---
title: "20 Best Face Recognition Datasets for ML in 2026"
description: "Your model won’t guess a face out of thin air. It learns. From pixels, patterns — and the datasets you feed it. And here's the…"
url: "https://unidata.pro/blog/best-ml-face-recognition-datasets/"
date_modified: "2026-05-26T12:51:59+03:00"
language: "en-US"
---
![Face Recognition Datasets](https://unidata.pro/wp-content/uploads/2025/07/face-recognition-datasets.webp)Your model won’t guess a face out of thin air. It learns. From pixels, patterns — and the [datasets](https://unidata.pro/datasets/) you feed it. And here's the catch: the quality of your data defines everything. Accuracy. Bias. Speed. Security.

So if you're working on face recognition in 2026 — whether it's unlocking doors, verifying selfies, or flagging deepfakes — these are the datasets that actually matter. No filler. No fluff. Just the ones used, cited, or benchmarked by teams who take face data seriously.

How to Tell If a Dataset’s Actually Worth It
--------------------------------------------

You don’t need 10 million blurry selfies. You need **signal** — not noise.

Before we hit the list, here’s how top teams size up a face recognition dataset. If yours doesn’t tick most of these boxes, it’s probably not ready for production.

**Type of Data  
Are we talking static photos, live video, 3D scans, or synthetic faces? Different tasks = different formats.

**Size & Diversity  
It's not just how many images — but how many people, poses, ethnicities, lighting conditions, devices. If it only knows sunny selfies from Silicon Valley, it’ll struggle in a rainy parking garage.

**Use Case Alignment  
Is the dataset built for detection, verification, spoofing defense, or fairness testing? Pick wrong, and your model learns the wrong game.

**Annotation Quality  
Keypoints, age, gender, liveness, emotion — these labels turn raw pixels into trainable structure.

**Access & Licensing  
Some are open. Some require NDA. Some are ISO‑certified for compliance-heavy industries. Know before you build.

Foundational Recognition Datasets
---------------------------------

These are the core datasets used to train and benchmark most modern face recognition models. They’re large, well-documented, and time-tested — ideal for pretraining, fine-tuning, and academic comparison.

### **1.**[ VGGFace2 ](https://www.robots.ox.ac.uk/~vgg/data/vgg_face2/)

![VGGFace2  dataset](https://unidata.pro/wp-content/uploads/2025/07/vggface2-scaled.webp)**Format:** Photo

**Volume:** 3.3 million images (9,131 identities)

**Access:** Free (academic use only)

Released by Oxford, VGGFace2 includes 3.3M face images from 9,131 identities with wide variation in age, pose, and lighting. It remains one of the most trusted datasets for recognition and bias testing, used in ArcFace, FaceNet, and other SOTA models.

