Exploring aesthetic judgments of modern sculptures in virtual reality: a multimodal machine learning approach

1 Introduction

Why are some objects perceived as aesthetically pleasing while others are not? How do we experience art? These questions lie at the core of empirical aesthetics, a field at the intersection of psychology and neuroscience. Over the past 2 decades, some advances have been made in understanding art perception using eye tracking and electroencephalography (EEG) to study emotional responses to visual stimuli (Pelowski et al., 2018; Rosenberg and Groner, 2022; Massaro et al., 2012; Quian Quiroga and Pedreira, 2011), as well as machine learning to predict creativity judgments in paintings (Spee et al., 2023; 2024). Recent advances in VR and multimodal computational methods have further expanded empirical aesthetics research. Immersive VR environments allow researchers to study aesthetic experiences under experimentally controlled conditions. Simultaneously, the integration of EEG and eye-tracking measures has emerged as a powerful multimodal approach for investigating the interplay between neural activity, visual attention, and affective responses during complex perceptual tasks (Rivas-Vidal et al., 2025). Recent developments in machine learning further demonstrate that aesthetic evaluations and emotional states can be predicted from combined visual, physiological, and behavioral signals. For example, Kumar et al. (2026) demonstrated that integrating EEG and eye-tracking features yields more accurate prediction of cognitive affective states than single -modality approaches. This further highlights the potential of multimodal data fusion for understanding aesthetic experience (Chien and Kuo, 2026; Kumar et al., 2026). Despite this progress, research has largely focused on two-dimensional (2D) media, leaving three -dimensional (3D) artworks rarely erexplored (Assembek et al., 2026; Tomasetig et al., 2026). Although 2D and 3D art share key principles such as composition, balance, and contrast, they differ fundamentally in how they engage perceptual, emotional, and neural mechanisms. Three-dimensional artworks possess spatial presence, material texture, and structural complexity, enabling dynamic viewing from multiple perspectives. Such interaction increases cognitive engagement, intensifies emotional responses, and enhances novelty (Chatterjee and Vartanian, 2014), factors known to shape the depth of aesthetic experience (Leder et al., 2004). This gap reflects both historical and practical constraints. Western art history has traditionally prioritized painting (Gombrich and Gombrich, 1995), a bias that extended into empirical research. In addition, paintings are easy to digitize and study in controlled settings, whereas sculptures require physical presence and dynamic viewing. Museum collections are also dominated by paintings, and sculpture-only exhibitions are rare outside retrospectives, limiting opportunities for large-scale empirical investigation.

1.1 Emotions in art perception

Emotions play a central role in aesthetic experience and art perception. Numerous studies havedemonstrated a strong link between emotional responses and aesthetic evaluation, particularly within the valence–arousal framework proposed by Russell (1980) . This framework characterizes emotional experience along two dimensions: valence, which reflects the positivity or negativity of the response, and arousal, which indexes its intensity. Both dimensions have been shown to be critical predictors of aesthetic preference and judgment (Scherer, 2005). Artworks that elicit positive valence and moderate arousal are typically rated as more aesthetically pleasing (Berlyne, 1974). To assess emotional responses to visual stimuli, including artworks, Bradley and Lang developed the Self-Assessment Manikin (SAM) scale (Bradley and Lang, 1994). SAM represents valence and arousal using pictorial symbols, thereby reducing reliance on language and minimizing cultural bias. This makes it a robust and widely used method for studying emotional experience. Emotional responses also shape cognitive evaluations. According to the affect-as-information hypothesis, individuals use their emotional state as input when making judgments (Clore et al., 2012). In the context of art perception, initial affective reactions can guide aesthetic evaluations, such that positive emotions lead to higher ratings of beauty and overall aesthetic value (Silvia, 2005).

1.2 Neurophysiological methods in art studies

Electroencephalography (EEG) is widely used in empirical aesthetics to investigate the cognitive processes underlying art perception. By recording electrical activity from the scalp, EEG provides insight into the neural mechanisms engaged during aesthetic experience (Barbieri, 2026). Specifically, EEG captures neural oscillations across multiple frequency bands, which have been associated with distinct cognitive and emotional processes during art viewing (Cela-Conde et al., 2011). Different frequency bands have been linked to specific aspects of aesthetic processing and appreciation. Increased alpha power has been associated with focused internal attention, whereas beta and gamma activity are often related to higher-order cognitive processing and arousal (Cela-Conde et al., 2011; Jacobsen, 2013). Delta and theta oscillations, in contrast, have been connected to emotional responses and aesthetic appreciation, with theta activity increasing during the viewing of pleasant artworks (Maglione et al., 2017). Another commonly used EEG metric is the alpha/beta ratio, which distinguishes between relaxed and cognitively engaged states: higher values indicate a more relaxed and less alert state, whereas lower values reflect increased mental effort, alertness, arousal, stress, or cognitive engagement (Wen and Aris, 2020). EEG studies further indicate that perceptions of beauty, novelty, and other aesthetic judgments are associated with distinct neural patterns (Munar et al., 2012). Thus, EEG offers a valuable tool for examining the temporal dynamics of art perception and evaluation. Although its spatial resolution is limited compared to neuroimaging methods such as fMRI, EEG is well suited for capturing neural activity during both naturalistic and controlled viewing conditions. In the present study, EEG is used to examine how neural dynamics relate to subjective aesthetic judgments of sculptures presented in a virtual reality environment.

Beside EEG, Eye tracking is a central method in art perception research used to examine how viewers visually engage with artworks. Key eye-tracking metrics, such as fixations (periods of relatively stable gaze) and saccades (rapid eye movements between fixations), provide insights into visual attention and cognitive processing during art viewing (Rosenberg and Groner, 2022). Early eye-tracking studies demonstrated that gaze behavior is influenced by both low-level visual features, such as color and contrast, and high-level factors, including expertise and personal preference (Quian Quiroga and Pedreira, 2011). Viewing strategies also differ between art experts and novices: experts tend to focus on specific regions of an artwork, whereas novices exhibit broader scanning patterns (Tallon et al., 2021). Pupil dilation is another important eye-tracking measure that reflects both cognitive and emotional processes during aesthetic experience. Pupil size is sensitive not only to changes in luminance but also to arousal and mental effort (Mathôt, 2018). In art perception, pupil dilation is particularly informative for assessing processing fluency, as artworks that are easier to process tend to elicit larger pupil dilations and higher aesthetic preference ratings (Kuchinke et al., 2009). Blink rate provides an additional index of visual engagement and attention, with reduced blinking associated with increased interest in visual stimuli (Ranti et al., 2020). Technological development in mobile and wearable eye-tracking systems have enabled a shift from tatic laboratory expepriments to studies conducted in naturalistic museum environments. The mode of artwork presentation strongly influences viewing behavior, with mobile eye-tracking studies showing increased engagement and longer viewing times in museums compared to controlled laboratory settings (Reitstätter et al., 2020; Warwick et al., 2025). Furthermore, large-scale initiatives such as the Museum Gaze project have emphasized the importance of investigating visual exploration under ecologically valid conditions and have demonstrated the potential of mobile eye tracking for understanding visitors’ interactions with artworks in real exhibition spaces (Reitstätter et al., 2025).

