Show HN: AI Search For Every Photo And Every Frame Of Video On macOS

TL;DR

Prime Big Deal Days · Oct 6–7Offer from Amazon

Get tech for your team delivered free — and shop member deals

• Fast, free delivery on millions of items
• Access to Prime Big Deal Days deals on October 6–7
• Prime Video, Amazon Music and more included

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

A project called SCM, introduced in a Show HN post and documented on GitHub, offers a macOS app for searching photos and video using local vision, OCR and speech transcripts. The project says media stays on the Mac after model weights are downloaded; its capabilities and privacy assurances are developer claims, not independently verified findings.

The developer of SCM (Screen Memories) has presented a macOS application for searching photos and video stored in local folders, using on-device vision, OCR and speech transcription. The GitHub project describes searches for images by meaning, individual video scenes, visible text and spoken words, but its performance and privacy claims have not been independently assessed in the supplied material.

SCM organizes searches into five modes: Files, Scenes, OCR, Dialogue and LLMs. Files mode ranks whole images and videos against a plain-language query using image embeddings, with additional filename and phrase signals, according to the project documentation. The app also displays a reason-for-match badge and score details. Scenes mode searches segments within videos and is designed to open a result at the matching timecode rather than only at the start of a file.

OCR mode looks for literal text detected in images and video frames, while Dialogue mode searches spoken words transcribed with Whisper. The project says OCR uses Tesseract and supports English plus 35 selectable languages. Dialogue results are grouped by how closely words match a transcript and can seek to the relevant line. These features rely on extracted text or transcripts rather than the vision search mode, the developer says.

The project describes SCM as local-first: inference runs on the Mac, and downloaded model weights are used offline afterward. The default CLIP model download is listed at about 435 MB. An optional chat feature can use a local llama.cpp sidecar to answer questions from information SCM has extracted, including dialogue, OCR and filenames; the documentation says this feature does not download or run until enabled. Listed chat models range from about 1.1 GB to 2 GB.

Searching Media Beyond Filenames

SCM targets a familiar problem for people with large personal media collections: finding a moment that is remembered visually or by its dialogue, but not by its filename. A query such as a description of a scene can search image content, while OCR and transcript modes offer separate routes to locate a sign, caption or spoken phrase. The scene-level approach could reduce the time spent opening and scrubbing through long videos, if it works reliably on a user’s library.

Its local-processing design also speaks to privacy concerns around uploading personal photos and recordings to a remote search service. The trade-off is that indexing and inference depend on the user’s Mac and downloaded models. The GitHub listing does not provide independent measurements of search accuracy, indexing speed, resource use or privacy behavior, so readers should treat these as described product features rather than verified results.

How SCM Organizes Local Search

The project combines several distinct search methods rather than treating every query as a single AI search. Files mode uses vision embeddings and can add filename and phrase signals; Scenes searches video segments; OCR performs literal matching against recognized text; and Dialogue retrieves words from Whisper transcripts. The documentation says OCR and Dialogue can operate while the AI engine is unavailable, because they do not require the vision model.

SCM also describes watched-folder importing, content-hash deduplication and background re-embedding when models are changed. Users can save searches as tabs, with up to 20 saved queries, and can narrow results to built-in categories such as Videos or optional Screenshots and Email tabs. The README lists Homebrew installation for Apple Silicon Macs running macOS 12 or later, while development uses Bun and an Electron-based app package. Those details describe the project’s published setup, not independent compatibility testing.

“Search like you think — describe a memory in plain language; a local vision model does the rest.”

— SCM project description on GitHub

Accuracy and Privacy Still Unverified

The supplied project material does not include independent tests or benchmarks showing how accurately SCM finds images, scenes, OCR text or spoken phrases across different media libraries. It also does not report typical indexing times, storage needs, battery or memory use, or how behavior varies across supported Macs. The developer says media stays on the machine, but the provided description does not include an external security or privacy audit.

It is also unclear from the supplied material how broadly the app handles video formats, very large libraries, or changes to files after import. The GitHub documentation lists features and installation requirements, but no user adoption figures, independent reviews or comparison with other media-search tools were provided.

Check the Release and Test Results

Readers can review the project’s GitHub repository for its current code, installation instructions and release details. The README identifies Homebrew as the simplest installation route for Apple Silicon Macs running macOS 12 or later, and describes a default model download of roughly 435 MB. Users considering installation should check the repository for the latest version and requirements before proceeding.

The next useful evidence would be hands-on testing across different photo and video libraries, including reported search accuracy, indexing time and resource use. Until such evidence is available, SCM’s published feature set and offline-processing description remain claims by the project’s developer.

Key Questions

What is SCM?

SCM, or Screen Memories, is a macOS project for searching locally stored photos and videos using vision search, scene matching, OCR and speech transcripts, according to its GitHub description.

Does SCM upload photos to the cloud?

The developer says inference runs on the Mac and media does not leave the machine. The project says model weights download once and subsequent use is offline. Those privacy statements are developer claims in the supplied material, not independently audited findings.

Can it search for a specific moment in a video?

SCM’s Scenes mode is designed to search segmented video scenes and open a result at its timecode. The project documentation describes the feature, but no independent accuracy testing was supplied.

What can OCR and Dialogue search find?

OCR searches text detected in images and video frames. Dialogue searches spoken words transcribed with Whisper, with results grouped by how closely the transcript matches the query.

What Mac does SCM require?

The project’s installation notes list Apple Silicon and macOS 12 or later for its Homebrew cask. Readers should check the repository for current compatibility and release information.

Source: hn