 **Dataset Spotlight (click to expand)** ```

name: COUGHVID
type: Cough recordings
format: Audio
volume: 25,000+ samples
access: Free (open access)
use_case: Disease detection, acoustic modeling, mobile diagnostics
  
```

### **2.** [CASIA-WebFace](https://www.kaggle.com/datasets/debarghamitraroy/casia-webface)

![CASIA-WebFace dataset](https://unidata.pro/wp-content/uploads/2025/07/casia-webface-scaled.webp)**Format:** Photo

**Volume:** 494,414 images (10,575 identities)

**Access:** Free (academic use only)

Collected from web celebrity photos, this dataset contains 494,414 images of over 10,000 subjects. While some noise is present, it’s still a widely used benchmark for face embedding models and baseline verification pipelines.

### **3.** [LFW (Labeled Faces in the Wild)](https://www.kaggle.com/datasets/jessicali9530/lfw-dataset)

![LFW (Labeled Faces in the Wild)](https://unidata.pro/wp-content/uploads/2025/07/lfw-scaled.webp)**Format:** Photo

**Volume:** 13,000+ labeled images

**Access:** Free (open access)

A historic benchmark of 13K+ labeled images captured in the wild, LFW introduced unconstrained face verification. It’s no longer state-of-the-art, but still useful for sanity checks and legacy model evaluation.

### **4.** [MegaFace](https://megaface.cs.washington.edu/)

![MegaFace](https://unidata.pro/wp-content/uploads/2025/07/megaface--scaled.webp)**Format:** Photo

**Volume:** 1 million distractor images

**Access:** Free (open access)

This dataset isn’t for training — it’s a challenge benchmark that includes over 1 million distractor images to stress-test face identification systems. Essential for measuring how your model scales under real-world noise.

### **5.** [WebFace42M](https://arxiv.org/abs/2103.04098)

![](https://unidata.pro/wp-content/uploads/2025/07/webface42m--scaled.webp)**Format:** Photo

**Volume:** 42 million images (2 million identities)

**Access:** Free (academic use only)

A recent large-scale dataset with 42 million images from 2 million subjects, offering strong diversity in age, pose, and lighting. Used for training high-capacity models that need to generalize across global populations.

**[6. MS-Celeb-1M (MS1M)](https://arxiv.org/abs/1607.08221)**

![](https://unidata.pro/wp-content/uploads/2025/09/ms-celeb-1m-1024x230.webp)**Format:** Photo

**Volume:** ~10 million images (100,000 celebrities)

**Access:** Free (academic use only; original dataset retired — cleaned versions such as MS1M-ArcFace available via InsightFace)

Originally released by Microsoft Research at ECCV 2016, MS-Celeb-1M set out to recognize one million celebrities from their face images using the web as a training source. Despite known label noise in the raw version, it became the de facto pretraining backbone for almost every major face recognition model published between 2017 and 2024 — including ArcFace, CosFace, and AdaFace. The original dataset was pulled for privacy reasons, but the community-cleaned subset MS1M-ArcFace (~5.8 million images, 85K identities) remains widely used and is available through the InsightFace repository. If your model has been pretrained on any published SOTA weights, it was almost certainly trained on this data.

Fairness & Attribute Modeling
---------------------------------

If your model performs worse on certain demographics or can’t handle facial attributes, you’ll want these datasets. They include rich labels for age, gender, race, expressions, and more — ideal for fairness audits and multi-task training.

### **7.** [CelebA](https://mmlab.ie.cuhk.edu.hk/projects/CelebA.html) 

![CelebA dataset](https://unidata.pro/wp-content/uploads/2025/07/celeba-scaled.webp)**Format:** Photo

**Volume:** 200,000+ images

**Access:** Free (academic use only)

One of the most widely used facial attribute datasets, CelebA features over 200K celebrity images labeled with 40 binary traits (like smiling, eyeglasses, gender). It’s great for training multi-label classifiers and testing model bias.

 **Dataset Spotlight (click to expand)** ```

name: CelebA
format: photo
volume: 200,000+ images
access: free
license: academic use only
use_case:
  - attribute modeling
  - bias testing
 
```

### **8.** [CelebA-HQ](https://www.kaggle.com/datasets/lamsimon/celebahq/data)

![CelebA-HQ dataset](https://unidata.pro/wp-content/uploads/2025/07/celeba-hq--scaled.webp)**Format:** Photo (1024×1024)

**Volume:** 30,000 images

**Access:** Free (academic use only)

A high-resolution (1024×1024) version of CelebA, reprocessed and aligned for use in GANs and high-fidelity facial modeling. While smaller in size (30K images), it’s favored in generative research.

### **9.** [AgeDB](https://ibug.doc.ic.ac.uk/resources/agedb/)

![AgeDB dataset](https://unidata.pro/wp-content/uploads/2025/07/agedb--scaled.webp)**Format:** Photo

**Volume:** 12,240 images (440 subjects)

**Access:** Free (academic use only)

Focused on age variation, AgeDB provides 12,240 face images from 440 subjects with both real and apparent age labels. It’s ideal for age-invariant face recognition and age estimation benchmarks.

### **10.** [UTKFace](https://susanqq.github.io/UTKFace/)

![UTKFace dataset](https://unidata.pro/wp-content/uploads/2025/07/utkface-scaled.webp)**Format:** Photo

**Volume:** 20,000+ images

**Access:** Free (open access)

This open dataset includes over 20K images labeled by age, gender, and ethnicity. Despite some label noise, it’s a go-to for training and evaluating demographic fairness in face models.

### **11.** [CALFW](http://whdeng.cn/CALFW/index.html?reload=true)

![CALFW dataset](https://unidata.pro/wp-content/uploads/2025/07/calfw-scaled.webp)**Format:** Photo

**Volume:** ~4,000 image pairs

**Access:** Free (academic use only)

A curated version of LFW where image pairs differ in age — designed to test how well face verification holds up across time. Small but targeted, CALFW is great for temporal robustness studies.

Anti-Spoofing & Deepfake Detection
--------------------------------------

Face recognition systems fail fast when they can't tell real from fake. These datasets cover spoof attempts, 3D masks, deepfakes, and face obfuscation — making them essential for liveness detection, security testing, and adversarial defense.