However, naturalistic studies also introduce confounding factors, including variable lighting conditions, heterogeneous display arrangements, and the presence of other visitors, all of which can compromise data quality. Moreover, mobile eye-tracking systems typically have lower spatial resolution and reduced sampling rates than laboratory - based devices (Dowiasch et al., 2020). Contextual factors further affect gaze behavior. The presence of artwork labels or explanatory text leads viewers to alternate their gaze between the artwork and the accompanying information, increasing gaze switching and altering natural viewing patterns (Rosenberg and Klein, 2015). Contextual information can also guide viewers’ attention toward specific regions of an artwork, thereby shaping exploration strategies (Ganczarek et al., 2022).

Recent studies have further demonstrated that eye-tracking measures can improve the prediction of aesthetic evaluations when combined with computational and machine-learning approaches, highlighting the value of gaze behavior as an objective indicator of aesthetic experience (Chien and Kuo, 2026). The recent project “The Museum Gaze” employed mobile eye-tracking to compare how visitors view sculptures and paintings in museum settings (Rosenberg and Groner, 2022). Preliminary findings suggest that sculpture viewing is more dynamic, whereas paintings are typically observed from a fixed vantage point, resulting in more static viewing patterns. At the same time, recent studies have highlighted methodological challenges in using eye tracking to study sculpture perception, including small sample sizes and the complexity of analyzing gaze behavior in three-dimensional space (Stein et al., 2022). Together, these findings underscore the need for new methodological approaches that combine ecological validity with experimental control to study sculpture perception. In the present study, we address these limitations by using a virtual reality art gallery, which offers the benefits of a controlled lab environment while providing a naturalistic museum-like experience as closely as possible. Some initial results of the eye-tracking data of this study have been already published (Andrusyak and Wriessnegger, 2025).

1.3 Research questions

The reviewed literature highlights several research gaps that the present study seeks to address. First, the majority of empirical aesthetics research has focused on two-dimensional artworks, leaving emotional and aesthetic responses to three-dimensional artworks, such as sculptures, comparatively underexplored. Second, although both EEG and eye tracking have been widely used in art perception research, their combined application in immersive three-dimensional environments remains limited.

To address these gaps, the present study investigates two main research questions using data collected from participants who viewed sculptures across three rounds and provided both aesthetic judgments and ratings of specific sculptural attributes.

• Which aesthetic and emotional judgments of sculptures can be explained by subjective evaluations of sculptural attributes (Visual Harmony, Complexity, Texture, Shape Complexity, Symmetry, Emotional Expressiveness, Realism, and Pose), and how do these relationships differ between initial impressions and post-exposure evaluations?
• Are there measurable physiological differences in visual engagement and neural activity between participants with contrasting aesthetic evaluations?

To answer these questions, we conducted an experiment in which 28 non-expert participants viewed the same 10 figurative sculptures across three rounds in a custom-built VR environment closely resembling an art gallery. During the experiment, eye-tracking, pupil dilation, and electroencephalography (EEG) data were collected, along with subjective ratings provided by participants based on likability and sculptural attributes. Importantly, the present work should be understood as an exploratory proof-of-concept study based on a limited and highly specific stimulus set of ten modern figurative sculptures. The aim was not to establish generalizable laws of sculpture perception, but rather to evaluate the feasibility of combining VR, interpretable machine learning, eye tracking, and EEG for studying aesthetic judgments in immersive three-dimensional environments.

2 Methods

2.1 Participants

A total of 28 healthy participants (14 male, 14 females, mean age 25.7, age range 18–33) were included in the final analysis. All participants were considered non-art experts, validated using the Vienna Art Interest and Art Knowledge Questionnaire (VAIAK) [20]. Results showed limited art knowledge (64% answering “I do not know” to identification questions) and minimal engagement with art (68% visiting museums less than twice per year). The study was approved by the local ethics committee of the Graz University of Technology.

2.2 Experimental design

The experiment consisted of three rounds. In each round, participants viewed same 10 sculptures. After viewing a sculpture, participants answered a set of questions. Each sculpture was presented individually for a fixed duration of 70 s. After the fixed viewing time ended, there was a 3-s black fadeout and a first question appeared. In all rounds questions were presented on a white canvas occupying the entire field of view of the participant. Questions could not be skipped or left unanswered. Participants were allowed to take as much time as they needed to answer the presented questions. The next question appeared only after the participant provided the answer. After viewing all sculptures for the round and answering the last question, participants were automatically transported to the break room for 30 s. During this time participants were offered water or a chair. For an entire duration of the experiment participants remained in the VR headset. Overview of the experimental paradigm can be seen in Figure 1.

2.2.1 First and third rounds

In the first and third rounds participants provided their initial aesthetic and emotional evaluations using the scale-set of art-judgments. The art judgment scale was adapted from Spee et al. (2024). The original scale had 13 art judgments (items for evaluation) across six different aspects: aesthetic, qualitative, epistemic, adverse, semantic and preference. We extended it by adding two additional image-based question on Valence and Arousal to cover the emotional aspect. To measure the emotional impact participants were presented Self-Assessment Manikin (SAM) figures on Valence and Arousal and were asked to provide their response by clicking the button under the image that best represented their state after viewing the sculpture. Participant answered the text-based questions using a seven-point Likert scale, where 1 was the minimum pole, having the ‘Not at all’ label near it and 7 was the maximum pole with ‘Very much’ label. The image-based questions were on a 5-point Likert scale, each point corresponding to the exact image from a SAM figure. Third round was identical to the first round to later allow for consistency assessment and context influence analysis. b) Second Round In the second round, participants were asked to evaluate the artwork based on different sculptural attributes (e.g., Visual Harmony, Complexity, Texture). After viewing the sculpture, the question screen showed the prompt ‘Evaluate the artwork based on [Attribute]’ with a 7-point scale with clickable buttons labeled 1-7 below. The negative and positive poles of each attribute (e.g., Chaotic - Harmonious for Visual Harmony) were shown near the respective ends of the scale. The art-attributes scale was adapted based on the study of Spee et al. (2024). The scale was modified to exclude the attributes applicable only to paintings and ensuring that all attributes are applicable to sculptures. Questions could not be skipped or left unanswered. The next question appeared only after the participant provided the answer to the presented question.Table 1 shows the set of attributes and pole labels used in Round 2 for sculpture evaluation.

