Ethnic Differences in Women’s Perception of Simulated Facial Aging over a 15-Year Horizon: A GAN-Based Model Approach

Abstract

This study assessed the accuracy of a long-term AI-based projection of signs of facial aging and their acceptance by consumers. Standardized photographs of 25 Chinese and 25 French women were first taken at T0 and graded using ethnic-specific skin aging atlases. An AI-based algorithm then aged the photographs by 10 (T10) and 15 (T15) years. A total of 246 women from China, France, and Thailand compared these images in pairs (T0 vs. T0 + 10 or T0 + 15) and provided feedback on their overall impressions, realism, and psychological acceptance via a questionnaire. Besides lower face ptosis (p < 0.01), the simulated images revealed that regardless of ethnicity, there were no significant differences in grading. Irrespective of ethnic background, 62–78% of overall panelists found the projections realistic and liked them, while 85–96% of panelists (Chinese and French) were willing to test them. A total of 47% of Thai panelists were reluctant to try, while 4–14% found it scary. This indicated some degree of cultural influence. This study confirms women’s acceptance of future facial appearance with some degree of cultural divergence. It also highlights a valid methodology to explore skin aging for a more realistic and personalized cosmetic improvement and innovation.

1. Introduction

Previous research has attempted to clinically assess and grade the changes in facial features that occur progressively with aging, either chronologically or induced by the sun [1,2,3,4,5,6,7,8,9,10,11,12]. These studies, conducted on both genders, different ethnicities, and different ages, followed different methodologies. With the help of skin aging atlases, which grade 15 to 20 facial features individually [13,14,15,16], these previous studies include a global clinical grading performed in vivo or on standardized photographs, zoom in on some facial features, and follow their alterations with age. More recently, an AI-based algorithm was developed which is capable of automatically grading photos taken by the subjects themselves with their smartphones.

This interesting approach, not influenced by artificial lighting conditions and validated by dermatologists/experts, made it possible to extend the study to thousands of subjects of different ethnicities [1], followed by an analysis under anonymous conditions by the algorithm within seconds.

In brief, irrespective of the methodology used, all studies converged in establishing that the slow progression (by decades) of facial aging showed (i) differences between gender and ethnicity, (ii) that each facial feature possesses its own rate of increased severity along the lifespan depending on gender and ethnicity, and (iii) a better understanding of the specific impacts of solar exposure, chronic aerial pollution, or stress on some facial signs of aging [16,17].

Irrespective of the adopted methodology, all these studies graded a given facial status at a specific point in time “t” (a snapshot) across subjects of different ages (e.g., 18–70 y) as a necessary step to assess the previous impacts of different individual factors (gender, ethnic, sun exposure frequency, lifestyle, food, etc.) on the aging process. The aspect of this individual “past history” led us to consider conducting this study by using an inverse approach, i.e., to illustrate the future facial status of young and middle-aged women, 10 or 15 years after the initial “snapshot”. Such an objective was achievable through a recently developed dedicated algorithm that integrates the references of the skin aging atlases on women of different ethnicities along their lifespan.

An exploratory study was therefore undertaken, taking standardized facial photographs of women from two small cohorts of 35–49-year-old East Asian and European women as the initial facial appearances. These snapshots were further artificially modified by an AI-based algorithm at a 10–15-year horizon (2034 to 2039) to simulate the facial aging that occurs during this period. Preference was given to a 10–15-year simulation as it is applicable to all future consumers regardless of their ages, to keep a reasonable age leap regarding cultural codes, and to avoid inducing photorealism bias in consumer perception.

Such a two-fold approach first involved an appraisal of the clinical accuracy of the simulation, assessed by 15 experts and dermatologists, and second, an appraisal of its feasibility, realism, and reception (like, dislike, contentment, scariness, etc.) by 246 women from three different countries, aged 20–60 y, considered faithful representants of consumers. The results of this pilot study are the objects of the present paper.

2. Materials and Methods

2.1. Full-Face Photographs

Fifty full-face standardized photographs of two cohorts of 25 healthy Chinese women and 25 French Caucasian women, aged 35–49 y, were taken in similar technical conditions in our local facilities (Shanghai, China, and Paris, France) using a Nexa^® device (Canfield Beauty, Parsippany, NJ, USA).

