Face Verification on CA-LFW
96.12AccuracyFaceLiVTv2-M
Evaluation results
| Method | Links | |
|---|---|---|
| FaceLiVTv2-MYear=-, Param (M)=7.02, FLOPs=258M, Training Dataset=Glint360K, Model Scale=Lightweight, Experimental Setup=Original, Latency (ms) on iPhone 15 Pro=0.65(16.7%↓)2026.04 | 96.12 | |
| SwiftFaceFormer-L3Year='24, Param (M)=28.0, FLOPs=2.0G, Training Dataset=MS1MV3, Model Scale=Large-Scale, Experimental Setup=Original, Latency (ms) on iPhone 15 Pro=2.592026.04 | 96.03 | |
| FaceLiVTv2-LYear=-, Param (M)=8.52, FLOPs=309M, Training Dataset=Glint360K, Model Scale=Lightweight, Experimental Setup=Original, Latency (ms) on iPhone 15 Pro=0.71(14.0× ↓)2026.04 | 96 | |
| EdgeFace-SYear='24, Param (M)=3.6, FLOPs=306M, Training Dataset=Glint360K, Model Scale=Lightweight, Experimental Setup=Retrained from scratch, Latency (ms) on iPhone 15 Pro=9.892026.04 | 95.95 | |
| GhostFaceNetV1-1Year='23, Param (M)=4.1, FLOPs=216M, Training Dataset=MS1MV3, Model Scale=Lightweight, Experimental Setup=Original, Latency (ms) on iPhone 15 Pro=0.782026.04 | 95.93 | |
| FaceLiVTv2-SYear=-, Param (M)=4.62, FLOPs=179M, Training Dataset=Glint360K, Model Scale=Lightweight, Experimental Setup=Original, Latency (ms) on iPhone 15 Pro=0.54(30.8%↓)2026.04 | 95.93 | |
| SwiftFaceFormer-L1Year='24, Param (M)=11.8, FLOPs=805M, Training Dataset=MS1MV3, Model Scale=Lightweight, Experimental Setup=Original, Latency (ms) on iPhone 15 Pro=1.202026.04 | 95.8 | |
| SwiftFaceFormer-SYear='24, Param (M)=6.0, FLOPs=485M, Training Dataset=MS1MV3, Model Scale=Lightweight, Experimental Setup=Original, Latency (ms) on iPhone 15 Pro=0.652026.04 | 95.78 | |
| FaceLiVTv1-MYear='25, Param (M)=9.8, FLOPs=386M, Training Dataset=Glint360K, Model Scale=Lightweight, Experimental Setup=Original, Latency (ms) on iPhone 15 Pro=0.672026.04 | 95.76 | |
| GhostFaceNetV2-2Year='23, Param (M)=6.8, FLOPs=77M, Training Dataset=MS1MV3, Model Scale=Lightweight, Experimental Setup=Original, Latency (ms) on iPhone 15 Pro=0.672026.04 | 95.73 | |
| EdgeFace - SMPARAMS=3.65, MFLOPS=306.11, gamma=0.52023.07 | 95.71 | |
| EdgeFace-SYear='24, Param (M)=3.6, FLOPs=306M, Training Dataset=WebFace12M, Model Scale=Lightweight, Experimental Setup=Original, Latency (ms) on iPhone 15 Pro=9.892026.04 | 95.71 | |
| PocketNetM-128MPARAMS=1.68, MFLOPS=1099.022023.07 | 95.67 | |
| PocketNetM128Year='22, Param (M)=1.68, FLOPs=1.1G, Training Dataset=CASIA-WF, Model Scale=Large-Scale, Experimental Setup=Original, Latency (ms) on iPhone 15 Pro=0.982026.04 | 95.67 | |
| PocketNetM-256MPARAMS=1.75, MFLOPS=1099.152023.07 | 95.63 | |
| PocketNetM256Year='22, Param (M)=1.75, FLOPs=1.1G, Training Dataset=CASIA-WF, Model Scale=Large-Scale, Experimental Setup=Original, Latency (ms) on iPhone 15 Pro=0.982026.04 | 95.63 | |
| FaceLiVTv1-SYear='25, Param (M)=5.89, FLOPs=237M, Training Dataset=Glint360K, Model Scale=Lightweight, Experimental Setup=Original, Latency (ms) on iPhone 15 Pro=0.472026.04 | 95.63 | |
