The challenge
What the system needed to solve
Provide high-quality visual generation and restoration without uploading private source images or relying on metered cloud GPU APIs.
The build
Engineering decisions
Installed and validated Fooocus with a current CUDA/PyTorch stack on the dedicated Windows GPU workstation.
Bound the inference service to the private network and placed its browser route behind two-factor authentication.
Documented model bundles, data roots, boot-time recovery, GPU compatibility checks, and acceptance prompts.
Added Upscayl as a complementary local enhancement workflow for existing images.
Architecture
How the pieces connect
- 1Authenticated browser
- 2Caddy access policy
- 3Private GPU service
- 4Local models and output storage
Results
What changed
- Generation and enhancement use owned hardware with predictable cost and private inputs.
- The GPU service survives workstation restarts through a verified boot-time task.
- Model and application state are separated from replaceable software components.