All subjects were selected through two local agencies. In addition to the age range being the primary criterion of inclusion, an absence of any facial disorder (scars, acne, melasma, seborrheic dermatitis, telangiectasia, etc.) was the other criterion. All subjects were informed (in written and oral forms) of the objective of the study and signed an informed consent form. The latter mentioned that these photographs were taken under strict anonymous conditions (eyes shut), blind-coded, and erased post-study. These photos were referred to as T0, whereas those artificially modified were coded as T0 + 10 and T0 + 15. For the photographs, subjects were asked to adopt the most neutral expression, with closed eyes and visible, uncovered hair.

2.2. Part I: Facial Signs and Grading Scales Integrated in the AI-Based System

Facial aging signs 8 and 7, previously illustrated and graded in the skin aging atlases, were retained for integration in the AI-based algorithm designed for East Asian and European subjects, respectively, as shown in

Table 1. The eight facial signs for East Asian women and their scale’s range used in the simulation.

Clinical Clusters	Facial Signs	Definition of Scored Observation	Scale	Visual
Wrinkles/Texture	Forehead wrinkles	Depth of the transverse wrinkles on the forehead.	0–8
Wrinkles/Texture	Glabellar wrinkles	Depth of vertical wrinkles between eyebrows.	0–6
Wrinkles/Texture	Inter-ocular wrinkles	Depth of horizontal folds between inner eye corners.	0–7
Wrinkles/Texture	Periorbital wrinkles	Depth of folds at malar area below crow’s feet, eye orbit excepted.	0–9
Wrinkles/Texture	Nasolabial fold	Depth of the fold present between the base of the nose and lips.	0–7
Pigmentation signs	Density of pigmentary spots	Number of spots per area unit on the cheek.	0–7
Vascular signs	Diffused redness	Diffused redness and micro-vessels visible, especially on cheeks.	0–4
Firmness/Sagging	Ptosis of lower part of the face	Sagging severity of the lower parts of the chin.	0–6

and

Table 2. Facial signs for European women and their scale’s range used in the simulation.

Clinical Clusters	Facial Signs	Definition of Scored Observation	Scale	Visual
Wrinkles/Texture	Glabellar wrinkles	Depth of vertical wrinkles between both eyebrows.	0–5
Wrinkles/Texture	Periorbital wrinkles (upper cheek area)	Depth of folds found at the malar zone below crow’s feet, eye orbit except.	0–5
Wrinkles/Texture	Nasolabial fold	Depth of the deepest fold present on the face between base of the nose and lips.	0–5
Wrinkles/Texture	Marionette lines	Depth of the deepest fold at the corner of lips.	0–6
Pigmentation Signs	Whole-face pigmentation	Density of pigmentation disorders on all the face.	0–5
Vascular Signs	Vascular disorders	All diffused redness and dilation of blood micro-vessels visible on the face.	0–7
Firmness/Sagging	Ptosis of the lower part of the face	Sagging severity of the lower part of the face on each side of the chin.	0–5

. Grading of forehead wrinkles was not included for European women as these were previously demonstrated to show erratic progress along the lifespan [13].

2.3. Simulation and Grading Algorithm

An AI-based algorithm for projective aging simulation comprised three building blocks:

(i) A convolutional neural network based on the ResNet (Residual Neuronal Network) architecture [18] trained to estimate clinical scores from a local image patch using the skin aging atlas scales as references [13,14,15]. For the seven or eight facial signs associated with each ethnicity, one dedicated model was used to estimate the grade from its corresponding area.

(ii) Once the scores were established, the second block, which was dedicated to aging knowledge tables, computed the progression rate of each clinical sign grade over 10 or 15 years for average facial aging. This method was previously described [16,17].

(iii) The last block of the AI-based system was a skin aging model derived from the Aging MapGAN [19] model. Using a modified StarGAN [20] architecture that performs image-to-image translation, each local crop was virtually “aged” to match the grade computed for either +10 or +15 years. To ensure the accuracy of these transformed crops, they were fed through the first block again to allow the output to reach the target ‘aged’ score.

