faceflash 0.1.0

Fast face retrieval via PCA+ITQ binary quantization — up to 96x less memory than HNSW at equal recall

⚡ FaceFlash

FaceFlash searches 1M faces in 61 MB of RAM. HNSW needs 2.9 GB, USearch needs 2.5 GB, FAISS needs 1.9 GB — for the same 100% recall.

FaceFlash is a Rust face search engine with Python bindings, built on PCA+ITQ binary quantization — a learned hash that preserves identity information with zero recall loss and no separate training phase.

• 100% Recall@1, 48–96× less memory. On the MS1MV2 benchmark (44,291 identities), FaceFlash held 100% Recall@1 at every scale from 100K to 1M while using 48× less index memory than HNSWLIB at the default 512-bit config (~42× vs USearch, ~32× vs FAISS-Flat) — and 96× less at the 256-bit compact config (100% recall, ~2× single-query latency).
• Faster than HNSW at 100K. On the same benchmark, FaceFlash single-query latency was 0.30ms vs HNSWLIB's 0.60ms (2× faster). Batched throughput: 27,661 qps vs 5,813 qps (4.8× faster). Both at 100% recall.
• AVX-512 VPOPCNTDQ + NEON. Hand-written SIMD kernels process one 512-bit face code per instruction (~3× faster than scalar). Multi-core batched search measured at 10–17× throughput vs single-query serial on the same hardware.
• Zero-config indexing. Add faces, they're indexed — PCA fits automatically after 1,024 samples, no hyperparameter tuning, no rebuilds as the gallery grows.
• Pure local. No managed service, no data leaving your machine. Pair with any ArcFace model for a fully air-gapped face search stack.

from faceflash import FaceFlash

ff = FaceFlash()
ff.register_folder("employees/")   # bulk enroll
ff.save("my_index/")

result = ff.search("visitor.jpg")
# {"matches": [{"name": "Alice", "confidence": 0.92}], "search_time_ms": 0.4}

---

The Problem

Most face search libraries force a trade-off:

• HNSW is fast and accurate — but consumes 2.9 GB of RAM at 1M faces
• ScaNN / USearch are blazing fast — but drop to 94–99% recall
• FAISS-Flat is exact — but is 10× slower at scale

FaceFlash breaks this trade-off. It compresses each face into a 64-byte binary fingerprint using PCA + ITQ quantization, scans them with a single AVX-512 VPOPCNTDQ instruction, then re-ranks only the top candidates with exact cosine similarity. The result: exact accuracy at a fraction of the memory.

---

At a Glance

	FaceFlash	HNSWLIB	USearch	ScaNN	FAISS-Flat
Recall@1	100%	100%	99.5%	98.3%	100%
Memory @ 100K	3.05 MB	293 MB	254 MB	12 MB	195 MB
Memory @ 500K	15.3 MB	1,465 MB	1,270 MB	61 MB	977 MB
Memory @ 1M	30.5 MB	2,930 MB	2,539 MB	122 MB	1,953 MB
Latency @ 100K	0.30ms	0.60ms	0.17ms	0.10ms	4.90ms
Batched QPS @ 100K	27,661	5,813	137,264	—	—
Index build	Auto (PCA fit)	Build graph	Build graph	Partition	None

Tested on MS1MV2 (44,291 identities, 645,019 embeddings). Hardware: AMD EPYC 9355, 128 threads, AVX-512 active.

Memory = binary index only. Float vectors for cosine reranking are mmap'd from disk after save()/load() — only ~100 candidate rows are paged per query. See Limitations.

_{Index Memory at 500K Faces (256-bit compact config): FaceFlash (15 MB) vs HNSW (1,465 MB) — 96× less RAM at 100% recall}

---

Install

pip install "faceflash[cpu] @ git+https://github.com/raghavenderreddygrudhanti/faceflash.git"

# With benchmark dependencies
pip install "faceflash[cpu,benchmark] @ git+https://github.com/raghavenderreddygrudhanti/faceflash.git"

Requires a Rust toolchain — it compiles the AVX-512/NEON backend automatically and falls back to NumPy if unavailable.