TABLE 1. Set of attributes and pole labels used in Round 2 for sculpture evaluation.

To minimize sequence effects of the viewing order, a script was implemented in the Unity environment to dynamically change the sculpture appearance order. Each participant saw a unique sculpture order for every round, and no participant shared the same order with any other. Statistical analysis confirmed effective randomization of sculpture appearance order.

2.3 Virtual reality gallery and sculpture dataset

The virtual gallery was created using the Unity engine to simulate real exhibition spaces while maintaining strict experimental control. The environment included a main exhibition room with neutral light-blue walls and a separate break room for rest periods between viewing rounds. After participants completed their evaluations of each sculpture, the display faded to black and the next sculpture was presented in the same spatial location, preserving a constant viewing orientation. Participants were able to move freely around each sculpture and observe it from multiple viewpoints and distances. For safety, sculptures were always placed at the center of the physical laboratory space, and participants were informed of spatial boundaries and instructed to move cautiously at the beginning of the experiment. All 3D sculpture models were obtained from Sketchfab under the Creative Commons Attribution (CC BY) license or the Sketchfab Free Standard License. The 3D scans were preprocessed in Blender to standardize their dimensions and ensure that they are scaled to reflect their real-world dimensions. Additionally, text or any other markings were removed, that could distract participants or bias the visual attention during the viewing. Each sculpture was then placed on a pedestal with a light oak texture designed to position it at eye level. This is standard practice in galleries for consistent presentation, regardless of the sculpture size.

The final stimulus set consisted of ten sculptures (see Table 2). The first selection criterion was that the works be classified as modern sculpture. Modern sculpture is historically defined as worksproduced from the late 19th century through the 1970s, beginning with Auguste Rodin (1840–1917) and extending to the emergence of Minimalism (Hughes, 1980). The second criterion was thatthe sculptures be figurative. Figurative sculpture refers to three-dimensional art that represents thehuman figure or other recognizable forms, maintaining a clear connection to the natural world (Chilvers, 2009). In the present study, this criterion was further restricted to works depicting thehuman body, excluding animals and inanimate objects. Limiting the stimuli to human figures ensuredgreater homogeneity within the dataset, which was necessary to reduce variability and allow formeaningful comparisons given the small sample size of ten sculptures. The third criterion was self-sufficiency. According to Self-sufficient works can be fullyappreciated as autonomous objects without reliance on historical or cultural context. This principle guided the experimental design, in which sculptures were presented without contextualinformation such as titles, artist names, dates, or descriptive labels. This decision was motivated byfindings from Casteau and Smith (2024), who demonstrated that contextual information substantiallyalters viewers’ aesthetic experience and gaze behavior. Finally, diversity was emphasized to ensure arepresentative stimulus set. The selected sculptures included both single and multiple figures, maleand female subjects, varied textures and sizes, and a range of thematic content (Table 2; Figure 2).

TABLE 2. List of sculptures used in the study, sorted by year.

All preprocessed sculptures can be seen in Figure 2. Figure 3 shows an example of one of the figurative sculptures (“Standing Woman” from Alberto Giacometti) in the VR environment, positioned on a pedestal.

2.4 Experimental procedure

For this study a custom Virtual Reality (VR) environment was developed. The VR space was created in the Unity engine to closely replicate the experience of a real-world exhibition space while optimizing it for controlled experimental conditions. The environment consisted of two rooms: main display room, where sculptures were presented and break room, where participants were placed in between rounds. Participants were able to physically walk around and view each sculpture from various angles and distances. Participants remained in the same physical lab throughout the experiment. After participants completed the evaluation questions for one sculpture, the system would fade to black and replace the current sculpture with the next one in the sequence. This transition design allowed to present all sculptures in the same virtual room, while maintaining consistent spatial orientation for participants. A black fade-out effect was implemented during these transitions to ensure smooth changes between sculptures. The virtual exhibition room was designed to be larger than the physical laboratory space in order to approximate a real gallery environment. No visual indicators of the physical room boundaries were included, as such cues could distract attention and compromise eye-tracking data quality. To ensure safe navigation and adequate spatial awareness, participants received verbal instructions before the experiment. Sculptures were consistently positioned at the center of the physical room, and participants were informed of a fixed orientation reference and the available walking space (approximately 1.5 m) around each sculpture. They were advised to move cautiously atthe beginning of the session and were assured that they would receive verbal warnings if they approached physical boundaries. After a brief adaptation period, participants moved confidently within the virtual space and explored the sculptures from multiple angles and distances. Participants were free to view each sculpture in any manner they preferred, including standing, walking, or sitting. No specific viewing strategy was imposed; the only restriction was that participants were instructed not to touch the sculpture in order to preserve immersion.

• Aesthetic and Emotional Judgments (Rounds 1 and 3): Participants rated sculptures using an extended art judgment scale adapted from Spee et al. (2024), covering six aspects. Aesthetic aspects included aesthetically moving and beauty ratings. Qualitative aspects covered good work of art, creativity, and technical skill evaluations. Epistemic aspects included fascinating, interesting, and thought-provoking ratings. Adverse aspects covered boring and disturbing or irritating evaluations. Semantic aspects included familiarity and understanding ratings. Preference aspects focused on personal liking. Emotional responses were measured using Self-Assessment Manikin (SAM) scales for Valence and Arousal (Bradley and Lang, 1994). Text-based questions used seven-point Likert scales ranging from one meaning “Not at all” to seven meaning “Very much”, while SAM questions used five-point scales.
• Sculptural Attributes (Round 2): Eight sculptural attributes were evaluated on seven-point scales with clearly defined poles. Visual harmony ranged from chaotic to harmonious. Complexity ranged from simple to intricate. Texture ranged from rough to smooth. Shape complexity ranged from simple to complex. Symmetry ranged from asymmetrical to symmetrical. Emotional expressiveness ranged from neutral to intense. Realism ranged from abstract to lifelike. Pose ranged from static to dynamic.

