Hugging Face's logo

iFurySt
/
fastface

Image Classification
ONNX
computer-vision
face-analysis
age-estimation
gender-classification
cpu
Model card Files
xet
 Community
fastface / model_runs.md
iFurySt's picture
iFurySt
Add FastFaceDetector release candidate
15240f0 verified
preview code
|
Raw
History Blame Contribute Delete
60.8 kB

Model Runs

This document records completed FastFace model runs and deployment measurements.

Current Completed Runs

Run Backbone Input Training Data Best Epoch Gender Balanced Acc Gender Acc Age MAE FP32 ONNX Static INT8 ONNX
efficientnet_v2_s_128_imdb_source_balanced_gender_priority_real_fairface_utkface EfficientNetV2-S 128 FairFace + UTKFace + IMDB-clean, source-limited train/val, gender-priority loss 38 0.98605 0.98605 4.84 79.5 MB 22.5 MB
mobilenetv3_large128_imdb_source_balanced_distill_gender_priority_efficientnet_v2_s_imdb_fairface_utkface MobileNetV3-Large 128 FairFace + UTKFace + IMDB-clean, source-limited train/val, gender-priority loss, IMDB V2-S gender teacher 24 0.97929 0.97955 5.71 13.4 MB 3.8 MB
mobilenetv3_small112_imdb_facecrop_real_fairface_utkface MobileNetV3-Small 112 FairFace + UTKFace + IMDB-clean, natural source mix 24 0.96994 0.97006 5.95 4.9 MB 1.5 MB
mobilenetv3_small112_imdb_source_balanced_distill_gender_priority_efficientnet_v2_s_imdb_fairface_utkface MobileNetV3-Small 112 FairFace + UTKFace + IMDB-clean, source-limited train/val, gender-priority loss, IMDB V2-S gender teacher 35 0.96800 0.96855 6.25 4.9 MB 1.5 MB
mobilenetv3_small112_imdb_pretrain_finetune_fairface_utkface MobileNetV3-Small 112 IMDB-expanded Small112 initialization, fine-tuned on FairFace + UTKFace 8 0.96784 0.96756 6.21 4.9 MB 1.5 MB
mobilenetv3_small112_imdb_source_balanced_facecrop_real_fairface_utkface MobileNetV3-Small 112 FairFace + UTKFace + IMDB-clean, source-limited checkpoint selection 25 0.96553 0.96633 6.28 4.9 MB 1.5 MB
efficientnet_v2_s_128_gender_priority_real_fairface_utkface EfficientNetV2-S 128 FairFace + UTKFace, gender-priority loss 34 0.94703 0.94678 5.28 79.5 MB 22.5 MB
efficientnet_v2_s_128_real_fairface_utkface EfficientNetV2-S 128 FairFace + UTKFace 22 0.94594 0.94574 4.87 79.5 MB 22.5 MB
efficientnet_b0_128_gender_priority_real_fairface_utkface EfficientNet-B0 128 FairFace + UTKFace, gender-priority loss 37 0.93910 0.93916 5.42 18.0 MB 5.4 MB
efficientnet_b0_128_real_fairface_utkface EfficientNet-B0 128 FairFace + UTKFace 34 0.93850 0.93834 5.26 18 MB 5.4 MB
resnet18_128_real_fairface_utkface ResNet18 128 FairFace + UTKFace 39 0.93467 0.93490 5.18 43.8 MB 11.1 MB
convnext_tiny_128_real_fairface_utkface ConvNeXt-Tiny 128 FairFace + UTKFace 32 0.93443 0.93408 5.13 107.9 MB 27.8 MB
mobilenetv3_large128_distill_gender_priority_efficientnet_v2_s_fairface_utkface MobileNetV3-Large 128 FairFace + UTKFace + EfficientNetV2-S gender teacher, gender-priority loss 28 0.93376 0.93408 5.61 13.4 MB 3.8 MB
mobilenetv3_large128_distill_gender_efficientnet_v2_s_fairface_utkface MobileNetV3-Large 128 FairFace + UTKFace + EfficientNetV2-S gender teacher 34 0.93230 0.93266 5.72 13.4 MB 3.8 MB
mobilenetv3_large128_distill_efficientnet_v2_s_fairface_utkface MobileNetV3-Large 128 FairFace + UTKFace + EfficientNetV2-S teacher 24 0.93092 0.93131 5.39 13.4 MB 3.8 MB
mobilenetv3_large128_distill_light_efficientnet_v2_s_fairface_utkface MobileNetV3-Large 128 FairFace + UTKFace + light EfficientNetV2-S teacher 33 0.92994 0.92996 5.41 13.4 MB 3.8 MB
mobilenetv3_real_fairface_utkface MobileNetV3-Large 128 FairFace + UTKFace 34 0.92965 0.92984 5.53 14 MB 3.9 MB
mobilenetv3_large112_distill_gender_efficientnet_v2_s_fairface_utkface MobileNetV3-Large 112 FairFace + UTKFace + EfficientNetV2-S gender teacher 30 0.92895 0.92892 5.72 13.4 MB 3.8 MB
mobilenetv3_large112_real_fairface_utkface MobileNetV3-Large 112 FairFace + UTKFace 31 0.92828 0.92835 5.68 14 MB 3.9 MB
mobilenetv3_large112_distill_efficientnet_v2_s_fairface_utkface MobileNetV3-Large 112 FairFace + UTKFace + EfficientNetV2-S teacher 23 0.92687 0.92660 5.46 13.4 MB 3.8 MB
mobilenetv3_small112_distill_efficientnet_b0_fairface_utkface MobileNetV3-Small 112 FairFace + UTKFace + EfficientNet-B0 teacher 25 0.91051 0.91061 5.90 4.9 MB 1.5 MB
mobilenetv3_small128_real_fairface_utkface MobileNetV3-Small 128 FairFace + UTKFace 27 0.91028 0.91068 5.95 4.9 MB 1.5 MB
mobilenetv3_small112_real_fairface_utkface MobileNetV3-Small 112 FairFace + UTKFace 39 0.90989 0.91049 6.03 4.9 MB 1.6 MB
mobilenetv3_small112_distill_large112_fairface_utkface MobileNetV3-Small 112 FairFace + UTKFace + Large112 teacher 31 0.90745 0.90765 5.97 4.9 MB 1.6 MB
mobilenetv3_small112_distill_efficientnet_v2_s_fairface_utkface MobileNetV3-Small 112 FairFace + UTKFace + EfficientNetV2-S teacher 27 0.90711 0.90694 5.89 4.9 MB 1.5 MB
mobilenetv3_small112_distill_gender_efficientnet_v2_s_fairface_utkface MobileNetV3-Small 112 FairFace + UTKFace + EfficientNetV2-S gender teacher 32 0.90677 0.90679 6.04 4.9 MB 1.5 MB
mobilenetv3_small112_distill_light_large112_fairface_utkface MobileNetV3-Small 112 FairFace + UTKFace + light Large112 teacher 28 0.90622 0.90646 6.02 4.9 MB 1.6 MB
mobilenetv3_small112_lagenda_real_fairface_utkface MobileNetV3-Small 112 FairFace + UTKFace + Lagenda-HF, no bbox crop 31 0.87875 0.87821 6.92 4.9 MB 1.6 MB
mobilenetv3_small112_lagenda_facecrop_real_fairface_utkface MobileNetV3-Small 112 FairFace + UTKFace + Lagenda-HF, bbox crop margin 0.2 34 0.90670 0.90691 6.12 4.9 MB 1.6 MB
swin_t_128_real_fairface_utkface Swin-T 128 FairFace + UTKFace 36 0.92206 0.92204 5.53 109.3 MB 29.9 MB

EfficientNetV2-S gender-priority remains the current public FairFace + UTKFace teacher candidate. The IMDB-expanded EfficientNetV2-S source-balanced run has the highest mixed aggregate score and is a useful IMDB-inclusive teacher/challenger, but it is not the CPU deployment default and does not replace the public-data teacher for FairFace robustness because its FairFace source slice is slightly lower. The IMDB-expanded MobileNetV3-Large distillation run is the strongest completed MobileNetV3 accuracy/CPU tradeoff candidate. The IMDB-expanded MobileNetV3-Small runs keep stronger CPU throughput, but none is promoted as the deployment default because the source-sliced FairFace metric remains modest. MobileNetV3-Small 112 FP32 remains the current high-throughput CPU family.

Current Active Runs

Run Backbone Input Training Data Status
None - - - No active training run is currently recorded after the MobileNetV3-Large IMDB source-balanced gender-distillation run finalized on 2026-07-30.

