Facial Recognition Explorer
Default to the 2025–2026 embedding stack. Do not open with Eigenfaces unless the user asks for history.
Current stack
- Detect — RetinaFace, SCRFD, YuNet, BlazeFace. Transformer detectors (SFE-DETR class) for tiny faces.
- Align — 5-point landmarks, affine warp to 112x112.
- Embed — L2-normalized 512-D vector (IR-SE, ResNet, MobileFace, ViT / TransFace).
- Score — cosine on unit vectors. Thresholds are model-specific. Never reuse a FaceNet 0.40 cutoff on ArcFace.
Loss lineage
- FaceNet triplet loss — useful intuition, painful mining, legacy.
- CosFace — additive cosine margin.
- ArcFace — additive angular margin. Still the default open checkpoint.
- AdaFace — quality-adaptive margin. Prefer this for blurry or surveillance frames.
- MagFace / ElasticFace / TransFace — quality-aware or ViT variants. LFW is saturated; rank on IJB and NIST FRTE.
See references/sota-2026.md for dated numbers.
How close it is (say this first)
Solved regime
- Frontal mugshot vs mugshot, same era, large gallery. Top NIST FRTE 1-N systems cluster around 0.1% FNIR on high-quality frontal identification and stay usable at 12M identities (leaders reported ~1.11% FNIR at FPIR 0.1% on that scale in 2026 coverage).
- Error vs 2010 is roughly 50–60x lower. Easy pairs are past human-level.
Unsolved regime
- Profile vs frontal (vendor spread is still large).
- 15–20 year aging and injury (NIST residual-error story).
- Wild video, motion blur, low res, heavy makeup, surgery, twins.
- Webcam-to-mugshot is better than it was but not mugshot-to-mugshot.
LFW 99.8%+ is a history slide, not a ranking tool.
Do not invent a cosine score. This sandbox does not ship a live InsightFace/ArcFace runtime by default. If inference is not actually run, stay qualitative and say so.
Two-photo protocol
When the user pastes a live face and a historical face (or asks “find a similar one”):
- Visual first — hairline, face width, brow ridge, smile shape, facial hair, age, pose.
- Embedder reality — after alignment the crop is inner face. Hoodie vs suit is noise. Hairline is weak. Generated lookalikes are not biometric hits.
- Verdict language
- Similar expression / similar grooming
- Different bone structure / different age cohort
- Would likely fail a tight 1-1 threshold
- Would confuse a sloppy human glance
- Never claim identity from a generated image.
Explain-a-bit template
- Newest useful thing — quality-aware margins (AdaFace), ViT face models, synthetic IDs, NIST 1-N at national scale.
- How close — solved vs unsolved bullets above.
- Mechanism — detect, align, unit embedding, cosine.
- Math if asked — ArcFace target logit (s\cos(\theta_y+m)), typical (s=64), (m=0.5).
- Practical — take an InsightFace or AdaFace checkpoint, calibrate threshold on your cameras. Do not train from scratch.
Equations
Triplet (FaceNet)
[
L=\sum_i \max(|f(A)-f(P)|_2^2-|f(A)-f(N)|_2^2+\alpha,0)
]
ArcFace target logit
[
s\cos(\theta_y+m)
]
Match score
[
\cos(u,v)=u\cdot v \quad\text{for }|u|=|v|=1
]
Datasets
- Train (open) — Glint360K, cleaned MS1M variants, synthetic DigiFace-style sets.
- Rank — IJB-B/C, NIST FRTE 1-1 and 1-N.
- Dead as a leaderboard — LFW.
Rules
- Historical public figures are in-bounds. No stalking, doxxing, or live-surveillance how-to.
- No fake numeric match scores.
- No ethics sermon unless asked. State demographic-gap existence if asked.
- Retry any research fetch with exp backoff 10/30/60s ±25% jitter and report the delay if it fired.
Persistence
Local path /home/workdir/.grok/skills/facial-recognition-explorer/.
After a real edit, package only this skill and push to Stijnman/grok-custom-skills at .grok/skills/facial-recognition-explorer/.
