This study involved 26 healthy adult Japanese participants with no history of neurological or psychiatric disorders or professional education in artistic disciplines. Three participants were excluded from the analysis: one due to unremovable artefacts (e.g., saturated signals) in the recorded MEG data from previous dental work, and two for failure to maintain dichoptic vision, leading to ineffective stimulus suppression with CFS. Data obtained from the remaining 23 participants (16 females, all right-handed, mean age 35.74 ± 10.35 years) were analysed. This study was conducted in accordance with the ethical principles of the Declaration of Helsinki and was approved by the Ethics Committee of Hokuto Hospital (approval number: #1108) and Osaka Metropolitan University Graduate School of Medicine (2022-034). Written informed consent was obtained from all participants prior to their involvement in the study.

A set of 5 portrait paintings by 10 artists was sourced from WikiArt¹ (Supplementary Table 1). The paintings were selected to achieve gender balance (women, men, and unknown) among the painted models portrayed by each artist, ensuring a difference in the number of female and male models of < 2 (e.g., three women and two men). All paintings were of Western origin, spanning various styles from the 16th to the 20th centuries, and explicitly depicted human faces (portraits) or contained identifiable keywords depicting the presence of human faces, such as ‘head’ and ‘man with a moustache.’ All paintings were cropped to focus solely on the facial areas (from the top of the head to the chin) in a square of 500 × 500 pixels, converted into greyscale, and equated for spatial frequencies and histograms using the SHINE toolbox (Willenbockel et al., 2010). Finally, to enhance the suppression effect, the image edges were blurred, blending them into a background colour, defined as the mean luminance (intensity) of all images. For this process, an additional 25 pixels with the background colour were added to each side of the image, which were blurred by filtering with a two-dimensional Gaussian smoothing kernel with a σ (standard deviation of the Gaussian distribution, indicating the filtering effect strength) determined as a normal probability density function relative to the location of the pixel to the edge. Essentially, pixels closer to the edge were more strongly blurred, and vice versa, ensuring a smooth transition between the image and background colour. The facial stimuli were sorted based on the painters and their era (Supplementary Table 1), with their representative images depicted in Figure 1A.

Prior to the MEG experiment, participants completed pre-rating questionnaires to assess the intensity of the aesthetic emotion aroused by each facial stimulus. The paper-based questionnaire contained five questions/ratings for each stimulus, covering ‘biologi-ness’ (‘How much do you find the painted objects biological?’), object saliency (‘How well can you recognise the painted object?’), facial saliency (‘How well can you recognise a face in the painting?’), liking (‘How much do you like the painting?’), and beauty (‘How much beauty do you feel from the painting?’). Object and facial saliency scores were obtained to assess the intensity of perceptual and cognitive processing, an influential factor in aesthetic emotion (Reber et al., 1998, 2004). The beauty score was the main target of the analysis, representing a direct measure of positive aesthetic emotion, while the liking score was used to measure the intensity of aesthetic pleasure predicted by aesthetic emotion (Wassiliwizky and Menninghaus, 2021). Participants used a visual analogue scale (VAS) to indicate their subjective experience by drawing a vertical line on a horizontal line ranging from each extreme (‘biologi-ness’: from ‘not biological at all’ to ‘very biological’; object saliency: from ‘I cannot recognise at all’ to ‘I can recognise clearly’; facial saliency: from ‘I cannot recognise at all’ to ‘I can recognise clearly’; liking: from ‘I do not like it at all’ to ‘I like it very much’; and beauty: from ‘not beautiful at all’ to ‘very beautiful’). The VAS scores were normalised, that is, converted from 0 (negative extreme) to 1 (positive extreme) by measuring the position of the vertical line (response) from the edge of the horizontal line and used for the analysis. Notably, paper-based pre-rating questionnaires were used to minimise the duration of the MEG experiment and reduce potential eye fatigue among participants while maintaining dichoptic vision with stereoscope, which would lead to poor suppression effects of CFS. However, this repeated exposure of the stimuli through the use of pre-rating may have influenced the intensity of aesthetic pleasures, a limitation we have acknowledged in section 4.5. To validate the pre-rating scales, particularly for the unique measure employed in this study (‘biologi-ness’), three participants underwent a retest after 1 year of the initial pre-rating date (range, 409–598 days). The test-retest correlation (Spearman’s rho) of the ‘biologi-ness’ rating was comparable to that of object and facial saliency scores, indicating good reliability and replicability of the measurement (Supplementary Table 2).

