Abstract
Faces are a rich source of information for humans and a substantial amount of behavioral science research uses face stimuli to assess person perception. Unfortunately, this body of research is limited by an overreliance on young, predominantly white faces normed on young adult perceivers. To address these limitations, we created an open-access database of AI-generated faces that represents the same individuals at three life stages (young adulthood, middle age, and older adulthood) including equal numbers of males and females. Using advanced generative algorithms, the approach digitally aged 62 young individuals, thus preserving identity-specific features while realistically portraying age-related changes. The resulting database comprises 186 images. Each image has been age-normed and validated for authenticity. Although the database will be useful for many research questions, the stimuli are especially well-suited for research on age comparisons because the same individuals can be presented at different ages.
Subject terms: Human behaviour, Sociology
Background & summary
Humans rapidly form impressions of other people based on appearance, which contributes to assumptions about warmth, trustworthiness, fitness, and dominance and signals approach or avoidance1–4. Face perception has been so central to survival throughout evolution, a dedicated brain region extracts facial information in milliseconds5.
With age, facial features change in ways that have important implications for impression formation. Faces of older individuals are generally perceived as less attractive and less healthy than younger individuals. Gender biases also have been documented; with age, men are perceived as relatively more intelligent, confident, and trustworthy than women6. Relatedly, age estimations and assessments of cognitive competence are less accurate for older faces7,8.
Age-related changes in facial appearance have significant social implications. Reduced facial appeal, such as less favorable fitness and competence, have been associated with hiring decisions9. Additionally, age-related facial changes make it more difficult for people to decode emotions in older faces, often leading to a bias where negative emotions like sadness are more readily attributed to older individuals10. People also tend to misinterpret neutral facial expressions of older adults for expressions of pain, reducing the accuracy of pain assessment11. Research on behavioral confirmation shows that attributed social behaviors associated with facial features change subsequent social interactions in ways that reinforce perceived characteristics. Negative age stereotypes based on facial cues may lead younger people to assume older people are unhappy and thus avoid speaking with them. Failing to see a smile, the associated behavior subsequently reinforces stereotypes about emotional disengagement12,13.
Current research in the field of impression formation based on faces is limited by three important constraints: an overreliance on young adult perceivers, limited age ranges and racially homogenous experimental stimuli, and a lack of data about changed face perceptions with age14. While there are many available databases of faces, most consist of young, predominantly white faces posing various emotions, or cross-sectional samples of different people at different ages15. The few databases containing age-progressed images of the same individuals have been sourced primarily from criminal conviction records, such as the MORPH database16, limiting generalizability. Until recently, creating controlled age-progressed face databases has been costly both in time and resources, particularly when standardizing pose, lighting and expression in face photographs across ages.
One database developed by Ebner and colleagues, featuring white European faces of different people of different ages expressing a variety of emotions, was published in 201017. Studies using this database have revealed that both younger and older adult perceivers rate older faces as less attractive compared to younger faces18. Another study found that older faces were perceived as younger when they displayed a happy expression compared to when they expressed a negative emotion7. While this resource enables comparison between perceptions of older and younger faces, auxiliary, confounding variables (hairstyle, gender, eye color, etc.) introduce potential confounds when comparing older and younger faces of different individuals. Previous research underscores that even subtle facial characteristics, such as eyes positioned closer together, can significantly influence social attributions by signaling lower trustworthiness but higher competence19. Consequently, distinguishing age-related impressions from faces from effects driven by auxiliary facial features remains challenging.
These persistent methodological constraints significantly limit the understanding of how face perception changes across the life span and contribute to impression formation from faces. To overcome these limitations and encourage more research on face perception across ages, we developed a diverse, open-access face database, featuring AI-generated faces that vary across gender, age and are more racially ambiguous. Rather than compiling photographs of different individuals who inevitably vary across multiple dimensions, we created images of individual faces that we digitally aged using AI-algorithms. This approach yields age-progressed images of the same individual while controlling for auxiliary factors that typically confound cross-sectional comparisons.
