R
ReflectIris
← Back

Mirror Metric v6

5-dimension sigmoid-mapped score — auto-computed from training — 100 = indistinguishable from a real 4K mirror

Live from 5 training machines · Best: reflect-ml-02 (exp_197_smirk_deform) · 32.3 dB Train / 0.3356 LPIPS
cart-01
14
24.8 dB Train · 0.425
CSIM 0.193 · 15.2K / 200K
cart-02
20
23.6 dB Train · 0.535
CSIM 0.654 · 15.2K / 200K
gpu
6
22.5 dB Train · 0.555
8.3K / 200K
ml-01
24
20.8 dB Train · 0.532
CSIM 0.552 · 24.5K / 200K
ml-02
31
32.3 dB Train · 0.336
CSIM 0.472 · 24.8K / 200K
Mirror Metric
0
/100+47 this period
Recognizable → next milestone at 40
peak 50avg 213.7d
Best MM (progress)
All runs
Milestones
0
Mirror Metric v6
Updated Apr 5, 9:32 AM CT
100 = patient cannot distinguish from a real 4K mirror
Legacy v4: 44
LPIPS
72
CSIM
361
Mouth
1884
MS-SSIM
112
Face PSNR
745
Dimension Balance
Balanced progress across all five dimensions leads to a higher MM
Component Breakdown — 5 Dimensions
LPIPS30%
72 / 30
0.3356 LPIPSTarget: 0.00

LPIPS perceptual distance — lower = more photorealistic. Sigmoid k=-20, midpoint=0.15. Inverted: 0.0 = perfect score.

Best render
Best ever: 0.045 LPIPS
exp_159_audit_baseline · cart · step 128650
→
CSIM25%
361 / 25
0.472 CSIM (live)Target: 0.95

ArcFace CSIM cosine similarity — does it look like the person? Sigmoid k=10, midpoint=0.65. 0.85+ = strong identity.

Mouth20%
1884 / 20
~29.3 dB (estimated from face PSNR)Target: 35

Mouth-region PSNR — dental quality is the product. Sigmoid k=0.30, midpoint=20 dB. The hardest region (teeth, tongue, mucosa).

MS-SSIM15%
112 / 15
0.7541 MS-SSIMTarget: 0.98

Multi-scale structural similarity — captures structural fidelity at multiple resolutions. Sigmoid k=20, midpoint=0.88.

Best render
Best ever: 0.903 SSIM
exp_184_cnn_only · ml-02 · step 48550
→
Face PSNR10%
745 / 10
32.3 dB face PSNRTarget: 36

Full-face masked PSNR — overall reconstruction quality baseline. Sigmoid k=0.25, midpoint=28 dB.

Best render
Best ever: 39.4 dB
exp_197_smirk_deform · ml-02 · step 24600
→
Mirror Metric v6 — How It Works

Mirror Metric v6 is a research-backed quality score with five dimensions, each mapped through a sigmoid function for smooth, non-linear scoring. Mouth quality and perceptual fidelity are weighted highest because the dental region and photorealism are what patients examine most closely in the mirror.

f(x) = 100 / (1 + exp(-k * (x - mid)))
face_psnr = 0.10 * f(face_psnr, k=0.25, mid=28) // sigmoid — full-face baseline [AUTO]
ms_ssim = 0.15 * f(ms_ssim, k=20, mid=0.88) // sigmoid — structural fidelity [AUTO]
lpips = 0.30 * f(lpips, k=-20, mid=0.15) // sigmoid inverted — perceptual [AUTO]
csim = 0.25 * f(csim, k=10, mid=0.65) // sigmoid — identity preservation [AUTO]
mouth = 0.20 * f(mouth_psnr, k=0.30, mid=20) // sigmoid — dental region quality [AUTO]
MM = round(face_psnr + ms_ssim + lpips + csim + mouth)

[AUTO] All five dimensions are computed automatically from validation_log.csv (validation runs every 2.5K steps) with fallback to training_log.csv estimates. Sigmoid mapping provides smooth scoring without floor/ceiling clamp artifacts.

LPIPS (30%) — Learned perceptual image patch similarity. Lower values mean more photorealistic renders. The sigmoid is inverted (k=-20) so that lower LPIPS scores map to higher MM contributions. Midpoint at 0.15.

CSIM (25%) — ArcFace cosine similarity. A patient looking in a mirror must see themselves. 0.85+ = strong identity preservation. Sigmoid k=10, midpoint 0.65.

Mouth PSNR (20%) — The dental region is the product. Mouth-region PSNR is computed on a masked crop centered on the oral cavity. This is the hardest region to reconstruct (teeth, tongue, mucosa) and the most important for patient acceptance. Sigmoid k=0.30, midpoint 20 dB.

MS-SSIM (15%) — Multi-scale structural similarity captures structural fidelity across multiple resolutions, complementing the pixel-level PSNR and perceptual LPIPS metrics. Sigmoid k=20, midpoint 0.88.

Face PSNR (10%) — Full-face masked PSNR provides a stable baseline for overall reconstruction quality. Not a primary quality gate (Blau & Michaeli 2018), but useful for tracking convergence and regression. Sigmoid k=0.25, midpoint 28 dB.

v5 → v6 changes: Switched from 4-dimension linear clamping to 5-dimension sigmoid mapping. Added MS-SSIM (15%). Replaced DISTS with LPIPS (30%). Reweighted: LPIPS 30%, CSIM 25%, Mouth 20%, MS-SSIM 15%, Face PSNR 10%. Sigmoid curves provide smoother scoring without hard floor/ceiling artifacts.
Sources: NeRSemble benchmark (CSIM), Arc2Avatar, Cafca SIGGRAPH Asia 2024, Apple 3DGS Optimization 2025.
Path to 100
0-20
Proof of Concept
Avatar renders in real-time. Mouth region barely recognizable. Not photorealistic.
20-40
RecognizableYOU ARE HERE
Patient can identify themselves. Dental region shows correct geometry. Still synthetic-looking.
40-60
Convincing
Photorealistic face. Mouth region renders teeth and gingiva accurately. Identity strong.
60-80
Mirror-Like
Full dental region indistinguishable. Expressions perfect. Only experts notice artifacts.
80-100
Indistinguishable
Patient cannot tell the difference from a real mirror. Production ready.