The MEG experiment was conducted within 5 days of completing the pre-ratings. First, participants received instructions about the tasks and familiarised themselves with dichoptic vision using a handmade, non-magnetic stereoscope (Figure 1B). Subsequently, the participants lay down in a supine position in a magnetically shielded room (Figure 1C), with the stereoscope installed on the MEG dewer (between the participant’s face and the screen). Through the stereoscope, participants viewed the left half of the screen with their left eye and vice versa. To achieve the best suppression effect in CFS, the positions of stimuli presented to the left and right eyes were individually adjusted whereby the participants were presented with two squares (8.0° × 8.0° in visual angle) on the screen, one for each visual field, and asked to move around and find positions where the two overlapped perfectly in subjective perception (dichoptic vision). All images and stimuli were projected onto a screen placed 28.7 cm in front of the participants from outside the magnetically shielded room using a projector (PROPIxx; VPixx Technologies, Saint-Bruno, Canada) through a mirror. MEG data were recorded during the CFS task, which comprised four sessions, each containing 30 trials. Of the 120 trials, 50 were categorised as face-face (FF), another 50 as Mondrian-face (MF), and the remaining 20 as Mondrian-background (MB). In the FF and MF conditions, 50 facial stimuli were randomly presented once for each condition. The order of the three conditions was pseudo-randomised to ensure that no more than three consecutive trials belonged to the same condition. Figure 1D depicts the differences between the conditions. In the FF condition, each trial started with the presentation of fixation (2.7° × 2.7° in visual angle; fixation phase, 500 ms). Then, the facial stimulus faded in (fading-in phase, 700 ms), remained fixed on the screen (face phase, 2,000 ms), faded out (fading-out phase, 700 ms), and its aftereffect was suppressed by presenting a noisy mask (masking phase, 500 ms). The noisy mask comprised an image filled with 10,000 greyscale dots (arranged in a 100 × 100 matrix within a square of 8.0° × 8.0° in visual angle) of random intensity. In this condition, identical content was presented to both eyes, allowing participants to clearly visualise and consciously process the content. In the MF condition, flashing Mondrian images were presented to the participant’s dominant eye instead of a facial stimulus. The Mondrian images were randomly generated images with 1,000 randomly sized ellipses coloured in one of eight primary colours: red ([255 0 0] in RGB), green ([0 255 0]), blue ([0 0 255]), cyan ([0 255 255]), magenta ([255 0 255]), yellow ([255 255 0]), black ([0 0 0]), and white ([255 255 255]). The presentation of Mondrian images commenced 100 ms before the fading-in phase (preparation phase), where the images were sequentially refreshed with a temporal frequency tuned at 6.25 Hz, as previous studies indicated peak effectiveness of the masking/suppression at approximately 6 Hz (Zhu et al., 2016; Drewes et al., 2018). Participants who successfully maintained dichoptic vision were aware of only the flashing Mondrian stimuli presented to the dominant eye and were unaware of the facial stimuli presented to the non-dominant eye. Thus, the facial stimuli were processed non-consciously in the MF condition. In the MB condition, Mondrian images were presented to the dominant eye, while a stable background-coloured square was presented to the non-dominant eye. Similar to the MF condition, participants were aware of only the flashing Mondrian stimuli and were unaware of the background-coloured image presented to the non-dominant eye.

After each trial, the participants were asked to respond to two-alternative forced choice questions: Q1: ‘Did you see a face?,’ which was asked by presenting ‘See?’ on the screen (response phase 1) and Q2: ‘Do you feel beauty from the painting?,’ which was asked by presenting ‘Beauty?’ on the screen (response phase 2). Participants indicated answers by pressing one of the two buttons corresponding to ‘yes’ and ‘no’ using their left and right thumbs, with positions counterbalanced between participants. Participants were instructed to respond ‘yes’ to the Q1 (‘see?’ question) not only when the whole facial stimulus was visible but also if any minute part of the stimulus was visible despite Mondrian suppression. Furthermore, participants were instructed to guess the answer to Q2 (‘beauty?’ question) when they did not (consciously) see any faces in the trial. Responses to Q1 were used solely for screening and rejecting trials where the facial stimuli were not perfectly masked in the MF condition. Each response phase lasted until the participants provided their responses, with an inter-trial interval set at 1 s. Each session lasted approximately 4–5 min, varying based on the response time. The entire experimental procedure, including preparation, lasted approximately 60–90 min.

Cortical activity during the CFS task was recorded using a 160-channel whole-head-type MEG system (MEG vision PQ1160C; Yokogawa Electric Co., Kanazawa, Japan). During the scan, participants were asked to remain calm in a supine position in a magnetically shielded room, with scanning conditions controlled for consistency and comfort. The sensor and reference coils were gradiometers of 15.5 mm in diameter and 50 mm in baseline. Each pair of sensor coils was separated by a distance of 23 mm. The sampling frequency was 2,000 Hz, with 500 Hz low-pass filtering during recording. To co-register the MEG data with the anatomical brain images, five fiducial magnetic marker coils were placed on each participant’s face (40 mm above the nasion, bilaterally 10 mm in front of the tragus, and bilateral pre-auricular points) before the MEG scan, and their spatial coordinates were measured immediately before each session. During the scan, participants were monitored by a technical staff member using a video camera installed in the magnetically shielded room. Among the 23 participants, the individual magnetic resonance imaging (MRI) data of 5 participants were obtained retrospectively from another study. These were anatomical T1-weighted MR images acquired using a 3.0-T scanner (SIGMA Excite 3.0T, GE Healthcare, Milwaukee, WI, USA) with a standard head coil and three fiducial markers (Medtronic Surgical Navigation Technologies Inc., Broomfield, CO, USA) positioned at the three magnetic marker coils placed on the forehead.

