Mitigating Demographic Bias in Facial Datasets with Style-Based Multi-attribute Transfer

Generative Adversarial Networks (GANs) (Goodfellow et al. 2014) are the driving force behind the recent success of deep generative models in image synthesis (Radford et al. 2015; Karras et al. 2019; Brock et al. 2018; Karras et al. 2017; Park et al. 2019) and image-to-image translation (Choi et al. 2018; Huang et al. 2018; Kim et al. 2020; Isola et al. 2017; Zhu et al. 2017; Tang et al. 2019). Multiple variations have been proposed, including different training objectives (Salimans et al. 2016; Arjovsky et al. 2017; Mao et al. 2017; Lim and Ye 2017) and architectural choices (Radford et al. 2015; Karras et al. 2019). Among these variations, style-based GANs draw inspiration from the style transfer literature (Gatys et al. 2015) and use Adaptive Instance Normalization (AdaIN) (Huang and Belongie 2017) to condition image generation. MUNIT (Huang et al. 2018) uses AdaIN in a GAN setting to perform image-to-image translation by injecting style content into an autoencoder-like network at the bottleneck layers. The seminal work of StyleGAN (Karras et al. 2019) proposed a generator architecture using AdaIN to modulate style content at multiple resolutions. The state-of-the-art image-to-image translation methods continue to adopt style-based approaches (Huang et al. 2018; Kim et al. 2020; Park et al. 2019; Ma et al. 2018), and we follow in this vein–treating demographic attributes as target styles.

Synthesizing faces of a specific target demographic has been studied extensively, albeit for individual demographic attributes. In particular, age progression refers to the task of rendering an aged or rejuvenated image of an input face (Fu et al. 2010; Ramanathan et al. 2009; Georgopoulos et al. 2018). While earlier works in age progression proposed simplistic prototype-based approaches that could not produce photorealistic results, a number of recently proposed GAN-based methods are capable of convincing face aging. Zhang et al. (2017) introduced a conditional adversarial autoencoder (CAAE) that models aging as the traversal of a low-dimensional manifold. GAN-based image translation approaches were also proposed in Wang et al. (2018), Yang et al. (2018), where pre-trained networks were used to facilitate identity preservation and aging accuracy, respectively. Besides age, generating faces with different genders is a standard benchmark of the CELEB-A dataset (Liu et al. 2015), and is successfully performed by numerous attribute editing approaches (Choi et al. 2018; He et al. 2019, 2017; Perarnau et al. 2016; Li et al. 2016). Lastly, in order to mitigate racial bias in face recognition, Yucer et al. (2020) proposed to use CycleGAN (Zhu et al. 2017) to transfer race. In our work, instead of focusing on a single attribute, we propose to simultaneously edit multiple demographic attributes. Unlike many stand-alone age progression methods (Wang et al. 2018; Yang et al. 2018) and style transfer-based GANs (Karras et al. 2019; Huang et al. 2018), we require no additional networks to obtain the conditioning information.

In light of the growing concerns regarding algorithmic discrimination, the field of fairness-aware machine learning has attracted interest from the research community. Most of the work focuses on tabular data [e.g. the UCI Adult income dataset Dua and Graff (2017)] and tackles the problem of fair classification with regards to a protected attribute [e.g., Edwards and Storkey (2015), Madras et al. (2018)]. In such cases, both the output variable and the protected attribute are binary. For instance, a common scenario in the UCI Adult dataset is the classification of the subjects into two categories based on their wage, while being fair with regards to gender. These works have given rise to different definitions of fairness (Mehrabi et al. 2019; Verma and Rubin 2018), the most common of which are: (i) demographic parity, (ii) equalized odds, and (iii) equality of opportunity. Demographic parity ensures that the predicted label \({\hat{Y}}\) is independent of the protected attribute S, i.e., \(P({\hat{Y}}=1 | S=1) = P({\hat{Y}}=1 | S=0)\). Unlike demographic parity, a predictor \({\hat{Y}}\) that satisfies equalized odds can depend on S, but only through the target variable Y (Hardt et al. 2016), that is: \(P({\hat{Y}}=1 | S=1, Y=y) = P({\hat{Y}}=1 | S=0, Y=y), y\in \{0,1\}\). This definition implies that both the true and false positive rates will be the same for each population. Equality of opportunity is a relaxed notion of equalized odds, that only requires the true positive rates to be equal (Hardt et al. 2016), formally: \(P({\hat{Y}}=1 | S=1, Y=1) = P({\hat{Y}}=1 | S=0, Y=1)\). In this work, we use equality of opportunity to quantitatively evaluate the fairness of our face analysis models as it one of the most commonly employed approaches.