2.5 Measurement devices

During the experiment, besides survey response data, physiological data of the participants were recorded. Measurements included eye-tracking data and electroencephalography (EEG). For an entire duration of the experiment participants were wearing an HP Reverb G2 Omnicept Virtual Reality headset and wireless IDUN Guardian in-ear EEG device, while holding an HP Omnicept’s motion controller in their right hand, which are described in greater detail in sections below. The VR headset was connected with a wired cable, which was secured along the wall using cable management hooks to prevent it from interfering with participants movement.

2.5.1 VR headset and controller

HP Reverb G2 Omnicept Virtual Reality headset was connected to Unity with the OpenXR plugin and Mixed Reality Portal, allowing participants to see the Unity output while wearing the headset. Detailed tutorial on how to integrate Omnicept framework inside the Unity engine and access the sensor data is available on HP’s Developer Portal. The VR headset has built-in eye tracking sensors, data from which were accessed using the Omnicept Framework. According to the HP documentation the eye data is recorded with two cameras pointing at left and right eye. The eye camera and eye-tracking on the HP Reverb G2 Omnicept edition is powered by Tobii. Full list of eye data parameters recorded by sensor scan be seen in Table 3.

TABLE 3. Eye-tracking data collected from the HP Reverb G2 Omnicept.

To log the answers for post-viewing surveys participants were given one of the HP Omnicept’s motion controller. Once the sculpture viewing time ended and the question canvas appeared, the controller was activated and a green beam appeared, indicating where the controller is pointed, making it easier for participants to select a desired answer button. Before the start of the experiment for each participant calibration was performed with the HP Omnicept Runtime Eye Tracking Calibration, which confirms that the headset is placed correctly on the participants head and that they can see clearly.

2.5.2 In-ear EEG

Electroencephalography (EEG) data were recorded using the IDUN Guardian single-channel in-ear system (Idun Technologies AG, Switzerland). Participants were instructed to remove Bluetooth-enabled wearables prior to testing to avoid interference, and EEG signals were transmitted via Bluetooth Low Energy (BLE). The system consists of two wireless earbuds with Dryode electrodes and records EEG as the biopotential difference between the left and right ears, without separate reference or ground electrodes. Previous validation studies have shown that in-ear EEG provides signal quality comparable to conventional scalp-based recordings (Mandekar et al., 2022). For each participant, appropriate wing and eartip sizes were selected to ensure a stable fit during movement. Impedance was checked before the experiment using the IDUN Console Web Application. The device was cleaned and charged after each session. EEG data were sampled at 250 Hz. For an entire duration of the experiment eye-tracking data from the HP Reverb G2 Omnicept headset and EEG data from IDUN Guardian were streamed and synchronized alongside the event markers sent from Unity. The event markers signified the start and end for each sculpture viewing during the experiment. All data streams were sent to and synchronized in the Lab Streaming Layer (LSL) open-source software.

3 Data analysis

To answer the first research question—Which aesthetic and emotional judgments of sculptures can be best explained by subjective evaluations of sculptural attributes, and how do these relationships change between initial impressions and post-exposure evaluations? —we applied a machine learning (ML) approach to the survey data. Specifically, Gradient Boosting Decision Trees (GBDT) were combined with SHAP (SHapley Additive exPlanations) values to assess feature importance. This approach enabled us to model how sculptural attribute ratings obtained in Round 2 predict aesthetic and emotional judgments collected in Rounds 1 and 3. The second research question - Are there measurable physiological differences in visual engagement and neural activity between participants with contrasting aesthetic evaluations? – have been addressed by conducting separate statistical comparisons using eye-tracking and electroencephalography (EEG) data.

3.1 Survey-based interpretability analysis

This analysis examined how specific sculptural attributes influence aesthetic judgments and emotional responses across three data-collection rounds. We modeled how attribute ratings from Round 2 predict aesthetic judgments from Round 1 (initial impressions) and Round 3 (post-exposure evaluations), and how these evaluations changed over time. Data were obtained from 28 participants who viewed 10 sculptures across three rounds, yielding 840 observations. For machine learning, individual-level data were aggregated to the sculpture level by computing mean attribute and judgment ratings for each sculpture. This resulted in a dataset of 10 rows (one per sculpture) with averaged ratings across participants, enabling the identification of attribute–judgment relationships at the collective level rather than individual variability.

3.2 Machine learning model and validation

To examine how sculptural attributes (Round 2) influence aesthetic and emotional judgments (Rounds 1 and 3), we used Gradient Boosting Decision Trees (GBDT) with SHAP values for feature importance analysis. GBDT is well suited for small, structured datasets, as it captures non-linear relationships while limiting overfitting. Our goal was not prediction but to identify attributes that consistently contribute to judgments across sculptures. To reduce overfitting, tree depth was limited to two. Modeling was performed at the sculpture level using averaged ratings, as the individual-level sample size was insufficient for generalizable conclusions. An initial Leave-One-Out Cross-Validation (LOOCV) approach proved sensitive to sculptures with extreme ratings, biasing feature importance. To address this, we implemented a leave-one-sculpture-out procedure. For each target variable, ten GBDT models were trained, each excluding one sculpture entirely. Within each subset, LOOCV was applied and SHAP values were computed from a model trained on the remaining nine sculptures. Averaging results across runs yielded stable performance estimates and robust identification of influential attributes.

3.3 Data preprocessing and feature extraction

3.3.1 Eye-tracking data

Eye-tracking data were preprocessed by interpolating all streams to 120 Hz and applying Savitzky–Golay filtering to gaze position signals. Velocities were derived from the smoothed positions, and fixations were identified using a three-dimensional velocity threshold of 30 units. Extracted features included fixation and saccade dynamics (e.g., counts, mean durations, and velocities). Gaze dispersion was quantified as the standard deviation of the X, Y, and Z coordinates, reflecting the spatial distribution of visual attention. Viewing progression was assessed by examining fixation distributions across early, middle, and late viewing phases. Additional physiological measures included pupil dilation and blink rate.