MEG data were analysed offline using RICOH MEG Analysis software (RICOH, Tokyo, Japan), MATLAB (MathWorks, MA, USA), Brainstorm, which is documented and freely available for download online under the GNU general public licence² (Tadel et al., 2011), and the FreeSurfer image analysis suite, which is documented and freely available for download online.³ First, continuous MEG signals were cleaned using a dual-signal subspace projection algorithm (Sekihara et al., 2016) available in the vendor-provided software (RICOH MEG Analysis), comparable to the temporally extended signal-space separation algorithm, with the only difference being the approximation of the signal subspace projector (Cai et al., 2019). Next, to remove the remaining artefacts, the signals were decomposed via independent component analysis using the infomax algorithm implemented in Brainstorm (Makeig et al., 1996). Each component from the independent component analysis was visually inspected, and those with cardiac, blinking, and other salient artefacts were rejected. The data were divided into 6,000 ms epochs, each starting 1,500 ms before the preparation phase (1,000 ms before the fixation phase) and lasting 1,000 ms after the fading-out phase. This epoch length was designed to avoid edge effects in the time window of interest (from the preparation to the fading-out phase) caused by the filtering process for computing the envelopes, as described later. The starting point of the time window of interest (the onset of the preparation phase) was considered to be 0 ms throughout the analysis. The offset was removed using baseline signals averaged across −1,000 to 0 ms for each sensor and epoch. Epochs (trials) were rejected and excluded from the following analyses based on three criteria: (i) epochs corresponding to trials in the MF condition where participants responded ‘yes’ to the Q1 (‘see?’ question), indicating imperfect masking of facial stimuli (ii) technical issues resulting in the last trial of the session, during which the MEG measurement was finished, not being recorded, and (iii) trials contaminated with remaining artefacts. Based on these criteria, the number of trials rejected was 1.17 ± 2.48 trials (with a maximum of 9 trials) per participant, 1 trial from two participants and 2 trials from one participant, and 3.57 ± 3.07 trials (with a maximum of 10 trials) per participant, respectively. Consequently, the number of trials considered for the analyses were: FF: 48.70 ± 1.40 trials (19.48 ± 0.59 trials for biological, and 19.35 ± 1.07 trials for non-biological stimuli), MF: 47.09 ± 2.97 trials (18.91 ± 1.28 trials for biological and 18.74 ± 1.48 trials for non-biological stimuli), and MB: 19.30 ± 0.88 trials.

The remaining artefact-free signals were projected onto the cortical source using the default parameters of the Brainstorm toolbox. For the structural MRI data, ICBM152, a template anatomical brain image prepared by Brainstorm, was used for analysis. The ICBM152 is a non-linear average of 152 MR images from different subjects (Fonov et al., 2009) and is provided along with the cortical segments. In cases where individual MRI data were obtained retrospectively from another study (5 out of 23 participants), the T1-weighted MR images were segmented using the FreeSurfer pipeline ‘recon-all’ (Fischl, 2012) and then imported into Brainstorm. The signal source was restricted to the cortex, which was segmented into 15,000 vertices. Each MEG session was co-registered with the anatomical image using the spatial coordinates of five fiducial points and the nasion, and the relationship between 160 MEG channels and 15,000 vertices (leadfield matrix) was modelled (forward modelling) using an overlapping sphere model, a recommended option for MEG data in the Brainstorm. Before computing the source signals, the characteristics of the MEG sensor noise were modelled as a covariance matrix for each pair of channels (noise covariance), computed from an empty room recording (≥ 5 min) measured before starting the MEG experiment (< 1 h) for each participant using an identical MEG machine with the same acquisition setting. Using the forward model and noise covariance matrix, the source signals of MEG data were computed using the weighted minimum norm estimation (wMNE) method (Lin et al., 2006), which restricts the sources of the inverse problem by minimising the energy (L2 norm), while weighting the deep sources to facilitate their detection. This algorithm was selected because (i) it was recommended as a default option in Brainstorm and (ii) it is suitable for cases where a template brain is used instead of individual MR images, which only returns rough approximations of the forward model and is unsuitable for other inversion algorithms (such as beamformer), which require a better model approximation than the wMNE method. The orientation of the neural sources was restricted to be normal to the cortex. The data obtained from the source reconstruction process consisted of epoch time-series signals for each of the 15,000 cortical vertices. High-dimensional data were limited to 68 anatomical regions of interest (ROIs) defined by the Desikan-Killiany atlas (Desikan et al., 2006) by averaging the signals of the vertices included in each anatomical region. During averaging, the signs of the signals were flipped in the vertices, where the normal orientation was opposite to the dominant orientation of the corresponding region. For visual inspection, the ROI time series were averaged across trials under the same conditions (FF, MF, and MB) and categories (biological and non-biological) and across all or occipital ROIs. Finally, to obtain the non-phase-locked induced oscillatory power of the MEG data, a time-frequency (TF) analysis was performed. TF data were computed as envelopes of the epoched ROI time series before averaging, which were extracted using the Hilbert transform. The signals were narrow-band filtered for delta (δ, 2–4 Hz), theta (θ, 5–7 Hz), alpha (α, 8–12 Hz), beta (β, 15–29 Hz), low-gamma (lγ, 30–48 Hz), and high-gamma (hγ, 52–90 Hz) frequency activities, and their envelopes were computed as power of the Hilbert transform for each frequency band. The frequency windows for gamma bands (lγ and hγ) were designed to avoid power line noise at 50 Hz. The envelopes were flattened by multiplying the amplitudes by frequency (1/f compensation), normalised against a baseline period from −1,000 to −500 ms (ITI), and used in the statistical analyses.