| GhostFaceNetV1-2Year='23, Param (M)=4.1, FLOPs=60M, Training Dataset=MS1MV3, Model Scale=Lightweight, Experimental Setup=Original, Latency (ms) on iPhone 15 Pro=0.602026.04 | 95.6 | |
| FaceLiVTv2-XSYear=-, Param (M)=2.9, FLOPs=90M, Training Dataset=Glint360K, Model Scale=Lightweight, Experimental Setup=Original, Latency (ms) on iPhone 15 Pro=0.43(35.8%↓)2026.04 | 95.58 | |
| KANFace-0.5Year='25, Param (M)=6.80, FLOPs=397M, Training Dataset=Glint360K, Model Scale=Lightweight, Experimental Setup=Retrained from scratch, Latency (ms) on iPhone 15 Pro=9.982026.04 | 95.52 | |
| PocketNetS-256MPARAMS=0.99, MFLOPS=587.242023.07 | 95.5 | |
| PocketNetS256Year='22, Param (M)=0.99, FLOPs=587M, Training Dataset=CASIA-WF, Model Scale=Lightweight, Experimental Setup=Original, Latency (ms) on iPhone 15 Pro=0.882026.04 | 95.5 | |
| PocketNetS-128MPARAMS=0.92, MFLOPS=587.112023.07 | 95.48 | |
| KANFace-0.5Year='25, Param (M)=6.80, FLOPs=397M, Training Dataset=WebFace12M, Model Scale=Lightweight, Experimental Setup=Original, Latency (ms) on iPhone 15 Pro=9.982026.04 | 95.48 | |
| PocketNetS128Year='22, Param (M)=0.92, FLOPs=587M, Training Dataset=CASIA-WF, Model Scale=Lightweight, Experimental Setup=Original, Latency (ms) on iPhone 15 Pro=0.882026.04 | 95.48 | |
| SwiftFaceFormer-XSYear='24, Param (M)=3.4, FLOPs=294M, Training Dataset=MS1MV3, Model Scale=Lightweight, Experimental Setup=Original, Latency (ms) on iPhone 15 Pro=0.592026.04 | 95.47 | |
| KANFace-0.6Year='25, Param (M)=4.74, FLOPs=240M, Training Dataset=WebFace12M, Model Scale=Lightweight, Experimental Setup=Original, Latency (ms) on iPhone 15 Pro=6.522026.04 | 95.32 | |
| EdgeFace - XSMPARAMS=1.77, MFLOPS=154, gamma=0.62023.07 | 95.28 | |
| EdgeFace-XSYear='24, Param (M)=1.77, FLOPs=154M, Training Dataset=WebFace12M, Model Scale=Lightweight, Experimental Setup=Original, Latency (ms) on iPhone 15 Pro=5.822026.04 | 95.28 | |
| MobileFaceNetMPARAMS=2.0, MFLOPS=9332023.07 | 95.2 | |
| MobileFaceNetYear='21, Param (M)=0.99, FLOPs=440M, Training Dataset=MS1MV2, Model Scale=Lightweight, Experimental Setup=Original, Latency (ms) on iPhone 15 Pro=0.772026.04 | 95.2 | |
| VarGFaceNetMPARAMS=5.0, MFLOPS=10222023.07 | 95.15 | |
| VarGFaceNetYear='19, Param (M)=5.0, FLOPs=1.0G, Training Dataset=MS1MV3, Model Scale=Large-Scale, Experimental Setup=Original, Latency (ms) on iPhone 15 Pro=0.832026.04 | 95.15 | |
| ShuffleFaceNet 1.5xMPARAMS=2.6, MFLOPS=577.52023.07 | 95.05 | |
| ShuffleFaceNet-1.5Year='21, Param (M)=2.6, FLOPs=577M, Training Dataset=MS1MV2, Model Scale=Lightweight, Experimental Setup=Original, Latency (ms) on iPhone 15 Pro=0.692026.04 | 95.05 | |
| WebFace-4M (real)Network=TinyFaR-A, Training=Classification2023.08 | 95.02 | |
| WebFace-4M (real)Network=TinyFaR-B, Training=Classification2023.08 | 94.73 | |
| SynthDistill (synthetic)Network=TinyFaR-A, Training=Knowledge Distillation2023.08 | 94.57 | |
| MobileFaceNetV1MPARAMS=3.4, MFLOPS=11002023.07 | 94.47 | |
| ElasticFace (CASIA-WebFace)Augmentation=false, Identities=10.5 K2023.08 | 93.93 | |