---

Quick Start

from faceflash import FaceFlash

ff = FaceFlash()  # downloads ArcFace model (~166 MB) on first run

# Register individual faces
ff.register("Alice", "alice.jpg")
ff.register("Bob", "bob.jpg")

# Identify a face
result = ff.search("query.jpg")
# {"matches": [{"name": "Alice", "confidence": 0.92}], "search_time_ms": 0.4}

# Verify two faces are the same person
ff.verify("photo1.jpg", "photo2.jpg")
# {"match": True, "confidence": 0.87}

# Bulk enroll from folder (expects folder/person_name/photo.jpg)
ff.register_folder("employees/")
ff.save("my_index/")
ff.load("my_index/")

# Manage the index
len(ff)                  # how many faces are registered
"Alice" in ff            # is this person registered?
ff.names()               # list registered people
ff.remove("Alice")       # unregister a person (returns count removed)

For best accuracy, use pre-aligned 112x112 face crops. 5-point alignment (SCRFD/RetinaFace) adds +1.28 accuracy points over a basic center-crop.

---

Gallery Management

from faceflash import FaceFlash

ff = FaceFlash()
ff.register("Alice", "alice.jpg")
ff.register("Bob", "bob.jpg")
ff.register("Alice", "alice2.jpg")  # multiple photos per person

# Check gallery state
len(ff)              # 3 (total face entries)
ff.names()           # ["Alice", "Bob"]
"Alice" in ff        # True
"Charlie" in ff      # False

# Remove a person (GDPR / right-to-be-forgotten)
ff.remove("Bob")     # removes all entries for Bob
len(ff)              # 2
ff.names()           # ["Alice"]

# Monitor index stats
ff.stats()
# {'count': 2, 'pca_fitted': True, 'rust_backend': True,
#  'binary_memory_mb': 0.0, 'resident_memory_mb': 0.0, ...}

Batch Identification

Process many query faces at once (4.8x faster than one-by-one):

import numpy as np
from faceflash import FaceFlash

ff = FaceFlash()
ff.register_folder("gallery/")
ff.save("my_index/")

# Low-level batch search (for bulk dedup, watchlists, video frames)
embeddings = np.load("query_embeddings.npy")  # (N, 512) float32
results = ff.index.search_batch(embeddings, k=1)
# results[i] = [(name, similarity, index), ...]

# Example: find all matches above threshold
for i, matches in enumerate(results):
    if matches and matches[0][1] > 0.5:
        print(f"Query {i}: {matches[0][0]} (confidence {matches[0][1]:.2f})")

---

Is FaceFlash Right for You?

Scenario	Why FaceFlash wins
Edge / mobile / IoT	3–30 MB vs 293–2,930 MB for HNSW — fits in device RAM
Multi-tenant servers	100 galleries x 30 MB = 3 GB. HNSW: 100 x 1.5 GB = 150 GB
Batch dedup / watchlists	4.8x faster than HNSW batched at 100K; 1.9x at 500K
100% recall is non-negotiable	FaceFlash hits 100% at every scale; USearch drops to 94-99%
Budget / offline / air-gapped	Runs on Raspberry Pi, cheap VPS, phones — no GPU, no network
10K-500K face databases	The sweet spot: faster AND less memory than HNSW

When HNSW is the better choice:

• You need <0.3ms single-query latency at >500K faces and have gigabytes of RAM to spare
• Your database exceeds 2M faces (HNSW's O(log N) pulls clearly ahead)
• You need 100K+ batched QPS regardless of memory (USearch wins there)

---

How It Works

Each face is compressed into a 64-byte binary fingerprint:

1. ArcFace extracts a 512-dimensional float embedding
2. PCA aligns the quantization with the axes where identity varies most
3. ITQ rotates bits to maximize information per bit (balanced marginals)
4. AVX-512 VPOPCNTDQ scans all binary codes in a single instruction per face
5. Cosine rerank runs exact similarity on only the top ~100 candidates

This is why 512 bits is the fastest setting — the entire code fits in one AVX-512 register.

---

Benchmark Methodology

FaceFlash's recall, latency, throughput, and memory are measured. Competitor recall and latency are also measured; competitor memory is estimated from the standard vectors + index-structure overhead formula (e.g. HNSW ≈ 1.5× raw vectors).