In this study, we subcategorised the stimuli into two subsets based on the ‘biologi-ness’ score of the pre-ratings by each participant. Despite controlling for low-level features, qualitative differences between the categories influenced the results. Overall, the biological category scored higher on all pre-ratings than the non-biological category (Table 1), supporting the processing fluency theory of aesthetic pleasure (Reber et al., 1998, 2004). This theory proposes that ‘the more fluently perceivers can process an object, the more positive their aesthetic response,’ where prototypicality is a variable influencing fluency. In the present study, the biological category scored higher in the ‘biologi-ness’ rating, where the faces were depicted more clearly (higher in object and facial saliency ratings). Hence, as these stimuli were more prototypical than the other facial stimuli, they were processed more fluently, leading to enhanced affective ratings, such as liking and beauty, as exemplified by positive correlations between the rating scores (Supplementary Table 3). However, in the non-biological category, the correlation revealed that the liking score was independent of ‘biologi-ness’, object saliency, and

facial saliency scores (Supplementary Table 4). This indicated that factors beyond processing fluency influenced the aesthetic liking scores because the non-biological stimuli were low in prototypicality and their fluent processes were interrupted (Reber et al., 1998, 2004). The pleasure-interest model of aesthetic liking (PIA model) (Graf and Landwehr, 2015) distinguishes aesthetic

liking based on its construction mechanisms. Pleasure-based liking is influenced by processing fluency and triggered by bottom-up stimulus-driven automatic processing, whereas interest-based liking is elicited by the reduction of disfluency during top-down perceiver-driven controlled processing (Graf and Landwehr, 2017). For non-biological stimuli, including paintings in abstract styles (such as cubism, abstract art, and surrealism), participants are more likely to engage in deeper active elaboration to resolve the disfluency of the abstract paintings, triggering interest-based aesthetic liking. Therefore, interest-based liking is assumed to be another contributor to liking scores for non-biological stimuli. Additionally, the integration of processing fluency and learning theories for making future predictions provides explanations of aesthetic processing (Brielmann and Dayan, 2022; Starr, 2023). Within this line of research, the role of predictive processing (PP) has recently been receiving attention as a mental function accounting for the aesthetic experience (Frascaroli et al., 2024). This framework suggests that encountering unpredictable stimuli (e.g., artworks), which deviates from the schema of the viewers, gives rise to prediction error and ‘uncertainty,’ whose resolution (i.e., change) drives a positive affective experience (Van De Cruys et al., 2024). When presenting the stream of portrait paintings in our study, participants may have naturally expected the presence of facial images, and the presentation of less typical facial stimuli, namely the ones in the non-biological category, may have increased prediction error and uncertainty. Contemplation or elaborations of such stimuli might have led to the resolution of uncertainty (e.g., identification of facial images or finding them to ‘make sense’), which consequently enhanced the aesthetic pleasure measured by the liking score. In addition to interest-based liking, this serves as a unique source of the liking score in non-biological stimuli. Notably, the PP framework complicates the processing fluency theory that hedonic experience driven by fluency is biassed by prior expectation to the stimuli or fluency itself (e.g., ‘more fluent than expected,’ ‘prefer to be kept in a state of puzzlement’) (Yoo et al., 2024). Furthermore, the prototypicality of the stimuli, which is a key driver of the aesthetic emotion in processing fluency theory, is formed by repeated PP throughout the life of the human being. As exemplified by the ‘other-race effect’ (Pesciarelli et al., 2021; Stelter and Schweinberger, 2023) and facial familiarity effects (Axelrod et al., 2015), the ‘biological’ visual features for carrying evolutionary important information vary individually, although they share some common characteristics, such as symmetry and good proportion. This indicates that the ‘personally’ formed prototypical representation of an object, which is updated daily through individual active inferences (i.e., PP), is considered to carry evolutionarily important information and has prioritised access to the emotional system. Therefore, the PP framework not only plays a role as a unique contributor to aesthetic emotion for non-biological stimuli but also serves as a generative function of prototypical images and influences fluency-driven aesthetic emotion. Upon examining the association between pre-ratings and behavioural responses during the CFS task (Figures 2B, C and Table 2), we found that all pre-rating scores, except the facial saliency score, differed between beauty and non-beauty responses in the FF (conscious)-biological condition (Figure 2B and Table 2A). Moreover, a weak link between the facial saliency and beauty ratings was demonstrated by the low correlation coefficients (Supplementary Table 3). This finding may be attributed to a ceiling effect on the facial saliency score in biological stimuli (Figure 2A). Owing to the strong correlation between facial saliency and ‘biologi-ness’ scores and selection criteria that only retained stimuli with very high ‘biologi-ness’ scores in the biological category, both scores hit the ceiling. Consequently, correlations between facial saliency and ‘biologi-ness’ scores with other pre-rating scores were lower for the biological category than those of the non-biological category, which did not experience this effect. Notably, the non-consciously presented stimuli modified the behavioural responses during the CFS task, such that the object saliency score was higher for beauty than for non-beauty responses in the MF-biological condition, while the ‘biologi-ness’ and object and facial saliency scores were higher for beauty than for non-beauty responses in the MF-non-biological condition (Table 2B). This implies that non-conscious neural inputs of salient stimuli may bias subsequent aesthetic labelling behaviour. Previous studies have revealed that non-consciously presented faces undergo not only emotional, but also perceptual and cognitive processes (Axelrod et al., 2015), such as discrimination of face vs. scrambled face (Jiang and He, 2006), upright vs. inverted faces (Jiang et al., 2007; Stein et al., 2012), and familiarity (Gobbini et al., 2013). As image saliency directly influences perceptual and cognitive processes, it may interfere with the perceptual and/or cognitive fluency of non-conscious processing of facial stimuli and change their gut-level feelings captured by guessing the aesthetic quality of covert stimuli. This aligns with a concept of ‘unfelt’ fluency, which occurs on a lower, perceptual, and sub-personal level, reflects the level of matches between perceptual information with basic visual expectations derived from our visual systems (i.e., prototypical patterns of neural activity, determined congenitally and formed by PP), and operates as a non-conscious process (Brouillet and Friston, 2023). As beauty and liking scores are the results of conscious aesthetic evaluations, they are dynamic reflections of various perceptual and cognitive processes, such as fluency, elaboration, and prediction. However, in the non-conscious presentation scenario, the top-down regulative process is inactive for stimuli with weak intensity (such as subliminal stimuli) (Baars, 1988; Dehaene et al., 1998), indicating that only limited perceptual and cognitive processes can contribute to aesthetic emotion. For example, the interest-based liking system and PP for resolving uncertainty require a top-down regulation of incoming information to drive aesthetic emotion; thus, these would not be involved in the processing of non-consciously presented stimuli. Therefore, in the present study, the pleasure-based liking system, driven by ‘unfelt’ processing fluency, was likely dominant for aesthetic processing in the MF condition. The behavioural responses to the non-consciously presented stimuli were modified by the saliency scores but not by the beauty or liking scores, indicating that non-consciously presented information accesses limited processes associated with the perceptual saliency of the stimuli, such as the ‘unfelt’ fluency; hence, the behavioural response observed was likely based upon such limited information. Therefore, the results of the aesthetic labelling based on limited information (behavioural responses to the non-consciously presented stimuli) would be distinctive from those based on comprehensive information (beauty and liking scores in the pre-rating).