The EfficientNetV2-S source-balanced IMDB run is the strongest completed mixed-manifest teacher/challenger. It reaches full-manifest mixed evaluation gender balanced accuracy 0.98605, driven by imdb-clean at 0.99138, while FairFace lands at 0.94386. That FairFace slice is slightly lower than the earlier public-data EfficientNetV2-S gender-priority run's FairFace slice 0.94586, so this checkpoint is useful for IMDB-inclusive pseudo-labeling and age/gender coverage, not as a clean replacement for the public FairFace robustness teacher.

The MobileNetV3-Large source-balanced IMDB gender-distillation run is the strongest completed MobileNetV3 accuracy/CPU tradeoff candidate so far. It reaches mixed evaluation gender balanced accuracy 0.97929, with FairFace 0.92877, IMDB-clean 0.98548, and UTKFace 0.95017. It is much stronger than the Small112 distillation candidate on accuracy, but tuned FP32 batch-128 throughput is 4,477.7 img/s, roughly 41% of the Small112 distillation run's 10,837.8 img/s and 31% of the original Small112 throughput candidate's 14,483.4 img/s. Use it when gender accuracy matters more than maximum CPU throughput.

The MobileNetV3-Small source-balanced IMDB gender-distillation run is the strongest completed Small-family IMDB-inclusive gender candidate so far. It uses the completed source-balanced IMDB EfficientNetV2-S checkpoint as a low-weight gender-only teacher, reaches mixed evaluation gender balanced accuracy 0.96800, and improves source-sliced FairFace to 0.90562, above the earlier Small112 FairFace + UTKFace slice 0.90319 and the source-balanced non-distilled IMDB slice 0.89950. It is still not the pure throughput winner: tuned FP32 batch-128 throughput is 10,837.8 img/s versus 14,483.4 img/s for the original Small112 FP32 throughput candidate.

The natural IMDB-expanded MobileNetV3-Small run is a useful completed data-expansion result, but not the current deployment default. It reaches mixed evaluation gender balanced accuracy 0.96994, mostly driven by imdb-clean validation at 0.97854; FairFace lands at only 0.90004, below the FairFace + UTKFace small baseline. The source-balanced rerun caps IMDB-clean during training and checkpoint-selection validation, but final full-manifest evaluation still shows FairFace at only 0.89950 while IMDB-clean is 0.97347. Source caps alone therefore do not fix the public FairFace regression. The IMDB-pretrained then FairFace + UTKFace fine-tuned run improves FairFace only to 0.90024, still below the original non-IMDB Small112 FairFace slice at 0.90319, so it is also not promoted.

The tested EfficientNetV2-S gender-priority challenger is the best completed public-data teacher candidate so far. It raises evaluation gender balanced accuracy from regular V2-S 0.94594 to 0.94703, while age MAE worsens from 4.87 to 5.28. It is not a CPU deployment default: default FP32 batch-128 throughput is only 885.1 img/s, and default static INT8 batch-128 throughput is 841.0 img/s.

The tested EfficientNet-B0 gender-priority challenger is the best completed B0 run for gender. It raises evaluation gender balanced accuracy from 0.93850 to 0.93910, but age MAE worsens from 5.26 to 5.42, and tuned FP32 batch-128 throughput stays around 2,513.0 img/s. It is a useful middle accuracy candidate, not the CPU throughput default and not a replacement for EfficientNetV2-S gender-priority as teacher.

The tested MobileNetV3-Small distillation and 128-input runs are not default candidates. The EfficientNet-B0 teacher run slightly improves MobileNetV3-Small gender balanced accuracy from 0.90979 evaluation aggregate to 0.91051, and improves age MAE from 6.03 to 5.90, but tuned FP32 batch-128 throughput drops from 14,483.4 to 11,386.4 img/s. The stronger EfficientNetV2-S teacher reduces the student to 0.90711, likely because its logits over-constrain the small model on this limited data mix. Disabling age distillation for the EfficientNetV2-S Small112 student does not help; the gender-only run lands at 0.90677. The MobileNetV3-Small 128 run reaches 0.91028, but tuned FP32 batch-128 throughput is only 10,320.6 img/s. MobileNetV3-Large EfficientNetV2-S teacher runs can improve gender accuracy over some non-distilled baselines; the Large128 gender-priority variant is the strongest completed MobileNetV3-Large gender run, while the Large112 variants do not become defaults. The Large112 standard V2-S distillation run improves age MAE to 5.46 but drops gender balanced accuracy to 0.92687.

The tested Lagenda-HF expansion is also not a default candidate yet. Directly resizing Lagenda raw images failed because the images are full scenes rather than aligned face crops. Applying manifest bbox_face cropping fixes Lagenda validation dramatically, but the combined gender balanced accuracy still lands at 0.90670, below the FairFace + UTKFace small baseline.

The tested Swin-T 128 transformer challenger is also not a default candidate. It reaches only 0.92206 gender balanced accuracy, below ConvNeXt-Tiny, ResNet18, EfficientNet-B0, EfficientNetV2-S, and the best MobileNetV3-Large variants. Its default FP32 CPU batch-128 throughput is only 105.9 img/s, and static INT8 drops to 29.6 img/s. The full thread sweep was intentionally stopped after model_fp32_threads1 had run for more than 8 minutes without completing; the default CPU benchmark is already enough to rule it out for CPU deployment.

Age MAE is measured on the mixed public validation split. Treat it as directional only because FairFace contributes age-range labels, while UTKFace contributes exact ages.