	Details
Dataset	MS1MV2 — 645,019 ArcFace embeddings, 44,291 distinct identities
Embedding	ArcFace ONNX (w600k_r50), 512 dimensions, L2-normalized
Hardware	AMD EPYC 9355 (32 cores / 128 threads), AVX-512 VPOPCNTDQ enabled
Competitors	HNSWLIB 0.8+, FAISS 1.7+, USearch 2.x, ScaNN (latest)
Ground truth	Exact brute-force cosine argmax (FAISS-Flat)
Timing	time.perf_counter() per query, 10 warmup excluded
Recall metric	Recall@1 — fraction of queries where the true nearest neighbor is rank-1
Memory metric	FaceFlash: measured binary index size in RAM (floats mmap'd after save/load). Competitors: estimated (vectors + index-structure overhead)
Batched timing	Wall-clock for the full query batch / number of queries
Reproducibility	bash scripts/runpod_ms1m.sh reproduces all results end-to-end

All single-query rows are single-threaded. Batched rows use all available cores. Every benchmark script validates correctness before reporting speed.

---

Performance

Scale Summary (100K-1M)

_{Batched Throughput (QPS): FaceFlash 100K→1M — 4.8× faster than HNSW at 100K}

_{Single-Query Latency: FaceFlash 0.30ms vs HNSW 0.60ms at 100K — both at 100% recall}

_{Recall vs Memory Pareto: FaceFlash sits at the frontier — 100% recall, 30 MB}

Scale	Recall	Single-query	Batched QPS	Memory	vs HNSW memory
100K	100%	0.30ms (2× faster than HNSW)	27,661	6.1 MB	48× less
200K	100%	0.57ms (tied with HNSW)	19,930	12.2 MB	48× less
300K	100%	0.84ms	15,147	18.3 MB	48× less
500K	100%	1.45ms	10,337	30.5 MB	48× less
1M	100%	2.95ms	5,403	61 MB	48× less

Numbers are the default 512-bit config (fastest, 100% recall). The 256-bit compact config holds 100% recall at half the memory — 96× less than HNSW — trading ~2× single-query latency.

FaceFlash dominates up to 300K on every axis. At 500K-1M, HNSW edges ahead on single-query latency (O(log N) vs O(N)), but FaceFlash still wins on batched throughput and always uses 48–96× less memory.

1:N Identification - 44,290 Distinct People

The hardest test: one photo per person in the gallery, identify them from a different photo.

_{1:N Identification on 44,290 Identities: FaceFlash matches FAISS-Flat accuracy at 32× less memory}

Method	Rank-1 Accuracy	Memory
FAISS-Flat (exact ceiling)	95.8%	86.5 MB
FaceFlash (512b / 100c)	95.8%	2.70 MB
FaceFlash (512b / 300c)	95.8%	2.70 MB
FaceFlash (256b / 100c)	95.6%	1.35 MB

FaceFlash ties exact search using 32× less memory (512 float32 → 512 bits = exactly 32× compression). Binary quantization is lossless at 512 bits.

Detailed per-scale benchmark tables

100K Faces (6,939 identities)

Method	Recall@1	Latency	QPS	Memory
FaceFlash (batched)	100%	0.036ms	27,661	6.1 MB
FaceFlash (512b/200c)	100%	0.30ms	3,310	6.1 MB
FaceFlash (512b/100c)	100%	0.43ms	2,344	6.1 MB
HNSWLIB (ef=128)	100%	0.60ms	1,671	293 MB
HNSWLIB batched	100%	0.172ms	5,813	293 MB
USearch batched	99.5%	0.007ms	137,264	254 MB
USearch	99.5%	0.17ms	--	254 MB
ScaNN	98.3%	0.10ms	--	12 MB
FAISS-Flat (exact)	100%	4.90ms	204	195 MB

200K Faces (13,749 identities)

Method	Recall@1	Latency	QPS	Memory
FaceFlash (batched)	100%	0.050ms	19,930	12.2 MB
FaceFlash (512b/200c)	100%	0.57ms	1,751	12.2 MB
HNSWLIB (ef=128)	99.9%	0.65ms	1,531	586 MB
HNSWLIB batched	99.9%	0.378ms	2,646	586 MB
USearch batched	99.1%	0.008ms	121,660	508 MB
ScaNN	97.2%	0.19ms	--	24 MB

300K Faces (20,615 identities)

Method	Recall@1	Latency	QPS	Memory
FaceFlash (batched)	100%	0.066ms	15,147	18.3 MB
FaceFlash (512b/200c)	100%	0.84ms	1,187	18.3 MB
HNSWLIB (ef=128)	99.9%	0.66ms	1,510	879 MB
HNSWLIB batched	99.7%	0.269ms	3,715	879 MB
USearch batched	98.7%	0.014ms	73,383	762 MB
ScaNN	97.8%	0.28ms	--	37 MB

