Releases · PhotoPrism

PhotoPrism's October release puts every AI model on the GPU

Classification now weighs the whole frame, NSFW detection is rebuilt on ONNX, every built-in model can run on a supported NVIDIA card, and new face clusters name themselves.

By BoneThursday 8 October 20262 min readReleases

entryread only

filedThursday 8 October 2026

length2 min

sectionReleases

subjectPhotoPrism

sourceson the page

PhotoPrism’s October release, out on the 7th, is an AI-features refresh with three upgrades worth knowing about. Image classification is reworked on a new ONNX-based model and now takes more of the entire picture into account, which matters most for wide photos where the subject is not centered, the case the previous center-weighted model handled worst. NSFW detection gets the same treatment, including better handling of borderline frames where the old model disagreed with itself. And the one hardware owners will feel: on a supported NVIDIA graphics card, all of the built-in AI models can now run on the GPU instead of the CPU, which turns library-wide scans from an overnight affair into something closer to a coffee break, on the machines the feature supports.

Faces, phones, and the rest

Face recognition gets friendlier: new clusters are named automatically where the matches are confident enough, and manual corrections made during recognition are better preserved instead of being steamrolled by the next pass. Insta360 users gain support for the X4 and X5, including their videos, proxy files and brackets, with improved stacking. Everything here still runs locally: the AI changes shift which hardware can do the scanning, not where your library lives. Database backups and WebDAV sync both pick up improvements, and there are the usual smaller fixes around them.

The upgrade notes carry a chore

The switch to ONNX comes with a one-time migration. Existing labels are preserved, but to replace them with results from the new model, the published procedure is two commands: reset the labels, then run them again with force, which is the kind of instruction worth actually following if you care about classification quality. Custom TensorFlow models configured in the vision file are no longer supported and must be replaced with ONNX versions; that model needs replacing, not re-pointing. The old environment switches have shrunk accordingly: disabling everything TensorFlow touches now only disables the TensorFlow-based models, and disabling classification entirely is a model choice rather than a flag. One default change quietly improves life behind a reverse proxy: the trusted proxy setting now includes the loopback addresses.

My read

ONNX plus GPU is the combination the library owner has been waiting for: real hardware behind the scan, and a classifier that looks at the photograph rather than the center crop. The label migration is a ten-minute chore with the commands handed to you, and the notes around it remain the standard other self-hosted projects are judged by. Run the two commands, flip the GPU on if you have one, and re-judge the classifier on the wide shots you know it used to miss, then run a scan and enjoy the shorter wait.

Source: the October release.

ReleasesHomebridge 2.5.0 is a Matter release: new device types, and fixes for identity and cacheHomebridge 2.5.0 adds Matter device types for solar, battery, EV chargers, meters, tariffs and water heaters, and spends the rest of its list on behaviour that used to go wrong: air quality reporting only two levels, thermostats registering late, restored accessories losing their uniqueId or owner, and cache reads that took the accessory with them. It also bumps @matter/main to v0.18.0 and @homebridge/hap-nodejs to v2.3.0.

ReleasesMealie 3.30.0 lets you own the AI prompt, time each step, and plan a whole weekMealie 3.30.0 moves AI prompt control into the app, so prompts can be appended to, replaced or extended with global instructions, and it retires the OPENAI_CUSTOM_PROMPT_DIR variable while importing existing prompts automatically. Instruction steps gain individual durations, meal plan entries can cover a full week, and a batch of fixes tightens import scoping, locale lookup and the permissions needed to manage units, tools, labels and recipe actions.

ReleasesMemos 26.10.0 signs API clients out, stops reading /etc/secrets and makes export an admin jobMemos 26.10.0 is the first release under calendar versioning, and its upgrade notes matter more than its features: old installs need a 0.31.0 waypoint, cached access tokens expire on upgrade, the /etc/secrets files from 0.31.0 are no longer read, member export disappears, and the old upload and memo-length settings are replaced without migration. The features are a rich editor by default, per-user usage limits and declarative configuration.