The database comprises 62 individuals, each depicted at three different life stages: young adulthood (18–35), middle-age (36–64) and older age (65+). Individuals depicted in the dataset span a range of skin tones from very pale to dark, including racially ambiguous faces. Gender is balanced with an equal distribution of male and female faces. Each image maintains identity-specific features while realistically portraying age-related facial changes. This controlled progression allows researchers to isolate the specific impact of aging on social perception while holding other factors constant.
This database has broad potential reuse in psychological and behavioral science research. Researchers can use these stimuli to compare faces of the same person at different ages in developmental research; study impression formation of faces at different ages in social psychological research; examine decision-making processes based on faces in behavioral economics research; examine population characteristics in cross-cultural research and investigate memory and recognition accuracy for younger and older faces in cognitive psychology research, among other research applications. With this resource we aim to accelerate research on age-comparisons.
Methods
Input data
The initial dataset of young adults was sourced from face photographs available at https://thispersondoesnotexist.com. This website generates random realistic faces using the StyleGAN2 algorithm20. Each page refresh generates a new, unique face. The selection process for inclusion in the present database adhered to the following criteria: (1) images featured young adults, aged approximately 18 to 35 years; (2) images did not include headwear, (sun)glasses or excessive make-up; (3) photographs featured neutral backgrounds, and (4) facial poses appeared to look directly into the camera. All individuals were smiling with and without visible teeth. We selected equal numbers of male and female faces and given the relatively complex and fluid nature of racial identities ensured faces were racially heterogeneous. While the StyleGAN2 algorithm inherits biases present in its training data (the NVIDIA Flickr-Face-HQ dataset)20,21, we carefully curated our selections to ensure faces represented a range of skin color and facial features. This approach provides researchers with a more diverse and representative set of facial stimuli that better reflects the diversity found in real-world populations. The final set comprises 62 images of individuals, equally balanced by gender. These images provided the foundation for our subsequent age-progression manipulations.
Computational processing
Sixty-two young adult faces were digitally aged to create versions of the same face at approximately 50 and 80 years old using the Style-based Age Manipulation (SAM) algorithm22. SAM is an image-to-image translation method that uses a pre-trained age regression network to guide the encoding of images. The encoder converts the original face image into the latent space of a pre-trained unconditional GAN, and subsequently latent codes associated with specific target ages are generated. The target ages can be passed as arguments to this encoder.
Although SAM produces realistic aged versions of young adult faces23, it does not automatically modify hair color - a significant limitation since hair typically changes with age. To address this, we added a style-mixing technique using reference images of older individuals with gray or white hair. Specifically, we modified the reference-guided inference script provided by the authors of SAM to selectively apply hair characteristics from reference images while maintaining the identity of the original face. Reference images were systematically obtained by feeding a selection of our AI-generated images to GPT-4 (using ChatGPT 4.0) using the prompt: “Can you make the person in this picture 80 years old with white/gray hair?”. Using GPT-4, we obtained 7 older male and 7 older female reference images of faces across a range of ethnicities with gray or white hair.
We ran our reference-inference script with each of the generated reference images (male images for male faces, and female images for female faces). We modified the alpha level and layers of the reference-guided inference script according to the desired age. For middle-aged faces, we set the alpha level at .04, denoting that it should mix the reference picture at a strength of .04 into the original picture. For the older adult faces, we set the alpha level at .06, indicating that it should mix the reference picture at a strength of .06 into the original picture. In both age conditions, we instructed the model to target only the middle layers of the model (7–9), where changes to the hair style and color are most impacted22. The final step was to remove unwanted earrings from pictures. SAM tends to generate earrings where there were none in the original picture, so we edited the middle-aged adult and older adult pictures manually to remove earrings when they appeared in one ear only and both ears were clearly visible.
We executed this aging pipeline (see Fig. 1) on Stanford’s High-Performance Computing (HPC) cluster using a single, 16GB GPU to process all 62 faces. The SAM code, publicly available on Github from the original authors22, was downloaded and implemented with minimal modifications to accommodate our specific reference-guided hair color transformation requirements. Figure 2 shows two examples of the output from this aging pipeline.
Fig. 1.
Face age-manipulation pipeline.
Fig. 2.
Two examples of age-progressed faces in the database.
