VideoMusic video pro
Drop in a song, or just say what it should be about; it writes, sings and shoots the video.
This is the full version. Music video light gives you a quick video cut on the beat from the models you have. Music video pro gives you a full-scale cinematic music video of the whole song: one lead performer who stays the same face in every shot — an Identity reviewer keeps the same face in every shot, frame by frame against her sheet — one singing voice from first line to last, one feel, several outfits, props drawn once and reused, motion graphics timed to the words, a plan you approve, and a team that works in reviewed waves until the last frame is in.
You drop in a song and one sentence about the look. No song? Just tell it what the song should be about: it writes lyrics with a real repeating chorus, chooses one voice, one key, one tempo and one feel, sings the whole thing through the music model your account reaches, plays it to you, and carries on as if you had brought it. Lyrics, reference images and must-keep characters are welcome, never required. First the song is listened to as code: sections (intro, verses, choruses, bridge, outro), beats and downbeats, energy, spectrum. You see that as an analysis board, opened in the surface. Then you see the look itself before anything is shot, and you sign off on it. The first gate, G0 Mood, is one screen: three or four style frames, the palette, the character lineup and one five-second motion test from the video model that will animate the shots — the test matters because stills do not show how a model moves. One question: approve the look? The Director researches visual ideas on the web and writes the plan: the one performer, the style bible (cinematic and photoreal unless your sentence says otherwise), every prop, the full-song storyboard, every prompt. It shows you the script as a one-page table — scene, action, the exact words sung, the exact words on screen, length — and asks the second question: approve the script? That is G1. Only you approve. The team reviews and recommends, but a gate opens on your words alone, saved in `approvals/<gate>.md`, and no stage starts without the file for the gate before it.
After G1 the identity sheets are made, and then — before any high-quality render — G2 Scene previews: one cheap, low-resolution clip per scene on one screen, each beside the frame you approved at G0 so drift shows in two seconds; you approve or comment on each scene. Approved scenes go to their full clips at once; a scene you comment on comes back changed on its own and holds up nothing else. With the script, before any clip, you are told once in one plain sentence what the clips come to at the listed rates — nothing to approve, and not said again until the receipt. Only then does the team work in waves: 2 assets first, then 3–4, then 4–8, then 6–10 until the song is covered. Each wave is a set of Workers, one per job. The identity sheets come first: one performer, front and three-quarter view, in each outfit, on green; every prop centred on green. Each shot is then animated from its keyframe by MiniMax Hailuo 3 — the one video model for every clip, singing or not, so the whole cut has one look and no shot is rendered differently from the one beside it — and the identity sheet keeps the same face in every clip. When the performer sings on camera, her mouth is animated from the actual vocal, so the words match. Those shots are made by MiniMax Hailuo 3, the lip-sync model, from the identity sheet and that shot's exact stretch of the song, and the same singer sings the same words in every one of them. The performer is cut out and composited over a plate with p5 motion graphics, and a Reviewer inspects every asset before the next wave starts. Beside it an Identity reviewer does one job only: it pulls the first, middle and last frame of every still and clip in the wave onto one contact sheet under the locked identity-sheet crops, opens it for you, and sends back any shot where the face, the hair, the outfit or a prop has drifted — so the same person, with the same things in her hands, is in every shot.
The words are timed by a transcriber that runs locally in the project, corrected against the lyrics you gave — or the lyrics it wrote for you, which are exact — so the type lands on the words. Everything is assembled with ffmpeg against the unbroken original song — the singing clips are silent and sit at their exact place on the timeline, so the mouths line up with the song — into one 1080p MP4, which is copied to media/generated/ and opened in Goatfish's player next to the analysis board. You see it twice: G3, the rough cut, and G4, the final — each one screen, each scene's approved frame next to the new one — and the working files are cleared only after your yes at G4. Once you approve a gate, what it covered is locked: the model, the character, the reference images, the script and the chosen takes. Changing a locked thing is a change request shown to you first — before and after side by side, the reason, and what has to be redone — never a quiet swap. Every note you give is saved word for word with an ID; the Director answers 'heard: X, I'll change only Y' and changes only that. If you step away, the team stops at the next gate and waits; it may line up two or three options for the morning, but it never moves past a gate you have not approved.
Under the hood this pack runs makevoid's open-source motion-graphics-music-video pipeline (MIT) — https://github.com/makevoid/motion-graphics-music-video-skill, pinned to commit 93248cc6117d05a2c108339fb5936bfb9c6f313e — with Goatfish's own image and video tools in place of the hosted models it was written for, and local transcription and stem separation in place of hosted ones. Only the parts a video needs are fetched into the project; you install nothing, connect nothing and set no keys. The About this pack page carries the license.
The team: Director, Worker, Editor, Reviewer, Identity reviewer.
Download GoatFish for free