The interesting claim is not "this is AI" and it is not "a human made this". It is the split, and the split is the same on every release:
Lyrics are human. Music, vocals, mix, master, images and video are AI-generated. Direction is human at every single step.
That last clause is the one that matters. No track, image or video on this label is published unchecked.
What a human does
A human sets the brief, the creative direction, the objectives and the reference material. Writes the lyrics, in full. Art-directs the scenography, wardrobe and photography. Authors the character consistency specification that every image and video of an artist is generated through. Then selects, rejects and sequences every output.
The rejection rate is the work. A generative model will produce an unlimited quantity of nearly-right material, and nearly-right is worthless. Choosing is the job.
What a machine does
| Step | Tool | What it produces |
|---|---|---|
| Composition | Suno | The track, and the exploration that precedes it |
| Mix and master | Suno | The finished audio |
| Voice | ElevenLabs | Spoken sections and ad-libs |
| Stills | Higgsfield, OpenArt | Photography, character reference sheets |
| Video | Seedance 2.x, Higgsfield | Motion |
Every one of these has an entry on /contributors, with the roles it is credited in. Models and tools sit in the same index as the humans, distinguished by a type field rather than hidden in a footnote.
Why the artists look the same in every image
Generative image models drift. Ask for the same person twice and you get cousins. A label whose artists exist only as images cannot tolerate that, so each artist has a face lock — a written specification of permanent identity markers, camera treatment, and a drift guard listing what must never appear.
Grio has two razor slits in the left brow only, and short freeform twists — not braids, not locs, not an afro. Abel has one small black cross on the left cheekbone, never mirrored. Piet has a thick blond chevron mustache and no beard.
These are not stylistic preferences. They are the reason three hundred images read as one person.
The prompt system
Shot prompts are not written by hand. Each project holds a set of composable blocks — identity lock, wardrobe, film look, lighting recipe, camera DNA, world, audio policy, drift guard, reference bindings, continuity pin — and a build step welds the selected blocks into every shot. Resolution runs shot, then scene, then project, so a single shot can override the film look without touching the other forty.
The constraints that shape the output are unglamorous and specific. Neither Seedance nor Higgsfield has a negative-prompt field, so drift guards are inlined. Identity is held by reference Elements and a verbatim identity block, never by seeds — Higgsfield has no seed at all. Coherence degrades past roughly ten seconds, so longer shots are staged rather than requested. Every reference needs both a job and an exclusion, or it leaks its own background into the frame.
Provenance metadata
Images and video carry C2PA provenance metadata where the pipeline supports it. Optimisation steps routinely strip that metadata, so an automated check runs on every build and fails it if provenance has been lost between the source file and the file this site serves.
EU AI Act Article 50 has applied since 2 August 2026. We would be doing this anyway.
What we will not claim
No release on this label is presented as human-made. No artist here is presented as a person. There is no ambiguity to resolve, no reveal being saved, and no version of this project in which the answer is withheld.
Every artist page carries its own provenance table. Read them.