It is only very recently that fairness-aware algorithms for face analysis have attracted similar attention. The experimental results in Buolamwini and Gebru (2018) indicate that commercial gender recognition systems demonstrate a significant difference in performance between lighter male and darker female faces. This study resulted in replies from the vendors that focused on the importance of diversity in the training datasets (Raji and Buolamwini 2019). Thereafter, a series of in-processing and pre-processing methods have been proposed for fair face analysis–the former targeting unfairness at the algorithm-level and the latter at the data-level. In the former category, Alvi et al. (2018) proposed a joint learning and un-learning framework, while Kim et al. (2019) learn a fair classifier by minimizing the mutual information between the intermediate representation and the bias. These methods employ techniques from domain adaptation to learn a representation that minimizes classification loss while being invariant to the sensitive attribute. In the latter category, Sattigeri et al. (2018) extend AC-GAN (Odena et al. 2017) to generate a fair dataset, while Quadrianto et al. (2018) use an autoencoder to remove sensitive information from images. In this work, we introduce an image-to-image translation model to augment the training set, and thus, our framework is most closely related to Quadrianto et al. (2018). However, contrary to Quadrianto et al. (2018), our neural style transfer approach is able to generate naturalistic faces by translating demographic attributes instead of removing them (e.g., gender-less faces).

In this section, we introduce the notation and definitions of the operations used throughout this paper. Concretely, we denote tensors by calligraphic letters, e.g., \({\mathcal {X}}\), matrices by uppercase boldface letters, e.g., \({\mathbf {X}}\) and vectors by lowercase boldface letters, e.g., \({\mathbf {x}}\). We refer to each element of a \(K^{th}\) order tensor \({\mathcal {X}}\) by K indices, i.e., \(({\mathcal {X}})_{i_{1} i_{2} \ldots i_{K}} \doteq x_{i_{1} i_{2} \ldots i_{K}}\). Similarly, we refer the elements of matrix \({\mathbf {X}}\) as \(x_{ij}\), while \({\mathbf {x}}_j\) denotes the j-th column of the matrix. The D-dimensional vector of ones is denoted as \({\mathbf {1}}_D\).

Hadamard product The Hadamard product of \({\mathbf {A}} \in \mathbb {R}^{I \times N}\) and \({\mathbf {B}} \in \mathbb {R}^{I \times N}\) is denoted by \({\mathbf {A}} * {\mathbf {B}}\) and is the element-wise multiplication of the two matrices.

Kronecker product The Kronecker product of two matrices \({\mathbf {A}} \in \mathbb {R}^{I \times J}\) and \({\mathbf {B}} \in \mathbb {R}^{K \times L}\) is denoted by \({\mathbf {A}} \otimes {\mathbf {B}} \in \mathbb {R}^{(IK)\times (JL)}\) and is defined as:Khatri-Rao product The Khatri-Rao product of two matrices \({\mathbf {A}} \in \mathbb {R}^{I \times N}\) and \({\mathbf {B}} \in \mathbb {R}^{J \times N}\) is denoted by \({\mathbf {A}} \odot {\mathbf {B}}\). The resulting matrix is of dimensions \((IJ)\times N\), and is defined as:Tensor unfolding The mode-k unfolding of a tensor \({\mathcal {X}} \in \mathbb {R}^{I_1 \times I_2 \times \ldots \times I_K}\) reorders the elements of \({\mathcal {X}}\) into a matrix \({\mathbf {X}}_{(k)} \in \mathbb {R}^{I_{k} \times {\bar{I}}_{k}}\) with \({\bar{I}}_{k}= \prod _{\begin{array}{c} t=1 \\ t \ne k \end{array}}^K I_t \).