Source-Sliced Validation

Run Source Count Gender Balanced Acc Gender Acc Age MAE Age CS@5
efficientnet_v2_s_128_imdb_source_balanced_gender_priority_real_fairface_utkface FairFace 10,954 0.94386 0.94376 5.32 0.57395
efficientnet_v2_s_128_imdb_source_balanced_gender_priority_real_fairface_utkface IMDB-clean 102,059 0.99138 0.99135 4.79 0.63523
efficientnet_v2_s_128_imdb_source_balanced_gender_priority_real_fairface_utkface UTKFace 2,425 0.95424 0.95423 4.57 0.66969
mobilenetv3_large128_imdb_source_balanced_distill_gender_priority_efficientnet_v2_s_imdb_fairface_utkface FairFace 10,954 0.92877 0.92879 5.94 0.54619
mobilenetv3_large128_imdb_source_balanced_distill_gender_priority_efficientnet_v2_s_imdb_fairface_utkface IMDB-clean 102,059 0.98548 0.98568 5.70 0.55329
mobilenetv3_large128_imdb_source_balanced_distill_gender_priority_efficientnet_v2_s_imdb_fairface_utkface UTKFace 2,425 0.95017 0.95052 4.96 0.63423
mobilenetv3_small112_imdb_source_balanced_distill_gender_priority_efficientnet_v2_s_imdb_fairface_utkface FairFace 10,954 0.90562 0.90588 6.48 0.50858
mobilenetv3_small112_imdb_source_balanced_distill_gender_priority_efficientnet_v2_s_imdb_fairface_utkface IMDB-clean 102,059 0.97542 0.97594 6.24 0.50751
mobilenetv3_small112_imdb_source_balanced_distill_gender_priority_efficientnet_v2_s_imdb_fairface_utkface UTKFace 2,425 0.94101 0.94103 5.37 0.61443
mobilenetv3_small112_imdb_facecrop_real_fairface_utkface FairFace 10,954 0.90004 0.90031 6.22 0.52894
mobilenetv3_small112_imdb_facecrop_real_fairface_utkface IMDB-clean 102,059 0.97854 0.97852 5.94 0.54274
mobilenetv3_small112_imdb_facecrop_real_fairface_utkface UTKFace 2,425 0.92936 0.92907 5.16 0.62887
mobilenetv3_small112_imdb_pretrain_finetune_fairface_utkface FairFace 10,954 0.90024 0.90004 6.15 0.53396
mobilenetv3_small112_imdb_pretrain_finetune_fairface_utkface IMDB-clean 102,059 0.97605 0.97565 6.24 0.52547
mobilenetv3_small112_imdb_pretrain_finetune_fairface_utkface UTKFace 2,425 0.93190 0.93196 5.11 0.64000
mobilenetv3_small112_imdb_source_balanced_facecrop_real_fairface_utkface FairFace 10,954 0.89950 0.90004 6.25 0.52145
mobilenetv3_small112_imdb_source_balanced_facecrop_real_fairface_utkface IMDB-clean 102,059 0.97347 0.97414 6.31 0.51614
mobilenetv3_small112_imdb_source_balanced_facecrop_real_fairface_utkface UTKFace 2,425 0.93607 0.93691 5.21 0.63381
efficientnet_v2_s_128_gender_priority_real_fairface_utkface FairFace 10,954 0.94586 0.94559 5.42 0.57066
efficientnet_v2_s_128_gender_priority_real_fairface_utkface UTKFace 2,425 0.95229 0.95216 4.63 0.66474
efficientnet_v2_s_128_real_fairface_utkface FairFace 10,954 0.94427 0.94404 4.97 0.57413
efficientnet_v2_s_128_real_fairface_utkface UTKFace 2,425 0.95351 0.95340 4.42 0.68330
efficientnet_b0_128_gender_priority_real_fairface_utkface FairFace 10,954 0.93648 0.93655 5.55 0.56637
efficientnet_b0_128_gender_priority_real_fairface_utkface UTKFace 2,425 0.95093 0.95093 4.82 0.64701
efficientnet_b0_128_real_fairface_utkface FairFace 10,954 0.93578 0.93555 5.39 0.55779
efficientnet_b0_128_real_fairface_utkface UTKFace 2,425 0.95079 0.95093 4.69 0.66186
resnet18_128_real_fairface_utkface FairFace 10,954 0.93229 0.93254 5.32 0.55770
resnet18_128_real_fairface_utkface UTKFace 2,425 0.94544 0.94557 4.58 0.66639
convnext_tiny_128_real_fairface_utkface FairFace 10,954 0.93166 0.93135 5.25 0.56436
convnext_tiny_128_real_fairface_utkface UTKFace 2,425 0.94692 0.94639 4.56 0.66722
swin_t_128_real_fairface_utkface FairFace 10,954 0.91832 0.91829 5.72 0.55624
swin_t_128_real_fairface_utkface UTKFace 2,425 0.93891 0.93897 4.64 0.66309
mobilenetv3_large128_distill_gender_priority_efficientnet_v2_s_fairface_utkface FairFace 10,954 0.93070 0.93108 5.79 0.55039
mobilenetv3_large128_distill_gender_priority_efficientnet_v2_s_fairface_utkface UTKFace 2,425 0.94753 0.94763 4.82 0.65237
mobilenetv3_large128_distill_gender_efficientnet_v2_s_fairface_utkface FairFace 10,954 0.92959 0.92998 5.90 0.52474
mobilenetv3_large128_distill_gender_efficientnet_v2_s_fairface_utkface UTKFace 2,425 0.94452 0.94474 4.91 0.64948
mobilenetv3_large128_distill_efficientnet_v2_s_fairface_utkface FairFace 10,954 0.92753 0.92797 5.53 0.53898
mobilenetv3_large128_distill_efficientnet_v2_s_fairface_utkface UTKFace 2,425 0.94622 0.94639 4.79 0.65031
mobilenetv3_large128_distill_light_efficientnet_v2_s_fairface_utkface FairFace 10,954 0.92638 0.92642 5.57 0.53779
mobilenetv3_large128_distill_light_efficientnet_v2_s_fairface_utkface UTKFace 2,425 0.94602 0.94598 4.69 0.65773
mobilenetv3_real_fairface_utkface FairFace 10,954 0.92724 0.92742 5.72 0.53396
mobilenetv3_real_fairface_utkface UTKFace 2,425 0.94077 0.94103 4.71 0.66722
mobilenetv3_large112_distill_gender_efficientnet_v2_s_fairface_utkface FairFace 10,954 0.92619 0.92615 5.91 0.52629
mobilenetv3_large112_distill_gender_efficientnet_v2_s_fairface_utkface UTKFace 2,425 0.94144 0.94144 4.86 0.63835
mobilenetv3_large112_real_fairface_utkface FairFace 10,954 0.92465 0.92478 5.88 0.52501
mobilenetv3_large112_real_fairface_utkface UTKFace 2,425 0.94450 0.94433 4.75 0.64742
mobilenetv3_large112_distill_efficientnet_v2_s_fairface_utkface FairFace 10,954 0.92410 0.92386 5.63 0.54619
mobilenetv3_large112_distill_efficientnet_v2_s_fairface_utkface UTKFace 2,425 0.93938 0.93897 4.70 0.66969
mobilenetv3_small112_distill_efficientnet_b0_fairface_utkface FairFace 10,954 0.90365 0.90387 6.08 0.53852
mobilenetv3_small112_distill_efficientnet_b0_fairface_utkface UTKFace 2,425 0.94147 0.94103 5.07 0.63876
mobilenetv3_small128_real_fairface_utkface FairFace 10,954 0.90442 0.90497 6.12 0.52921
mobilenetv3_small128_real_fairface_utkface UTKFace 2,425 0.93671 0.93649 5.20 0.62474
mobilenetv3_small112_real_fairface_utkface FairFace 10,954 0.90319 0.90387 6.23 0.52657
mobilenetv3_small112_real_fairface_utkface UTKFace 2,425 0.93960 0.93979 5.12 0.63134
mobilenetv3_small112_distill_efficientnet_v2_s_fairface_utkface FairFace 10,954 0.90058 0.90049 6.08 0.53661
mobilenetv3_small112_distill_efficientnet_v2_s_fairface_utkface UTKFace 2,425 0.93656 0.93608 5.04 0.63052
mobilenetv3_small112_distill_gender_efficientnet_v2_s_fairface_utkface FairFace 10,954 0.90002 0.90013 6.23 0.52885
mobilenetv3_small112_distill_gender_efficientnet_v2_s_fairface_utkface UTKFace 2,425 0.93729 0.93691 5.19 0.62474
mobilenetv3_small112_distill_large112_fairface_utkface FairFace 10,954 0.90257 0.90278 6.14 0.53232
mobilenetv3_small112_distill_large112_fairface_utkface UTKFace 2,425 0.92928 0.92948 5.19 0.63876
mobilenetv3_small112_distill_light_large112_fairface_utkface FairFace 10,954 0.90042 0.90068 6.21 0.52912
mobilenetv3_small112_distill_light_large112_fairface_utkface UTKFace 2,425 0.93308 0.93320 5.13 0.62309
mobilenetv3_small112_lagenda_real_fairface_utkface FairFace 10,954 0.89729 0.89702 6.32 0.52109
mobilenetv3_small112_lagenda_real_fairface_utkface Lagenda 1,282 0.62406 0.62324 15.28 0.25429
mobilenetv3_small112_lagenda_real_fairface_utkface UTKFace 2,425 0.92867 0.92825 5.20 0.62845
mobilenetv3_small112_lagenda_facecrop_real_fairface_utkface FairFace 10,954 0.90197 0.90232 6.19 0.53496
mobilenetv3_small112_lagenda_facecrop_real_fairface_utkface Lagenda 1,282 0.88850 0.88846 7.54 0.49454
mobilenetv3_small112_lagenda_facecrop_real_fairface_utkface UTKFace 2,425 0.93735 0.93732 5.08 0.64000

CPU Benchmark

ONNX Runtime CPUExecutionProvider on <remote-gpu-host>.

Default-thread results:

Run Model Batch 1 Batch 8 Batch 32 Batch 128
efficientnet_v2_s_128_imdb_source_balanced_gender_priority_real_fairface_utkface FP32 44.4 img/s 282.3 img/s 638.8 img/s 1,044.1 img/s
efficientnet_v2_s_128_imdb_source_balanced_gender_priority_real_fairface_utkface Static INT8 102.6 img/s 387.5 img/s 634.7 img/s 990.5 img/s
mobilenetv3_large128_imdb_source_balanced_distill_gender_priority_efficientnet_v2_s_imdb_fairface_utkface FP32 164.0 img/s 938.1 img/s 2,705.2 img/s 4,401.1 img/s
mobilenetv3_large128_imdb_source_balanced_distill_gender_priority_efficientnet_v2_s_imdb_fairface_utkface Static INT8 385.1 img/s 693.3 img/s 963.1 img/s 1,081.7 img/s
mobilenetv3_small112_imdb_source_balanced_distill_gender_priority_efficientnet_v2_s_imdb_fairface_utkface FP32 215.4 img/s 1,042.9 img/s 4,332.8 img/s 9,540.9 img/s
mobilenetv3_small112_imdb_source_balanced_distill_gender_priority_efficientnet_v2_s_imdb_fairface_utkface Static INT8 768.7 img/s 1,298.1 img/s 1,882.6 img/s 2,352.7 img/s
mobilenetv3_small112_imdb_facecrop_real_fairface_utkface FP32 215.3 img/s 1,057.9 img/s 4,357.1 img/s 9,523.0 img/s
mobilenetv3_small112_imdb_facecrop_real_fairface_utkface Static INT8 797.7 img/s 1,302.4 img/s 1,912.4 img/s 2,252.2 img/s
mobilenetv3_small112_imdb_pretrain_finetune_fairface_utkface FP32 212.5 img/s 1,060.8 img/s 4,417.5 img/s 9,440.9 img/s
mobilenetv3_small112_imdb_pretrain_finetune_fairface_utkface Static INT8 767.1 img/s 1,301.6 img/s 1,892.4 img/s 2,228.4 img/s
mobilenetv3_small112_imdb_source_balanced_facecrop_real_fairface_utkface FP32 215.6 img/s 1,060.9 img/s 4,417.4 img/s 9,505.4 img/s
mobilenetv3_small112_imdb_source_balanced_facecrop_real_fairface_utkface Static INT8 772.8 img/s 1,295.9 img/s 1,881.2 img/s 2,249.1 img/s
efficientnet_v2_s_128_gender_priority_real_fairface_utkface FP32 45.4 img/s 240.3 img/s 536.1 img/s 885.1 img/s
efficientnet_v2_s_128_gender_priority_real_fairface_utkface Static INT8 101.2 img/s 381.7 img/s 637.9 img/s 841.0 img/s
efficientnet_v2_s_128_real_fairface_utkface FP32 42.5 img/s 269.7 img/s 556.6 img/s 1,062.4 img/s
efficientnet_v2_s_128_real_fairface_utkface Static INT8 103.3 img/s 386.4 img/s 637.1 img/s 962.8 img/s
efficientnet_b0_128_gender_priority_real_fairface_utkface FP32 115.9 img/s 650.5 img/s 1,494.2 img/s 2,402.9 img/s
efficientnet_b0_128_gender_priority_real_fairface_utkface Static INT8 205.0 img/s 557.3 img/s 890.7 img/s 1,140.0 img/s
efficientnet_b0_128_real_fairface_utkface FP32 115.8 img/s 667.3 img/s 1,487.4 img/s 2,424.5 img/s
efficientnet_b0_128_real_fairface_utkface Static INT8 205.9 img/s 556.0 img/s 896.2 img/s 1,049.5 img/s
resnet18_128_real_fairface_utkface FP32 557.8 img/s 1,957.0 img/s 3,037.8 img/s 4,082.1 img/s
resnet18_128_real_fairface_utkface Static INT8 607.0 img/s 1,882.2 img/s 2,406.6 img/s 2,679.1 img/s
convnext_tiny_128_real_fairface_utkface FP32 78.6 img/s 210.8 img/s 264.1 img/s 307.9 img/s
convnext_tiny_128_real_fairface_utkface Static INT8 74.4 img/s 151.2 img/s 175.2 img/s 219.7 img/s
swin_t_128_real_fairface_utkface FP32 37.5 img/s 81.2 img/s 106.8 img/s 105.9 img/s
swin_t_128_real_fairface_utkface Static INT8 21.8 img/s 29.2 img/s 30.1 img/s 29.6 img/s
mobilenetv3_large128_distill_gender_priority_efficientnet_v2_s_fairface_utkface FP32 162.3 img/s 927.9 img/s 2,681.4 img/s 4,416.7 img/s
mobilenetv3_large128_distill_gender_priority_efficientnet_v2_s_fairface_utkface Static INT8 448.4 img/s 935.6 img/s 1,092.4 img/s 1,341.2 img/s
mobilenetv3_large128_distill_gender_efficientnet_v2_s_fairface_utkface FP32 164.5 img/s 937.1 img/s 2,952.4 img/s 4,275.3 img/s
mobilenetv3_large128_distill_gender_efficientnet_v2_s_fairface_utkface Static INT8 431.0 img/s 806.3 img/s 1,200.4 img/s 1,362.1 img/s
mobilenetv3_large128_distill_efficientnet_v2_s_fairface_utkface FP32 163.8 img/s 911.4 img/s 2,631.8 img/s 4,214.4 img/s
mobilenetv3_large128_distill_efficientnet_v2_s_fairface_utkface Static INT8 381.8 img/s 691.3 img/s 969.6 img/s 1,102.6 img/s
mobilenetv3_large128_distill_light_efficientnet_v2_s_fairface_utkface FP32 165.1 img/s 934.4 img/s 2,695.8 img/s 4,286.1 img/s
mobilenetv3_large128_distill_light_efficientnet_v2_s_fairface_utkface Static INT8 383.0 img/s 691.6 img/s 963.9 img/s 1,164.2 img/s
mobilenetv3_real_fairface_utkface FP32 196.4 img/s 1,131.2 img/s 2,986.3 img/s 4,633.1 img/s
mobilenetv3_real_fairface_utkface Static INT8 438.5 img/s 952.5 img/s 1,315.3 img/s 1,350.2 img/s
mobilenetv3_large112_distill_gender_efficientnet_v2_s_fairface_utkface FP32 207.7 img/s 1,222.6 img/s 3,313.0 img/s 5,251.1 img/s
mobilenetv3_large112_distill_gender_efficientnet_v2_s_fairface_utkface Static INT8 437.2 img/s 783.8 img/s 1,461.9 img/s 1,287.4 img/s
mobilenetv3_large112_real_fairface_utkface FP32 217.6 img/s 1,179.6 img/s 3,106.5 img/s 5,139.9 img/s
mobilenetv3_large112_real_fairface_utkface Static INT8 509.9 img/s 966.7 img/s 1,351.5 img/s 1,629.2 img/s
mobilenetv3_large112_distill_efficientnet_v2_s_fairface_utkface FP32 168.8 img/s 946.4 img/s 3,285.2 img/s 4,980.5 img/s
mobilenetv3_large112_distill_efficientnet_v2_s_fairface_utkface Static INT8 433.8 img/s 777.1 img/s 1,111.8 img/s 1,292.4 img/s
mobilenetv3_small112_distill_efficientnet_b0_fairface_utkface FP32 212.3 img/s 1,039.3 img/s 4,342.1 img/s 9,479.8 img/s
mobilenetv3_small112_distill_efficientnet_b0_fairface_utkface Static INT8 764.6 img/s 1,294.7 img/s 2,017.5 img/s 2,440.6 img/s
mobilenetv3_small128_real_fairface_utkface FP32 207.7 img/s 1,082.5 img/s 4,676.0 img/s 8,432.3 img/s
mobilenetv3_small128_real_fairface_utkface Static INT8 673.1 img/s 1,168.2 img/s 1,636.6 img/s 1,944.7 img/s
mobilenetv3_small112_real_fairface_utkface FP32 255.5 img/s 1,369.4 img/s 5,539.2 img/s 10,286.6 img/s
mobilenetv3_small112_real_fairface_utkface Static INT8 897.1 img/s 1,677.0 img/s 2,778.8 img/s 3,097.7 img/s
mobilenetv3_small112_distill_efficientnet_v2_s_fairface_utkface FP32 214.6 img/s 1,057.4 img/s 4,363.6 img/s 9,521.9 img/s
mobilenetv3_small112_distill_efficientnet_v2_s_fairface_utkface Static INT8 912.9 img/s 1,570.8 img/s 2,402.8 img/s 2,983.9 img/s
mobilenetv3_small112_distill_gender_efficientnet_v2_s_fairface_utkface FP32 214.7 img/s 1,058.1 img/s 4,363.8 img/s 9,425.5 img/s
mobilenetv3_small112_distill_gender_efficientnet_v2_s_fairface_utkface Static INT8 766.6 img/s 1,302.3 img/s 1,878.3 img/s 2,244.0 img/s
mobilenetv3_small112_distill_large112_fairface_utkface FP32 267.7 img/s 1,338.8 img/s 5,144.0 img/s 10,003.3 img/s
mobilenetv3_small112_distill_large112_fairface_utkface Static INT8 1,200.1 img/s 1,955.7 img/s 2,261.2 img/s 2,273.2 img/s
mobilenetv3_small112_distill_light_large112_fairface_utkface FP32 268.6 img/s 1,471.4 img/s 5,407.8 img/s 10,254.8 img/s
mobilenetv3_small112_distill_light_large112_fairface_utkface Static INT8 920.3 img/s 1,595.9 img/s 2,472.9 img/s 3,177.1 img/s
mobilenetv3_small112_lagenda_real_fairface_utkface FP32 268.6 img/s 1,470.3 img/s 5,480.0 img/s 9,530.8 img/s
mobilenetv3_small112_lagenda_real_fairface_utkface Static INT8 915.7 img/s 1,593.8 img/s 2,490.9 img/s 2,499.2 img/s
mobilenetv3_small112_lagenda_facecrop_real_fairface_utkface FP32 217.2 img/s 1,063.2 img/s 4,436.6 img/s 9,496.8 img/s
mobilenetv3_small112_lagenda_facecrop_real_fairface_utkface Static INT8 893.4 img/s 1,554.8 img/s 2,422.3 img/s 3,052.0 img/s

Best results from intra-op thread sweep with inter_op_num_threads=1 and sequential execution:

Run Model Batch 1 Batch 8 Batch 32 Batch 128
mobilenetv3_large128_imdb_source_balanced_distill_gender_priority_efficientnet_v2_s_imdb_fairface_utkface FP32 527.0 img/s @ 8 threads 1,859.5 img/s @ 16 threads 3,452.2 img/s @ 28 threads 4,477.7 img/s @ 28 threads
mobilenetv3_large128_imdb_source_balanced_distill_gender_priority_efficientnet_v2_s_imdb_fairface_utkface Static INT8 742.5 img/s @ 8 threads 1,345.4 img/s @ 28 threads 1,527.4 img/s @ 28 threads 1,517.0 img/s @ 16 threads
mobilenetv3_small112_imdb_source_balanced_distill_gender_priority_efficientnet_v2_s_imdb_fairface_utkface FP32 1,353.7 img/s @ 2 threads 2,873.5 img/s @ 8 threads 8,063.9 img/s @ 8 threads 10,837.8 img/s @ 28 threads
mobilenetv3_small112_imdb_source_balanced_distill_gender_priority_efficientnet_v2_s_imdb_fairface_utkface Static INT8 1,436.0 img/s @ 8 threads 2,777.1 img/s @ 8 threads 3,446.4 img/s @ 28 threads 3,598.3 img/s @ 28 threads
mobilenetv3_small112_imdb_facecrop_real_fairface_utkface FP32 1,413.5 img/s @ 4 threads 3,360.9 img/s @ 4 threads 7,904.4 img/s @ 28 threads 12,748.3 img/s @ 28 threads
mobilenetv3_small112_imdb_facecrop_real_fairface_utkface Static INT8 1,382.9 img/s @ 4 threads 2,546.3 img/s @ 4 threads 3,152.7 img/s @ 8 threads 3,332.3 img/s @ 8 threads
mobilenetv3_small112_imdb_pretrain_finetune_fairface_utkface FP32 1,303.7 img/s @ 1 thread 2,824.2 img/s @ 28 threads 9,196.2 img/s @ 28 threads 13,050.6 img/s @ 28 threads
mobilenetv3_small112_imdb_pretrain_finetune_fairface_utkface Static INT8 1,418.7 img/s @ 4 threads 2,525.5 img/s @ 4 threads 3,100.2 img/s @ 28 threads 3,278.9 img/s @ 16 threads
mobilenetv3_small112_imdb_source_balanced_facecrop_real_fairface_utkface FP32 1,314.7 img/s @ 1 thread 3,116.5 img/s @ 16 threads 8,626.6 img/s @ 16 threads 12,579.9 img/s @ 28 threads
mobilenetv3_small112_imdb_source_balanced_facecrop_real_fairface_utkface Static INT8 1,352.0 img/s @ 8 threads 2,480.1 img/s @ 8 threads 3,100.8 img/s @ 8 threads 3,251.8 img/s @ 16 threads
efficientnet_v2_s_128_real_fairface_utkface FP32 125.1 img/s @ 28 threads 427.0 img/s @ 28 threads 760.1 img/s @ 28 threads 1,000.6 img/s @ 28 threads
efficientnet_v2_s_128_real_fairface_utkface Static INT8 181.8 img/s @ 16 threads 565.5 img/s @ 16 threads 799.5 img/s @ 28 threads 1,051.8 img/s @ 56 threads
efficientnet_b0_128_gender_priority_real_fairface_utkface FP32 500.8 img/s @ 8 threads 1,374.3 img/s @ 28 threads 2,013.7 img/s @ 28 threads 2,513.0 img/s @ 28 threads
efficientnet_b0_128_gender_priority_real_fairface_utkface Static INT8 373.7 img/s @ 8 threads 985.7 img/s @ 28 threads 1,295.1 img/s @ 28 threads 1,509.3 img/s @ 28 threads
efficientnet_b0_128_real_fairface_utkface FP32 515.7 img/s @ 8 threads 1,433.8 img/s @ 16 threads 1,973.2 img/s @ 28 threads 2,549.7 img/s @ 56 threads
efficientnet_b0_128_real_fairface_utkface Static INT8 389.3 img/s @ 8 threads 832.4 img/s @ 16 threads 1,326.9 img/s @ 28 threads 1,533.6 img/s @ 28 threads
resnet18_128_real_fairface_utkface FP32 936.7 img/s @ 16 threads 1,760.1 img/s @ 28 threads 2,284.2 img/s @ 112 threads 3,129.4 img/s @ 56 threads
resnet18_128_real_fairface_utkface Static INT8 1,156.1 img/s @ 28 threads 3,323.3 img/s @ 28 threads 4,084.8 img/s @ 28 threads 4,163.2 img/s @ 28 threads
convnext_tiny_128_real_fairface_utkface FP32 152.7 img/s @ 16 threads 237.3 img/s @ 28 threads 277.8 img/s @ 28 threads 298.4 img/s @ 28 threads
convnext_tiny_128_real_fairface_utkface Static INT8 128.7 img/s @ 16 threads 195.2 img/s @ 28 threads 231.3 img/s @ 28 threads 228.3 img/s @ 28 threads
mobilenetv3_large128_distill_gender_priority_efficientnet_v2_s_fairface_utkface FP32 747.8 img/s @ 8 threads 2,218.3 img/s @ 8 threads 3,787.6 img/s @ 28 threads 4,376.9 img/s @ 28 threads
mobilenetv3_large128_distill_gender_priority_efficientnet_v2_s_fairface_utkface Static INT8 578.4 img/s @ 8 threads 1,102.1 img/s @ 16 threads 1,366.0 img/s @ 16 threads 1,505.9 img/s @ 8 threads
mobilenetv3_large128_distill_gender_efficientnet_v2_s_fairface_utkface FP32 697.5 img/s @ 4 threads 1,878.6 img/s @ 8 threads 3,964.9 img/s @ 28 threads 4,365.8 img/s @ 28 threads
mobilenetv3_large128_distill_gender_efficientnet_v2_s_fairface_utkface Static INT8 611.5 img/s @ 28 threads 1,444.1 img/s @ 8 threads 1,353.3 img/s @ 8 threads 1,489.7 img/s @ 16 threads
mobilenetv3_large128_distill_efficientnet_v2_s_fairface_utkface FP32 737.6 img/s @ 4 threads 1,941.7 img/s @ 16 threads 3,359.0 img/s @ 16 threads 4,555.4 img/s @ 28 threads
mobilenetv3_large128_distill_efficientnet_v2_s_fairface_utkface Static INT8 610.6 img/s @ 2 threads 1,038.1 img/s @ 8 threads 1,364.7 img/s @ 16 threads 1,515.1 img/s @ 16 threads
mobilenetv3_large128_distill_light_efficientnet_v2_s_fairface_utkface FP32 770.3 img/s @ 8 threads 2,517.4 img/s @ 16 threads 3,955.7 img/s @ 16 threads 4,500.8 img/s @ 28 threads
mobilenetv3_large128_distill_light_efficientnet_v2_s_fairface_utkface Static INT8 705.1 img/s @ 4 threads 1,380.8 img/s @ 16 threads 1,624.0 img/s @ 28 threads 1,468.2 img/s @ 16 threads
mobilenetv3_real_fairface_utkface FP32 620.2 img/s @ 16 threads 2,138.1 img/s @ 28 threads 4,600.4 img/s @ 28 threads 4,570.7 img/s @ 28 threads
mobilenetv3_real_fairface_utkface Static INT8 577.7 img/s @ 8 threads 1,086.9 img/s @ 8 threads 1,358.4 img/s @ 8 threads 1,463.2 img/s @ 16 threads
mobilenetv3_large112_distill_gender_efficientnet_v2_s_fairface_utkface FP32 856.2 img/s @ 8 threads 2,502.6 img/s @ 16 threads 4,585.6 img/s @ 16 threads 5,345.9 img/s @ 28 threads
mobilenetv3_large112_distill_gender_efficientnet_v2_s_fairface_utkface Static INT8 766.7 img/s @ 28 threads 1,572.5 img/s @ 28 threads 1,919.5 img/s @ 28 threads 1,839.8 img/s @ 28 threads
mobilenetv3_large112_real_fairface_utkface FP32 598.2 img/s @ 8 threads 2,050.4 img/s @ 8 threads 3,989.5 img/s @ 16 threads 5,418.1 img/s @ 56 threads
mobilenetv3_large112_real_fairface_utkface Static INT8 844.3 img/s @ 8 threads 1,715.2 img/s @ 8 threads 2,172.9 img/s @ 8 threads 1,883.1 img/s @ 8 threads
mobilenetv3_large112_distill_efficientnet_v2_s_fairface_utkface FP32 823.5 img/s @ 4 threads 2,672.7 img/s @ 16 threads 4,487.6 img/s @ 16 threads 5,342.3 img/s @ 28 threads
mobilenetv3_large112_distill_efficientnet_v2_s_fairface_utkface Static INT8 794.7 img/s @ 16 threads 1,314.1 img/s @ 16 threads 1,608.7 img/s @ 8 threads 1,778.7 img/s @ 16 threads
mobilenetv3_small112_distill_efficientnet_b0_fairface_utkface FP32 1,460.2 img/s @ 4 threads 3,326.9 img/s @ 4 threads 9,130.1 img/s @ 16 threads 11,386.4 img/s @ 28 threads
mobilenetv3_small112_distill_efficientnet_b0_fairface_utkface Static INT8 1,408.0 img/s @ 4 threads 2,811.0 img/s @ 16 threads 3,615.8 img/s @ 16 threads 3,483.2 img/s @ 16 threads
mobilenetv3_small128_real_fairface_utkface FP32 1,258.2 img/s @ 8 threads 3,554.0 img/s @ 8 threads 8,392.6 img/s @ 28 threads 10,320.6 img/s @ 28 threads
mobilenetv3_small128_real_fairface_utkface Static INT8 1,274.2 img/s @ 8 threads 2,442.0 img/s @ 8 threads 2,453.3 img/s @ 28 threads 2,767.3 img/s @ 28 threads
mobilenetv3_small112_real_fairface_utkface FP32 1,468.3 img/s @ 4 threads 3,461.0 img/s @ 4 threads 11,273.4 img/s @ 16 threads 14,483.4 img/s @ 16 threads
mobilenetv3_small112_real_fairface_utkface Static INT8 1,323.6 img/s @ 28 threads 2,347.9 img/s @ 28 threads 3,431.4 img/s @ 28 threads 3,118.0 img/s @ 8 threads
mobilenetv3_small112_distill_efficientnet_v2_s_fairface_utkface FP32 1,367.3 img/s @ 8 threads 3,607.2 img/s @ 8 threads 11,321.9 img/s @ 16 threads 12,153.4 img/s @ 28 threads
mobilenetv3_small112_distill_efficientnet_v2_s_fairface_utkface Static INT8 1,310.1 img/s @ 2 threads 2,408.9 img/s @ 28 threads 2,900.3 img/s @ 16 threads 3,325.8 img/s @ 16 threads
mobilenetv3_small112_distill_gender_efficientnet_v2_s_fairface_utkface FP32 1,351.2 img/s @ 2 threads 2,587.2 img/s @ 8 threads 6,961.0 img/s @ 16 threads 10,611.2 img/s @ 28 threads
mobilenetv3_small112_distill_gender_efficientnet_v2_s_fairface_utkface Static INT8 1,327.2 img/s @ 28 threads 2,364.5 img/s @ 28 threads 3,223.9 img/s @ 28 threads 3,239.5 img/s @ 16 threads
mobilenetv3_small112_distill_large112_fairface_utkface FP32 1,407.2 img/s @ 4 threads 3,321.3 img/s @ 4 threads 8,434.1 img/s @ 16 threads 12,005.3 img/s @ 28 threads
mobilenetv3_small112_distill_large112_fairface_utkface Static INT8 1,399.7 img/s @ 8 threads 2,719.1 img/s @ 8 threads 3,934.3 img/s @ 8 threads 3,040.7 img/s @ 16 threads
mobilenetv3_small112_distill_light_large112_fairface_utkface FP32 1,297.2 img/s @ 1 thread 2,668.3 img/s @ 8 threads 7,212.9 img/s @ 16 threads 11,308.5 img/s @ 28 threads
mobilenetv3_small112_distill_light_large112_fairface_utkface Static INT8 1,368.6 img/s @ 16 threads 2,803.4 img/s @ 16 threads 4,097.9 img/s @ 16 threads 3,145.1 img/s @ 8 threads
mobilenetv3_small112_lagenda_real_fairface_utkface FP32 1,331.7 img/s @ 1 thread 3,413.7 img/s @ 16 threads 9,625.2 img/s @ 16 threads 12,300.6 img/s @ 28 threads
mobilenetv3_small112_lagenda_real_fairface_utkface Static INT8 1,315.2 img/s @ 28 threads 2,539.5 img/s @ 4 threads 2,816.6 img/s @ 8 threads 3,271.9 img/s @ 8 threads
mobilenetv3_small112_lagenda_facecrop_real_fairface_utkface FP32 1,290.2 img/s @ 1 thread 2,898.1 img/s @ 8 threads 7,727.2 img/s @ 28 threads 12,675.4 img/s @ 28 threads
mobilenetv3_small112_lagenda_facecrop_real_fairface_utkface Static INT8 1,379.9 img/s @ 16 threads 2,690.8 img/s @ 16 threads 3,032.9 img/s @ 28 threads 3,191.3 img/s @ 16 threads

