Performance
- Batch GPU inference: CLIP workers dequeue requests from all scan threads and process them as batches, keeping the GPU fully utilised. Scans are significantly faster on large libraries.
- Concurrent thumbnail fetching: assets are fetched in parallel with a configurable thread pool (
SCAN_WORKERS, auto-derived asGPU_WORKERS × 32). - YOLO runs on CUDA: the YOLO detector now loads onto the GPU when available.
Features
- AMD/ROCm support: Docker image supports NVIDIA (default), AMD/ROCm, and CPU-only via a build arg.
- In-UI getting started guide: the main panel shows a 6-step workflow on first open. The
ibutton in the sidebar header brings it back at any time. - Pet folder keys use person ID: pet data folders are keyed by Immich person UUID instead of name, avoiding issues with special characters.