Finally, each transformed crop was stitched back on the original full-face image (T0) using feathering to soften the edges and provide a more natural and realistic reconstruction.

2.4. Grading of Facial Skin Aging by the Expert Panel

Fifteen experts and dermatologists evaluated the facial signs presented in Table 1 and Table 2 using resized and reframed standardized pictures that were edited using Photoshop version 10^®. This process ensured that only one facial area of interest was displayed on the screen without any interference from the other area. Each picture was presented to every expert in a random manner to eliminate any bias. To guarantee the robustness of this process, several pictures were presented twice during this evaluation process. The evaluation of the pictures was performed under standardized conditions of lighting, expert position, and equipment calibration. For a given facial sign on each subject, the score attributed was the average graded value of all fifteen experts.

2.5. Part II: Appraising Naïve Panel

A total of 246 women aged 20–60 years were recruited locally from three different countries (China, n = 80, France, n = 81, and Thailand n = 85) to make up a well-balanced consumer panel of different cultures and age distributions. The Thai population was representative of different ethnicities from Central and North-Eastern Thailand.

Inclusion criteria of the consumer panel included good visual acuity and ease in handling a computer. The subjects in this panel signed an informed consent form and were given instructions on the task to complete. During their visit to the three local facilities located in Shanghai, Paris and Bangkok, the consumer panel were tasked to fill a questionnaire (see Section 2.6) after visualizing the photographs presented in pairs. These pairs of photographs systematically included T0 and were presented randomly (e.g., T0 vs. T + 10 or T0 vs. T + 15) as shown in Figure 1.

The eyes visible on the photographs were first masked with a black rectangle to maintain anonymity. These were then projected on a calibrated screen in comparable conditions of lighting. The consumers were all seated at an equal distance from the screen, and to reduce ocular fatigue, all examinations were limited to 30 sessions followed by a 30-min resting period.

To ensure the robustness of this process, several pairs of photographs were presented twice at random to the panelists during the perception recording. For each of the three countries, the answers were pooled. The flow of the study is depicted in Figure 2.

2.6. Questionnaire

The questionnaire provided was split into three categories described below in order to assess the perception of the naïve consumer panel on the extrapolating procedure:

(i) Realism: “How realistic do you find the simulation?” with a 5-point scale: “not realistic at all”, “rather not realistic”, “I do not know”, “rather realistic”, “very realistic”.

(ii) Liking: “How much do you like this simulation?” with a 5-point scale: “no, not at all”, “no, rather not”, “I do not know”, “yes, rather”, “yes, a lot”.

(iii) Overall opinion: Gather a global perception and establish an improvements axis of the system using three questions:

First, “would you like to test this kind of artificial intelligence on yourself?” (answers being either yes; no; or I do not know).

Second, “Overall, do you find these simulations consistent?” (answers being yes, both simulations; yes, only the one that predicts 10 years later; yes, only the one that predicts 15 years later; or no, neither of the two simulations and why).

Third, “Does anything disturb you?” where the answers were either nothing disturbs me; wrinkles are not increased enough; wrinkles are not homogeneous on the whole face; pigmentary spots are not increased enough; the result is not natural; the result is scary; the sagging of the skin is not increased enough; changes in skin texture are not noticeable; changes in eye contour are not noticeable enough; hair color is unchanged; other, please specify).

2.7. Statistics

For the validation of the accuracy of the simulation, a statistical comparison of (T0 − T0 + 10 y) knowledge and (T0 − T0 + 10 y) simulation as well as (T0 − T0 + 15 y) knowledge and (T0 − T0 + 15 y) simulation in three 5-year age clusters was made with the statistical software Xlstat 2023 (AddinsoftTM, Bordeaux, France). For each facial sign averaged from the 15 experts’ grading, a comparison of both knowledge and simulation was performed using a Variance Analysis on the difference between T0 and T0 + 10 years or between T0 and T0 + 15 years (2 factors: Model/Age-cluster) − 5% level of risk.

3. Results

3.1. Robustness of Simulation

Table 3. For East Asian women, the table shows the statistical comparison between the order simulated by the AI-based algorithm and experts’ grading on simulated images, i.e., if there are non-significant differences, the simulation is considered accurate and robust. Significant differences are represented by p < 0.05.