Data Records
The database24 is stored in its own repository on the Open Science Framework with the following file structure: a folder with three subfolders containing.png images of all 62 individuals at younger, middle-aged and older ages, a folder to replicate our aging pipeline, including two subfolders with reference images to replicate our results (male, female) as well as the Python code to run the aging pipeline, and a .csv file with output from the technical validation: the mean estimated ages, hit ratios and confidence scores for each face.
Technical Validation
In order to make the database useful for research on age comparisons, we conducted a validation study targeting age-estimates for each image and a validation study to assess whether images could be reliably distinguished from photographs of real people.
Ethics and institutional review
Both technical validation studies were approved by the Stanford University Ethics committee (protocol number 76753). Participants were recruited through Prolific, an online survey platform that automatically deidentifies participants by assigning each a unique participant ID rather than collecting personal identifiers. Prior to beginning the study, participants provided written informed consent for taking part in the study and for having their anonymized data made available for analysis and publication. For both studies, only aggregated data are reported, maintaining participant anonymity.
Age norming
To obtain age-estimates for each image, we recruited 300 U.S.-based participants on Prolific, an online survey platform. To ensure our participant sample included a diverse range of ages, we recruited 100 participants into three age groups: 18–34 representing young adults, 35–64 representing middle-aged adults, and 65 + representing older adults. While the cutoff score for older adults is clearly defined in the NIA guidelines25, the cutoff of 35 for middle-aged adults is somewhat arbitrary, as research looking to compare age groups typically places the cutoff somewhere between 30 and 40 years. We broadly define younger adults to be between 18 and 34 years old and middle-aged adults between 35 and 64 years old. While we restricted our sample to participants residing in the U.S., we note that the U.S. is highly culturally and ethnically diverse26. Nonetheless, future studies with international samples would further strengthen the utility of the images for cross-cultural research. Each participant estimated the ages (in numbers) for 62 pictures out of 186 total pictures; a random selection of one age for each individual. In total, each of the 186 AI-generated images received on average 100 ratings.
Results
Six participants were excluded from the dataset because they failed one or more attention checks, leaving 294 participants in our final sample. Participants’ mean age was 48.4 (SD = 18.17), 47% of participants identified as female and 53% as male and 68% of the sample was White. Table 1 below summarizes mean age estimates for all three AI-generated face age groups.
Table 1.
Means, standard deviations and 95% CIs of age estimates for all faces at three different ages.
| M | SD | 95% CI | |
|---|---|---|---|
| Younger faces | 31.58 | 7.85 | [31,38, 31.77] |
| Middle-aged faces | 52.69 | 8.52 | [52.48, 52.91] |
| Older faces | 70.35 | 8.89 | [70.13, 70.58] |
To assess whether ratings reliably differed by targeted age group (younger, middle-aged and older), gender of the face and rater age, we conducted a multilevel model predicting the age estimate from the targeted age group of a face, face gender, and the age of the rater, including a random intercept for rater (since raters rated several faces) as well as a random intercept for individual (as some faces of individuals may be easier or more difficult to estimate across ages). Analyses were executed using R base, lme4 and marginaleffects packages27–29. Table 2 shows the model output.
Table 2.
Output of fixed and random effects of two multilevel models predicting age estimates.
| Model 1 | Model 2 | |||
|---|---|---|---|---|
| b (SE) | p | b (SE) | p | |
| Fixed effects | ||||
| Intercept | 51.56 (0.38) | <0.001 | 31.64 (0.65) | <0.001 |
| Rater age (middle vs. young) | –0.89 (0.60) | 0.139 | ||
| Rater age (old vs. young) | –0.94 (0.63) | 0.133 | ||
| Face age group (middle vs. young) | 22.65 (0.25) | <0.001 | ||
| Face age group (old vs. young) | 42.05 (0.25) | <0.001 | ||
| Face gender (male vs. female) | 1.65 (0.59) | 0.006 | ||
| Rater gender (male vs. female) | –0.61 (0.48) | 0.209 | ||
| Rater age × Face age group (middle × middle vs. young) | 0.76 (0.31) | .013 | ||
| Rater age × Face age group (old × middle vs. young) | 0.20 (0.32) | 0.531 | ||
| Rater age × Face age group (middle × old vs. young) | 0.03 (0.31) | 0.912 | ||
| Rater age × Face age group (old × old vs. young) | –0.34 (0.32) | 0.281 | ||
| Face age group × Face gender (middle × male vs. female) | –3.58 (0.25) | <0.001 | ||
| Face age group × Face gender (old × male vs. female) | –6.23 (0.25) | <0.001 | ||
| Random effects | Variance | SD | Variance | SD |
| Residual | 302.98 | 17.41 | 48.22 | 6.94 |
| Rater (intercept) | 14.62 | 3.82 | 16.02 | 4.00 |
| Individual face (intercept) | 4.82 | 2.20 | 4.82 | 6.94 |
Model 1 is the intercept-only model with the following formula: age estimate ~ 1 + (1 | rater) + (1| individual face). Model 2 is the full model with the following formula: age estimate ~ face age group x rater age group + face age group x face gender + rater gender + (1 | rater) + (1| individual face). Significant effects are highlighted in bold.
