Out of the box age estimation through facial imagery: A Comprehensive Benchmark of Vision-Language Models vs. out-of-the-box Traditional Architectures

Simiao Ren; Xingyu Shen; Ankit Raj; Albert Dai; Caroline; Zhang; Yuan Xu; Zexi Chen

cs.CVFeb 8, 2026

Out of the box age estimation through facial imagery: A Comprehensive Benchmark of Vision-Language Models vs. out-of-the-box Traditional Architectures

Authors: Simiao Ren, Xingyu Shen, Ankit Raj, Albert Dai, Caroline, Zhang, Yuan Xu, Zexi Chen,

Organizations: Manlin · Reality Inc. · Duke University · New York University ·

Abstract

Facial age estimation plays a critical role in content moderation, age verification, and deepfake detection. However, no prior benchmark has systematically compared modern vision-language models (VLMs) with specialized age estimation architectures. We present the first large-scale cross-paradigm benchmark, evaluating 34 models - 22 specialized architectures with publicly available pretrained weights and 12 general-purpose VLMs - across eight standard datasets (UTKFace, IMDB-WIKI, MORPH, AFAD, CACD, FG-NET, APPA-REAL, and AgeDB), totaling 1,100 test images per model. Our key finding is striking: zero-shot VLMs significantly outperform most specialized models, achieving an average mean absolute error (MAE) of 5.65 years compared to 9.88 years for non-LLM models. The best-performing VLM (Gemini 3 Flash Preview, MAE 4.32) surpasses the strongest non-LLM model (MiVOLO, MAE 5.10) by 15%. MiVOLO - unique in combining face and body features using Vision Transformers - is the only specialized model that remains competitive with VLMs. We further analyze age verification at the 18-year threshold and find that most non-LLM models exhibit false adult rates between 39% and 100% for minors, whereas VLMs reduce this to 16%-29%. Additionally, coarse age binning (8-9 classes) consistently increases MAE beyond 13 years. Stratified analysis across 14 age groups reveals that all models struggle most at extreme ages (under 5 and over 65). Overall, these findings challenge the assumption that task-specific architectures are necessary for high-performance age estimation and suggest that future work should focus on distilling VLM capabilities into efficient specialized models.

Figures & tables

Category	Representative Models	#
Specialized (Non-LLM) Models
CNN-Regression	Yu4u, BoyuanJiang, Py-Agender	10
CNN-Expected Value	DEX (VGG-16)	1
Vision Transformer	MiVOLO (ViT + YOLOv8)	1
Multi-task	DeepFace, InsightFace, Muno-AI	4
Lightweight	SSR-Net	1

Table 1: Model taxonomy. We evaluate 22 specialized (non-LLM) architectures—restricted to those with publicly available pretrained weights—across 7 categories and 12 general-purpose VLMs in a zero-shot setting.

Dataset	Size	Ages	Description
UTKFace [ 41 ]	24K	0–116	In-the-wild faces
IMDB-WIKI [ 15 ]	213K	0–100	Celebrity images
MORPH [ 42 ]	50K	16–77	Mugshot-style
AFAD [ 16 ]	165K	15–72	Asian Face Age
CACD [ 43 ]	155K	14–62	Cross-Age Celebrity
FG-NET [ 44 ]	1K	0–69	Fine-grained aging

Table 2: Datasets used in our benchmark. We randomly sample 100 images per dataset (seed 42), except AgeDB which uses 400 images.

#	Model	UTK	IMDB	MORPH	AFAD	CACD	FG-NET	APPA	AgeDB	Avg.
1	Gemini 3 Flash Prev.	4.62	4.43	3.85	5.23	4.67	2.36	4.09	5.32	4.32
2	Gemini 2.5 Flash	4.81	5.74	4.75	5.22	5.38	3.42	4.67	5.80	4.97
3	MiVOLO	5.07	4.32	4.02	4.77	5.65	5.09	6.00	5.88	5.10
4	GPT-5 Nano	4.80	5.96	3.75	5.22	6.15	3.41	5.40	6.90	5.20
5	GPT-5.2	4.94	5.81	4.74	6.29	5.41	3.66	4.54	6.26	5.21
6	Qwen3-VL 235B	5.32	6.09	5.50	5.90	5.19	3.53	4.80	6.97	5.41

Table 3: Main results: MAE (years,

\downarrow

) across 8 datasets. Models sorted by average MAE. Orange = VLM, Blue = non-LLM. Bold = best per column; underline = second best. Only non-LLM models with publicly available pretrained weights are included.

Category	#	Mean MAE	Std.
VLM (zero-shot)	12	5.65	0.72
ViT (MiVOLO)	1	5.10	—
CNN-Regression	10	8.52	1.36
CNN-Expected Value	1	7.19	—
Multi-task	4	9.94	1.62
Retrieval (MWR)	1	10.04	—

Table 4: Architecture category analysis. Mean and standard deviation of average MAE within each category. Coarse-bin models (8–9 classes) consistently produce the highest errors. “Non-LLM excl. coarse” excludes the 4 coarse-bin models.

Figure 1: All 34 models ranked by average MAE. Orange = VLM, blue = non-LLM. VLMs occupy 12 of the top 13 positions.

Figure 2: Distribution of average MAE by model type. The VLM interquartile range falls entirely below the non-LLM median.

