Needs a start image; 720p |
| A still and a vocal line | A lip sync model | Makes a new clip from the still; does not edit footage |
| A filmed performance | Keep it real | There is no mouth re-sync on existing footage |
Where it breaks
- Drift across takes. Every generation redraws the face. Re-run takes that drift; failed generations are refunded automatically, but a take that rendered and looks wrong is a normal take.
- Outfit changes. Jackets lose a zip, a print changes. Keep the outfit in the references and check it per shot against the full-length sheet still.
- Hands on instruments. Fingers on frets are a weak spot of video models. Frame the hands loosely, or film them.
- Lip sync length. Pruna Avatar and OmniHuman stop at 35 seconds, Hailuo H3 Max Lip Sync clips the audio at about 14.8 seconds. Sing in lines, not verses.
- No video upscaler. A 720p take stays 720p. Choose 1080p settings for the shots that matter.
- Montage has no sound. Cut on the Video Editor timeline, where the song is.
FAQ
How do I keep the same singer in every AI shot?
Make a reference sheet of approved stills first and use it as the reference in every shot. Reference-to-video models (Happy Horse with 1 to 9 references, Kling O1 Reference) place the singer in new scenes; a start-frame model like Kling v3 Standard keeps the most of one approved still. Check each take and re-run the ones that drift.
Can AI make the artist lip sync to my song?
Yes, from a still: a lip sync model takes one portrait and an audio clip and returns a new clip of that face singing it. On 8frame that is Pruna Avatar, OmniHuman v2 or Hailuo H3 Max Lip Sync. None of them changes the mouth in a video you already filmed.
Do I need the artist's permission?
For a real person, yes: written consent that names the video, its teasers and where they will run. YouTube also asks for a disclosure when content makes a real person appear to say or do something they did not.
Which part of the video should stay filmed?
Anything where a real performance is the point: live vocals in close-up, an instrument solo, the band together. Use AI for places you cannot shoot, wide shots and inserts, and lip sync lines for the virtual or stylised parts.
What does a performance video cost on 8frame?
It depends on how many shots you generate and retake, so there is no fixed number. Every node has a fixed price: Happy Horse 95 credits for 5 s at 720p, Kling O1 Reference 73 for 5 s, Kling v3 Standard 86 for 5 s with sound. Plans start at $19 a month for 1,000 credits.
Sources
- YouTube Help, Disclosing use of altered or synthetic content ("Makes a real person appear to say or do something they didn't do"), checked 2026-11-09
- 8frame prices, defaults and limits: the 8frame app, checked 2026-11-09
Try it without signing up: the 8frame canvas opens on a real board. The canvas is free and unlimited; generation is paid from $19/month.