| IDPERTURBData Generation=Diffusion Model, DGMs Train Dataset=C-WF, #imgs (IDs x imgs/ID)=1.0M(20k × 50)2026.02 | 93.85 | |
| SynthDistill (synthetic)Network=TinyFaR-B, Training=Knowledge Distillation2023.08 | 93.78 | |
| C-WFData Generation=Authentic, #imgs (IDs x imgs/ID)=0.49M(∼10.5k × 47)2026.02 | 93.78 | |
| IDPERTURBData Generation=Diffusion Model, DGMs Train Dataset=C-WF, #imgs (IDs x imgs/ID)=0.5M(10k × 50)2026.02 | 93.5 | |
| Arc2FaceData Generation=Diffusion Model, DGMs Train Dataset=WF4M + FFHQ + CelebA, #imgs (IDs x imgs/ID)=1.2M(20k × 50 + 40k × 5)2026.02 | 93.33 | |
| UIFaceData Generation=Diffusion Model, DGMs Train Dataset=C-WF, #imgs (IDs x imgs/ID)=1.0M(20k × 50)2026.02 | 93.18 | |
| WebFace-4M (real)Network=TinyFaR-C, Training=Classification2023.08 | 93.08 | |
| CemiFaceData Generation=Diffusion Model, DGMs Train Dataset=C-WF, #imgs (IDs x imgs/ID)=1.0M(20k × 50)2026.02 | 93.01 | |
| DCFaceData Generation=Diffusion Model, DGMs Train Dataset=FFHQ + C-WF, #imgs (IDs x imgs/ID)=1.2M(20k × 50 + 40k × 5)2026.02 | 92.82 | |
| Arc2FaceData Generation=Diffusion Model, DGMs Train Dataset=WF4M + FFHQ + CelebA, #imgs (IDs x imgs/ID)=0.5M(10k × 50)2026.02 | 92.63 | |
| ProxylessFaceNASMPARAMS=3.2, MFLOPS=9002023.07 | 92.55 | |
| CemiFaceData Generation=Diffusion Model, DGMs Train Dataset=C-WF, #imgs (IDs x imgs/ID)=0.5M(10k × 50)2026.02 | 92.42 | |
| UIFaceData Generation=Diffusion Model, DGMs Train Dataset=C-WF, #imgs (IDs x imgs/ID)=0.5M(10k × 50)2026.02 | 92.25 | |
| SynthDistill (synthetic)Network=TinyFaR-C, Training=Knowledge Distillation2023.08 | 91.8 | |
| NegFaceDiffData Generation=Diffusion Model, DGMs Train Dataset=C-WF, #imgs (IDs x imgs/ID)=0.5M(10k × 50)2026.02 | 91.65 | |
| DCFaceTraining set=FFHQ+CASIA, Backbone=LResNet50-IR2024.09 | 91.6 | |
| DCFaceData Generation=Diffusion Model, DGMs Train Dataset=FFHQ + C-WF, #imgs (IDs x imgs/ID)=0.5M(10k × 50)2026.02 | 91.6 | |
| IDPERTURBData Generation=Diffusion Model, DGMs Train Dataset=FFHQ, #imgs (IDs x imgs/ID)=0.5M(10k × 50)2026.02 | 91.42 | |
| ID³Training set=CASIA, Backbone=LResNet50-IR2024.09 | 90.73 | |
| ID3Data Generation=Diffusion Model, DGMs Train Dataset=C-WF, #imgs (IDs x imgs/ID)=0.5M(10k × 50)2026.02 | 90.73 | |
| IDiff-Face CPD25 (Uniform) + RandAugmentAugmentation=true, Identities=10 K, Images per identity=50, Total training images=500 K2023.08 | 90.65 | |
| IDiff-FaceData Generation=Diffusion Model, DGMs Train Dataset=FFHQ, #imgs (IDs x imgs/ID)=0.5M(10k × 50)2026.02 | 90.65 | |
| NegFaceDiffData Generation=Diffusion Model, DGMs Train Dataset=FFHQ, #imgs (IDs x imgs/ID)=0.5M(10k × 50)2026.02 | 90.28 | |
| DCFace (synthetic)Network=TinyFaR-A, Training=Classification2023.08 | 90.08 | |
| DCFace (synthetic)Network=TinyFaR-B, Training=Classification2023.08 | 89.62 | |
| HyperFaceData Generation=Diffusion Model, DGMs Train Dataset=WF4M + FFHQ + CelebA, #imgs (IDs x imgs/ID)=0.5M(10k × 50)2026.02 | 89.4 | |
| ID³Training set=FFHQ, Backbone=LResNet50-IR2024.09 | 89.3 | |
| ID3Data Generation=Diffusion Model, DGMs Train Dataset=FFHQ, #imgs (IDs x imgs/ID)=0.5M(10k × 50)2026.02 | 89.3 | |
