Two cards, no cloud

Generative video is sold as a cloud product with a per-frame meter running. Used daily, that meter is the whole cost structure, and every render ships your inputs to someone else's machine. I run the pipeline at home instead. The model loads, the queue runs, the output writes to disk — no API key, no per-frame bill, nothing leaving the house.

The interesting part is GPU routing. The workstation has two RTX 5060 Ti cards plus an RTX 4080 reserved for gaming. The app pins generation to the two 5060 Ti cards, watches VRAM headroom per card, and dispatches the next batch to whichever has room. The 4080 stays untouched on purpose, so a long render never steals frames from anything running on the gaming card. Two mid-range cards that stay out of each other's way beat one big card that is always being preempted by something rendering a game.

The surface is deliberately plain. A Gradio frontend exposes it over the local network, prompt presets cover the common framings — slow zoom, parallax pan, ambient drift — and a batch queue lets a long session run unattended while a dashboard shows per-job progress, ETA, and which card is doing the work. It runs as the production tool behind a fleet of Discord bots: requests arrive over webhook, hit the queue, and write back to the requesting channel when the render finishes.

Boring infrastructure, interesting outputs. The reason to own the hardware is not thrift — it is that an unmetered pipeline changes what you are willing to generate. When a render costs electricity instead of credits, you stop rationing and start experimenting, and the experiments are where the good outputs come from.