The Conviction
Privacy is not a feature. It is a right. Every major video editor sends your footage to the cloud for “AI processing.” LivCut does not. Every AI engine, every compute kernel, every export runs locally on the Apple Neural Engine and Metal GPU. Zero cloud. Zero telemetry. Zero compromise.
The industry normalized uploading your personal videos to someone else's server. They call it a convenience. It is a trade: your content, your faces, your locations, in exchange for a filter. LivCut proves the trade is unnecessary. Professional-grade editing, 66 Metal compute kernels, 11 on-device AI engines. All of it running on the hardware you already own.
66
Metal Compute Kernels
11
On-Device AI Engines
382
Tests
209
Swift Files
66 Metal Compute Kernels
The rendering engine is built on Metal compute kernels. Not wrappers around Core Image. Not pre-baked LUTs. Custom GPU kernels written in the Metal Shading Language, dispatched directly to the GPU for every frame. Color grading, transitions, particle effects, beauty retouching: each one a dedicated compute pipeline running at native resolution.
Metal Performance Shaders handle the heavy lifting where Apple's optimized implementations outperform hand-rolled code: convolutions, morphological operations, histogram equalization. The pipeline composes them with custom kernels seamlessly. One unified render graph, no texture copies between stages.
Design Note: MSL as String Literals
LivCut compiles Metal Shading Language at runtime from string literals rather than precompiled .metal files. This is an intentional architectural choice: runtime MSL compilation enables dynamic kernel composition (chaining effects with varying parameter counts), hot-reload during development without Xcode rebuilds, and parameterized shader variants generated from user-facing controls. The trade-off is a brief first-launch compilation cost, amortized by Metal's pipeline cache on subsequent runs.
Export goes through VideoToolbox for hardware-accelerated encoding. ProRes, HEVC, H.264, HDR HLG. The hardware encoder runs independently of the GPU render pipeline, so you get 60fps preview with real-time scrubbing while a background export writes to disk. No watermark. No waiting.
11 On-Device AI Engines
Every AI capability in LivCut runs through CoreML on the Apple Neural Engine. No server round-trips. Results in milliseconds, not minutes.
Scene Detection
Automatic cut point identification across timelines
Silence Removal
Audio analysis and gap trimming in a single pass
Smart Cut
Content-aware editing that preserves narrative flow
Filler Word Removal
Speech recognition targeting ums, uhs, and repeats
Beauty Retouching
Vision framework + Core Image for real-time face enhancement
Auto-Captions
On-device speech-to-text with styled subtitle generation
Object Tracking
Frame-to-frame tracking for motion graphics and effects
Background Removal
Segmentation masks at preview framerate, no green screen
All 11 engines share a common inference pipeline. Models load lazily, run on the Neural Engine when available, and fall back to GPU compute when the ANE is saturated. The user never sees a spinner. They see results.
Development Progress
LivCut tracks 124 features across six competitors: CapCut, InShot, VN, KineMaster, iMovie, and Final Cut Pro iPad. A core subset of features is implemented and tested; the majority are in progress. The focus is on shipping a solid foundation before claiming parity.
Where LivCut Is Designed to Exceed the Field
- ProRes Export ... professional codec support, no desktop required
- HDR Export (HLG) ... true high dynamic range output for modern displays
- No Watermark (Free Tier) ... your content, unsigned and unbranded
- Background Export ... encode while you keep editing, no lockout
- RGB Color Curves ... per-channel control that rivals desktop NLEs
- 3D Transitions ... Metal-accelerated volumetric effects between clips
- Ripple Edit ... timeline-aware editing that closes gaps automatically
Five Packages, One Engine
The codebase is organized as five Swift Package Manager libraries. Each package owns a single domain. Dependencies flow in one direction. Built with Swift 6.0, targeting iOS 17+ and macOS 14+.
Cross-module type definitions, protocols, and value types that every package imports.
AVFoundation + Metal pipeline. Timeline model, composition builder, real-time preview, hardware-accelerated export.
66 Metal compute kernels and Metal Performance Shaders. Color grading, transitions, particles, beauty filters.
CoreML model management, Apple Neural Engine dispatch, inference scheduling across all 11 AI engines.
Creator templates with parameterized timelines. Powers the marketplace and preset system.
What's Next
The Rust core. swift-bridge for iOS, UniFFI for Android. One rendering engine, two platforms, the same privacy guarantee. The compute kernels translate to platform-native GPU APIs. The AI models stay on-device regardless of operating system. Cross-platform without compromise.
Beyond the port: the Creator Bill of Rights. A public commitment to how user data is handled (it is not), how revenue is shared (80/20 in the creator's favor), and how the editor evolves (openly). The ambition is clear: DaVinci-level color science on a phone. Professional tools that respect the people who use them.
April 2026 . -JJ