Mode-k vector product The mode-k vector product of a tensor \({\mathcal {X}}\in \mathbb {R}^{I_1\times I_2\times \dots \times I_K}\) and a vector \({\mathbf {v}}\in \mathbb {R}^{I_k}\) is denoted by \({\mathcal {X}}\times _k {\mathbf {v}}\in \mathbb {R}^{I_1\times \ldots \times I_{k-1} \times I_{k+1} \times \ldots \times I_K}\), defined elementwise as:CP decomposition The CANDECOMP/PARAFAC (CP) tensor decomposition (Carroll and Chang 1970; Harshman 1970) factorizes a tensor into a sum of rank-one tensors. Let \({\mathcal {X}}\) be a tensor of rank R. Its CP decomposition is:where \(\circ \) denotes the vector outer product. The matrices \(\big \{ {\mathbf {U}}_{[{k}]} = [{\mathbf {u}}_1^{(k)},{\mathbf {u}}_2^{(k)}, \cdots , {\mathbf {u}}_R^{(k)} ] \in \mathbb {R}^{I_k \times R} \big \}_{k=1}^{K}\) consist of the rank-one components. The CP decomposition of a third order tensor \({\mathcal {X}}\) is written in matrix form as Kolda and Bader (2009):

Style-conditioned Generator For simplicity of presentation, we consider the case of two attributes, namely gender and age. The generator of our framework is an autoencoder G that learns a mapping from a face \({\mathcal {X}}_{g, a}\) of gender g and age a to a synthesized face \(\mathcal {{\tilde{X}}}_{g', a'}\) of gender \(g'\) and age \(a'\). To achieve this, G uses the gender features of a target image \({\mathcal {X}}'_{g'}\) and the age features of target image \({\mathcal {X}}'_{a'}\), which are extracted using the corresponding discriminators \(D^{gen}\) and \(D^{age}\). We consider these features to capture the demographic styles \({\mathbf {s}}_g\) and \({\mathbf {s}}_a\). The styles are obtained from the target images as the activations at different layers of the discriminator networks. The age and gender styles are then fused and injected into the decoder of the generator using the proposed multi-attribute AdaIN, which is described in detail below.

Multi-attribute AdaIN The standard approach to style transfer is conditioning the generation on a single style representation. This is achieved by using AdaIN at different layers of the generator. The AdaIN operator scales and shifts the normalized activations at each layer of the decoder, in order to match a target style. We denote \({\mathcal {Z}}^i \in \mathbb {R}^{d_h^i \times d_w^i \times d_c^i}\) the activation of the i-th layer of the generator and its corresponding target style \(s^i\) which is represented by a set of two vectors, i.e., \(s^i \doteq \{ \varvec{\mu }^i, \varvec{\sigma }^i\}\). In this work, the vectors \(\varvec{\mu }^i\) and \(\varvec{\sigma }^i\) are obtained from a target image as the channel-wise first and second order moments of the activations at each layer of the discriminators. We cast vectors \(\varvec{\mu }^i\) and \(\varvec{\sigma }^i\) to have the same dimensions as \({\mathcal {Z}}\):where \({\mathcal {M}}^i, {\mathcal {S}}^i \in \mathbb {R}^{d_h^i \times d_w^i \times d_c^i}\). Then, the AdaIN operator is defined as:where \(\mu ({\mathcal {Z}}^i)\) and \(\sigma ({\mathcal {Z}}^i)\) denote the tensorized channel-wise mean and variance of the activations \({\mathcal {Z}}^i\). In this work, we aim to transfer a collection of styles (i.e., attributes) that are captured by different activations. To this end, we introduce a multi-linear style mixing model that models the multiplicative interactions (Jayakumar et al. 2020) between the activations for each attribute, by assuming a tensorial mixing structure. Concretely, given target styles for age \(s^i_a \doteq \{\varvec{\mu }^i_a, \varvec{\sigma }^i_a\}\) and gender \(s^i_g \doteq \{\varvec{\mu }^i_g, \varvec{\sigma }^i_g\}\) at layer i of the discriminators, we propose the following factorization:which can be written (as shown in Kolda (2006), Kolda and Bader (2009)) as:where \(\mathbf {{W}}^{i}_{(1)}\) is the mode-1 unfolding of tensor \(\mathbf {{\mathcal {W}}}^{i}\). The fused second order statistics are calculated from \(\varvec{\sigma }^i_a\) and \(\varvec{\sigma }^i_g\) in a similar fashion:The multiplicative interactions between the features are captured by tensors \(\{{\mathcal {W}}^{i}, {\mathcal {H}}^{i}\} \in \mathbb {R}^{d_c^i\times d_c^i \times d_c^i}\). Due to the dimensionality of the higher order tensors (\(d_{c}^{i}{3}\) parameters) the number of parameters of the fusion model scales exponentially. Indicatively, a convolutional layer of the discriminator with 256 filters would require a tensor of 16M parameters to fuse the activations (statistics) of two attributes. To avoid this, tensors \({\mathcal {W}}^i\) and \({\mathcal {H}}^i\) assume a CP decomposition, and Eqs. (10) and (11) become:where matrices \({{\mathbf {U}}^i_{[j]}, {\mathbf {V}}^i_{[j]}}\) for \(j \in \{1\dots 3\}\) consist of the rank-1 tensor of the CP decompositions of tensors \({\mathcal {W}}^i\) and \({\mathcal {H}}^i\). The resulting \(\varvec{\mu }^i\) and \(\varvec{\sigma }^i\) are then used in Eqs. (6) and (8) to modulate the style in the layers of the generator.