### **12.** [MFR2](https://sites.google.com/view/masktheface/mfr2-dataset) 

![MFR2 ](https://unidata.pro/wp-content/uploads/2025/07/mfr2--scaled.webp)**Format:** Photo

**Volume:** 53,000 images

**Access:** Free (academic use only)

Created for real-world masked face recognition, MFR2 includes 53,000 images of subjects wearing medical masks under natural lighting. It reflects post-COVID challenges and helps models learn to see behind occlusion.

 **Dataset Spotlight (click to expand)** ```

name: MFR2
format: photo
volume: 53,000 images
access: free
license: academic use only
use_case:
  - masked face recognition
 
```

### **13.** [3DMAD](https://www.idiap.ch/en/scientific-research/data/3dmad)

![3DMAD dataset](https://unidata.pro/wp-content/uploads/2025/07/3dmad-scaled.webp)**Format:** Video

**Volume:** 17 subjects (multiple sessions)

**Access:** Free (academic use only)

A high-quality spoofing dataset focused on 3D mask attacks, captured with depth sensors across 17 subjects. It’s often used for evaluating liveness models under presentation attack conditions.

### **14.** [FaceForensics++](https://github.com/ondyari/FaceForensics)

![FaceForensics++](https://unidata.pro/wp-content/uploads/2025/07/faceforensics-scaled.webp)**Format:** Video

**Volume:** 1.8M+ frames

**Access:** Free (academic use only)

A benchmark for manipulated video detection with 1.8M+ frames covering four types of facial tampering. Includes compression levels and ground truth masks for precise training and evaluation.

### **15.** [DFDC](https://ai.meta.com/datasets/dfdc/)

![Deepfake detection challenge dataset](https://unidata.pro/wp-content/uploads/2025/07/dfdc-scaled.webp)**Format:** Video

**Volume:** 100,000+ video clips

**Access:** Free (academic use, non-commercial)

Released by Meta, this dataset includes over 100,000 video clips of real and fake faces under different lighting and compression. It’s noisy but realistic — useful for building robust adversarial defenses.

### **16.** [FFHQ](https://github.com/NVlabs/ffhq-dataset)

![Flickr Faces HQ dataset](https://unidata.pro/wp-content/uploads/2025/07/ffhq-scaled.webp)**Format:** Photo (1024×1024)

**Volume:** 70,000 images

**Access:** Free (open access)

70,000 high-resolution images scraped from Flickr, designed for training GANs and testing generative bias. Includes wide variation in age, skin tone, and image background — perfect for fairness augmentation.

Additional Free Datasets
------------------------

Not every dataset fits neatly into detection or verification. These five give you broader tools: emotion tags, 3D structure, surveillance realism, and extreme poses — great for edge cases and specialized training.

### **17.** [AffectNet](https://arxiv.org/abs/1708.03985)

![AffectNet](https://unidata.pro/wp-content/uploads/2025/07/affectnet-scaled.webp)**Format:** Photo

**Volume:** 1,000,000+ images

**Access:** Free (academic use only)

A massive dataset with over 1M images labeled for facial expressions, arousal, and valence. Ideal for emotion classification, but also includes 68 facial landmarks for multi-task training.

 **Dataset Spotlight (click to expand)** ```

name: AffectNet
format: photo
volume: 1,000,000+ images
access: free
license: academic use only
use_case:
  - emotion recognition
  - facial keypoint detection
 
```

### 18. [SCface](https://www.scface.org/)

![Surveillance cameras face dataset](https://unidata.pro/wp-content/uploads/2025/07/scface-scaled.webp)**Format:** Photo (RGB + IR)

**Volume:** 4,000+ images (130 subjects)

**Access:** Free (academic use, via request)

Simulates low-resolution surveillance scenarios using visible and infrared cameras. Includes 4,000+ images from 130 subjects at multiple distances and angles — great for testing robustness.

### **19.** [WIDER FACE](http://shuoyang1213.me/WIDERFACE/)

![WIDER FACE dataset](https://unidata.pro/wp-content/uploads/2025/07/wider-face-scaled.webp)**Format:** Photo

**Volume:** 32,000 images (400,000+ faces)

**Access:** Free (academic use only)

One of the largest datasets for face detection in complex scenes. It contains 32K images with nearly 400K labeled faces, many in extreme poses, small sizes, and crowded settings.

### **20.** [EURECOM Kinect Face](https://rgb-d.eurecom.fr/)

![](https://unidata.pro/wp-content/uploads/2025/07/eurecom-kinect-face-scaled.webp)**Format:** RGB + Depth

**Volume:** 52 subjects (multiple settings)

**Access:** Free (academic use only)

RGB-D dataset with RGB and depth images from 52 subjects under various occlusion and expression settings. Useful for 3D-aware face recognition and pose variation handling.