No main effects by rater age group were observed, suggesting that younger, middle-aged and older raters did not differ significantly in their age estimates. We observed a main effect of targeted age group, as we expected, in which younger, middle-aged and older faces had very different age estimates (b = 22.65, p < 0.001 for middle-aged faces and b = 42.05, p < 0.001 for older faces). We also observed a main effect of face gender, whereby male faces were rated significantly higher in age compared to female faces (b = 1.65, p < 0.01). We observed a significant interaction between rater age and face age group, whereby middle-aged raters, compared to younger raters, rated middle-aged faces significantly older than younger raters (b = 0.76, p < 0.05). We did not observe a reliable interaction between rater age and face age group for other rater ages, suggesting that overall, participants’ age did not influence the age estimates for younger, middle-aged and older faces.
We also observed a significant interaction effect between face age group and face gender, whereby middle-aged male and older male faces were rated significantly younger than their same-age female counterparts. Figure 3 below plots the predictions for the interaction between face age group and face gender. Younger male faces were rated older than young female faces. Previous research indicates that estimating age from faces is more difficult the older the face is and that on average, male faces receive older estimates than female faces7. Our contrasting finding that middle-aged and older female faces were perceived as older than male faces may be explained by the absence of appearance-enhancing factors in our stimuli. In daily life, women often utilize skincare routines and cosmetics that can mask visible signs of aging. Since our model did not incorporate these common aging-masking effects, the female faces may have appeared older than what participants in our sample may be used to from observing older faces in everyday life. Figure 4 below illustrates the mean age-estimates for each AI-generated face, at young, middle-age and older ages and for male and female faces.
Fig. 3.
Predicted age estimates of the interaction between targeted face age group and face gender predicting age in Model 2. Colors represent the gender of the faces.
Fig. 4.
Perceived age for each face per targeted age group (younger, middle-aged, and older faces). Colors represent the gender of the individual face. The boxplots show the mean distribution of age estimates for each face age group and gender.
Authenticity validation
To validate the pictures for authenticity, we asked a new sample of participants on Prolific (N = 300) to judge whether a presented image was a photograph of a real person or an AI-generated image. In addition, we asked participants to rate how confident they felt in their judgement.
We recruited 300 U.S.-based participants on Prolific. To ensure the participant sample included participants from different ages, we recruited participants along three age groups: young adults (18–34), middle-aged adults (35–64) and older adults (65+). Participants were tasked with identifying whether a face presented on the screen was of a real human or AI-generated. After making their decision, participants were asked how confident they were in their judgment, which they indicated on a scale from 0 (not confident at all) to 100 (highly confident). Participants were presented with a random selection of 31 images from our database, supplemented with 31 images of real people. These images were obtained from the NVIDIA Flickr-Faces-HQ (FFHQ) dataset20, which is a set of images of human faces sourced from the website Flickr that contain licenses that allow free use, redistribution, and adaptation for non-commercial purposes. We removed the background from all images to minimize background cues that might signal that an image is of a real person. In addition to the human faces and images from our database, we supplemented the study with 18 images that featured implausible facial characteristics (such as an extra eye, extra ear, excessively long earlobe, etc.), that can only be rendered by AI. These images served as attention checks for our study, to make sure that participants paid equal attention to all images.