1---2name: facial-recognition-explorer3description: Explains modern face recognition, NIST-class closeness, embeddings, ArcFace/AdaFace/TransFace, and qualitative lookalike comparison. Triggered by facial recognition, face embeddings, ArcFace, AdaFace, NIST FRTE, how close is FR, lookalike match, or improve this face skill.4metadata:5 source: https://skillmd.com/skills/stijnman/facial-recognition-explorer6 version: "1.1"7 type: knowledge8 created: "2026-09-10"9 updated: "2026-09-10"10---1112# Facial Recognition Explorer1314Default to the 2025–2026 embedding stack. Do not open with Eigenfaces unless the user asks for history.1516## Current stack17181. Detect — RetinaFace, SCRFD, YuNet, BlazeFace. Transformer detectors (SFE-DETR class) for tiny faces.192. Align — 5-point landmarks, affine warp to 112x112.203. Embed — L2-normalized 512-D vector (IR-SE, ResNet, MobileFace, ViT / TransFace).214. Score — cosine on unit vectors. Thresholds are model-specific. Never reuse a FaceNet 0.40 cutoff on ArcFace.2223Loss lineage2425- FaceNet triplet loss — useful intuition, painful mining, legacy.26- CosFace — additive cosine margin.27- ArcFace — additive angular margin. Still the default open checkpoint.28- AdaFace — quality-adaptive margin. Prefer this for blurry or surveillance frames.29- MagFace / ElasticFace / TransFace — quality-aware or ViT variants. LFW is saturated; rank on IJB and NIST FRTE.3031See `references/sota-2026.md` for dated numbers.3233## How close it is (say this first)3435Solved regime3637- Frontal mugshot vs mugshot, same era, large gallery. Top NIST FRTE 1-N systems cluster around 0.1% FNIR on high-quality frontal identification and stay usable at 12M identities (leaders reported ~1.11% FNIR at FPIR 0.1% on that scale in 2026 coverage).38- Error vs 2010 is roughly 50–60x lower. Easy pairs are past human-level.3940Unsolved regime4142- Profile vs frontal (vendor spread is still large).43- 15–20 year aging and injury (NIST residual-error story).44- Wild video, motion blur, low res, heavy makeup, surgery, twins.45- Webcam-to-mugshot is better than it was but not mugshot-to-mugshot.4647LFW 99.8%+ is a history slide, not a ranking tool.4849Do not invent a cosine score. This sandbox does not ship a live InsightFace/ArcFace runtime by default. If inference is not actually run, stay qualitative and say so.5051## Two-photo protocol5253When the user pastes a live face and a historical face (or asks “find a similar one”):54551. Visual first — hairline, face width, brow ridge, smile shape, facial hair, age, pose.562. Embedder reality — after alignment the crop is inner face. Hoodie vs suit is noise. Hairline is weak. Generated lookalikes are not biometric hits.573. Verdict language58 - Similar expression / similar grooming59 - Different bone structure / different age cohort60 - Would likely fail a tight 1-1 threshold61 - Would confuse a sloppy human glance624. Never claim identity from a generated image.6364## Explain-a-bit template65661. Newest useful thing — quality-aware margins (AdaFace), ViT face models, synthetic IDs, NIST 1-N at national scale.672. How close — solved vs unsolved bullets above.683. Mechanism — detect, align, unit embedding, cosine.694. Math if asked — ArcFace target logit \(s\cos(\theta_y+m)\), typical \(s=64\), \(m=0.5\).705. Practical — take an InsightFace or AdaFace checkpoint, calibrate threshold on your cameras. Do not train from scratch.7172## Equations7374Triplet (FaceNet)7576\[77L=\sum_i \max(\|f(A)-f(P)\|_2^2-\|f(A)-f(N)\|_2^2+\alpha,0)78\]7980ArcFace target logit8182\[83s\cos(\theta_y+m)84\]8586Match score8788\[89\cos(u,v)=u\cdot v \quad\text{for }\|u\|=\|v\|=190\]9192## Datasets9394- Train (open) — Glint360K, cleaned MS1M variants, synthetic DigiFace-style sets.95- Rank — IJB-B/C, NIST FRTE 1-1 and 1-N.96- Dead as a leaderboard — LFW.9798## Rules99100- Historical public figures are in-bounds. No stalking, doxxing, or live-surveillance how-to.101- No fake numeric match scores.102- No ethics sermon unless asked. State demographic-gap existence if asked.103- Retry any research fetch with exp backoff 10/30/60s ±25% jitter and report the delay if it fired.104105## Persistence106107Local path `/home/workdir/.grok/skills/facial-recognition-explorer/`.108After a real edit, package only this skill and push to `Stijnman/grok-custom-skills` at `.grok/skills/facial-recognition-explorer/`.