3.3.2 EEG data

To analyze the EEG data, we adopted the same approach as used for the eye-tracking data. We selected a sculpture with highly polarized Beauty ratings (Alberto Giacometti’s Standing Woman) and compared EEG patterns between the two aesthetic judgment groups. However, due to data quality issues in the EEG recordings, the final group sizes differed. After excluding participants with missing or incomplete EEG data, the resulting split consisted of 10 participants in the beautiful group and 6 participants in the Not Beautiful group. Raw EEG data and Unity event markers were loaded from the XDF files for each participant. Recordings with missing streams or empty data were identified and excluded from further analysis (4 recordings at 0.0000 Hz). Three recordings had reversed timestamps (end timestamp value was earlier than the start timestamp). For these cases, timestamps and EEG data were reversed to restore the signal to the chronological order. All time stamps were reconstructed using a fixed 250 Hz sampling rate. All EEG signal processing was performed in Python using the MNE, numpy, pandas, and scipy libraries. The continuous EEG signal was segmented into viewing epochs based on StartViewing and EndViewing markers from the Unity stream. Each marker contained a sculpture ID (e.g., StartViewing_4 indicated the start of viewing sculpture #4, which corresponds to Standing Woman by Giacometti). Marker timestamps were aligned to the EEG time base. To remove low-frequency drifts and high-frequency noise, a second-order Butterworth bandpass filter with cutoff frequencies at 0.5 Hz and 45 Hz was applied to each epoch. To detect dropouts for each epoch, the number of EEG samples was counted, and those with fewer than 50 samples (0.2 s of data at 250 Hz) were flagged as missing. Missing samples were likely due to temporary Bluetooth disconnections or hardware signal interference, which can occur with wireless EEG systems, especially when participants are moving in VR. We reconstructed all timestamps using a fixed 250 Hz sampling rate and aligned sculpture viewing epochs based on Unity markers indicating Start and End viewing of the sculpture. Even with occasional dropouts, synchronization remained reliable because Unity event markers were aligned to the reconstructed EEG timeline rather than the original timestamps. Participants were excluded from further analysis if they had fewer than 10 valid EEG segments across all rounds. Dropout analysis showed that all 24 participants had complete EEG data in Round 1, 19 had complete data in Rounds 1 and 2, and only 15 had complete data across all three rounds.

EEG features were computed for each 70-s sculpture-viewing epoch defined by Unity event markers. Signals were recorded from a single in-ear channel and bandpass filtered (0.5–45 Hz, second-order Butterworth) prior to feature extraction. Frequency-domain features were obtained by estimating Power Spectral Density (PSD) using Welch’s method (sampling rate: 250 Hz; segment length adapted to data length, up to 1024 samples). Mean power was calculated for five standard bands (delta: 0.5–4 Hz, theta: 4–8 Hz, alpha: 8–13 Hz, beta: 13–30 Hz, gamma: 30–45 Hz), and the beta/alpha ratio was computed as an index of cognitive arousal. Time-domain features were captured using Hjorth parameters: Activity (signal variance), Mobility (rate of change), and Complexity (frequency variation). All features were derived from the same filtered single-channel signal for each epoch. To remove artifact-related outliers, z-scores were computed for features prone to noise (activity, standard deviation, maximum amplitude, gamma power, complexity, and beta/alpha ratio). Epochs with any feature exceeding |z| > 2 were excluded.

3.3.3 Statistical analysis

3.3.3.1 Eye-tracking data analysis

Statistical analysis was performed on the preprocessed eye-tracking data comparing two groups: those who rated the sculpture as ‘Beautiful’ (n = 12) and those who rated it as ‘Not Beautiful’ (n = 9). The main objective of this analysis was to identify meaningful differences in visual engagement patterns between these groups. Eye-tracking features for statistical comparisons were selected based on the established methodologies in eye-tracking and aesthetic perception studies [45, 46]. Before conducting statistical comparisons, all selected features were tested for normality with the Shapiro- Wilk test to assess data distribution. Based on these results, suitable comparison methods were chosen (parametric independent t-tests or non-parametric Mann-Whitney U tests). Mann-Whitney U tests were primarily used due to the small sample size and non-normal distributions of some metrics. To quantify how big are the differences between the two groups, Cohen’s d effect sizes were calculated for all comparisons. These effect sizes were categorized according to conventional thresholds: large (d ≥ 0.8), medium (0.5 ≤ d < 0.8), small (0.2 ≤ d < 0.5), and negligible (d < 0.2). Afterword, percentage differences were calculated between group means for interpretation. Results from all statistical comparisons were ranked by effect size magnitude to identify the features showing the strongest and most meaningful differences between two groups.

3.3.3.2 EEG data analysis

The statistical analysis of the EEG data follows the same structure as the eye-tracking analysis. Statistical analysis was done on the data of participants who rated the sculpture as ‘Beautiful’ (n = 10) and those who rated it as ‘Not Beautiful’ (n = 6). First, the feature-wise normality testing using the Shapiro-Wilk test was done, followed by group comparisons using either independent t-tests or Mann- Whitney U tests depending on the distribution characteristics. For each EEG feature, Cohen’s d was computed to quantify the effect size of group differences.

4 Results

4.1 Survey results

4.1.1 Initial impressions (round 1)

To understand participants’ initial aesthetic judgments, we used ratings from Round 1 as the target and Round 2 perceptual attribute ratings as predictors. Given that models were trained on only 10 data points, these R values reflect within-sample descriptive fit of attribute-judgment relationships rather than out-of-sample predictive generalizability. All results should be interpreted as exploratory and hypothesis-generating. Within this exploratory single-stimulus-set analysis, Beauty, Valence, and Technical Skill showed the strongest descriptive attribute–judgment relationships, whereas the remaining judgments showed low or negative explained variance. For the best-predicted judgments, model performance was as follows: Beauty: R² = 0.699 (±0.097), indicating that approximately 70% of variance in Beauty ratings could be explained by perceptual attributes. Valence: R² = 0.684 (±0.163) indicating that the model descriptively captured approximately 70% of variance within this specific stimulus set. Technical Skill: R ² = 0.468 (±0.193), showing moderate but meaningful exploratory power. The SHAP values feature importance analysis revealed the contribution of each perceptual attribute to the prediction of aesthetic judgments. The analysis calculated both absolute SHAP importance (size of influence) and signed SHAP values (directional influence) across all cross-validation folds. For Beauty, the strongest predictors were Visual Harmony (0.301), Texture (0.251), and Emotional Expressiveness (0.039). Average signed SHAP values showed that higher ratings of Texture (+0.003; higher ratings indicate smoother texture) and Visual Harmony (+0.006; higher ratings indicate more visual harmony) were associated with higher perceived beauty. Furthermore, higher ratings of smoothness and visual harmony were descriptively associated with higher beauty ratings. Meanwhile, Emotional Expressiveness showed a weak negative effect (−0.003; higher ratings correspond to more emotionally expressive sculptures), suggesting that emotionally expressive sculptures were slightly less likely to be rated as beautiful compared to more neutral ones. For Valence, the top predictive features were Visual Harmony (0.179), Texture (0.095), and Em oti onal Expressiveness (0.029). Average signed SHAP values showed that higher ratings of Visual Harmony (+0.0009) were weakly associated with more positive emotional responses, while Texture (−0.001) and Emotional Expressiveness (−0.002) had slight negative effects. This indicates that sculptures perceived as visually harmonious were generally associated with more positive emotional responses. Meanwhile, the small negative signed SHAP values for Texture (−0.001) and Emotional Expressiveness (−0.002) indicate that sculptures with smooth texture and more emotionally intense sculptures tend to have less positive evaluations For Technical Skill, the strongest positive contributors were Emotional Expressiveness (0.103) and Pose (0.081), with signed SHAP values of +0.014 and +0.003, respectively. Sculptures rated as more emotionally expressive and more dynamic in pose were perceived as more technically skillful. In contrast, Shape Complexity (0.116) had a signed SHAP value of −0.009, indicating that sculptures with a complex shape were rated as less technically skillful. The average signed SHAP values for the three best-predicted judgments Beauty, Valence, and Technical Skill, can be seen in Figure 4.