Thread tuning changes the deployment recommendation:

  • MobileNetV3-Small 112 FP32 is the current CPU throughput winner for all measured batch sizes.
  • The source-balanced IMDB gender-distilled MobileNetV3-Large 128 FP32 run is the strongest completed MobileNetV3 accuracy candidate: FairFace 0.92877, IMDB-clean 0.98548, UTKFace 0.95017, and tuned FP32 batch-128 throughput 4,477.7 img/s.
  • The natural IMDB-expanded MobileNetV3-Small 112 FP32 run keeps the same high-throughput deployment class and reaches 12,748.3 img/s at tuned batch 128, but it is not promoted because FairFace source-sliced validation regresses.
  • The source-balanced IMDB-expanded MobileNetV3-Small 112 FP32 run also stays in the high-throughput class and reaches 12,579.9 img/s at tuned batch 128, but it does not improve FairFace source robustness; FairFace gender balanced accuracy is 0.89950.
  • The IMDB-pretrained FairFace + UTKFace fine-tune reaches 13,050.6 img/s at tuned FP32 batch 128, but still does not restore FairFace source robustness; FairFace gender balanced accuracy is 0.90024.
  • The IMDB V2-S gender-distilled, source-balanced MobileNetV3-Small 112 FP32 run is the strongest completed Small-family IMDB-inclusive gender candidate so far. It reaches FairFace 0.90562, IMDB-clean 0.97542, UTKFace 0.94101, and tuned FP32 batch-128 throughput 10,837.8 img/s. It improves source robustness versus prior Small IMDB runs but does not beat the original Small112 throughput result.
  • The IMDB-expanded EfficientNetV2-S 128 source-balanced gender-priority run is the strongest mixed-manifest teacher/challenger so far at aggregate 0.98605, but default FP32 batch-128 throughput is only 1,044.1 img/s and the FairFace slice 0.94386 does not beat the earlier public-data V2-S gender-priority FairFace slice 0.94586.
  • EfficientNetV2-S 128 gender-priority is the current public-validation accuracy winner and teacher candidate, but it is much slower than MobileNetV3-Small and MobileNetV3-Large on CPU.
  • The full thread sweeps for EfficientNetV2-S 128 gender-priority and the IMDB-expanded EfficientNetV2-S 128 source-balanced gender-priority run were intentionally skipped after default CPU benchmarks and model-card generation, because the regular V2-S run already has a same-architecture full sweep and these slow challengers are not CPU deployment defaults.
  • EfficientNet-B0 128 FP32 remains a useful smaller teacher/accuracy baseline, but it is no longer the top public-validation accuracy run. The gender-priority B0 variant is the best B0 gender result so far (0.93910), at the cost of age MAE and with essentially the same CPU throughput class.
  • ResNet18 128 lands between MobileNetV3-Large and EfficientNet-B0 on gender accuracy. Its INT8 throughput is better than EfficientNet-B0 but still far behind MobileNetV3-Small FP32 at high batch.
  • ConvNeXt-Tiny 128 is not a useful deployment or teacher default on the current data mix. It does not beat EfficientNetV2-S, EfficientNet-B0, or ResNet18 on gender accuracy, and its CPU throughput is far below every MobileNet/ResNet deployment candidate.
  • Swin-T 128 is not a useful deployment or teacher default on the current data mix. It lands at only 0.92206 gender balanced accuracy and is much slower than ConvNeXt-Tiny on CPU.
  • MobileNetV3-Large 128 with gender-priority EfficientNetV2-S distillation is the best completed MobileNetV3-Large gender-accuracy variant so far (0.93376), while age MAE lands at 5.61 and FP32 throughput remains Large-class.
  • The earlier Large128 gender-only EfficientNetV2-S distillation run remains useful but is superseded for the gender-first target: it reaches 0.93230 gender balanced accuracy and 5.72 age MAE.
  • The standard EfficientNetV2-S distillation run is the better balanced Large 128 distillation run so far: it reaches 0.93092 gender balanced accuracy and 5.39 age MAE, while the lower-weight variant lands at 0.92994.
  • The EfficientNetV2-S gender-priority teacher challenger is the best public-data gender run so far: it reaches 0.94703 evaluation gender balanced accuracy, improving over regular V2-S 0.94594, while age MAE regresses from 4.87 to 5.28.
  • The natural IMDB-expanded MobileNetV3-Small run is not a clean default despite its mixed aggregate 0.96994. It overfits toward the IMDB-clean celebrity domain: FairFace gender balanced accuracy is only 0.90004, while IMDB-clean is 0.97854.
  • The source-balanced IMDB-expanded MobileNetV3-Small rerun is also not a clean default. Its checkpoint-selection validation used capped IMDB-clean rows, but final full-manifest evaluation still lands at FairFace 0.89950, IMDB-clean 0.97347, and UTKFace 0.93607.
  • The IMDB-pretrained FairFace + UTKFace fine-tune is also not a default. It tests whether IMDB initialization plus public-domain fine-tuning can recover the FairFace slice, but lands at FairFace 0.90024, still below the original non-IMDB Small112 FairFace slice 0.90319.
  • MobileNetV3-Large 128 FP32 remains the better balanced accuracy/CPU-throughput family than EfficientNet-B0.
  • MobileNetV3-Large 112 with EfficientNetV2-S gender-only distillation slightly improves Large112 gender balanced accuracy (0.92895 versus 0.92828) and default FP32 batch-128 throughput (5,251.1 versus 5,139.9 img/s), but tuned FP32 batch-128 throughput remains effectively similar and it does not beat Large128 gender-priority accuracy.
  • MobileNetV3-Large 112 with standard EfficientNetV2-S distillation is not useful for the gender-first target. It improves age MAE to 5.46, but drops gender balanced accuracy to 0.92687.
  • MobileNetV3-Large 112 FP32 is not a clear default. It keeps most of the large accuracy and improves batch-128 throughput, but does not beat large-128 at batch 1, 8, or 32 in the tuned sweep.
  • The EfficientNet-B0 teacher distillation run is the best MobileNetV3-Small accuracy variant so far, but the improvement is tiny and FP32 throughput trails the non-distilled small model after tuning.
  • The EfficientNetV2-S teacher distillation run is not a default. Despite the stronger teacher, the MobileNetV3-Small student drops to 0.90711 gender balanced accuracy.
  • The EfficientNetV2-S gender-only distillation run is also not a MobileNetV3-Small default. It lands at 0.90677, below both the regular V2-S distillation run and the non-distilled Small112 baseline.
  • The lower-weight EfficientNetV2-S distillation run for MobileNetV3-Large 128 is not useful; it lands at 0.92994, below the standard-weight V2-S distillation run.
  • MobileNetV3-Small 128 FP32 does not justify replacing MobileNetV3-Small 112 FP32: it gains only about 0.00039 gender balanced accuracy and loses high-batch tuned throughput.
  • The earlier MobileNetV3-Large teacher distillation runs are not defaults. They improve some INT8 measurements, but gender accuracy drops and FP32 throughput trails the non-distilled small model after tuning.
  • Lagenda-HF requires bbox cropping. Without it, Lagenda validation collapses; with it, Lagenda becomes usable but still does not improve the primary mixed validation metric.
  • Static INT8 is not the default path. It improves some low-latency cases but loses to FP32 for sustained throughput after thread tuning and additional variant tuning.

