Maple

Cross-platform desktop app for importing, browsing, and organising photos. UI built with Slint — no GTK/libadwaita runtime required.

Features: recursive photo import with dedup, a fast library browser with full-text and semantic search, face detection/recognition, AI-generated image descriptions, burst/near-duplicate stack detection, and collections.

Prerequisites

• Rust (stable, 2021 edition) — install via rustup
• A C toolchain — rusqlite (bundled), sqlite-vec, and webp all compile vendored C sources via the cc crate:
  • Linux: build-essential (sudo apt install build-essential)
  • macOS: Xcode Command Line Tools (xcode-select --install)
  • Windows: MSVC Build Tools

No system GTK headers are needed — Slint ships its own renderer.

Maple builds in two variants depending on which ONNX Runtime execution backend you want: CPU (default, simplest) or GPU (CUDA/TensorRT). The backend is fixed at compile time — there's no runtime fallback between the two within a single binary.

Installation — CPU build (default)

This is the simplest option and what you want unless you specifically need GPU-accelerated inference (face detection, AI tagging, stack detection, semantic search all still work on CPU, just slower).

cargo build --release -p maple

ort (the ONNX Runtime crate) downloads a platform-matched prebuilt onnxruntime binary at build time and links it into the binary. This means:

• Network access is required the first time you build (to fetch the onnxruntime binary).
• The result is CPU-only — no CUDA/TensorRT execution providers are available. If settings.toml has device = "cuda:0" or "tensorrt:0" under [face], [stacks], or [semantic], it silently falls back to CPU (this is ort's normal behavior when an execution provider is unavailable, not a bug).

Run it with:

cargo run --release -p maple

Installation — GPU build (CUDA / TensorRT)

Opt in explicitly, one package at a time:

cargo build --release -p maple --no-default-features --features gpu

The GPU variant uses ort's load-dynamic mode instead of linking a binary at build time:

• No network access is needed to build.
• The resulting binary loads onnxruntime at runtime instead, and needs a real onnxruntime shared library to even start — this applies even if you only intend to run CPU inference on a gpu build, not just for GPU use.

You need to supply a CUDA/TensorRT-enabled onnxruntime release, e.g. Microsoft's official onnxruntime-gpu build, either by:

1. Setting the ORT_DYLIB_PATH environment variable to point at the shared library, e.g.:

   ORT_DYLIB_PATH=/path/to/onnxruntime/lib/libonnxruntime.so \
     cargo run --release -p maple --no-default-features --features gpu

2. Or placing the library where the OS's dynamic loader finds it by default:
   • Linux: a directory on LD_LIBRARY_PATH, or set an rpath
   • Windows: next to the .exe
   • macOS: a DYLD_* library path

Then set device = "cuda:0" (or "tensorrt:0") under the relevant section ([face], [stacks], [semantic]) in settings.toml.

Ship both variants as separate release artifacts per platform if you want to offer GPU acceleration — pick one backend per build.

Model setup

Maple's AI features need ONNX models. Some are downloaded automatically; face detection models must be supplied manually.

• Face detection/recognition ([face] in settings.toml) — download manually and point detector_model (and optionally embedder_model) at the file:
  • Detector: atksh/onnx-facial-lmk-detector (default, single-pass detection + landmarks + aligned crops) or InsightFace SCRFD
  • Embedder (optional, enables person grouping): an InsightFace ArcFace model
• Stack detection ([stacks]) — downloads the DINOv2 ONNX export (model_repo = "onnx-community/dinov2-small") from Hugging Face Hub automatically on first use.
• Semantic search ([semantic]) — downloads a sentence-transformer ONNX export (default sentence-transformers/all-MiniLM-L6-v2) from Hugging Face Hub automatically on first use, cached under ~/.config/maple/models/.
• AI image descriptions ([ai]) — no local model; talks to an OpenAI-compatible server (e.g. LM Studio) over HTTP, configured via server_url and model.

All settings live in ~/.config/maple/settings.toml, written on first launch from maple-state/defaults.toml.

Build & Test

cargo build --workspace
cargo test --workspace
cargo clippy --workspace

Workspace Crates

Crate	Purpose
maple	Binary entry point
maple-ui	Slint UI: windows, views, widgets
maple-state	Settings (settings.toml), Session (session.json), SeenSet (bloom filter)
maple-import	Recursive image scanner, BLAKE3 hasher, file copier, raw file support
maple-db	SQLite library database, background scanner, EXIF, AI tagging, face detection