500K Faces (34,328 identities)

Method	Recall@1	Latency	QPS	Memory
FaceFlash (batched)	100%	0.097ms	10,337	30.5 MB
FaceFlash (512b/200c)	100%	1.45ms	692	30.5 MB
HNSWLIB (ef=128)	100%	0.71ms	1,416	1,465 MB
HNSWLIB batched	99.9%	0.179ms	5,577	1,465 MB
USearch batched	98.4%	0.013ms	76,150	1,270 MB
ScaNN	97.6%	0.45ms	--	61 MB

1M Faces (44,291 identities)

Method	Recall@1	Latency	QPS	Memory
FaceFlash (batched)	100%	0.185ms	5,403	61 MB
FaceFlash (512b/100c)	100%	2.92ms	342	61 MB
HNSWLIB (ef=128)	100%	0.66ms	1,523	2,930 MB
HNSWLIB batched	100%	0.178ms	5,621	2,930 MB
USearch batched	94.1%	0.013ms	77,266	2,539 MB
ScaNN	98.2%	0.86ms	--	122 MB

---

Tuning

Pick a config that matches your deployment:

Deployment	Config	Recall@1	Memory/face	Notes
Ultra-compact (mobile/IoT)	n_bits=128, n_candidates=500	99.4%	16 bytes	Minimum RAM
Balanced	n_bits=256, n_candidates=100	100%	32 bytes	Good default
Default (fastest)	n_bits=512, n_candidates=100	100%	64 bytes	One AVX-512 instruction
Edge (minimize disk reads)	n_bits=512, n_candidates=50	99.5%	64 bytes	--

ff = FaceFlash(n_bits=128, n_candidates=500)   # mobile/IoT
ff = FaceFlash(n_bits=256, n_candidates=100)   # balanced
ff = FaceFlash(n_bits=512, n_candidates=100)   # default: fastest

ff.search("query.jpg", n_candidates=200)       # per-query override

Speed up large indexes with clustering

_{IVF Clustering Speedup: 5–8× faster at 500K+ with configurable recall trade-off}

ff.index.build_clusters(n_probe=16)

Scale	n_probe	Recall@1	Latency	Speedup
100K	16	96.1%	0.12ms	2.6x
100K	32	98.5%	0.17ms	1.8x
500K	16	87.9%	0.31ms	5.0x
500K	32	92.8%	0.56ms	2.8x

Clustering is mainly useful at 500K+ where it delivers 5-8x speedup at ~88-93% recall.

---

Architecture

faceflash/                          # Python package
├── engine.py         ◀─ High-level API (register, search, verify)
├── detect.py         ◀─ Face detection (SCRFD + Haar fallback)
├── align.py          ◀─ 5-point alignment to ArcFace template
├── embed.py          ◀─ ArcFace ONNX embedding (512-dim, auto-download)
├── index.py          ◀─ Binary index + batched search
└── pca_quantize.py   ◀─ PCA+ITQ quantizer (the core algorithm)

rust/                               # Rust backend (PyO3 + Rayon)
├── src/lib.rs        ◀─ AVX-512 VPOPCNTDQ / NEON / scalar POPCNT
└── Cargo.toml

Why PCA+ITQ? ArcFace embeddings concentrate identity information along principal axes. PCA aligns quantization with those axes; ITQ rotates bits for balanced marginals. The result is lossless compression at 512 bits.

Why not HNSW internally? HNSW stores a graph on top of full float vectors — about 1.5x raw memory. FaceFlash stores 32–64 bytes per face. Float vectors are memory-mapped from disk and paged only for the top ~100 candidates per query. Trade-off: higher single-query latency at 500K+, but 48–96× less memory.

Why Rust + AVX-512? AVX-512 VPOPCNTDQ processes an entire 512-bit code in one instruction (~3× faster than scalar POPCNT). Combined with Rayon multi-core parallelism (and cache-blocked batching to keep the scan cache-friendly), batched search reaches 10-17× throughput versus single-query serial. Runtime-detected — no user configuration needed.