Gender Disagreement Review

Use scripts/compare-gender-models.sh for a fixed comparison between the current converged candidates, public FairFace-ONNX, and MiVOLO. The default review set is FairFace validation, UTKFace validation, and a seed-stable IMDB-clean validation sample capped to the FairFace validation size.

The original 2026-07-31 public-FairFace review output is outputs/analysis/gender-comparison-current. The MiVOLO-inclusive review output is outputs/analysis/gender-comparison-mivolo-current. It compares:

  • our_large128_imdb_distill
  • our_small112_imdb_distill
  • teacher_v2s_imdb
  • public_fairface_onnx
  • mivolo_imdb_face

On 24,333 selected validation rows, aggregate gender balanced accuracy is 0.96638 for teacher_v2s_imdb, 0.95618 for our_large128_imdb_distill, 0.94658 for public_fairface_onnx, 0.94461 for mivolo_imdb_face, and 0.94059 for our_small112_imdb_distill. The MiVOLO baseline is the official face-only IMDB-clean age+gender checkpoint, staged at third_party/mivolo/weights/model_imdb_face_4.22_99.38.pth.tar.

MiVOLO source-sliced gender balanced accuracy:

Dataset Gender Balanced Acc Gender Acc
FairFace 0.89753 0.89858
IMDB-clean 0.99614 0.99607
UTKFace 0.92825 0.92701

Pairwise against the main CPU candidate, our_large128_imdb_distill and MiVOLO disagree on 1,532 rows. FastFace Large is correct by public labels on 901 of those rows, and MiVOLO is correct on 631. MiVOLO and public FairFace-ONNX disagree on 1,753 rows.

Important review files:

outputs/analysis/gender-comparison-mivolo-current/summary.json
outputs/analysis/gender-comparison-mivolo-current/predictions.jsonl
outputs/analysis/gender-comparison-mivolo-current/gender_disagreements.csv
outputs/analysis/gender-comparison-mivolo-current/gender_disagreements_top.jpg
outputs/analysis/gender-comparison-mivolo-current/focused/public_vs_our_large.csv
outputs/analysis/gender-comparison-mivolo-current/focused/teacher_vs_our_large.csv
outputs/analysis/gender-comparison-mivolo-current/focused/mivolo_vs_our_large.csv
outputs/analysis/gender-comparison-mivolo-current/focused/mivolo_vs_public_fairface.csv

For manual labeling, run scripts/build-manual-gender-review.sh after the comparison. The default manual review now uses only focused/public_vs_our_large.csv, because rows where every model agrees are not useful for deciding whether FastFace or the public baseline is wrong. The current output is outputs/analysis/manual-public-gender-review-current, with manual_gender_review.xlsx, manual_gender_review.csv, images/, and manual_gender_review_package.zip. The workbook embeds one face-crop thumbnail per row, replaces image_path with image, adds explicit our_large_gender and public_fairface_gender columns, and leaves manual_gender blank for human labels. The current public-vs-our review has 1,301 rows.

INT8 Variant Tuning

Additional static INT8 tuning was run on the current throughput candidate, mobilenetv3_small112_real_fairface_utkface, with 1,024 calibration samples and the same thread sweep as the main benchmark.

Best result per model and batch:

Variant Batch 1 Batch 8 Batch 32 Batch 128
FP32 tuned baseline 1,468.3 img/s @ 4 threads 3,461.0 img/s @ 4 threads 11,273.4 img/s @ 16 threads 14,483.4 img/s @ 16 threads
Existing static INT8 QDQ U8/S8 per-channel 1,323.6 img/s @ 28 threads 2,347.9 img/s @ 28 threads 3,431.4 img/s @ 28 threads 3,118.0 img/s @ 8 threads
QDQ U8/S8 per-channel 1,403.5 img/s @ 4 threads 2,810.4 img/s @ 8 threads 2,991.1 img/s @ 8 threads 3,255.6 img/s @ 16 threads
QDQ U8/S8 tensor-wise 1,279.5 img/s @ 8 threads 2,498.0 img/s @ 4 threads 3,004.9 img/s @ 28 threads 3,301.3 img/s @ 8 threads
QOperator U8/S8 tensor-wise 1,093.9 img/s @ 1 thread 1,741.6 img/s @ 4 threads 1,707.3 img/s @ 4 threads 1,602.9 img/s @ 4 threads
QDQ S8/S8 per-channel 762.3 img/s @ 1 thread 1,051.0 img/s @ 1 thread 1,179.8 img/s @ 4 threads 1,244.5 img/s @ 4 threads
Preprocessed QDQ U8/S8 per-channel 1,266.9 img/s @ 2 threads 2,243.7 img/s @ 4 threads 3,071.3 img/s @ 8 threads 3,519.5 img/s @ 28 threads
Preprocessed QDQ U8/S8 tensor-wise 1,308.4 img/s @ 16 threads 2,443.7 img/s @ 4 threads 2,883.6 img/s @ 8 threads 3,476.6 img/s @ 8 threads
Preprocessed QOperator U8/S8 tensor-wise 1,422.4 img/s @ 8 threads 2,564.3 img/s @ 16 threads 3,758.9 img/s @ 8 threads 3,993.4 img/s @ 8 threads
Preprocessed QDQ S8/S8 per-channel 402.8 img/s @ 4 threads 677.1 img/s @ 28 threads 834.2 img/s @ 16 threads 920.7 img/s @ 16 threads

QOperator per-channel failed in ONNX Runtime quantization for this graph with a per-channel weight broadcast error. Preprocessing improved the QOperator tensor-wise high-batch result, but it still reached only 3,993.4 img/s at batch 128, far below the FP32 baseline. Keep MobileNetV3-Small 112 FP32 as the CPU throughput default.