| IDiff-Face CPD50 (Two-Stage) + RandAugmentAugmentation=true, Identities=10 K, Images per identity=50, Total training images=500 K2023.08 | 88.52 | |
| DCFace (synthetic)Network=TinyFaR-C, Training=Classification2023.08 | 88.48 | |
| IDiff-FaceTraining set=FFHQ, Backbone=LResNet50-IR2024.09 | 86.32 | |
| SFace2+Data Generation=GAN-Based, DGMs Train Dataset=C-WF, #imgs (IDs x imgs/ID)=0.63M(10.5k × 60)2026.02 | 83.4 | |
| ExFaceGAN(Con)Data Generation=GAN-Based, DGMs Train Dataset=FFHQ, #imgs (IDs x imgs/ID)=0.5M(10k × 50)2026.02 | 82.98 | |
| DigiFaceData Generation=Digital Rendering, #imgs (IDs x imgs/ID)=1.2M(10k × 72 + 100k × 5)2026.02 | 82.55 | |
| Reinforcement-guided synthetic data generation frameworkData Generation=DM, Training Set=small-scale CASIA-WebFace [54] subset2026.04 | 81.68 | |
| IDnetData Generation=GAN-Based, DGMs Train Dataset=C-WF, #imgs (IDs x imgs/ID)=0.53M(10.5k × 50)2026.02 | 79.9 | |
| Authentic CASIA-Webface [54] subsetData Generation=Authentic, Training Set=small-scale CASIA-WebFace [54] subset2026.04 | 78.78 | |
| NegFaceDiffData Generation=DM, Training Set=small-scale CASIA-WebFace [54] subset2026.04 | 78.67 | |
| DigiFaceData Generation=Digital Rendering, #imgs (IDs x imgs/ID)=0.5M(10k × 50)2026.02 | 78.62 | |
| SFaceData Generation=GAN-Based, DGMs Train Dataset=C-WF, #imgs (IDs x imgs/ID)=0.63M(10.5k × 60)2026.02 | 77.93 | |
| USynthFaceData Generation=GAN-Based, DGMs Train Dataset=FFHQ, #imgs (IDs x imgs/ID)=0.4M(400k × 1)2026.02 | 77.05 | |
| SFaceTraining set=FFHQ, Backbone=LResNet50-IR2024.09 | 76.92 | |
| SFace (synthetic)Network=TinyFaR-C, Training=Classification2023.08 | 76.7 | |
| DCFaceData Generation=GAN, Training Set=small-scale CASIA-WebFace [54] subset2026.04 | 76.53 | |
| SFace (synthetic)Network=TinyFaR-B, Training=Classification2023.08 | 76.15 | |
| SFace (synthetic)Network=TinyFaR-A, Training=Classification2023.08 | 75.48 | |
| IDiff-FaceData Generation=DM, Training Set=small-scale CASIA-WebFace [54] subset2026.04 | 75.42 | |
| DigiFaceData Generation=Digital Rendering, #imgs (IDs x imgs/ID)=0.5M(10k × 50), Reproduced=true2026.02 | 75.3 | |
| SynFaceTraining set=FFHQ, Backbone=LResNet50-IR2024.09 | 74.73 | |
| SynFace (w/IM)Data Generation=GAN-Based, DGMs Train Dataset=FFHQ, #imgs (IDs x imgs/ID)=0.5M(10k × 50)2026.02 | 74.73 | |
| IDnetData Generation=GAN, Training Set=small-scale CASIA-WebFace [54] subset2026.04 | 72.67 | |
| SFace2Data Generation=GAN, Training Set=small-scale CASIA-WebFace [54] subset2026.04 | 72.35 | |
| DigiFace (synthetic)Network=TinyFaR-A, Training=Classification2023.08 | 71.65 | |
| ID-NetTraining set=FFHQ, Backbone=LResNet50-IR2024.09 | 71.5 | |
| SFaceData Generation=GAN, Training Set=small-scale CASIA-WebFace [54] subset2026.04 | 71.5 | |
| DigiFace (synthetic)Network=TinyFaR-B, Training=Classification2023.08 | 70.27 | |
| DigiFaceTraining set=FFHQ, Backbone=LResNet50-IR2024.09 | 69.23 | |
| DigiFace (synthetic)Network=TinyFaR-C, Training=Classification2023.08 | 69.18 |