rvFACE

Rust + WebAssembly face recognition — a complete port of the Faceplugin Open-Source-Face-Recognition-SDK (Python/PyTorch) to Rust, running natively and in the browser on WebGPU or CPU, with a web UI.

Analyze: detection · 68 landmarks · pose	1:1 compare: score gauge · threshold-75 verdict

Pipeline

image ─► slim-320 SSD detector ─► 68-pt MobileFaceNet landmarks ─► head pose
                                        │
                                        ▼
                          eyes-level alignment (128×128)
                                        │
                                        ▼
                       embedding CNN ─► L2-normalized feature
                                        │
                                        ▼
                     similarity = (dot + 1) × 50   (match > 75)

Workspace

Path	What
crates/rvface-core	Framework-free pipeline math (priors, NMS, alignment, pose, similarity, image ops)
crates/rvface-models	Burn ports of the three CNNs (CPU: ndarray · WebGPU: wgpu)
crates/rvface-cli	Native CLI (rvface detect, rvface compare)
crates/rvface-wasm	Browser bindings (wasm-bindgen)
web/	Web UI (Vite + TS): upload/webcam, overlays, 1:1 compare, backend toggle
tools/	Python: weight conversion → safetensors, golden parity fixtures
docs/adrs/	Architecture decision records (start at 0001)

Quick start

# native
cd rvface
python3 tools/fetch_and_convert.py          # download + convert weights → models/
cargo run -p rvface-cli --release -- compare a.jpg b.png

# browser
cd web && npm install && npm run dev        # weights served from web/public/models/

Status

Complete. All three networks ported to Burn with PyTorch golden-parity green (max|Δ| ~1e-7 on real weights), the full pipeline reproduces the upstream demo verdict on its own test images (score 78.2 → same person), and the browser runs the identical engine (wasm, 1.42 MB gzipped, CPU with SIMD128 or WebGPU with automatic CPU fallback) — no mocks anywhere.

• 66 Rust tests: unit math, seven PyTorch parity fixtures, end-to-end on the upstream test images (validation strategy)
• Benchmarks: native analyze 176 ms, browser ~0.5 s (CPU; includes the 5× denormal-weight fix)
• See ADR-0003 (+ addendum) for the weight licensing story. Two weight files are properly licensed and ship with the repo + demo: the detector (MIT lineage) and the default embedder (foamliu/MobileFaceNet, Apache-2.0 — notices in models/LICENSES.md), so the web demo runs live face detection out of the box. The landmark checkpoint has no upstream LICENSE file (models/README.md) and is never redistributed: the tooling fetches it locally and the web demo collects that one file via a drop-zone to unlock landmarks/pose/compare. See also how to drop in the exact upstream IRN-50 embedder via --irn50.

License & responsible use

Code is MIT. Only properly licensed model weights are redistributed — the MIT-lineage detector and the Apache-2.0 foamliu embedder, with notices in models/LICENSES.md. The fetch tooling downloads the remaining third-party checkpoints locally (SHA-256-pinned); their licensing is documented per-file in models/README.md and ADR-0003 — review it before any commercial use.

This is face-recognition software, i.e. biometric processing. It runs entirely locally (no telemetry, no network calls at inference). It is intended for consent-based applications — authentication, personal photo tooling, research. Do not use it for surveillance, tracking, or identification of people who have not consented, and check the biometric-data laws that apply in your jurisdiction (e.g. GDPR Art. 9, BIPA) before deployment.