Average Aging	Δ T10-T0 (35–39 y)	Δ T15-T0 (35–39 y)	Δ T10-T0 (40–44 y)	Δ T15-T0 (40–44 y)	Δ T10-T0 (45–49 y)	Δ T15-T0 (45–49 y)
Nb Vol	10	10	9	9	6	6
Forehead wrinkles scored	0.65	1.32	0.86	1.41	0.8	1.08
Forehead wrinkles known	0.85	1.45	0.90	1.35	1.05	1.50
p-values	0.09	0.33	0.81	0.78	0.16	0.09
Glabellar wrinkles scored	0.43	0.81	0.39	0.69	0.52	0.89
Glabellar wrinkles known	0.38	0.68	0.40	0.65	0.55	0.80
p-value	0.63	0.14	0.89	0.54	0.67	0.45
Inter-ocular wrinkles scored	0.57	1.05	0.74	1.39	1.28	1.64
Inter-ocular wrinkles known	0.61	1.11	0.70	1.30	1.10	1.70
p-value	0.79	0.76	0.67	0.43	0.15	0.43
Nasolabial fold scored	0.18	0.59	0.72	1.08	0.56	0.81
Nasolabial fold known	0.32	0.62	0.55	0.95	0.70	1.10
p-value	0.15	0.89	0.19	0.42	0.13	0.02
Periorbital wrinkles scored	0.32	0.96	0.87	0.85	0.92	1.10
Periorbital wrinkles known	0.30	0.60	0.59	0.93	0.64	0.98
p-value	0.81	0.06	0.02	0.71	0.33	0.56
Density of spots scored	0.20	0.32	0.39	0.80	0.62	0.58
Density of spots known	0.11	0.29	0.28	0.58	0.48	0.78
p-value	0.26	0.79	0.14	0.07	0.03	0.05
Diffused redness scored	0.29	0.45	0.27	0.43	0.35	0.36
Diffused redness known	0.41	0.67	0.28	0.46	0.44	0.62
p-value	0.19	0.02	0.90	0.81	0.24	<0.00
Ptosis scored	0.06	0.23	0.08	0.18	0.04	0.31
Ptosis known	0.35	0.6	0.34	0.42	0.33	0.41
p-value	<0.00	<0.00	<0.00	<0.01	<0.00	0.38

and

Table 4. For European women, the table shows the statistical comparison between the order simulated by the AI-based algorithm and experts’ grading on simulated images, i.e., if there are non-significant differences, the simulation is considered accurate and robust. Significant differences are represented by p < 0.05.

Average Aging	Δ T10-T0 (35–39 y)	Δ T15-T0 (35–39 y)	Δ T10-T0 (40–44 y)	Δ T15-T0 (40–44 y)	Δ T10-T0 (45–49 y)	Δ T15-T0 (45–49 y)
Nb Vol	9	9	9	9	7	7
Glabellar wrinkles scored	0.55	0.6	0.38	0.42	0.12	0.31
Glabellar wrinkles known	0.60	0.70	0.40	0.50	0.20	0.30
p-value	0.64	0.34	0.76	0.25	0.15	0.92
Nasolabial fold scored	0.56	0.56	0.31	0.36	0.27	0.40
Nasolabial fold known	0.40	0.70	0.50	0.60	0.40	0.50
p-value	0.13	0.30	<0.01	<0.01	0.19	0.55
Periorbital wrinkles scored	0.55	0.75	0.41	0.71	0.93	1.12
Periorbital wrinkles known	0.50	0.80	0.50	0.90	0.70	1.10
p-value	0.55	0.66	0.51	0.21	0.10	0.86
Marionette lines scored	0.60	1.15	0.37	0.85	0.60	0.65
Marionette lines known	0.50	0.70	0.40	0.70	0.50	0.80
p-value	0.51	0.01	0.73	0.04	0.54	0.38
Whole-face pigmentation scored	0.57	0.64	0.53	0.48	0.20	0.26
Whole-face pigmentation known	0.40	0.50	0.30	0.40	0.20	0.30
p-value	0.24	0.35	0.16	0.44	0.99	0.75
Vascular disorders scored	0.21	0.25	0.20	0.24	0.10	0.21
Vascular disorders known	0.30	0.40	0.20	0.30	0.20	0.30
p-value	0.11	0.04	0.99	0.43	0.21	0.40
Ptosis scored	0.08	0.29	0.26	0.36	0.31	0.43
Ptosis known	0.40	0.80	0.60	0.90	0.70	1.00
p-value	<0.00	<0.00	<0.00	<0.00	<0.01	<0.01

