Perché abbiamo creato heybop: il videoclip non dovrebbe essere la parte difficile

A bedroom producer in 2026 can write, arrange, mix and master a track in a weekend. The same producer, asked for a music video, usually posts a static cover art with a waveform bouncing on top of it. Not because the video does not matter — short-form platforms reward video above everything else — but because the video is still priced like it is 2014: a shooting day, a crew, an editor, a month of back and forth.

Generative video models changed the cost of a shot. They did not change the cost of a video. A shot is six seconds; a video is thirty of them that cut on the right beat, keep the same person in frame, and end when the song ends. The first wave of AI music-video tools sold a prompt box and a progress bar, and the results looked exactly like that: pretty, incoherent, and impossible to fix.

What a music video actually needs

Before writing a line of code we watched about two hundred low-budget music videos and wrote down what made the good ones work. The list was shorter than we expected.

• Cuts that land on downbeats. Not near them. On them. A cut that arrives 120 ms late reads as sloppy even to people who cannot say why.
• A visual arc that follows the song's structure: something changes when the chorus hits, and the bridge looks different from the verse.
• Continuity. If there is a person, it is the same person in every shot.
• A strong first three seconds, because that is the only part most viewers will see.

None of these are model problems. They are analysis and editing problems. So heybop is built as a studio around the models, not as a wrapper over them.

Four decisions that shape heybop

Decision one: the beat grid is the source of truth. When you upload a track, heybop runs onset detection and section analysis before anything else. Every scene is planned against that grid, every generated clip is trimmed to it, and the export is cut to it. The Director AI is not allowed to invent a shot list that ignores the music.

Decision two: every scene is a card you can edit. A storyboard in heybop is a list of scenes, each with its own prompt, camera move, duration and reference image. If scene seven is wrong, you rewrite scene seven and regenerate scene seven — for the price of one scene, shown on the button before you click.

Decision three: the price is always visible. Credits are simple (one credit is roughly one US cent), every action shows its exact cost, and failed generations are refunded automatically. If you would rather bill MiniMax directly, paste your own key and heybop charges a flat 20% platform fee on every job instead of the full price.

Decision four: songs, voices and video live in one place. Google Lyria 3.5 writes and sings a track in Music Lab; it lands in the MV timeline. Speech 2.8 records a voice-over in Voice Lab; it lands on the same timeline as a second audio track. You should not need three subscriptions and a folder of downloads to finish one release.

Why you never pick an engine

We evaluated every major video engine on the same twenty prompts, scored for motion coherence, prompt adherence and cost per usable second. No single engine won everything: one was clearly best on the shots where the camera lingers, another was nearly as good in motion at a fraction of the cost, and a third was the only one that kept a real person's face from a photo. So heybop does not show you a model picker. It reads the storyboard and routes each shot to the engine that fits its job, and you pay one stable per-clip price whichever engine ran it. Building on one vendor's stack also keeps latency and billing predictable.

The goal is not to replace a director. It is to give every artist a first cut worth showing to one.

What is next

Lyric-synced captions, a shared review link with timestamped comments, and per-scene motion strength controls are in progress. Follow the changelog for what ships each week, and tell us what is missing — the studio is shaped by the people releasing music with it.