Figure 3: Complete MAE heatmap across all 34 models and 8 datasets. Models sorted by average MAE (top = best). Orange labels = VLM, blue labels = non-LLM. Darker cells indicate higher error. The top tier is dominated by VLMs with uniformly low errors across all datasets.

Figure 4: MAE by architecture category. Each dot is one model; black bars show category means. Coarse-bin models are uniformly poor. VLMs show the tightest cluster with lowest mean.

Figure 5: Error rates at the 18-year threshold. Left: False Adult Rate (minors missed); Right: False Minor Rate. Models sorted by FAR. VLM model names in orange .

Model	FAR $\downarrow$	FMR $\downarrow$	Type
Seed-1.6	16.3	3.5	VLM
Gemini 3 Flash Prev.	19.2	1.4	VLM
Qwen3-VL 235B	19.2	2.7	VLM
GPT-5 Nano	20.4	1.7	VLM
Mistral Small 3.2	21.2	2.6	VLM
MiVOLO	21.2	2.5	Non-LLM

Table 5: Age verification at the 18-year threshold across all datasets. FAR = False Adult Rate (% of actual minors predicted

\geq

18); FMR = False Minor Rate (% of actual adults predicted

<

18). Lower FAR/FMR is better. Selected models with

\geq

5 minor samples shown.

Figure 6: Per-age-group MAE heatmap for top 15 models. Darker = higher error. All models struggle at extreme ages, but VLMs ( orange labels) degrade more gracefully.

Large Vision-Language Models

Large vision-language models (LVLMs) connect image and video perception with language generation, reasoning, and interactive tasks.

See all 404 papers

VLAgeBench: Benchmarking Large Vision-Language Models for Zero-Shot Human Age Estimation
Mar 27, 2026Rakib Hossain Sajib, Md Kishor Morol, Rajan Das Gupta +2Large Vision-Language Models VLM Evaluation
Department of Computer Science and Engineering, Begum Rokeya University, Rangpur, Rangpur, Bangladesh · EliteLab.AI, Queens, New York, United States · Department of Computer Science, American International University - Bangladesh, Dhaka, Bangladesh
Can a Teenager Fool an AI? Evaluating Low-Cost Cosmetic Attacks on Age Estimation Systems
Feb 23, 2026Xingyu Shen, Tommy Duong, Xiaodong An +6Adversarial Attacks Physical Adversarial Attacks
Reality Inc. · Duke University · UC Berkeley +3
Toward Ethical Facial Age Estimation: A Generalized Zero-Shot Benchmark Without Training on Children's Data
May 28, 2026Caio Petrucci, Leo Sampaio Ferraz Ribeiro, Sandra AvilaZero-Shot Learning OOD Generalization
1Instituto de Computac¸˜ao, Universidade Estadual de Campinas (UNICAMP), Campinas, Brazil · 2Instituto de Ciˆencias Matem´aticas e de Computac¸˜ao, Universidade de S˜ao Paulo (USP), S˜ao Carlos, Brazil

Mar 27, 2026J/K move · Enter open · S save

VLAgeBench: Benchmarking Large Vision-Language Models for Zero-Shot Human Age Estimation

Rakib Hossain Sajib, Md Kishor Morol, Rajan Das Gupta, Mohammad Sakib Mahmood+1

Department of Computer Science and Engineering, Begum Rokeya University, Rangpur, Rangpur, Bangladesh · EliteLab.AI, Queens, New York, United States · Department of Computer Science, American International University - Bangladesh, Dhaka, Bangladesh

Large Vision-Language Models VLM Evaluation Zero-Shot Learning

Open paper AskPDF arXiv

Human age estimation from facial images represents a challenging computer vision task with significant applications in biometrics, healthcare, and human-computer interaction. While traditional deep learning approaches require extensive labeled datasets and domain-specific training, recent advances in large vision-language models (LVLMs) offer the potential for zero-shot age estimation. This study presents a comprehensive zero-shot evaluation of state-of-the-art Large Vision-Language Models (LVLMs) for facial age estimation, a task traditionally dominated by domain-specific convolutional networks and supervised learning. We assess the performance of GPT-4o, Claude 3.5 Sonnet, and LLaMA 3.2 Vision on two benchmark datasets, UTKFace and FG-NET, without any fine-tuning or task-specific adaptation. Using eight evaluation metrics, including MAE, MSE, RMSE, MAPE, MBE,

R^2

, CCC, and

\pm

5-year accuracy, we demonstrate that general-purpose LVLMs can deliver competitive performance in zero-shot settings. Our findings highlight the emergent capabilities of LVLMs for accurate biometric age estimation and position these models as promising tools for real-world applications. Additionally, we highlight performance disparities linked to image quality and demographic subgroups, underscoring the need for fairness-aware multimodal inference. This work introduces a reproducible benchmark and positions LVLMs as promising tools for real-world applications in forensic science, healthcare monitoring, and human-computer interaction. The benchmark focuses on strict zero-shot inference without fine-tuning and highlights remaining challenges related to prompt sensitivity, interpretability, computational cost, and demographic fairness.

Abstract

Large Vision-Language Models

Explore similar work

VLAgeBench: Benchmarking Large Vision-Language Models for Zero-Shot Human Age Estimation

Can a Teenager Fool an AI? Evaluating Low-Cost Cosmetic Attacks on Age Estimation Systems

Toward Ethical Facial Age Estimation: A Generalized Zero-Shot Benchmark Without Training on Children's Data