show the results of the comparison between what is already known and published on the AI-based system in the image simulation (+10 years or +15 years) and what the fifteen experts graded. The comparison was conducted for both East Asian and European cohorts.

With the exception of ptosis of the lower face, the majority of the values graded on the skin aging atlases [13,14,15] yielded a non-significant p-value which supports the accuracy and robustness of this experience. When significant differences are noticed, the simulation grading is lower than order.

3.2. Perception of Simulation Experience

Positive answers were pooled from

Table 5. The general perception of the appraisal consumer panel of the simulation according to countries. All criteria are expressed as percentages (%) in rounded figures.

Country	France		China		Thailand
Number of records	n = 81		n = 80		n = 85
Simulation	+10 y	+15 y	+10 y	+15 y	+10 y	+15 y
Realism (% answers 4 + 5)	66%	64%	72%	78%	64%	68%
Liking (% answers 4 + 5)	64%	62%	75%	78%	67%	65%
Test on myself (% yes for both timings)	85%		96%		53%

and expressed as a rounded percentage (%). Specifically, rows 4 and 5 of Table 5 indicated that the photographic simulations were widely perceived as being realistic and liked by the three ethnic groups (more so by the Chinese panel) irrespective of the simulated age severity (T0 + 10 years or T0 + 15 years).

Interestingly, the consumer panelists found the realism to be similar for both simulated timelines (T0 + 10 years or T0 + 15 years). With regard to their acceptance level, all panelists and Chinese women in particular were agreeable to be personally tested for both simulated timelines of T0 + 10 years or T0 + 15 years, whereas 47% of Thai women were reluctant.

Evaluating the coherence/consistency of the simulation involved analyzing the responses expressed as percentages from the three panels to the “Does anything disturb you?” section of the questionnaire. Given that 80% of all consumer panelists found some disturbances,

Table 6. List of the items of the resulting simulation judged as rather weak, expressed by the three appraising panels, in % (rounded figures).

Questions/Panels	Chinese	French	Thai
Nothing disturbs me	20%	20%	16%
Wrinkles are not increased enough	30%	27%	46%
Wrinkles are not homogeneous on the whole face	29%	25%	41%
Pigmentary spots are not increased enough	31%	9%	38%
The result is not natural	6%	12%	38%
The result is scary	14%	4%	9%
The sagging of the skin is not increased enough	53%	37%	47%
Changes in skin texture are not noticeable enough	31%	36%	39%
Changes in eye contour are not noticeable enough	33%	22%	19%
Hair color is unchanged	19%	22%	26%

reports some elements of the simulating process to which improvements could be brought. These improvements mostly consisted of more accentuated or homogeneous wrinkles, sagging and skin texture. However, this term is too global or vague and may be too redundant for wrinkling and sagging.

A striking difference between East Asian women and European women (31% and 38% vs. 9%, respectively) was found on pigmentary spots, likely explained by the fact that these are a much more specific trait of East Asian skin aging [21]. Similarly, the eye contour seems to be another element to which improvements should be made.

Interestingly, the simulation process appeared scary for a few subjects and rather natural. Thai subjects, however, differed from Chinese women on certain aspects, reporting the result as more natural but with less pronounced and more uneven wrinkles, particularly around the eye contour. Of note, Thai women clearly differ from the two other groups in their will to test this procedure on themselves despite their rather good appreciation of the realism of the process (Table 5).