### **21.** [PUT Face Database](https://www.researchgate.net/publication/232085001_The_PUT_face_database)

![PUT Face Database](https://unidata.pro/wp-content/uploads/2025/07/put-face-database--scaled.webp)**Format:** Photo

**Volume:** 100 subjects (varied head pose, lighting, expression)

**Access:** Free (academic use only)

Includes controlled images from 100 subjects with variation in head pose, illumination, and expression. Good for training classical models or augmenting lightweight pipelines.

**Production-Grade Datasets (Commercial Use)**
----------------------------------------------

Open data can take you far — but when real-world compliance, spoofing risks, or ID verification come into play, you need structured, certified data. These Unidata datasets are designed for live deployment environments: mobile onboarding, ISO certification, and biometric KYC flows.

### **1.** [iBeta Level 1](https://unidata.pro/datasets/ibeta-level-1-video-attacks/)

![iBeta Level 1 Unidata](https://unidata.pro/wp-content/uploads/2025/07/ibeta-scaled.webp)**Format:** Video

**Volume:** 35,800+ videos

**Access:** Paid (commercial, ISO-ready)

A liveness dataset built for ISO 30107-3 Level 1 testing. Includes 35,800+ high-quality videos of both genuine access attempts and spoof attacks, recorded under varied lighting and device setups. Ideal for validating liveness models before certification.

 **Dataset Spotlight (click to expand)** ```

name: iBeta Level 1
format: video
volume: 35,800+ videos
access: paid
license: commercial, ISO-ready
use_case:
  - liveness detection
  - compliance testing
 
```

### **2.** [Liveness Detection Video Dataset](https://unidata.pro/datasets/liveness-detection-video-dataset/)

![Unidata Liveness Detection Video Dataset](https://unidata.pro/wp-content/uploads/2025/07/liveness-detection-scaled.webp)**Format:** Video + Image

**Volume:** 248,000 video clips + 66,000 stills (1,056 subjects)

**Access:** Paid (commercial license)

One of the largest commercial face video datasets for spoof prevention. Includes 248K liveness video clips and 66K stills from 1,056 subjects across diverse settings — indoor, outdoor, various devices, and skin tones. Useful for both training and evaluating production-ready anti-spoofing models.

 **Dataset Spotlight (click to expand)** ```

name: Liveness Detection Video Dataset
format: video + image
volume: 248,000 video clips + 66,000 stills (1,056 subjects)
access: paid
license: commercial
use_case:
  - spoof detection
  - biometric security
 
```

### **3.** [Selfie with ID Dataset ](https://unidata.pro/datasets/selfie-with-id/) 

![Unidata Selfie with ID Dataset  ](https://unidata.pro/wp-content/uploads/2025/07/selfie-dataset-scaled.webp)**Format:** Photo

**Volume:** 65,000+ selfie–ID pairs (5,000+ users)

**Access:** Paid (commercial license)

Over 65,000 paired selfie and document images from 5,000+ users across 40+ countries. Each sample includes diverse lighting, document types, and real-world capture conditions — ideal for testing document-face match models in KYC workflows.

