CelebA, short for CelebFaces Attributes Dataset, is a large-scale face dataset introduced in 2015 by researchers at the Chinese University of Hong Kong. It contains 202,599 face images of 10,177 unique celebrity identities, with each image annotated with 40 binary attribute labels such as 'Smiling', 'Eyeglasses', 'Male', and 'Young'. The dataset was created to support research in facial attribute recognition, face detection, and face generation, and it has become a standard benchmark in the fields of Machine learning and Deep learning.
The images in CelebA are sourced from celebrity photographs collected from the internet, then aligned and cropped to a centered face region. The annotations were produced through a combination of automated labeling and human verification, ensuring high consistency across the 40 attributes. The dataset is split into three partitions: 162,770 images for training, 19,867 for validation, and 19,962 for testing, a standard split that facilitates reproducible comparisons across different models.
Attribute Prediction and Recognition
CelebA is most commonly used for attribute prediction tasks, where a model must classify one or more of the 40 binary attributes from a face image. This task has driven advances in Neural network architectures, particularly convolutional neural networks (CNNs), which have achieved high accuracy on the benchmark. The dataset also supports multi-task learning, where a single model predicts all attributes simultaneously, a common approach in modern facial analysis systems. Researchers often use CelebA to evaluate generalization across diverse demographics, as the celebrity identities cover a wide range of ages, ethnicities, and facial expressions.
Face Generation and Synthesis
Beyond classification, CelebA has become a foundational dataset for generative models. It is frequently used to train generative adversarial networks (GANs) and variational autoencoders (VAEs) to produce realistic synthetic faces. The dataset's attribute labels enable conditional generation, where a model can generate a face with specific traits, such as adding a smile or changing hair color. This capability has been instrumental in advancing Generative AI techniques, including style transfer and face editing. Many early GAN papers, including those on progressive growing and style-based architectures, used CelebA as a primary evaluation dataset.
Benchmarking and Evaluation
CelebA serves as a standard benchmark for comparing the performance of different Artificial intelligence models. Its large size and rich annotations allow for rigorous testing of algorithms in face recognition, landmark detection, and attribute editing. The dataset is often paired with metrics such as classification accuracy, F1 score, and Fréchet Inception Distance (FID) for generated images. Because of its widespread adoption, results on CelebA are directly comparable across studies, making it a de facto standard in the computer vision community. The dataset has also been used to study fairness and bias in facial analysis systems, as the attribute labels can reveal demographic disparities in model performance.
Technical Details and Access
The dataset is distributed as a set of aligned and cropped images, each 178x218 pixels, along with a text file listing the attribute annotations. The original release also includes landmark locations for each face, though these are less commonly used in modern applications. CelebA is publicly available for academic research, and its popularity has led to integration into major frameworks like PyTorch and TensorFlow, which provide built-in data loaders for the dataset. The dataset's creators also released a companion dataset, CelebA-HQ, which provides higher-resolution versions of the images (up to 1024x1024) for tasks requiring finer detail.
Impact and Limitations
CelebA has had a significant impact on the development of facial analysis technologies, influencing everything from social media filters to security systems. However, it also has known limitations. The dataset is biased toward celebrity faces, which may not represent the general population, and the binary attribute labels can oversimplify complex facial characteristics. Additionally, the images are relatively low-resolution, which can limit performance on tasks requiring fine-grained detail. Despite these issues, CelebA remains a widely used resource, and its design has inspired subsequent datasets like FFHQ and VGGFace2. As of 2025, it continues to be a standard reference point for evaluating new methods in face-related Deep learning research.