PhotoPrism, the open-source, AI-powered photo management application that lets users organize and browse their photo libraries on a self-hosted server or NAS, has released its October 2026 update.
The biggest change is a reworked image classification system based on a new ONNX model. Unlike before, it considers more of the entire picture, improving recognition accuracy, especially for wide photos where important objects are not centered.
The update replaces the existing NSFW detection system with a new ONNX-based model. Users with supported NVIDIA graphics cards can now run all of PhotoPrism’s built-in AI models on the GPU instead of relying entirely on CPU processing.
Face recognition has received attention as well. PhotoPrism can now automatically name newly detected face clusters when all matches agree, reducing the need for manual identification. Corrections made while recognition is running are also better preserved.
For users with 360-degree cameras, the release expands support for Insta360 X4 and X5 videos, proxy files, and brackets. File stacking is improved to better handle related Insta360 files indexed at different times, and rescanning no longer restores previously archived Insta360 pictures. Other improvements include faster preview lookups, corrected orientation for previews extracted from RAW files, and better hardware video transcoding that respects quality and bitrate settings. The update also fixes playback issues affecting animated WebP images with unusual dimensions or transparency.
On the server side, MariaDB backups now use consistent snapshots without locking. Restore operations can preserve other database rows when an individual statement fails. WebDAV synchronization is improved so failed transfers no longer delay other downloads.
Additional changes include support for Let’s Encrypt certificates using TLS-ALPN-01, an upload retry button, new CLI commands for managing cameras and lenses, and improvements to low-memory mode.
Lastly, one important upgrade consideration is that image classification and NSFW detection now use ONNX models. Existing image labels are preserved, but administrators can regenerate them using the new model. Custom TensorFlow models configured through vision.yml are no longer supported and must be replaced with ONNX alternatives.
For more details, see the release notes. The developers recommend using the project’s Docker images for installations on private servers and NAS devices.