Imuhar

Imuhar: Free Up Space
Cull by intent, not pixels.
Privacy-first photo cleanup for iOS 18+ — on-device models, explainable Keep/Toss, deletes go to Recently Deleted.

---

The problem

Your camera roll is full of utility junk (receipts, whiteboard shots, screenshots of screenshots), social misses (blinks, looking away), and weaker frames from every burst. Apple’s duplicate finder is pixel-matching. Most “cleaner” apps are freemium traps that want your photos in the cloud.

Imuhar asks a different question: what was the human trying to do with this photo? Then it explains why something should go.

What it finds

Find	How it knows
Duplicates & better shots	Near-dup clusters; aesthetic + face quality + sharpness − miss penalty
Utility junk	Vision text/barcode/document + SigLIP zero-shot + optional VLM
Social misses	Face landmarks — blink, looking away (mouth-open experimental)
Blurry / exposure	Laplacian sharpness + exposure histogram

You pick categories before a scan. Review is swipe cards with reason chips — not a black-box score.

Privacy, for real

• Photos never leave the device
• No accounts, no ads, no analytics
• Deletes → Recently Deleted (30-day undo via Apple)
• One optional network path: checksum-verified VLM weights you opt into in Settings

Details: PRIVACY.md

How it works

Cheap signals on every photo; expensive models only on the uncertain subset — and heavy VLM work prefers charging.

SwiftUI  →  PhotosKit / Vision / Core ML / MLX
    ↕ UniFFI
Rust     →  signals · fusion · clustering · ranking · SQLite · thermal scheduler

Cascade: metadata → Rust pixels → Vision + Core ML → optional MLX VLM → near-dup clusters → ranked review queue.

Full story: ARCHITECTURE.md

Requirements

• macOS + Xcode 16+
• Rust stable (rustup)
• iOS 18+ device (A15+ recommended; VLM needs ≥6 GB RAM)
• XcodeGen: brew install xcodegen

Build & run

# 1. Rust core → XCFramework + Swift bindings
./scripts/build-xcframework.sh

# 2. Generate & open the iOS project
cd ios && xcodegen generate
open RustCleaner.xcodeproj

Set your Development Team in Xcode, then run on a real device — PhotoKit needs a real library to feel like the product.

Rust-only (fast iteration)

cd core
cargo test --workspace
cargo run -p cleaner-cli -- weights

Models

Bundled Core ML (SigLIP + aesthetic head) plus optional on-device VLM. Licenses, sizes, and conversion scripts: models/MODELS.md.

Contributing

See CONTRIBUTING.md. Fusion weights are TOML-tunable without touching Rust:

cargo run -p cleaner-cli -- weights > my-weights.toml
# edit, then load via CleanerEngine.setWeightsToml in the app

License

MIT © 2026 Gidi Hanoch