Selects

Cull thousands of travel photos down to your keepers — locally, privately, on your own machine.

Point it at a folder of photos and videos. It indexes, scores, clusters, and groups them into day-by-day stories, surfaces the best shots, and gets out of your way. Nothing is uploaded anywhere. (The Map view is an exception: it loads OpenStreetMap tiles from the internet, and reverse geocoding of GPS coordinates into place names is optional and can be disabled.)

Why Selects

I built Selects for the post-trip mess: 3,000 photos, five versions of the same mountain, ten blurry food shots, and one frame where everyone actually has their eyes open.

The app is meant to do the boring first pass locally, then leave the final taste call to you.

The only network call is an optional place-name lookup for geotagged shots.

Features

Install

One line (recommended) — installs via uv, which brings its own Python, so there is nothing to install first and no downloaded app for macOS Gatekeeper or Windows SmartScreen to flag. It opens the web UI; AI models download from the app's first-run setup screen.

# macOS / Linux
curl -LsSf https://bihanikeshav.github.io/selects/install.sh | sh

# Windows (PowerShell)
irm https://bihanikeshav.github.io/selects/install.ps1 | iex

Desktop bundle — prefer a self-contained download? Grab the bundle for your OS from Releases and run it. No Python required; it downloads its AI models on first launch. (Unsigned, so the OS shows a one-time "unverified" prompt — right-click → Open on macOS, or More info → Run anyway on Windows.)

Via pip / uv (Python 3.11+):

uv tool install "selects[ml]"  # or: pip install "selects[ml]"
pip install selects            # base app + web GUI + CLI, no on-device AI
pip install "selects[ml,desktop]"  # AI + a native desktop window instead of a browser tab
selects serve                  # open the web UI (port 8000; honors SELECTS_WEB_PORT)
selects index /path/to/trip    # or run headless from the CLI

RAM++ is the ONNX graph in the selects-onnx bundle — it is already part of [ml]. No extra pip install and no recognize-anything git checkout.

Platform support

ONNX Runtime does not need NVIDIA CUDA to use a GPU. Windows [ml] ships DirectML (any DX12 GPU). macOS can use CoreML. CUDA is used only if you install onnxruntime-gpu on NVIDIA Linux.

SigLIP, CLIP-IQA, and RAM++ still run on CPU everywhere — those graphs are not DirectML-safe. Whisper can use DirectML/CoreML/CUDA when the provider exists. InsightFace uses CUDA if present, otherwise CPU. selects doctor and the Models page show which provider this machine selected.

Architecture

API-first: a FastAPI backend does all the work; every client — web UI, CLI, future mobile — is just another consumer of the same /api surface. State lives in a per-library SQLite DB inside the photo folder (<folder>/.selects/), so a library is self-contained and portable.

flowchart TD
    Clients["Web UI · CLI · Mobile (planned)"]
    API["FastAPI + WebSocket API"]
    Pipeline["Pipeline orchestrator"]
    Classical["Classical scoring<br/>(blur, exposure, faces)"]
    ML["ML stages<br/>(SigLIP, CLIP-IQA, ArcFace, RAM++, Whisper)"]
    Files[("Photos & videos<br/>local disk")]
    DB[("Per-library SQLite<br/>&lt;folder&gt;/.selects/")]

    Clients --> API
    API --> Pipeline
    API --> DB
    Pipeline --> Classical
    Pipeline --> ML
    Pipeline --> Files
    Pipeline --> DB

Each stage reads/writes <folder>/.selects/index.db and is independently re-runnable via selects index <folder> --pass <stage>:

speed_mode=fast skips ram_tag, smart_tag, face_embed, persons, and video Whisper / InsightFace enrichment. Aesthetic curation ranks on CLIP-IQA (Embedding.aesthetic_iqa in [0, 1]) with configurable per-scope and library-wide percentile thresholds (see Configuration).

