AI Photo Search Comes to the Mac
SCM is an AI photo search project that turns a personal media library into a collection searchable with natural language. The open-source macOS app runs vision models locally and indexes photos and videos, so users can describe a scene instead of remembering a filename. The project drew strong attention after its Hacker News launch and is hosted on GitHub.
SCM Searches Beyond Whole Files
SCM separates the task into complementary modes. Visual search matches photos or videos by meaning, Scenes segments videos and points to a timecode, OCR searches text that is visibly present, and Dialogue uses Whisper to find spoken words. Filenames and phrases can influence some rankings, while results show why they matched. Queries can also be saved as tabs with a defined scope.
The app also includes library-oriented maintenance features. Watched folders can import automatically, content hashes help detect duplicates after renames, and changing models triggers background re-embedding. OCR supports English plus optional additional languages, while video search offers sampling choices that trade speed and disk usage. An optional local LLM can answer questions from extracted dialogue, OCR, and filename evidence, with citations.
Local Privacy Still Has Practical Costs
SCM’s central privacy claim is straightforward: media stays on the Mac. After model weights are downloaded, processing can continue offline without an account or cloud upload. That does not make the experience cost-free, however. The default vision model requires roughly 435 MB, chat models require about 1.1 GB or more, and initial indexing, video segmentation, Whisper transcription, and model changes can consume substantial CPU time, storage, and waiting time. The app is currently macOS-focused, and search quality depends on model choice, sampling density, OCR recognition, and transcription accuracy. For large video libraries, the open question is whether local indexing remains affordable to maintain while returning results quickly across different Mac configurations.