For each image, we calculated a hit ratio, i.e. the number of hits divided by the number of hits and misses. In addition, we calculated a mean confidence rating for each image, as well as a correlation between correctness and confidence score for all human trials and a correlation between correctness and confidence for all trials of the AI-generated images, excluding the attention checks.
Results
We removed 50 participants from the dataset because they failed to identify two or more attention checks as AI-generated images. Participants’ mean age was 47.82 (SD = 16.99). A little over half of our participant sample identified as female (58%) and 72% of the sample was White.
Across all AI-generated images from our database, the majority had hit ratios below 0.5 (chance level), meaning participants incorrectly identified these AI faces as real more often than correctly identifying them as AI-generated. In contrast, participants were more accurate at identifying real people as authentic (hit ratios above 0.5). This pattern demonstrates that our AI-generated faces achieve high perceptual authenticity, as participants frequently mistook them for photographs of real individuals. This is in line with previous work that suggests that people cannot distinguish between a face generated by the StyleGAN algorithm and a picture of a real face30,31. We add to this literature by observing that age-modifications of AI-generated images are deemed equally realistic.
Figure 5 below shows the hit ratios for each picture across three age groups; younger, middle-age, and older age. While most AI faces were misidentified more often than chance as real across all age versions, 13 AI-generated individuals had at least one age version with a hit ratio above 0.5, indicating that participants correctly identified those specific images as AI-generated more often than chance. This variability suggests that while our AI faces are generally highly realistic, some individual faces or age modifications may contain subtle artifacts that make them more detectable as AI. Full statistics (hit ratio with 95% CIs and confidence judgement) for each image can be obtained from the .csv file in the database repository24.
Fig. 5.
Hit ratios for all pictures. The top panel shows hit ratios for the faces of real people and the bottom panel shows hit ratios for the 62 AI-generated individuals. Each individual is labeled on the x-axis. The colors and shapes represent the age group of the individual, so that each AI-generated individual has one line on the x-axis with three datapoints, representing the individual at young (dots), middle (triangles) and older age (squares). The dashed red line indicates the 0.5 hit ratio mark. Values below the 0.5 line indicate a low hit ratio, meaning a face was most often not identified as real or AI correctly. Values above the 0.5 line indicate a high hit ratio, meaning a face was most often correctly identified as real or AI.
On average, people are fairly confident in their judgments of real and AI-generated images. The average confidence for real faces is 69.0 (SD = 21.5), and for our AI-generated faces in the database it is 70.1 (SD = 21.0) and for the attention checks confidence is 92.1 (SD = 15.3). Table 3 below lists the confidence scores for each image type and image age group.
Table 3.
Mean confidence scores and standard deviations for each image group, AI-generated and Real images, across three image age groups.
| Age | M | SD | |
|---|---|---|---|
| AI-generated | Younger | 71.67 | 20.74 |
| Middle-age | 71.15 | 20.81 | |
| Older | 69.48 | 21.50 | |
| Real | Younger | 68.78 | 21.71 |
| Middle-age | 68.85 | 21.35 | |
| Older | 69.52 | 21.23 | |
| Attention checks | — | 92.13 | 15.26 |
Attention check images, confidence scores and hit ratios are highly correlated (r = 0.96***, 95% CI [0.89, 0.98]). For real faces, there is no significant correlation between hit ratio and confidence score (neither for young, middle, and old). For young and middle-aged AI-generated faces, there is a significant negative correlation between hit ratio and confidence score, suggesting that people are more confident when they are mislabeling AI-generated faces as real than when they are labeling AI-generated faces as AI (young: r = −0.52***, 95% CI [−0.68, −0.30], middle: r = −0.27*, 95% CI [−0.49, −0.02]). For older AI-generated faces, there is no significant relationship between hit ratio and confidence scores (r = −0.06, p = 0.60, 95% CI [−0.31, 0.18]). This finding has been observed in the literature previously; participants are more confident in hits (correctly identified) than misses (incorrectly identified) images32.
Acknowledgements
This research was supported by a grant from the National Institute on Aging (R37AG00881630) to the second author.
Author contributions
A.P. (Conceptualization, Methodology, Stimuli development, Experimental analysis, Writing), L.L.C. (Conceptualization, Methodology, Funding acquisition, Writing).
