Local-first AI Culling

Local-first · AI culling

Keep the best,
on your own machine

Burst shots are grouped automatically; a two-stage funnel hands you the candidates. Photos never leave your machine, and the machine never makes the final cut — you do, in the A/B arena.

↓ Go to downloadsOther platforms

Free · macOS / Windows

Two principles

Two things we won't bend on

Photos stay local

Originals are only ever read on your machine; only low-res previews for scoring go to the cloud — used once, then discarded.

You make the final cut

The machine only hands each group a good-enough shortlist. What to keep, or whether to drop the whole group, is your call.

How it works

Four steps, from bursts to picks

Each funnel uses the same rule: everything ≥60 passes; if short of the quota, top up by score — never cutting anything good enough.

  1. 01

    Group

    Similar bursts are clustered into one 'moment' by semantics, time and faces.

  2. 02

    Local pre-screen

    On your machine: technical scoring — sharpness, exposure, faces, aesthetics; penalties for closed eyes / out-of-focus.

  3. 03

    AI review

    Online AI scores aesthetics, expression, composition and semantics, with explainable keep/cut reasons.

  4. 04

    A/B arena

    You decide head-to-head in the arena — the final cut is always yours.

Features

Why Keeper

Local model scoring

Several local AI models (quality / aesthetics / faces) run on your machine, offline-capable.

Cloud LLM aesthetics

Volcano Ark for a second, taste-aware pass.

RAW · HEIC support

Native decoding of common camera RAW and HEIC, with thumbnail caching.

Face-aware grouping

Same scene, different people get split apart; group photos by 'same crew or not'.

Explainable picks

Every photo gets a keep/cut reason — not a black box.

Resumable projects

Culling is project-based; every step is persisted, quit and resume anytime.

A look inside

A darkroom built for culling

Same design language as this site — warm, dark, restrained, letting photos glow.

Group detail: local + AI scores and verdicts
Group detail: local + AI scores and verdicts
A/B arena: head-to-head final pick
A/B arena: head-to-head final pick

Requirements

  • macOS 11+ (Apple Silicon) or Windows 10+ (x64)
  • First launch downloads ~1.6 GB of local AI models to the app cache directory; the local pipeline then runs offline, while AI review still needs the network
  • AI review needs your own Volcano Ark API key (entered in-app)

Download

Pick your platform

Free download, installers served from Cloudflare's global network.

macOS · Apple Silicon

M-series (aarch64) · .dmg

Unavailable

Windows

x64 · .exe

Unavailable

First open needs manual approval (currently unsigned)

macOS: if it says 'damaged' or 'unverified developer', go to System Settings → Privacy & Security and click 'Open Anyway'; or run xattr -dr com.apple.quarantine /Applications/Keeper.app

Windows: when SmartScreen blocks it, click 'More info → Run anyway'.

FAQ

You might be wondering

Is Keeper free?

The current MVP is free. A commercial cloud-relay/billing version is planned.

Are my photos uploaded?

Originals never leave your machine. Only low-res previews for LLM scoring are uploaded then discarded; geocoding sends GPS coordinates only.

Do I need internet?

First launch downloads local models (~1.6 GB). After that the local pipeline is fully offline; only the AI review needs the network.

How do I open an unsigned build?

macOS: 'Open Anyway' in Privacy & Security, or strip the quarantine attr with xattr. Windows: 'Run anyway' in SmartScreen.

Do I need an API key?

The AI review needs a Volcano Ark API key, entered in Settings; it's stored locally and never committed.

Can I use it commercially?

The bundled face-recognition model is for non-commercial research only; replace or license it before commercial use.