Discriminators In the proposed framework we utilize multiple discriminator networks–one for each attribute. Similarly to Liu et al. (2019), each of the discriminators are trained to distinguish between real and fake images of each class. Since each discriminator is responsible for one attribute, the resulting representations are also discriminative with respect to that attribute. These representations are used to condition the style transfer in the decoder of the generator at each layer. Hence, we design \(D^{gen}\) and \(D^{age}\) as mirrored decoders of the generator.

Our full objective function comprises of two terms, namely the adversarial and the reconstruction loss. The losses are described as follows.

Adversarial loss To train the generator to synthesize photorealistic images with characteristics of the desired class, we use an adversarial loss. Given an input image \({\mathcal {X}}_{g,a}\) and its translation \(G({\mathcal {X}}_{g,a},{\mathbf {s}}_{g'},{\mathbf {s}}_{a'} )\) to age \(a'\) and gender class \(g'\), conditioned on demographic styles \(s_{g'}\) and \(s_{a'}\) (obtained for each layer from \(D^{gen}({\mathcal {X}}'_{g'})\) and \(D^{age}({\mathcal {X}}'_{a'})\) respectively), we compute an adversarial loss term for each attribute y as:where \(D^y\) is the discriminator for attribute y. The combined adversarial loss for both age and gender is:Rather than having the generator minimize \({\mathcal {L}}_{adv}\), we instead adopt a feature-matching loss (Salimans et al. 2016) in order to train the generator to match the attribute patterns of particular target images. The feature-matching loss is defined as:Reconstruction loss Our framework is trained to edit the selected demographic attributes, while preserving the remaining input information. To this end, we use a cycle constraint (Zhu et al. 2017) to ensure that the synthetic images can be translated back to the original input:where a, g are the input labels for \({\mathcal {X}}\)’s age and gender respectively.

Full objective The full objective functions for D and G are:where \(\lambda _{rec}\) is the hyper-parameters for the reconstruction loss terms.

We adopt the same generator architecture as StarGAN (Choi et al. 2018), with 6 layers in the encoder-decoder and 6 residual blocks (He et al. 2016) in the bottleneck. We depart from their choice of architecture for our discriminators however, and use a simple 3 layered CNN to match the dimensionality of the decoder (more details on the model architectures can be found in the supplementary material). Following recent progress in stabilising the training of GANs (Mescheder et al. 2018), we use the \(R_1\) gradient penalty loss term in the discriminator’s objective:where \({\mathcal {X}}\) denotes the real images sampled from the true data distribution.

To further encourage the translated images to contain class-relevant attribute patterns, we find it beneficial to use the style parameters extracted from the synthetic images as the target styles in the reconstruction term in Eq. (19). Concretely, rather than translating the synthetic images back to the input images using the style parameters \(s_g,s_a\) from the input images, we instead use \({\tilde{s}}_{g}\) from \(D^{gen}(\mathcal {{\tilde{X}}}_{g,a})\) and \({\tilde{s}}_{a}\) from \(D^{age}(\mathcal {{\tilde{X}}}_{g,a})\). By requiring the synthetic images to retain sufficient style information to reconstruct the original images, the reconstruction term thus also aids in the attribute transfer.

We train our networks with the hyperparameters outlined in Liu et al. (2019): \(\lambda _{rec}=0.01\), \(\lambda _{gp}=10.0\), and opt to train our networks end-to-end with Kingma and Ba (2014), with a learning rate of \(10^{-4}\), and \(\beta _1=0.5,\beta _2=0.99\). For the multiplicative fusion layers, we set the rank of the tensors in the CP decomposition to be equal to half the number of filters at each layer.

