Infrastructure direction
Where should the work live?
Hosting is a set of separate decisions: compute, model files, storage, networking, secrets, backups, and the interface people use. Make each layer explicit before moving a local experiment online.
Private workstation
Best for sensitive inputs, frequent iteration, and predictable ownership. Limited by hardware, electricity, uptime, and local networking.
Managed service
Fastest route to a usable interface. Check data retention, export, model choice, rate limits, and whether the service supports your real workflow.
Hetzner-style server
Good for a coordinator, API, database, dashboard, or always-on service. You own more operations: updates, firewall, backups, access, and monitoring.
RunPod or marketplace
Useful for bursts of image or video generation. Treat storage, idle time, image persistence, and deletion behavior as part of the cost.
A simple architecture map
| Layer | Question | Failure to prevent |
|---|---|---|
| Interface | Who can reach it, and what can they do? | An accidentally public admin panel |
| Coordinator | What receives jobs and reports status? | Lost jobs or invisible retries |
| Model runtime | Which model runs, with what limits? | Unexpected quality, license, or GPU cost |
| Storage | What persists, for how long, and who can delete it? | Secrets or user media in an unprotected volume |
| Operations | How are updates, backups, logs, and alerts handled? | A service that works until the first restart |
How to compare providers
Practical starting point
Keep the handbook and small coordinator local. Put only the expensive, bursty generation step on a rented GPU. Move an always-on public service to a managed or self-managed host only after the local workflow has a measured baseline.