Artifact Paths 

runs/mobilenetv3_real_fairface_utkface/
  best.pt
  model_fp32.onnx
  model_int8_static.onnx
  evaluation_val.json
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  cpu-thread-sweep-summary.json
  model_card.md

runs/mobilenetv3_large128_distill_efficientnet_v2_s_fairface_utkface/
  best.pt
  model_fp32.onnx
  model_int8_static.onnx
  evaluation_val.json
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  cpu-thread-sweep-summary.json
  model_card.md

runs/mobilenetv3_large128_distill_gender_efficientnet_v2_s_fairface_utkface/
  best.pt
  model_fp32.onnx
  model_int8_static.onnx
  evaluation_val.json
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  cpu-thread-sweep-summary.json
  model_card.md

runs/mobilenetv3_large128_distill_gender_priority_efficientnet_v2_s_fairface_utkface/
  best.pt
  model_fp32.onnx
  model_int8_static.onnx
  evaluation_val.json
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  cpu-thread-sweep-summary.json
  model_card.md

runs/mobilenetv3_large128_distill_light_efficientnet_v2_s_fairface_utkface/
  best.pt
  model_fp32.onnx
  model_int8_static.onnx
  evaluation_val.json
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  cpu-thread-sweep-summary.json
  model_card.md

runs/mobilenetv3_large112_real_fairface_utkface/
  best.pt
  model_fp32.onnx
  model_int8_static.onnx
  evaluation_val.json
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  cpu-thread-sweep-summary.json
  model_card.md

runs/mobilenetv3_large112_distill_gender_efficientnet_v2_s_fairface_utkface/
  best.pt
  model_fp32.onnx
  model_int8_static.onnx
  evaluation_val.json
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  cpu-thread-sweep-summary.json
  model_card.md

runs/mobilenetv3_large112_distill_efficientnet_v2_s_fairface_utkface/
  best.pt
  model_fp32.onnx
  model_int8_static.onnx
  evaluation_val.json
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  cpu-thread-sweep-summary.json
  model_card.md

runs/mobilenetv3_small112_real_fairface_utkface/
  best.pt
  model_fp32.onnx
  model_int8_static.onnx
  evaluation_val.json
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  cpu-thread-sweep-summary.json
  int8-tuning/cpu-thread-sweep-summary.json
  int8-tuning-preprocessed/cpu-thread-sweep-summary.json
  model_card.md

runs/mobilenetv3_small112_imdb_facecrop_real_fairface_utkface/
  best.pt
  last.pt
  config.resolved.yaml
  metrics.jsonl
  evaluation_val.json
  model_fp32.onnx
  model_int8_static.onnx
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  cpu-thread-sweep-summary.json
  model_card.md

runs/mobilenetv3_small112_imdb_source_balanced_facecrop_real_fairface_utkface/
  best.pt
  last.pt
  config.resolved.yaml
  metrics.jsonl
  evaluation_val.json
  model_fp32.onnx
  model_int8_static.onnx
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  cpu-thread-sweep-summary.json
  model_card.md

runs/mobilenetv3_small112_imdb_pretrain_finetune_fairface_utkface/
  best.pt
  last.pt
  config.resolved.yaml
  metrics.jsonl
  evaluation_val.json
  model_fp32.onnx
  model_int8_static.onnx
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  cpu-thread-sweep-summary.json
  model_card.md

runs/efficientnet_v2_s_128_imdb_source_balanced_gender_priority_real_fairface_utkface/
  best.pt
  last.pt
  config.resolved.yaml
  metrics.jsonl
  evaluation_val.json
  model_fp32.onnx
  model_int8_static.onnx
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  model_card.md

runs/mobilenetv3_small112_imdb_source_balanced_distill_gender_priority_efficientnet_v2_s_imdb_fairface_utkface/
  best.pt
  last.pt
  config.resolved.yaml
  metrics.jsonl
  evaluation_val.json
  model_fp32.onnx
  model_int8_static.onnx
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  cpu-thread-sweep-summary.json
  model_card.md

runs/mobilenetv3_large128_imdb_source_balanced_distill_gender_priority_efficientnet_v2_s_imdb_fairface_utkface/
  best.pt
  last.pt
  config.resolved.yaml
  metrics.jsonl
  evaluation_val.json
  model_fp32.onnx
  model_int8_static.onnx
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  cpu-thread-sweep-summary.json
  model_card.md

runs/mobilenetv3_small112_distill_efficientnet_b0_fairface_utkface/
  best.pt
  model_fp32.onnx
  model_int8_static.onnx
  evaluation_val.json
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  cpu-thread-sweep-summary.json
  model_card.md

runs/mobilenetv3_small112_distill_efficientnet_v2_s_fairface_utkface/
  best.pt
  model_fp32.onnx
  model_int8_static.onnx
  evaluation_val.json
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  cpu-thread-sweep-summary.json
  model_card.md

runs/mobilenetv3_small112_distill_gender_efficientnet_v2_s_fairface_utkface/
  best.pt
  model_fp32.onnx
  model_int8_static.onnx
  evaluation_val.json
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  cpu-thread-sweep-summary.json
  model_card.md

runs/mobilenetv3_small128_real_fairface_utkface/
  best.pt
  model_fp32.onnx
  model_int8_static.onnx
  evaluation_val.json
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  cpu-thread-sweep-summary.json
  model_card.md

runs/mobilenetv3_small112_distill_large112_fairface_utkface/
  best.pt
  model_fp32.onnx
  model_int8_static.onnx
  evaluation_val.json
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  cpu-thread-sweep-summary.json
  model_card.md

runs/mobilenetv3_small112_distill_light_large112_fairface_utkface/
  best.pt
  model_fp32.onnx
  model_int8_static.onnx
  evaluation_val.json
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  cpu-thread-sweep-summary.json
  model_card.md

runs/mobilenetv3_small112_lagenda_real_fairface_utkface/
  best.pt
  model_fp32.onnx
  model_int8_static.onnx
  evaluation_val.json
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  cpu-thread-sweep-summary.json
  model_card.md

runs/mobilenetv3_small112_lagenda_facecrop_real_fairface_utkface/
  best.pt
  model_fp32.onnx
  model_int8_static.onnx
  evaluation_val.json
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  cpu-thread-sweep-summary.json
  model_card.md

runs/efficientnet_b0_128_real_fairface_utkface/
  best.pt
  last.pt
  config.resolved.yaml
  metrics.jsonl
  evaluation_val.json
  model_fp32.onnx
  model_int8_static.onnx
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  cpu-thread-sweep-summary.json
  model_card.md

runs/efficientnet_b0_128_gender_priority_real_fairface_utkface/
  best.pt
  last.pt
  config.resolved.yaml
  metrics.jsonl
  evaluation_val.json
  model_fp32.onnx
  model_int8_static.onnx
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  cpu-thread-sweep-summary.json
  model_card.md

runs/efficientnet_v2_s_128_gender_priority_real_fairface_utkface/
  best.pt
  last.pt
  config.resolved.yaml
  metrics.jsonl
  evaluation_val.json
  model_fp32.onnx
  model_int8_static.onnx
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  model_card.md

runs/efficientnet_v2_s_128_real_fairface_utkface/
  best.pt
  last.pt
  config.resolved.yaml
  metrics.jsonl
  evaluation_val.json
  model_fp32.onnx
  model_int8_static.onnx
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  cpu-thread-sweep-summary.json
  model_card.md

runs/resnet18_128_real_fairface_utkface/
  best.pt
  last.pt
  config.resolved.yaml
  metrics.jsonl
  evaluation_val.json
  model_fp32.onnx
  model_int8_static.onnx
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  cpu-thread-sweep-summary.json
  model_card.md

runs/convnext_tiny_128_real_fairface_utkface/
  best.pt
  last.pt
  config.resolved.yaml
  metrics.jsonl
  evaluation_val.json
  model_fp32.onnx
  model_int8_static.onnx
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  cpu-thread-sweep-summary.json
  model_card.md

runs/swin_t_128_real_fairface_utkface/
  best.pt
  model_fp32.onnx
  model_int8_static.onnx
  evaluation_val.json
  benchmark_fp32_cpu.json
  benchmark_int8_static_cpu.json
  model_card.md