We conduct experiments on a number of popular databases with demographic attribute annotations. Concretely, we adopt the MORPH (Ricanek and Tesafaye 2006), CACD (Chen et al. 2014), KANFace (Georgopoulos et al. 2020), and UTKFace (Zhang et al. 2017) datasets. MORPH’s second album features 55,134 near-frontal facial images of 13,618 identities–captured in a controlled setting and demonstrate little variation in expressions, illumination, or background. For our experiments on bias mitigation, we consider all three of the given annotated attributes: ‘age’, ‘gender’, and ‘race’. We use images with the attribute skin tone labelled as either dark-skinned or light-skinned, which amounts to 53,140 facial images, covering more than \(96\%\) of the dataset. In contrast, the CACD dataset features images captured in-the-wild, collected from Google Images. It contains over 160,000 images from 2000 celebrities. The dataset is annotated with regards to age. We manually annotate gender, using the provided identities. Given these two annotated attributes, we model both age and gender for our experiments on CACD. In Sect. 4.6, we train classifiers in the augmented MORPH and KANFace datasets. KANFace is the largest manually annotated video dataset of faces captured in-the-wild. We utilize the static version of KANFace that consists of 40 K images that are annotated with regards to ‘identity’, ‘age’, ‘gender’, and ‘kinship’. Similarly to CACD, we utilize only the age and gender labels. Finally, the UTKFace dataset consists of over 20,000 images of individuals aged between 0 and 116 years old. The dataset is labelled with ‘age’, ‘gender’, and ‘ethnicity’ labels. Similar to MORPH, we utilize ‘skin tone’ labels instead of ‘ethnicity’. For all datasets, we use 80% of the images for training and keep 20% for testing. The faces are grouped into 5 distinct classes: 0–18, 19–30, 31–45, 46–60, 61+. This age split is more fine-grained than the one commonly used in face aging literature (Yang et al. 2018; Wang et al. 2016; Yang et al. 2016) (i.e., 0–30, 31–40, 41–50, 51+), utilized to uncover the biases against faces under 18 and over 60 years old.

We benchmark our model against two strong baselines for multi-attribute image translation, namely StarGAN (Choi et al. 2018) and AttGAN (He et al. 2019). Additionally, since face aging has been in itself investigated in the literature, we compare against two standalone face aging GANs, namely IPCGAN (Wang et al. 2018) and CAAE (Zhang et al. 2017). Both StarGAN and AttGAN translate an input image to multiple target domains by conditioning on a one-hot encoded label vector. StarGAN differs from AttGAN in that it employs a cycle-constraint reconstruction loss, while AttGAN utilizes a so-called attribute classification constraint to encourage correct attribute classification in synthesized images. On the other hand, IPCGAN uses a separate pre-trained age classifier to enforce the aging features of a particular target age class on the translated faces, while CAAE learns face aging by traversing a manifold. We benchmark our model against these 4 baselines both qualitatively and quantitatively in the sections that follow. The baselines were implemented using the authors’ publicly released code where available.²

In this section, we propose to benchmark the diversity-enhancing capabilities of our model with regards to the available demographic attributes. In particular, we translate a test-set of 1000 faces to all attribute classes and calculate the diversity metrics of the augmented set, as proposed in Merler et al. (2019). We measure the Shannon H (ShH) and Simpson D (SiD) diversity indices and the Shannon E (ShE) and Simpson E (SiE) evenness indices. The metrics are calculated as follows:where S denotes the number of classes and \(p_i\) is the probability of each class. Higher diversity indices indicate a more diverse dataset, while evenness indices closer to 1 indicate a label distribution that is closer to uniform. The age and gender labels for each image are obtained using the Face++ public API,³ while the skin tone labels are obtained using the Clarifai⁴ API.

In the case of the age attribute, we adopt the standard protocol as described in the age progression literature (Yang et al. 2018; Wang et al. 2016; Yang et al. 2016), and translate only faces from the youngest age group (under 18 years old) to the rest of the age groups; that is, we perform face aging.

We compare the diversity metrics of the augmented sets to those of the original test images (GT) and present the results in Tables 1, 2 and 3. Along with our method, both StarGAN and AttGAN are capable of synthesizing a dataset with a distribution of labels close to uniform in the case of the binary attributes skin tone and gender. For the more challenging age attribute however, the superiority of our method and modelling choice is evident, with our augmented datasets’ age labels being much more evenly spread than those of the baselines, especially the age progression ones.