 **Dataset Spotlight (click to expand)** ```

name: Selfie with ID Dataset
format: photo
volume: 65,000+ selfie–ID pairs (5,000+ users)
access: paid
license: commercial
use_case:
  - ID verification
  - KYC
  - document matching
 
```

Final Takeaway
--------------

You don’t need to pay to get started with face recognition. These 20 free datasets cover the essentials — from clean baselines to deepfake defense. Pair them smartly, benchmark often, and know when it’s time to level up to production-grade data. Because great models start with the right faces.

Cheat Sheet: All 23 Face Recognition Datasets
---------------------------------------------

 **Dataset Spotlight (click to expand)** ```

- name: VGGFace2
  format: photo
  volume: 3.3 million images (9,131 identities)
  access: free
  license: academic use only
  use_case:
    - face recognition
    - bias testing

- name: CASIA-WebFace
  format: photo
  volume: 494,414 images (10,575 identities)
  access: free
  license: academic use only
  use_case:
    - face embeddings
    - pretraining

- name: LFW
  format: photo
  volume: 13,000+ images
  access: free
  license: open access
  use_case:
    - face verification
    - legacy model evaluation

- name: MegaFace
  format: photo
  volume: 1 million distractor images
  access: free
  license: open access
  use_case:
    - 1:N benchmarking
    - false positive testing

- name: WebFace42M
  format: photo
  volume: 42 million images (2 million identities)
  access: free
  license: academic use only
  use_case:
    - large-scale pretraining
    - fairness analysis

- name: CelebA
  format: photo
  volume: 200,000+ images
  access: free
  license: academic use only
  use_case:
    - attribute modeling
    - bias testing

- name: CelebA-HQ
  format: photo (1024×1024)
  volume: 30,000 images
  access: free
  license: academic use only
  use_case:
    - generative modeling
    - style transfer

- name: AgeDB
  format: photo
  volume: 12,240 images (440 subjects)
  access: free
  license: academic use only
  use_case:
    - age estimation
    - robustness testing

- name: UTKFace
  format: photo
  volume: 20,000+ images
  access: free
  license: open access
  use_case:
    - bias analysis
    - subgroup accuracy

- name: CALFW
  format: photo
  volume: ~4,000 image pairs
  access: free
  license: academic use only
  use_case:
    - long-term face verification
    - temporal robustness

- name: MFR2
  format: photo
  volume: 53,000 images
  access: free
  license: academic use only
  use_case:
    - masked face recognition

- name: 3DMAD
  format: video
  volume: 17 subjects (multiple sessions)
  access: free
  license: academic use only
  use_case:
    - 3D anti-spoofing
    - liveness evaluation

- name: FaceForensics++
  format: video
  volume: 1.8M+ frames
  access: free
  license: academic use only
  use_case:
    - deepfake detection
    - forgery classification

- name: DFDC
  format: video
  volume: 100,000+ video clips
  access: free
  license: academic use (non-commercial)
  use_case:
    - deepfake training
    - adversarial robustness

- name: FFHQ
  format: photo (1024×1024)
  volume: 70,000 images
  access: free
  license: open access
  use_case:
    - generative modeling
    - dataset balance

- name: AffectNet
  format: photo
  volume: 1,000,000+ images
  access: free
  license: academic use only
  use_case:
    - emotion recognition
    - facial keypoint detection

- name: SCface
  format: photo (RGB + IR)
  volume: 4,000+ images (130 subjects)
  access: free
  license: academic use (via request)
  use_case:
    - low-res face recognition
    - surveillance

- name: WIDER FACE
  format: photo
  volume: 32,000 images (400,000+ faces)
  access: free
  license: academic use only
  use_case:
    - face detection
    - occlusion and crowd handling

- name: EURECOM Kinect Face
  format: RGB + depth
  volume: 52 subjects (multiple settings)
  access: free
  license: academic use only
  use_case:
    - depth-based recognition
    - expression robustness

- name: PUT Face Database
  format: photo
  volume: 100 subjects (varied head pose, lighting, expression)
  access: free
  license: academic use only
  use_case:
    - pose robustness
    - face preprocessing

- name: iBeta Level 1
  format: video
  volume: 35,800+ videos
  access: paid
  license: commercial, ISO-ready
  use_case:
    - liveness detection
    - compliance testing

- name: Liveness Detection Video Dataset
  format: video + image
  volume: 248,000 video clips + 66,000 stills (1,056 subjects)
  access: paid
  license: commercial
  use_case:
    - spoof detection
    - biometric security

- name: Selfie with ID Dataset
  format: photo
  volume: 65,000+ selfie–ID pairs (5,000+ users)
  access: paid
  license: commercial
  use_case:
    - ID verification
    - KYC
    - document matching
 
```

## Additional Modules

### FAQ Section

**Questions and Answers:**

- **Question:** What are the best free datasets for training face recognition models? — **Answer:** Top free options include VGGFace2, WebFace42M, CASIA-WebFace, and CelebA. These cover large-scale recognition, attribute prediction, and demographic diversity for academic use.
- **Question:** Which dataset is best for deepfake detection or liveness training? — **Answer:** Use FaceForensics++, DFDC, and MFR2 for spoof detection and video forgery. For more advanced liveness scenarios, 3DMAD (3D masks) and commercial sets like iBeta Level 1 are recommended.
- **Question:** Can I use these datasets for commercial applications? — **Answer:** Most open datasets are limited to academic or non-commercial research. For production use, datasets like Selfie with ID, iBeta Level 1, and Unidata’s liveness video sets offer commercial licenses.
- **Question:** What’s the difference between face verification and identification datasets? — **Answer:** Verification datasets (like LFW or CALFW) test pairwise similarity. Identification datasets (like VGGFace2 or MegaFace) are used for 1:N recognition across large galleries.
- **Question:** Are there face datasets with age, gender, and ethnicity labels? — **Answer:** Yes — UTKFace, FFHQ, CelebA, and IJB-C include demographic metadata. These are ideal for fairness audits and training bias-aware models.

[Full list of this site's AI-readable pages](https://unidata.pro/llms.txt)