4. Discussion

Virtually imaging the human face (or hair) to illustrate likely and provisory modifications is not a new procedure and is widely used in retail or beauty salons (make-up, hair dying, perms, etc.), supported by increasingly sophisticated technologies. The work presented here aimed at exploring a different aspect, less “neutral” in the simulation process as it simulated the long-term transformations naturally brought to the facial traits of a given individual. While other studies have previously investigated the broader use of AI in the image-based fields [22,23], this study remains, to our knowledge, the first attempt to project facial changes that will likely occur on an individual in the long term. This objective was grounded by the vast amount of data provided by the validated skin aging atlases established on thousands of subjects, of different genders, environments and ethnicities.

A previous study investigated how generated visuals of AI models represented women, with a focus on racial and body type inclusivity [24]. It explored two main themes: how women are depicted in AI-generated imagery using beauty-related prompts, and how women perceive these representations. It highlighted 42% of women’s concerns with underrepresentation across age groups and misrepresentation related to ethnicity [24]. The projections of facial changes likely to occur long-term in an individual depicted here were an unprecedented realistic approach as it included women of all ages but also included women of different nationalities. This marked a shift towards a more inclusive representation.

It was initially believed that the present “Dorian Gray” approach might be rejected by consumers for many intimate or cultural reasons, in brief being judged as too scary. However, this pilot study suggested that such is not the case. Besides its recognized inherent realism for a majority of subjects, the simulation process employed here was globally “liked” by more than two thirds of the appraisal panel, taken as a faithful representation of consumers.

The results obtained here should nonetheless be seen as qualitative and taken with caution, given the relatively low number of appraising subjects (n = 246) compared to the millions of inhabitants of the three countries. However, they remain valuable in guiding the improvements of the simulation process. To further add to the robustness of this study, more images of women representative of South-East Asia could be included in the analysis such as Indonesian women. Furthermore, the present work highlights the accuracy and robustness in the simulation process with respect to the input values provided to the AI-based system.

Despite its strengths and weaknesses, the goal of this simulation methodology was nevertheless well accepted by a majority of the appraising participants. This was likely due to mixed and intimate reasons such as curiosity, aesthetic exigence and/or cultural influence. The latter is partly evidenced by differences in appraisals recorded between Chinese and Thai women (wrinkled features, unnatural results, etc.). Despite this, both groups are often broadly categorized under the same ‘East Asian’ cluster, which in reality encompasses a diverse mosaic of cultural imprints. To summarize, from a technical viewpoint, the GAN approach used here has proven its efficacy along with ease of processing. As a preliminary step, the present work laid the groundwork for more complex studies incorporating larger sample sizes, additional ethnicities, gender diversity, etc., once the suggested improvements are implemented and validated.

We acknowledge that the approach adopted in this study goes beyond the superficial aspect of facial skin. It touches on a more complex and intimate self-appreciation of physical appearance and its broader social implications. As such, communicating the results requires strict adherence to ethical guidelines, with a clear explanation that this virtual extrapolation illustrates the most probable long-term changes and is by no means a certainty. It should be emphasized that lifestyle factors such as physical activity, sleep patterns, diet, smoking, cosmetic habits, sun exposure, and air pollution, collectively referred to as the exposome, have a strong bearing on the skin aging process [3,11,25,26,27].

The validated and robust GAN solution employed here plays a major role in the beauty and skincare field, benefiting consumers, dermatologists, surgeons, and beauticians. This is particularly relevant in these fields where the environmental and lifestyle choices have a slow and insidious impact on the skin. Therefore, building awareness and trust through more personalized diagnostics and services is crucial. In addition, emerging trends around longevity are shaping how to best prepare the skin of tomorrow and access potential simulated effects of formulae, experiences, or procedures.

This present work opens the door to a projective cosmetic concept that is worthy of being explored further as it carries a real, intimate and personalized characteristic. With regard to these latter aspects, this novel procedure demands full respect for confidentiality, along with appropriate communication tailored to a culturally diverse audience.

This study evaluated the accuracy of an AI-based facial aging simulation system in predicting long-term (10–15 years) facial aging signs at the individual level, as well as its acceptance by consumers. Standardized photographs were taken at baseline (T0) and assessed using ethnic-specific skin aging atlases to generate projected images at T10 and T15 through the AI system. Although some cultural variations were observed, this pilot study confirms overall female acceptance of their projected future appearance and underscores the potential of AI-driven tools towards more realistic, personalized, and durable cosmetic innovations.