After highlighting the behavioural differences between the biological and non-biological categories in section 4.1 we discussed the interaction between biological vs. non-biological and conscious vs. non-conscious contrasts in the correlation analysis between the pre-ratings and TF data. Here, we have focused on the results in the most representative FF (conscious)-biological condition (Figures 5A–D and Tables 5A–D) and structured the following paragraphs to sequentially discuss spatial, oscillatory, and temporal dimensions for clarity.

Regarding the spatial (regional) dimension of the results, the saliency scores were positively correlated with low-frequency activities (from the delta to alpha bands) in parietal and temporal regions (Figures 5A, B). These results aligned with those of previous studies indicating that emotional faces induce delta and theta band activities, particularly in occipitotemporal regions (Güntekin and Başar, 2014). Moreover, the level of processing fluency in facial stimuli was associated with neural responses in occipitotemporal and parietal regions (Natu and O’Toole, 2011). Multiple face-selective areas in the occipital and temporal cortices include the occipital face area (Haxby et al., 1999; Gauthier et al., 2000), fusiform face area (Kanwisher et al., 1997), posterior part of the superior temporal sulcus (Kanwisher et al., 1997; Hoffman and Haxby, 2000), anterior temporal lobe (Tsao et al., 2008; Rajimehr et al., 2009), and anterior superior temporal sulcus (Pitcher et al., 2011), which border the clusters identified relevant to the saliency scores. Stimuli with higher saliency scores were processed more fluently and exhibited enhanced slow band activities in these regions, signifying involvement of bottom-up perceptual and cognitive processes related to visual images. In contrast, affective processing of the stimuli, quantified by liking and beauty scores, negatively correlated with low-frequency activities mainly in frontal regions, such as the lOFC and superior frontal gyrus (Figures 5C, D). Previous studies have demonstrated that prefrontal regions respond to aesthetic ratings of various stimuli (Nakamura et al., 1998; Kawabata and Zeki, 2004; Di Dio et al., 2007; Winston et al., 2007; Kirk, 2008; Chatterjee et al., 2009; Ishizu and Zeki, 2011; Kedia et al., 2014; Martín-Loeches et al., 2014; Ferrari et al., 2015, 2017; Chuan-Peng et al., 2020). This indicates that reductions in low-frequency activities in the ‘emotional brain’ coincide with the arousal of aesthetic emotions.