4.1.2 Post-exposure evaluations (round 3)

To understand participants’ post-exposure aesthetic judgments, we used ratings from Round 3 as the target and Round 2 perceptual attribute ratings as predictors. The most reliably predicted judgments were Valence, Beauty, and Technical Skill. All other aesthetic and emotional judgments showed low or even negative explained variance. For the best-predicted judgments, model performance was as follows: Valence: R ² = 0.883 (±0.041), indicating that nearly 88% of the variance in Valence ratings could be explained by perceptual attributes. Beauty: R ² = 0.710 (±0.090), indicating that approximately 71% of the variance in Beauty ratings could be explained by perceptual attributes. Technical Skill: R ² = 0.678 (±0.132), showing strong explanatory power within this stimulus set. Though given the n = 10 sample, this should be interpreted as a descriptive fit rather than a generalizable prediction. The SHAP values feature importance analysis revealed the contribution of each perceptual attribute to the prediction of aesthetic judgments based on Round 3 ratings (Table 3). The analysis calculated both absolute SHAP importance (size of influence) and signed SHAP values (directional influence) across all cross-validation folds.

For Valence the results suggest that sculptures rated as more visually harmonious and smoother in texture are linked to more positive emotional responses, while symmetrical sculptures slightly reduce the positive valence. For Beauty results indicate that smoother and more visually harmonious sculptures are perceived as more beautiful, while highly realistic lifelike sculptures are rated as less beautiful than abstract ones. For Technical Skill results show that sculptures judged as more intricate (composition, detail density) were associated with higher perceived technical skill, while higher in realism and shape complexity showed small negative effects. The average signed SHAP values for the three best-predicted judgments—Valence, Beauty, and Technical Skill for Round 3 can be seen in Figure 5.

4.1.3 Evolution of aesthetic judgments

To understand evolution of aesthetic judgments from initial impressions to post-exposure evaluations we compared model performance and feature influences between Round 1 and 3. While the same three judgments (Beauty, Valence, and Technical Skill) remained the most predictable, predictive accuracy improved and there was a significant shift in influencing attributes. A summary of model performance changes between Round 1 and 3 can be seen in Table 4.

TABLE 4. Model performance changes from round one to round three.

Overall, predictive accuracy increased across all top judgments in Round 3: Valence improved from R² = 0.684 (±0.163) to R² = 0.883 (±0.041), Beauty improved from R² = 0.699 (±0.097) to R² = 0.710 (±0.090) and Technical Skill improved from R² = 0.468 (±0.193) to R² = 0.678 (±0.132). In addition to improved model performance, the feature importance and direction of influential attributes also changed between rounds.

4.2 Eye tracking results

Eye-tracking features were compared between participants who rated the sculpture Standing Woman as ‘Beautiful’ (n = 12) and those who rated it as ‘Not Beautiful’ (n = 9).

The “Beautiful” group demonstrated lower fixation count (−24.26%), saccade count (−25.66%), and blink rate (−25.45%) (Figures 6, 7). Vertical gaze dispersion (gaze_y_std) was also lower (−6.12%) (Figure 8).

Gaze depth variability (gaze_z_std) was the only feature showing a statistically significant group difference, with a medium effect size (d = −0.79, p = 0.030). Except for gaze depth variability, the reported eye-tracking differences did not reach statistical significance and therefore cannot be interpreted as evidence for reliable group effects. These observations are reported descriptively to illustrate potential patterns informing future studies with larger samples.

Given the exploratory nature of the study and the limited sample size, no correction for multiple comparisons was applied, in line with recommendations that such corrections in exploratory analyses may obscure potentially meaningful patterns (Bender and Lange, 2001).

Participants who rated the sculpture as beautiful exhibited more stable depth gaze than those who did not (Figure 8), indicating less dispersed viewing behavior. Several additional features showed medium-sized numerical differences between groups; however, because these comparisons were non-significant, no inferential conclusions can be drawn from them.

Blink rate was also 25.45% lower in the ‘Beautiful’ group (Figure 7). Vertical gaze dispersion. (gaze_y_std) was lower in the ‘Beautiful’ group, with a 6.12% difference (Figure 8). However, these differences are reported descriptively only and should not be interpreted as evidence of reliable group differences.

Temporal viewing measures showed longer mean fixation durations in the “Beautiful” group (+38.58%), while early fixation count differed minimally between groups (+0.95%) (Figure 9).

4.3 EEG results

EEG features were compared between participants who rated the sculpture Standing Woman as Beautiful’ (n = 10) and those who rated it as ‘Not Beautiful’ (n = 6). Table 5 summarizes the statistical results.

TABLE 5. Summary of statistical comparisons between ’Beautiful’ (B) and ’Not Beautiful’ (NB) groups for selected EEG features.

Power-related features and Hjorth activity are log10-transformed to improve interpretability.

The biggest difference between the two groups was observed in gamma power, with a statistically significant group difference with a large effect size (p = 0.042, d = −1.46). Participants who rated the sculpture as ‘Beautiful’ had lower gamma power compared to those who rated the sculpture as ‘Not Beautiful’ (0.33 vs. 1.21) (see Figure 10). Several additional EEG features showed medium-to-large effect sizes numerically; however, because these comparisons did not reach statistical significance, they should be interpreted only as exploratory descriptive observations. Delta power was lower in the beautiful group (4.36 vs. 5.02, d = −0.94, p = 0.056), as well as beta power (1.26 vs. 2.00, d = −0.84, p = 0.056), and alpha power (2.12 vs. 2.83, d = −0.81, p = 0.056) (see Figure 10).

