Skip to main content

AI management

Photoly runs SigLIP2 (image and text) on ONNX Runtime. It runs either in the server's ai-worker child process or in the user's browser. Photoly downloads nothing until you enable AI.

Where each model runs​

Go to Admin > AI Management > Where AI runs. Set the placement for each model: photo indexing and search.

  • Server
  • Browser
  • Let users choose. Each user then picks under Account > AI on this device.

Server placement​

  • The server loads a model only while a user needs it there. It frees the model when all users choose the browser.
  • The memory meter shows what fits.
  • If a model may not fit, Photoly asks before you choose Server (low-memory). Load Anyway forces it.
  • Photoly restarts a dead worker on the next request (worker-died).
  • Unload frees a loaded model at once. It reloads on demand.

Browser placement​

  • Photoly serves browser models from /models/<revision>/.
  • The browser caches them. Users can remove them under Account > AI on this device.
  • The browser search model is small (int8, 318 MB) or full (1.13 GB).

For indexing and search behavior, see Search and similar media.

Operation​

Commands:

  • photoly serve
  • photoly migrate
  • photoly ai-worker (serve starts it)
  • photoly verify
  • photoly health

Storage and runtime:

  • Weights are in PHOTOLY_DATA. This is the photoly_models volume.
  • ONNXRUNTIME_LIB points to the ONNX Runtime library.
  • CPU works everywhere. The CUDA provider needs a CUDA build of the image.

Sizes and upgrade notes are in Media tools, Extract Text and Smart Tag operations.