Upon exploring the oscillatory characteristics of the induced response in the FF-biological condition, perceptual/cognitive (object and facial saliency) and affective (liking and beauty) processes exhibited correlations in the same low-frequency activities, but in different regions and opposite directions. While the former exhibited positive relationships (Figures 5A, B), the latter exhibited negative relationships (Figures 5C, D). Similar contradictory results have been reported in studies on emotional faces, with some reporting augmented or induced event-related synchronisation (ERS) in delta and theta band activities (Aftanas et al., 2001, 2004; Balconi and Pozzoli, 2007, 2009; Başar et al., 2008; Balconi et al., 2009b,a; Bamidis et al., 2009; Güntekin and Başar, 2009, 2014; Knyazev et al., 2009), while others reporting their attenuation or event-related desynchronisation (ERD) (Balconi and Lucchiari, 2006; Balconi and Pozzoli, 2009). This inconsistency between ERS and ERD during emotional picture processing has also been documented in alpha and lower-beta bands (Schubring and Schupp, 2021). These discrepancies may be explained by the physiological antagonism between cognitive and affective processes in fluency-driven aesthetic processing (Hoshi and Menninghaus, 2018). A previous study demonstrated that highly affective texts predict larger pupil dilations, whereas highly fluent texts predict smaller pupil dilations (Hoshi and Menninghaus, 2018), indicating that the cognitive and affective processes of aesthetically appealing stimuli cause antagonistic reactions to the autonomic nervous system. Briefly, the balance between sympathetic and parasympathetic reactions drives pupillary dilation, with excitement in sympathetic activity and inhibition of parasympathetic activity evoking dilation, and vice versa (van der Wel and van Steenbergen, 2018). Although studies focusing on the relationship between task-related autonomic nervous system reactions and cortical oscillatory activity are limited, one study revealed that autonomic behaviour is associated with delta and theta band responses to emotional stimuli (Balconi et al., 2015). These findings imply that fluency-driven aesthetic processes induce cognitive and affective responses in antagonistic directions in delta and theta band activities in the central nervous system. Previous studies on visual emotional processing have revealed inconsistent results between ERS and ERD using standard picture groups, such as ‘Pictures of Facial Affect’ (Ekman and Friesen, 1976) and ‘International Affective Picture System’ (Lang et al., 1997). Researchers typically assumed that these standardised sets would be equally processed by all individuals in both cognitive and affective senses. However, the potential variabilities in cognitive and affective processes were often overlooked in these studies, which may have led some studies to emphasise cognitive processing and others to emphasise affective processing induced by the stimulus set, resulting in contradictory findings.

Upon examining the temporal dynamics in the FF-biological condition, we found that neural processes occur in multiple time windows. Remarkably, for the beauty score, a negative cluster was identified in a very early time frame in the left lOFC (preparation and fading-in phases, 0–1000 ms) (Figure 5D). Previous neuroaesthetics studies have used electrophysiological measurements, such as electroencephalography (EEG) (Jacobsen and Höfel, 2003; Muñoz and Martín-Loeches, 2015; Sarasso et al., 2020) and MEG (Cela-Conde et al., 2004, 2013), to study the temporal dimensions of aesthetic processing. Among these, two studies have suggested that aesthetic processing comprises two temporally distinct steps. One EEG-based study (Jacobsen and Höfel, 2003) examined event-related potential and found that graphical patterns judged as ‘not beautiful’ evoked early frontal negativity after 300 ms, whereas judgements of their symmetry evoked sustained posterior negativity approximately 600 ms after stimulus onset. Another MEG-based study (Cela-Conde et al., 2013) using paintings as stimuli reported distinct connectivity patterns between two time windows: an early window (250–750 ms), characterised by dense local connections within the occipital region and extending links to orbitofrontal regions for making quick judgements regarding beauty, and the later window (1,000–1,500 ms), with distributed global connection, which was active only for ‘beautiful’ stimuli and involved in appraisals of detailed aspects of beauty. Despite their different timeframes, these two studies reported temporally distinct processes: early processes related to aesthetic emotion for making quick impressions and later processes for detailed evaluations of stimuli. This two-step process corresponds to the temporal dynamics observed in our study, where early low-frequency activities in the OFC were associated with the affective dimension of aesthetic appeal. As the OFC is a core region for the subjective experience of aesthetic emotion, namely beauty (Ishizu and Zeki, 2011), the early processes might be related to aesthetic emotion for making quick impressions. Subsequently, low-frequency activities were induced by cognitive and affective processes in the other brain regions, where the stimuli could be aesthetically evaluated in detail.

In contrast to the FF (conscious)-biological condition discussed in section 4.2 we focused on the results of the correlation analysis in the MF (non-conscious)-biological condition (Figures 5E–H and Tables 5E–H).

Before discussing the results in the non-conscious condition, it is worth noting that the differences between conscious and non-conscious conditions were strongly biassed by frequency-tagged responses (steady-state visually evoked field) (Parkkonen et al., 2008; Vialatte et al., 2010) to Mondrian images, which entail robust visual evoked responses at a frequency corresponding to the refreshing rate of the presented stimuli. To optimise suppression effects (Zhu et al., 2016; Drewes et al., 2018), we set the refresh rate of the Mondrian images to 6.25 Hz (section 2.4). Visual inspection of the ROI time series (Figure 3) revealed clear ERFs at approximately 6 Hz under MF and MB conditions. The results of the TF data should be interpreted with a consideration of frequency-tagged responses in the theta band (5–7 Hz). As Mondrian images were presented throughout the time window of interest (0–3,500 ms), the sustained activity increased in the theta band, which was commonly found for both categories (biological and non-biological), should be considered as frequency-tagged responses to the Mondrian images. Sustained augmentation of theta band activity was found in the bilateral pericalcarine, lateral occipital, and inferior parietal cortices, lingual gyri, cuneus, and right parahippocampal cortex in the MF conditions (MF-biological and MF-non-biological) (Figure 4). The stimulus-related activities found in these regions must be contaminated by frequency-tagged responses and should be interpreted with caution.

