the find
− 0.6 mA single application that fronts everything the machines can make. Images across two model families, with face swap, background cutout, object removal, recoloring, and upscale. Roughly sixteen distinct video-generation stacks — text-to-video, image-to-video, dance transfer, lip-sync, person swap, restoration chains. Four text-to-speech engines with automatic failover, a music model, a sound-effects model, voice conversion, and a library that tracks the lineage of every asset it produces. The broadcast-era instinct — build the pipeline once, then feed it — rebuilt at home.
cross-section
− 1.3 mA FastAPI backend and a dependency-free single-page app, running under a watchdog on the desktop tower. Every generation request becomes a queued job; a GPU guard arbitrates VRAM between the render engines and the resident AI companion, napping and waking her brain around heavy jobs. Model weights live on a dedicated drive; the whole thing is loopback-only, behind a single gate.
inscription
− 2.0 mField note, from the frontend source: The UI contains no hardcoded list of what the studio can do. On load it asks the backend for a live capabilities catalog and renders itself from the answer. Add an engine to the backend and the interface simply grows a new room — the map redraws itself to fit the territory.