How we made a 2-minute AI music video in 7 days

We made a 2-minute animated music video for an original track, with Ava and Jordan recurring from start to finish. The full production ran for 7 calendar days, from 10-16 August 2026.

The funnel was the real story. We produced 205 generations, got 20 minutes of footage, and kept 1 minute 58 seconds in the final edit, which worked out to roughly a 10:1 ratio.

That ratio, not the quality of any one clip, was what people underestimate when they make a music video with AI. A good AI music video generator can give strong clips, but the finished video still came from selection, timing, reuse, and cuts.

We did not animate anything until the look existed as still images. Concept art fixed the palette, wardrobe, and both faces before we moved into shot production.

After that, every shot got its own starting frame generated in Nano Banana 2 from the concept art. A portion of those frames were refined afterward in an image editor, and the project used 151 frames total.

This order mattered because image-to-video inherited composition, lighting, and character from the still. A weak frame could not be rescued by a good motion prompt.

We also gave the characters names, Ava and Jordan, and used those names in the prompts. Twenty-two prompts referred to them by name, and that carried identity better than re-describing hair and clothes every time; it became one of our clearest character consistency AI video lessons.

Concept art, made before a single shot was animated. It fixed the palette, the wardrobe and both faces — everything downstream had something to match.

For video generation, 204 of 205 runs used Seedance 2.0 image-to-video at 1080p. Within that work, 155 were 5-second clips at 15 credits each, and 35 were 10-second clips at 30 credits each.

The average render took 94 seconds, the longest took 19 minutes, and the project used 5.3 hours of machine time in total. For the video ledger, the number was exact: 3,613 credits charged, 30 refunded for two failed runs, and 3,583 net credits.

The full project came to roughly 4,500 credits, but the two halves came from different sources. The video number was measured from the account ledger; the frames and lip sync were priced from the public rate card because those tools were not billed on the account we used.

Nano Banana 2 images cost 2 credits from a prompt or 4 through the editor, and ours were a mix of both, so 151 frames came to about 453 credits. Kling 2.6 lip sync billed 2 credits per second of audio: the 27 lip-synced clips we kept measured 209 billed seconds, which came to 402 credits, while all 33 lip-sync runs worked out to roughly 491 credits.

This was the single most repeated fight. When we asked for a chorus performance, the model kept drifting toward its default: wide-eyed, sincere, and slightly sad, while we wanted cool, self-aware, and a bit amused.

Positive description alone did not move it enough. What helped was writing the expression as an arc inside one shot, so the face changed over time instead of holding one adjective for ten seconds; 26 of our prompts used that kind of arc.

The other fix was naming the unwanted default and banning it outright. We used phrases like "no deer eyes, no innocent wide-eyed expression, no ballad sadness," and 19 prompts carried a ban like it; when a model had a strong prior, a negative was a stronger instrument than an adjective.

Once a location existed, returning to it for a second angle became one of the most expensive parts of the workflow. Rewriting the prompt often changed too much, so we got more value from re-rolling the identical prompt.

In total, 21 prompts were submitted more than once byte-for-byte, creating 29 extra generations. We also locked both ends of a move with a first frame and a last frame, which was what the single Kling 2.6 generation in the project was for.

Depth of field helped cover the remaining mismatch. Thirty-four prompts pushed the background soft and secondary, so a re-generated set did not need to match perfectly in every detail.

Our order was simple: generate the shot first, then apply lip sync in Kling 2.6, driven by the vocal from the finished track. We made 33 lip-synced clips that way.

The counter-intuitive part was the motion prompt. It had to keep mouth movement subtle and relaxed because the mouth got replaced later; 10 prompts said this explicitly, and shots with expressive mouth articulation fought the lip sync applied on top.

The strongest lesson of the project was that generation skill without editing skill produced a showreel, not a music video. The track ran at 130 BPM, and the finished video had 142 shots in 118 seconds.

The median shot was 0.70 seconds, which matched a dotted quarter. Eighty-five shots were exactly 1.5 beats long, 31 were exactly 2 beats, and 117 of 141 cuts landed within half a frame of the half-beat grid.

The density also came from where the cuts went. Eighty of 141 cuts jumped to a different location, so the video cross-cut between school, cafe, bedroom, and party instead of playing each scene as a block.

We reused material on purpose: 82 source clips covered 123 shots, and some clips appeared up to four times. The tempo was also built into generation, not only the edit, with 19 prompts naming a BPM and 3 scripting choreography beat by beat; the two long holds around 77-84 seconds were the only breather, and they were deliberate.

Two failure modes still broke through. On long, near-static holds, the model could hallucinate extra figures, with a second person fading in out of nothing; lettering also degraded whenever the camera moved, so WESTFIELD on the cheer uniforms drifted into nonsense across a shot.

The practical mitigations were not glamorous, but they helped. We kept holds short, which the fast cut already did, kept readable text out of frame or out of focus, and re-rolled rather than rewrote when a clip was close.

This was the sequence we would repeat, in order.

1. 1Finish the track first. The edit is built on its tempo, so you need final audio before you generate a single shot.
2. 2Design the look on paper. Concept art that fixes palette, wardrobe and both faces gives every later prompt something to match.
3. 3Name your characters and use the names in every prompt — "Ava" and "Jordan" held identity across generations better than physical description alone.
4. 4Generate one frame per shot in Nano Banana 2, working from the concept art rather than from scratch.
5. 5Animate each frame in Seedance 2.0. Five seconds at 1080p is the workhorse; reach for ten only when the action genuinely needs it.
6. 6Script motion in beats at the track tempo, so generated movement lands on the same grid the edit will use.
7. 7Say what you do not want. Wrong emotional defaults are the most common reason a technically fine shot reads wrong.
8. 8Add lip sync last, in Kling 2.6, and only on the shots that hold the camera on a face.
9. 9Cut on the beat, and cross-cut between locations instead of playing each scene as a block.