In this section, the results of the MF-biological condition are discussed for saliency scores (object and facial saliency scores) and affective scores (liking and beauty). Saliency scores were negatively associated with low-frequency activity in the fronto-temporal regions (Figures 5E, F), such as the mOFC, rACC, parahippocampal, superior temporal, and inferior parietal cortices. The temporal regions overlapped with the face-selective areas, such as the anterior and posterior parts of the superior temporal sulcus (Duchaine and Yovel, 2015), indicating the involvement of bottom-up visual processing related to the facial images. However, the correlation direction was negative, which is opposite from that of temporal and parietal clusters identified in the FF-biological condition (Figures 5A, B). Additionally, more negative clusters were exclusively found in the frontal and limbic regions in the MF-biological condition, such as the mOFC and rACC (Figure 5E). Despite being designed to capture perceptual and cognitive processing of the stimuli, the saliency score was sensitive to affective processing (Supplementary Table 3). Furthermore, as demonstrated by the behavioural data (section 4.1), the pleasure-based liking system, driven by ‘unfelt’ processing fluency, was considered dominant for aesthetic processing in the MF condition. As saliency scores were directly linked to perceptual fluency, the association between the saliency score and affective processing may have been more emphasised in the MF-biological condition than in the FF-biological condition, leading the distinct correlation patterns between the two conditions. The negative correlations in the frontal and limbic regions found for saliency scores in the MF-biological condition (Figure 5E) were similar to the results for affective ratings in the FF-biological condition (Figures 5C, D). For example, the ACC, a part of the reward circuit (Haber and Knutson, 2009), is responsive to aesthetic emotions (Kawabata and Zeki, 2004; Vartanian and Goel, 2004; Brown et al., 2011; Tsukiura and Cabeza, 2011; Boccia et al., 2016), and is often coactivated with the OFC. Taken together, we can infer that the saliency scores likely became more strongly associated with affective processing during aesthetic labelling of non-consciously presented images, which resulted in negative clusters in the frontal and limbic regions and modifications of behavioural responses.

Regarding the affective ratings (liking and beauty) in the MF-biological condition, the results showed two types of correlations: (i) negative correlations in low-frequency activities in the cACC and precentral and right lingual gyri, and (ii) positive correlations in low-frequency activities in the temporal region, left lingual gyrus, and frontal pole (Figures 5G, H). For the negative correlations, the cluster in ACC may capture the affective processing of the covert stimuli, akin to the overt stimuli (Figure 5C). Additionally, the negative cluster in the precentral gyrus may also be related to the affective response as the presentation of non-beautiful, ‘ugly,’ or aversive stimuli is associated with neural responses in the precentral and postcentral gyri (Kawabata and Zeki, 2004; Ishizu and Zeki, 2011; Hayes and Northoff, 2012). Traditionally, it is assumed that the non-conscious processing of emotional information follows a rapid subcortical pathway, bypassing the slower cortical route to directly relay information to the amygdala and ‘emotion system’ (LeDoux, 1996; Tamietto and De Gelder, 2010). Therefore, it would be reasonable that the non-consciously presented biological stimuli with lower beauty scores induced fast and low-frequency activities in the precentral gyrus during the very early time window, approximately 500 ms after stimulus onset. For the positive correlations, clusters were primarily found in the occipitotemporal regions at the theta band. Although the significant clusters were temporally limited, the TF maps (Figures 5G, H) revealed sustained theta band activities throughout the stimulus presentation period. This requires cautious interpretation, given that the pattern indicates frequency-tagged responses to the Mondrian images used in the CFS paradigm. While the exact reason for the correlations between affective ratings and frequency-tagged responses remains unknown, we can infer that the properties of the masked stimuli can influence the masking effect of the CFS, including the paradigm known as ‘breaking CFS’ (Jiang et al., 2007; Stein et al., 2011). Thus, the affective properties of the biological stimuli, which are associated with low-frequency brain activities (Figures 5C, D), may modulate the frequency-tagged responses induced by CFS in the similar frequency band (theta).