This is particularly pronounced in the case of the biased MORPH test-set (Table 1), where the difference between the ground truth images and the synthetic ones is significant. Our model consistently outperforms the baselines for all attributes in both datasets.

In this section, we investigate the effect that a lack of diversity entails with respect to model performance. To this end, we fine-tune and test a state-of-the-art gender recognition model (Rothe et al. 2018) on the MORPH and KANFace datasets. By measuring the True Positive Rate (TPR) for each demographic subpopulation, we are able to uncover the bias of the model with regards to age and skin tone. In particular, Figs. 7 and 8 show that the model trained on MOPRH is biased against dark-skinned females (TPR of 40% for dark-skinned females over 60 years old), while the model trained on KANFace is biased against male faces under 18 (TPR of 64%). Age and skin tone bias has been studied for the task of face recognition in Nagpal et al. (2019), as well as in human perception in psychology literature (Schaich et al. 2016; Bothwell et al. 1989).

We propose to mitigate the bias in both models by augmenting the training set using the proposed image translation method. In particular, we train the models on the diverse augmented sets and evaluate the fairness of the trained classifiers. The MORPH dataset is augmented using the models trained on MORPH in Sect. 4.4.1. For the KANFace dataset we use the models trained on CACD.

Of the various fairness metrics discussed in Sect. 2.3, we opt to quantify fairness using Equality of Opportunity (EO) (Hardt et al. 2016), which is defined as the difference in TPRs between the subpopulations.In particular, we report the EO score between each (age, gender) class for KANFace and (age, gender, skin tone) for MORPH.

We present the TPRs (\(\uparrow \)) and EO (\(\downarrow \)) on the ground-truth and synthetic test-sets in Figs. 8 and 7 for MORPH and KANFace respectively. Despite StarGAN’s images having strong class-discriminative features as established in Sect. 4.4, we demonstrate here that training on their artifact-ridden images leads to even more biased classifiers. Similarly, whilst training on AttGAN-generated training sets can improve the TPR for dark-skinned women over 60 years old by \(20\%\) for MORPH, the dataset bias is exacerbated in KANFace when training on augmentations from this model (compared to our augmentations that lead to almost half the EO for males under 18).

Overall, the results indicate that the proposed framework is the only method able to mitigate all of the demonstrated dataset biases in generating more diverse data that can be used to train fairer classifiers.

Due to the generative nature of the proposed framework, our method suffers from its dependence to external data. We investigate this inherent shortcoming of our method by conducting two experiments, which are presented as follows.

Training on limited data Firstly, we investigate the impact of the size of the training set of the GAN on the bias of the trained classifier. Concretely, we show in Fig. 10 the results of training a classifier with our synthetic data, when our GAN is trained on training subsets of different sizes. For each model, we report the equality of opportunity for the most underrepresented class, i.e., male faces under 18 on KANFace. We notice that when our model is trained on less than 25% of the training images, the bias of the trained classifier is even magnified. It should be noted that this problem is not specific to our method, but one that will plague all generative model-based bias mitigation methods via augmentation.

Augmentation with real data Since the use of the proposed framework assumes the existence of an external training set, we explore the option of augmenting the training set of the biased classifier using real images from the external set. In particular, we use the CACD dataset as the external set, and try to flatten the distribution of the training set of the classifier (i.e., KANFace). However, the support for each class in the external set is not sufficient to do so in the extreme cases of people under 18 and over 60 years old. The resulting training distributions along with equality of opportunity scores are presented in Fig. 11. The results highlight the advantage of using the proposed data augmentation method, instead of directly training on the external data.

Learning the real distribution of face images is a task of high data and computational complexity (Arora and Zhang 2017). Therefore, generative models have to rely on their inductive bias in order to generalize on unobserved modes of variation. However, any finite dataset is inherently biased, and training a generative model on a biased dataset can even exacerbate this bias. Indicatively, GANs and VAEs are not able to generate images of unobserved attribute combinations (Zhao et al. 2018). Different approaches for debiased generation have been proposed in the literature and include importance reweighting (Grover et al. 2019a, b) and modeling the multiplicative interactions (Georgopoulos et al. 2020). Therefore, the proposed and any generative model-based approach to data augmentation should be utilized with caution regarding the distribution of the training set.