Data availability
The database24 is available in its own repository on the Open Science Framework at https://osf.io/vfn5c/. The repository includes face images at three ages (younger adults, middle-aged adults and older adults), Python code for the aging pipeline, and a .csv file with all validation data.
Code availability
The code used to generate the images in the database was adapted from Alaluf and colleagues22. Their code can be obtained from https://github.com/yuval-alaluf/SAM.
A Python script with our modifications to their code is available in the database repository24. The script is dependent on the SAM Github project22 to work.
Competing interests
The authors declare no competing interests.
Footnotes
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
- 1.Ambady, N. & Rosenthal, R. Thin slices of expressive behavior as predictors of interpersonal consequences: A meta-analysis. Psychological Bulletin111, 256–274, 10.1037/0033-2909.111.2.256 (1992). [Google Scholar]
- 2.Todorov, A. The social perception of faces. In The SAGE Handbook of Social Cognition 96–114, 10.4135/9781446247631.n6 (SAGE Publications Ltd, 2012).
- 3.Zebrowitz, L. A. First impressions from faces. Curr Dir Psychol Sci26, 237–242, 10.1177/0963721416683996 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Zebrowitz, L. A. & Montepare, J. M. Social psychological face perception: Why appearance matters. Social and Personality Psychology Compass2, 1497–1517, 10.1111/j.1751-9004.2008.00109.x (2008). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Kanwisher, N. & Yovel, G. The fusiform face area: a cortical region specialized for the perception of faces. Philos Trans R Soc Lond B Biol Sci361, 2109–2128, 10.1098/rstb.2006.1934 (2006). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Sutherland, C. A. M. & Young, A. W. Youth and Beauty: Age-Based Trait Impressions from Faces. In Emotion Communication by the Aging Face and Body (eds. Hess, U., Adams, Jr., R. B. & Kleck, R. E.) 171–196, 10.1017/9781009209656.008 (Cambridge University Press, 2023).
- 7.Voelkle, M. C., Ebner, N. C., Lindenberger, U. & Riediger, M. Let me guess how old you are: Effects of age, gender, and facial expression on perceptions of age. Psychology and Aging27, 265–277, 10.1037/a0025065 (2012). [DOI] [PubMed] [Google Scholar]
- 8.Zebrowitz, L. A. et al. Older and younger adults’ accuracy in discerning health and competence in older and younger faces. Psychology and Aging29, 454–468, 10.1037/a0036255 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Kaufmann, M. C., Krings, F., Zebrowitz, L. A. & Sczesny, S. Age bias in selection decisions: The role of facial appearance and fitness impressions. Front. Psychol. 8, 10.3389/fpsyg.2017.02065 (2017). [DOI] [PMC free article] [PubMed]
- 10.Folster, M., Hess, U. & Werheid, K. Facial age affects emotional expression decoding. Front. Psychol. 5, 10.3389/fpsyg.2014.00030 (2014). [DOI] [PMC free article] [PubMed]
- 11.Matheson, D. H. The painful truth: Interpretation of facial expressions of pain in older adults. Journal of Nonverbal Behavior21, 223–238, 10.1023/A:1024973615079 (1997). [Google Scholar]
- 12.Sabik, N. J. Is social engagement linked to body image and depression among aging women? Journal of Women & Aging29, 405–416, 10.1080/08952841.2016.1213106 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Barnett, M. D., Moore, J. M. & Edzards, S. M. Body image satisfaction and loneliness among young adult and older adult age cohorts. Archives of Gerontology and Geriatrics89, 104088, 10.1016/j.archger.2020.104088 (2020). [DOI] [PubMed] [Google Scholar]
- 14.Mondloch, C. J., Twele, A. C. & Thierry, S. M. We need to move beyond rating scales, white faces and adult perceivers: Invited Commentary on Sutherland & Young (2022), understanding trait impressions from faces. British Journal of Psychology114, 504–507, 10.1111/bjop.12619 (2023). [DOI] [PubMed] [Google Scholar]
- 15.Princeton University. Face Image Database. at <https://libguides.princeton.edu/facedatabases> (2022).