---

Limitations

• Single-query at 1M+ — O(N) linear scan; HNSW is 4.4x faster per single query at 1M. Batched path ties.
• Memory during build — holds all float vectors in RAM. The 48–96× savings apply after save() / load().
• AVX-512 VPOPCNTDQ — the ~3× kernel speedup requires Ice Lake / Zen 4+ / EPYC 9004+. Older CPUs fall back to scalar POPCNT automatically.
• Rerank I/O — pages ~100 float rows from disk per query. Invisible on NVMe; adds latency on slow storage.

---

Reproduce the Benchmarks

# Local (LFW + VGGFace2 100K)
python scripts/extract_lfw_embeddings.py
python benchmarks/bench_ann_comparison.py --scales 100K --queries 500

# RunPod (full suite)
export GITHUB_TOKEN=<token>
bash scripts/runpod_ms1m.sh   # FORCE_EXTRACT=1 for full 85K extraction

---

Roadmap

v0.1.0 (current)

• PCA+ITQ binary quantization + Rust search backend
• High-level API: register, search, verify
• Benchmarked against FAISS, HNSWLIB, USearch, ScaNN at 100K-1M
• 1:N identification on 44,290 distinct identities (MS1MV2)
• 5-point alignment via SCRFD/RetinaFace — 99.85% LFW accuracy

v0.2.0 (done)

• Prebuilt wheels (pip install faceflash)
• Full 85K-identity benchmark (76,872 identities extracted, 44,291 with sufficient data)
• On-device memory measurement (3.05 MB binary index @100K with 256b)

v0.3.0 (done)

• IVF coarse clustering (2.7-4.9x speedup at scale)
• AVX-512 VPOPCNTDQ — native 512-bit popcount (~3x faster than scalar)
• Batched search — 17x throughput at 500K-1M (multi-core + VPOPCNTDQ; cache-blocked)
• NEON kernels — ARM-optimized (vcntq_u8)

v1.0.0 — next

• Stable public API (no breaking changes)
• DiskANN comparison
• Mobile deployment (ONNX + CoreML)
• Streaming insertion (add faces without refitting PCA)

---

Contributing

# Dev setup (one command)
git clone https://github.com/raghavenderreddygrudhanti/faceflash.git
cd faceflash && python -m venv .venv && source .venv/bin/activate
pip install -e ".[cpu,benchmark]" && maturin develop --release
python -m pytest tests/  # 33 tests, ~17s

Open areas for contribution:

Area	Difficulty	Impact
DiskANN comparison	Medium	High — the one competitor missing
Mobile deployment (ONNX + CoreML)	Medium	High — iOS/Android face search
Streaming insertion (no PCA refit)	Hard	High — online learning
GPU batched search (CUDA)	Hard	Medium — 10M+ galleries
Raspberry Pi / Jetson benchmarks	Easy	Medium — edge credibility
WebAssembly build	Medium	Medium — browser face search

See CONTRIBUTING.md for coding guidelines.

---

Credits & References

FaceFlash does not introduce a new embedding model or hashing algorithm. It combines proven, published techniques into a CPU-efficient retrieval system — with hand-written Rust AVX-512/NEON kernels, cache-blocked batching, and rigorous, reproducible benchmarks. The contribution is the system (exact-recall face search in a megabyte-scale footprint) and the honest measurement of it, not the underlying math.

It builds directly on:

Component	Reference
Binary quantization (PCA + ITQ — the core method)	Gong & Lazebnik, "Iterative Quantization: A Procrustean Approach to Learning Binary Codes," CVPR 2011
Face embeddings	Deng, Guo, Xue & Zafeiriou, "ArcFace: Additive Angular Margin Loss for Deep Face Recognition," CVPR 2019 — weights from InsightFace
Detection & 5-point alignment	RetinaFace (CVPR 2020) / SCRFD (ICLR 2022), InsightFace
Dataset	Guo et al., "MS-Celeb-1M," ECCV 2016 (MS1MV2 = ArcFace's refined/cleaned version)
Verification benchmark	Huang et al., "Labeled Faces in the Wild (LFW)," UMass TR 2007
Baselines compared	FAISS (Johnson et al.), HNSW (Malkov & Yashunin, TPAMI 2018), ScaNN (Guo et al., ICML 2020), USearch

PCA dates to Pearson (1901) / Hotelling (1933); the Hamming distance to Hamming (1950); POPCNT / VPOPCNTDQ are Intel/AMD hardware instructions. FaceFlash's value is in how these are combined and implemented — not in inventing them.

---

License

MIT — see LICENSE.

If FaceFlash is useful to you, a star helps others find it.