Roadmap

Shipped (v0.1)

• Discovery search, tagging, people, face-aware culling
• Stories, aesthetic curation, duplicate finder
• Keyboard culling + compare, taste learning
• Export (copy/zip + XMP), trip recap, watch folder
• Video highlights, clip/moment cull, local transcripts, Models page
• CPU desktop builds for Windows, macOS, Linux + PyPI package
• DirectML / CoreML GPU path (no CUDA required); SigLIP/RAM++ still CPU

Planned

• SigLIP/RAM++ on GPU — re-export graphs that DirectML/CoreML can run
• Android companion — LAN remote that drives the desktop backend from your phone
• Android standalone — on-device culling for small libraries (no desktop needed)
• Cursor-based pagination, auto-tuned aesthetic/burst thresholds, iOS parity

Quickstart (from source)

Requires Python 3.11+ and Node 18+.

pip install -e ".[ml]"        # ML stack (onnxruntime, insightface, sklearn, …); omit [ml] for classical-only
selects serve /path/to/photos # backend + web UI (indexes in the background)

cd frontend && npm install && npm run dev   # hot-reloading UI (separate terminal)

selects serve opens the web UI (--no-browser to skip) and indexes in the background (--no-background to skip). With the frontend built once (npm run build), the backend serves the UI same-origin — no npm run dev needed. With no folder argument it opens the active library, or onboarding if none exists. Drive stages directly with selects index <folder> [--pass <stage>] and check hardware with selects doctor.

macOS (Apple Silicon)

No Docker or other services are needed; the models run in-process on ONNX Runtime. Use a native arm64 Python. An Intel Homebrew install (/usr/local) runs as x86_64 under Rosetta, and there umap-learn's llvmlite dependency has no wheel, so it tries to build from source and fails. uv can fetch a native interpreter:

uv venv --managed-python --python cpython-3.11-macos-aarch64 .venv
uv pip install --python .venv/bin/python -e ".[ml,dev]"
.venv/bin/python -c "import platform; print(platform.machine())"   # expect: arm64
.venv/bin/selects serve /path/to/photos

The first ML run (or the Models page) downloads weights into ~/.cache/selects/models/ (SELECTS_MODELS_DIR to override): selects-onnx/ (~3.1 GB), buffalo_l/ (~330 MB), and whisper-small/ (~250–500 MB). Older InsightFace installs under ~/.insightface/models/ still count as present. Downloads use plain HTTP because Selects sets HF_HUB_DISABLE_XET=1 by default; the Hub's Xet downloader can hang mid-file. Set HF_HUB_DISABLE_XET=0 to opt back in, and HF_TOKEN for faster, less rate-limited downloads.

Video-only folders show up under Videos in the sidebar. The photo views (Cull, Curated, People, Map) stay empty for them.

Configuration

selects serve binds 8000 by default. If --port is omitted, it honors SELECTS_WEB_PORT. Per-folder settings via pydantic-settings; override any field with a SELECTS_-prefixed env var (or .env). See selects/config.py.

Derived paths under <folder>/.selects/: index.db, thumbs/, previews/.

Per-trip customization — drop optional JSON into <folder>/.selects/ (missing/malformed falls back to defaults); see examples/ladakh/:

Development

pip install -e ".[dev]" && pytest && ruff check .
cd frontend && npm run lint && npm run e2e   # ESLint + Playwright smoke test

npm run e2e needs Chromium once (npx playwright install chromium). It builds the SPA into selects/server/static/, indexes a throwaway six-photo library in the temp directory and drives the real server on port 8765. Point SELECTS_PYTHON at the interpreter that has selects installed if it is not the one on PATH. Use an absolute path, because the fixture runs it from the repo root (e.g. SELECTS_PYTHON="$PWD/../.venv/bin/python" npm run e2e from frontend/ on macOS/Linux, or ...\.venv\Scripts\python.exe on Windows).

For the native desktop window (pywebview), install selects[desktop] (or selects[ml,desktop] for AI + the desktop window).

Schema is managed with Alembic; migrations ship in selects/db/migrations/ (no alembic.ini) and init_db() upgrades each library's DB to head on open. After editing selects/db/models.py, autogenerate a revision against a throwaway SQLite URL and review it — SQLite ALTERs go through render_as_batch (enabled in env.py).

Layout

Desktop build — pip install "pyinstaller>=6.6" then python packaging/build.py [--ml]. It builds the frontend into selects/server/static/ (same-origin UI) and runs PyInstaller (onedir) via packaging/selects.spec into dist/selects/.

Contributing

1. Fork, branch, pip install -e ".[dev]", add tests under tests/.
2. pytest and ruff check . must be green.
3. Open a PR — the architecture and roadmap are the best places to find direction.

Good first areas: GPU execution providers, the mobile client (the API already exists), export formats, and tuning aesthetic/burst defaults for different shooting styles.

Known limitations

• Aesthetic/burst thresholds were tuned on a single trip; may need adjustment for other styles/gear.
• List endpoints use offset/limit, not cursor-based, pagination.

License

MIT — see LICENSE.