- 16.Ricanek, K. & Tesafaye, T. MORPH: a longitudinal image database of normal adult age-progression. in 7th International Conference on Automatic Face and Gesture Recognition (FGR06) 341–345, 10.1109/FGR.2006.78 (2006).
- 17.Ebner, N. C., Riediger, M. & Lindenberger, U. FACES–a database of facial expressions in young, middle-aged, and older women and men: development and validation. Behav Res Methods42, 351–362, 10.3758/BRM.42.1.351 (2010). [DOI] [PubMed] [Google Scholar]
- 18.Ebner, N. C. et al An adult developmental approach to perceived facial attractiveness and distinctiveness. Front. Psychol. 9, 10.3389/fpsyg.2018.00561 (2018). [DOI] [PMC free article] [PubMed]
- 19.Olivola, C. Y., Funk, F. & Todorov, A. Social attributions from faces bias human choices. Trends in Cognitive Sciences18, 566–570, 10.1016/j.tics.2014.09.007 (2014). [DOI] [PubMed] [Google Scholar]
- 20.Karras, T. et al Analyzing and improving the image quality of StyleGAN. in Proc. CVPR, 10.1109/CVPR42600.2020.00813 (2020).
- 21.Huber, M., Luu, A. T., Boutros, F., Kuijper, A. & Damer, N. Bias and diversity in synthetic-based face recognition. in 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) 6203–6214, 10.1109/WACV57701.2024.00610 (2024).
- 22.Alaluf, Y., Patashnik, O. & Cohen-Or, D. Only a matter of style: age transformation using a style-based regression model. ACM Trans. Graph.40, 45:1–45:12, 10.1145/3450626.3459805 (2021). [Google Scholar]
- 23.Han, S. et al. A Chinese face dataset with dynamic expressions and diverse ages synthesized by deep learning. Sci Data10, 878, 10.1038/s41597-023-02701-2 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Pot, A. & Carstensen, L. L. A Generated Image Repository of Aging Faces (GIRAF), 10.17605/osf.io/vfn5c (2025). [DOI] [PMC free article] [PubMed]
- 25.Age | National Institutes of Health (NIH). at <https://www.nih.gov/nih-style-guide/age> (2025).
- 26.Vespa, J., Medina, L. & Armstrong, D. M. Demographic Turning Points for the United States: Population Projections for 2020 to 2060. 25–1144 at <https://www.census.gov/content/dam/Census/library/publications/2020/demo/p25-1144.pdf> (U.S. Census Bureau, 2020).
- 27.R Core Team. R: A language and environment for statistical computing. at <https://www.r-project.org/> (2021).
- 28.Bates, D., Mächler, M., Bolker, B. & Walker, S. Fitting linear mixed-effects models using lme4. Journal of Statistical Software67, 1–48, 10.18637/jss.v067.i01 (2015). [Google Scholar]
- 29.Arel-Bundock, V., Greifer, N. & Heiss, A. How to Interpret Statistical Models Using marginaleffects for R and Python. Journal of Statistical Software111, 10.18637/jss.v111.i09 (2024).
- 30.Bozkir, E., Riedmiller, C., Skodras, A. N., Kasneci, G. & Kasneci, E. Can you tell real from fake face images? Perception of computer-generated faces by humans. ACM Trans. Appl. Percept.22, 6:1–6:23, 10.1145/3696667 (2024). [Google Scholar]
- 31.Nightingale, S. J. & Farid, H. AI-synthesized faces are indistinguishable from real faces and more trustworthy. Proceedings of the National Academy of Sciences119, e2120481119, 10.1073/pnas.2120481119 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Miller, E. J. et al. AI hyperrealism: Why AI faces are perceived as more real than human ones. Psychol Sci34, 1390–1403, 10.1177/09567976231207095 (2023). [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
The database24 is available in its own repository on the Open Science Framework at https://osf.io/vfn5c/. The repository includes face images at three ages (younger adults, middle-aged adults and older adults), Python code for the aging pipeline, and a .csv file with all validation data.
The code used to generate the images in the database was adapted from Alaluf and colleagues22. Their code can be obtained from https://github.com/yuval-alaluf/SAM.
A Python script with our modifications to their code is available in the database repository24. The script is dependent on the SAM Github project22 to work.