The beta-to-alpha ratio was higher in the ‘Beautiful’ group (d = 1.09, p = 0.053; 0.32 vs. 0.18 in the ‘Not Beautiful’ group) (see Figure 11).

EEG activity was lower in the ‘Beautiful’ group, with values of 5.83 versus 6.23, showing a large effect size (d = −1.00, p = 0.093). Theta power also showed a medium effect (3.09 vs. 3.67, d = −0.74, p = 0.093), fitting the pattern of lower power across most frequency bands in participants who found the sculpture. beautiful (see Figure 12).

EEG complexity showed a medium effect size (d = −0.59, p = 0.073), with the ‘Beautiful’ group having lower complexity values (5.14 vs. 5.76). EEG mobility had a smaller effect in the opposite direction (d = 0.46, p = 0.093), with slightly higher values in the ‘Beautiful’ group (0.14 vs. 0.10) (see Figure 13).

5 Discussion

5.1 Survey results

The survey-based analysis directly addresses the first research question, which examines which aesthetic and emotional judgments can be explained by sculptural attributes and how these relationships change over time. Across both initial and post-exposure evaluations, only three outcomes—Beauty, Valence, and Technical Skill—were meaningfully predictable from the rated attributes. Predictive performance increased from Round one to Round 3, particularly for Valence (variance increased from 68% to 88%), suggesting that participants incorporated attribute-based reasoning more consistently after being explicitly prompted to evaluate these features. These findings indicate that aesthetic judgments are not fixed but can become more structured when viewers are guided to attend specific perceptual dimensions. However, given the small number of stimuli and the aggregation of data at the sculpture level, these patterns should be interpreted as indicative rather than generalizable. The observed relationships may reflect properties of the specific stimulus set rather than stable principles of sculpture perception.

The Valence model showed the largest improvement, with Visual Harmony and Texture emerging as stable positive predictors across rounds. This finding aligns with previous research on Western art. Spee et al. (2023), for example, reported that visual harmony and emotional intensity play key roles in predicting creativity and liking judgments in paintings. The perceptual attributes identified in the present study are therefore consistent with earlier work.

The shift in Texture from a negative to a positive predictor suggests that participants developed a preference for smoother surfaces after exposure to attribute descriptions. Judgments of Technical Skill also changed: whereas initial evaluations were explained by Emotional Expressiveness and Pose Dynamics, later evaluations were primarily driven by Complexity. In Round 1, Emotional Expressiveness and Pose contributed most strongly to the model, whereas in Round 3, Complexity emerged as the dominant contributor. This shift may suggest that participants initially relied more on intuitive and affective impressions, but after explicitly evaluating sculptural attributes in Round 2, later judgments became more structurally and analytically oriented. Participants in this study were explicitly non-experts and evaluated sculptures without contextual information such as artistic intent, historical background, or production techniques. Under these conditions, it is unlikely that ratings of “technical skill” reflect expert-level assessments of craftsmanship in an art-historical sense. However, because of the exploratory design and limited stimulus set, this interpretation remains speculative and requires replication. Instead, the construct likely functioned as a perceptual proxy, capturing surface-level impressions such as visual polish, perceived effort, structural coherence, or familiarity with realistic representation. The finding that attributes such as Emotional Expressiveness, Pose, and Complexity contributed to Technical Skill ratings supports this interpretation: participants may have equated dynamic, expressive, or intricate forms with higher “skill,” rather than evaluating technical mastery per se. Thus, within the present paradigm, Technical Skill should be understood as a subjective, perception-based judgment rather than an objective measure of artistic expertise. Future studies could address this limitation by including expert participants, providing contextual information, or explicitly disentangling different components of perceived skill (e.g., realism vs. craftsmanship vs. complexity).

Overall, these findings propose that aesthetic judgments are flexible and can evolve with increased perceptual awareness. However, they should be interpreted cautiously due to the limited diversity of the stimulus set and the small sample size. The ten sculptures represent a narrow selection of modern figurative works, making it difficult to determine whether the observed effects reflect general preferences for attributes such as texture or are driven by responses to specific artworks. Consequently, it remains unclear whether similar patterns would emerge for more abstract, culturally diverse, or conceptually complex sculptures.

5.2 Eye-tracking results

The eye-tracking results suggest differences in visual engagement between participants who found the sculpture beautiful and those who did not. Although most effects did not reach statistical significance due to the small sample size, consistent trends emerged that warrant further investigation. Participants who rated Standing Woman as beautiful showed lower gaze depth variability, indicating more stable depth focus. Given the sculpture’s limited depth variation, this likely reflects sustained attention to the sculpture rather than shifts toward background elements. The “Beautiful” group also exhibited fewer fixations and saccades, longer mean fixation durations, and lower blink rates. Together, these patterns suggest reduced exploratory scanning and more sustained, focused viewing. These trends align with prior findings in two-dimensional art, where positive aesthetic judgments have been linked to fewer fixations, longer fixation durations, and lower blink rates (Massaro et al., 2012; Brieber et al., 2014; Ranti et al., 2020). Lower vertical gaze dispersion further supports the interpretation of more constrained and focused attention, consistent with earlier work (Vogt and Magnussen, 2007).

These trends are broadly consistent with prior work linking aesthetic preference to more focused visual engagement. However, given the limited sample size and the absence of correction for multiple comparisons, these findings should be interpreted as preliminary patterns rather than robust effects. Larger-scale studies are needed to confirm whether these visual engagement patterns reliably predict aesthetic judgments in three-dimensional art contexts.

5.3 EEG-results

The EEG analysis revealed distinct neural activity patterns between participants who found Standing Woman beautiful and those who did not. The “Beautiful” group showed lower gamma power, a band associated with high-level cognitive processing and attention (Jensen et al., 2007). This suggests that perceiving the sculpture as beautiful required less effortful processing, consistent with processing fluency theory (Reber et al., 2004) and prior findings that aesthetically pleasing stimuli evoke reduced high-frequency activity (Cela-Conde et al., 2011). However, because no electrooculogram (EOG) artifact correction was applied, gamma-band effects may partly reflect eye-movement artifacts rather than neural activity [63]. Future work should therefore include ocular artifact removal. The “Beautiful” group also exhibited a higher beta-to-alpha ratio, indicating greater cognitive arousal and focused engagement (Knyazev, 2007).

