Why we built boppi: the music video should not be the hard part

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 boppi is built as a studio around the models, not as a wrapper over them.

Four decisions that shape boppi

Decision one: the beat grid is the source of truth. When you upload a track, boppi 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 boppi 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 boppi charges a flat 20% platform fee on every job instead of the full price.

Decision four: songs, voices and video live in one place. Music 3.0 writes 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 MiniMax

We evaluated every major video model on the same twenty prompts, scored for motion coherence, prompt adherence and cost per usable second. Hailuo 2.3 gave us the best ratio of quality to price for six- and ten-second shots, and its fast variant is good enough for drafting that we can afford to let you iterate. Music 3.0 produces full songs with vocals up to six minutes, and Speech 2.8 covers forty-plus languages with emotion control. 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.