Building the studio's 24/7 render pipeline: queues, budgets and vision-model QC

Our render farm has no racks and no fans. It is a job queue, a credit ledger and a set of provider adapters — and it is the closest thing the studio has to a superpower.

Jobs, not clicks

Nothing renders synchronously. A brief enters the queue with a priority and a budget; workers pick it up, call the right provider, and write the result back with full lineage — who asked, what prompt, which model, what cost. Reproducibility is not a feature, it is the default.

Failure is a state, not a surprise

Providers time out, moderate, and hallucinate. Every job has an explicit state machine: queued, rendering, ready, failed — with automatic refunds of credits on failure. The pipeline treats a flaky API the way a set treats weather: annoying, planned for, never fatal.

The inspector never blinks

Each visual render is reviewed by a vision model against its brief and scored out of 100. Below the bar, the system rewrites the prompt — sharper subject, cleaner framing — and re-queues automatically, a bounded number of times. Humans see the survivors and the interesting corpses.

Why we wrote about it

Because this layer, not any single model, is what turned generative tools from demos into a dependable production department.