In addition, power was lower across several bands (delta, alpha, and beta), replicating earlier findings that aesthetic appreciation is associated with reduced neural activation (Knyazev, 2012; Munar et al., 2012). Lower delta power suggests sustained external attention rather than internal processing (Knyazev, 2012). Overall EEG amplitude was reduced in the “Beautiful” group, indicating more efficient neural processing during aesthetic appreciation. This may reflect reduced neural activation during aesthetic appreciation. Though this interpretation should be treated with caution given the single-channel setup and absence of EOG artefact correction. Together, these results characterize a neural profile of beauty perception marked by lower gamma, alpha, beta, and delta power alongside a higher beta/alpha ratio. This pattern supports the view that artworks are experienced as beautiful when they can be processed fluently and with sustained attention, rather than with increased cognitive effort, in line with processing fluency theory and prior neuroaesthetic research.

Both EEG and eye-tracking results reveal the same pattern: aesthetic appreciation of sculpture involves flowing efficient focused processing, rather than effortful one. The eye-tracking data showed that participants who found the sculpture beautiful exhibited 24.26% fewer fixations and 25.66% fewer saccades, 38.58% longer fixation durations, and significantly lower gaze depth variability, indicating focused attention rather than extensive visual search. EEG findings revealed that participants who found the sculpture beautiful had significantly lower gamma power and lower overall neural activity, indicating more efficient cognitive processing with less neural effort. These EEG and eye-tracking findings characterize a state of engaged, but at the same time effortless processing of a sculpture in a group that found the sculpture beautiful. Participants maintained focused attention (higher beta/alpha ratio of 0.32 vs. 0.18, longer fixations) while requiring less cognitive effort (lower gamma power, fewer eye movements). The shared pattern across two completely independent physiological analyses suggests convergent patterns across two independent physiological modalities, which should be further investigated in larger samples” … Despite a small sample size and the challenges of collecting clean physiological data in a VR environment, both neural and visual engagement patterns showed the same conclusion. These findings offer preliminary/exploratory answers to our second research question about whether there are measurable physiological differences in visual engagement and neural activity between participants with contrasting aesthetic evaluations.

Taken together, the results of this study are best understood within an exploratory framework. The combination of a small sample, a limited stimulus set (10 sculptures), single-sculpture physiological analysis, and technical constraints in EEG acquisition limits the extent to which strong or general conclusions can be drawn. Therefore, EEG findings should be interpreted cautiously and primarily serve as an initial exploration of neural correlates of aesthetic judgments in immersive 3D environments. Rather than providing definitive answers about the mechanisms of aesthetic experience, the study identifies promising patterns (1) Certain perceptual attributes (e.g., Visual Harmony, Texture): appear consistently related to aesthetic judgments. (2) Aesthetic evaluations become more structured after guided attention to attributes (3) Multimodal physiological measures show converging, though preliminary, differences between evaluation groups.

6 Limitations

Several limitations should be considered when interpreting the findings. The most important constraint is the small and homogeneous sample (28 non-expert participants aged 18–33), which limits generalizability to older adults and art experts and results in limited statistical power. The machine learning analysis also faced constraints. Aggregating data at the sculpture level yielded only 10 data points for models with 15 input features, increasing the risk of overfitting despite cross-validation. Consequently, the identified attribute–judgment relationships may not generalize beyond this dataset. However, the goal was exploratory—to identify which sculptural attributes best explained judgments within this specific collection—and the consistency of attribute rankings across judgments suggests meaningful internal patterns. The stimulus set was limited to 10 modern Western figurative sculptures, excluding contemporary, non-Western, and abstract works. Although high-quality 3D scans were used, digital reproductions may differ from real-world viewing conditions. Data quality was also affected by technical issues: the single-channel in-ear EEG was sensitive to movement artifacts, and eye-tracking accuracy varied with headset fit and calibration. The absence of head-position tracking further limited spatially precise gaze analysis.

A further limitation addressed the selected sculptures which might have introduced an aesthetic bias. As the artworks were professionally created sculptures, participants may have been predisposed to evaluate them positively. Following that future studies should include a broader range of stimuli, including artworks varying more strongly in aesthetic appeal, style, and quality, to increase variability in aesthetic evaluations. Furthermore, the artwork of less known artists could be also included in a future study.

Ecological validity was constrained by the VR setting and fixed viewing duration (70 s per sculpture), which may have altered natural viewing behavior. Finally, participant fatigue likely influenced later responses: the session lasted about 1 hour, and repeated exposure to the same sculptures across three rounds led to reports of boredom and disengagement. Despite these limitations, the study provides useful insights into VR-based sculpture aesthetics and establishes a methodological framework for future work with larger, more diverse samples and improved technical controls.

7 Conclusion

This study addresses a gap in research on sculptural aesthetics by examining how 28 art novices evaluated ten modern figurative sculptures across three rounds in a custom virtual reality gallery. Participants provided aesthetic, emotional, and attribute ratings while eye-tracking and in-ear EEG were recorded, enabling complementary survey-based and physiological analyses. Only three judgments were consistently predictable from sculptural attributes: Beauty, Valence, and Technical Skill. Visual Harmony and Texture emerged as the strongest predictors, with smoother and more harmonious sculptures preferred. Judgments became more systematic over time, and Valence predictability increased from 68% to 88% explained variance. Physiological results converged with these findings: participants who rated sculptures as beautiful showed fewer fixations and saccades, lower gamma power, and more stable visual attention, indicating efficient and focused processing. The results suggest that VR is a viable platform for controlled aesthetic research on sculpture. The convergence between survey-based and physiological outcomes, together with consistency with prior 2D and real-world studies, supports VR’s validity for empirical aesthetics. By combining interpretable machine learning with multimodal physiological measures, this work provides a preliminary framework for studying three-dimensional aesthetic experience and offers empirical support for processing fluency theory in immersive environments. Summarizing, this study provides initial evidence that aesthetic judgments of sculptures in VR can be partially explained by perceptual attributes and are accompanied by measurable differences in visual and physiological responses. However, these findings should be interpreted cautiously. The results do not establish definitive mechanisms but instead highlight potential relationships and methodological directions for future research. Larger, more rigorously controlled studies will be required to validate and extend these observations.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. Generative AI was used to proof english.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/frvir.2026.1804634/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Xianxun Zhu, Shanghai University, China

Reviewed by: Takeharu Seno, Kyushu University, Japan

Sunny Bairisal, National Institute of Fashion Technology, India