The results for non-biological stimuli were different from those for biological stimuli. In the FF (conscious)-non-biological condition, the object saliency score was negatively correlated with the low-frequency activities in the mOFC (Figures 6A, B), suggesting associations between perceptual/cognitive saliency and affective process. As discussed in section 4.1 fluency, as well as the interest-based liking system and PP framework, contribute to driving the aesthetic emotion aroused by non-biological stimuli. Our results revealed traces of these processes in non-biological stimuli. For example, the bankssts (STS), where the negative correlation was found with the facial saliency score (Figure 6B), is involved in detecting social information in various stimuli, such as biological motion and eye gaze, and relaying the information to other regions for further processing (Iidaka, 2014; Deen et al., 2015). Similarly, another study revealed that the STS innervates the amygdala to deliver emotional information about facial stimuli (Pitcher et al., 2017). This indicates that the bankssts were involved in the interest-based liking system and PP, contributing to resolving disfluency or uncertainty in stimuli by detecting social and emotional cues in the abstract, non-biological images and relaying the information to the other regions to modify emotional responses. Additionally, there is substantial evidence that awareness of ambiguous perceptual images modulates the activities in V1 (Tong, 2003; Zeki, 2004). The positive correlations between saliency scores and low-frequency activity in the pericalcarine cortex (V1) (Figures 6A, B) may reflect the relationship between perceptual saliency and awareness (i.e., resolving disfluent and uncertain states). Similarly, the superior parietal cortex, where the facial saliency score was positively correlated to the theta/alpha-band activities, has been associated with the disambiguation process of bistable images (Kanai et al., 2011). Furthermore, the inferior temporal cortex is famously associated with object recognition (Logothetis and Sheinberg, 1996; Conway, 2018); thus, the negative cluster identified in this region for facial saliency score would indicate the contribution of the recognition process to disambiguate the visual information to drive affective processing. Although the neurological mechanisms underlying the interest-based liking system and PP framework in the context of aesthetic emotion remain unknown, these regions may play a pivotal role in the processes. Regarding the affective scores, the liking score correlated negatively with the low-frequency activity in the rACC (Figure 6C), consistent with biological stimuli (Figure 5C), supporting the link between ACC activity and aesthetic pleasure. The beauty score correlated with theta-band activity in the occipital regions, such as the lateral occipital cortex and cuneus (Figure 6D), reflecting that the level of aesthetic emotion was associated with the visual information processing, such as the fluency, elaboration, and prediction. Notably, the role of cortical areas in such visual information processing cannot be specified because we did not measure detailed behavioural information for speculating the dominant mental processes underlying aesthetic emotion, which may have varied stimulus-wise (e.g., processing fluency, interest-based liking, or PP). Therefore, our interpretations of the MEG data do not exclude other possibilities; for example, the involvement of the inferior temporal cortex may not reflect its contribution to the disambiguation process (related top-down interest-based liking system and PP framework), but rather the level of prototypicality of the stimuli (related to bottom-up fluency processes).

In the MF (non-conscious)-non-biological condition, facial saliency score was negatively correlated with low-frequency activity in the frontal and limbic regions, including the cACC, precentral, and rostral middle frontal gyri (Figure 6F), similar to negative clusters observed in the frontal and limbic regions for biological stimuli (MF-biological condition) (Figure 5E). As discussed in section 4.3 the negative correlations likely indicate the increased link between saliency scores and affective processing during the aesthetic labelling of non-consciously presented images. This suggests that, for non-consciously presented stimuli, regardless of their categories, the pleasure-based liking system driven by ‘unfelt’ processing fluency was likely dominant for aesthetic processing, and the other ‘top-down’ frameworks were largely suppressed. The liking and beauty scores in the MF-non-biological condition demonstrated negative low-frequency clusters, mainly in the fronto-temporal and central regions (Figures 6G, H). Negative clusters in the central region were also found for the biological category under the same condition (Figure 5H), which would be indicative of aversive responses to the ‘non-beautiful’ stimuli. The neural processing for the non-consciously presented aversive stimuli may share similar neural systems across categories. Negative clusters were also found in the caudal middle frontal gyrus and insular cortex (Figure 6H). The dorsolateral prefrontal cortex, including the middle frontal gyrus, has been shown to be responsive to aesthetic emotion (Cela-Conde et al., 2004; Cattaneo et al., 2014). Similarly, the insula is considered to be a region responsive to emotional (Singer et al., 2009; Gasquoine, 2014; Uddin et al., 2017) and aesthetically appealing stimuli (Di Dio et al., 2007; Di Dio and Vittorio, 2009; Brown et al., 2011; Chatterjee and Vartanian, 2014; Boccia et al., 2016). These findings indicate that the non-conscious presentation of non-biological stimuli induced neural responses related to aesthetic emotion, observed as negative correlations in the low-frequency bands.

This study has four main limitations. First, the CFS method for non-conscious presentation imposed technical constraints; the stimuli faded in and out to enhance the suppression effects, allowing us to study only the induced components (oscillations) of the MEG signals, which were contaminated by the signals induced by the Mondrian mask at 6.25 Hz. Future studies investigating this topic should consider using time-locked evoked activities. Second, owing to environmental constraints, the number of stimuli (trials) and pre-rating scales were limited. The MEG system was installed in a hospital setting and regularly used for clinical examinations; thus, the experimental procedures were designed to be completed within a limited time slot, restricting the number of trials in the CFS task, variations in stimuli, and pre-ratings. Although the stimuli were carefully selected (section 2.2), they did not cover before the Renaissance periods or modern art styles. In addition, owing to the limited number of pre-rating scales, the results may be limited to specific positively valenced aesthetic emotions (liking and beauty), preventing generalisability to other negatively valenced aesthetic emotions (Istók et al., 2009; Augustin et al., 2012; Knoop et al., 2016). Third, the ratings and neural responses were obtained non-simultaneously and separated by a gap of up to 5 days between the pre-rating and MEG experiment, because of the aforementioned environmental factors and to reduce the physical burden among participants (section 2.3). Repeated exposure is known to alter the hedonic experience to the stimuli (Montoya et al., 2017). Moreover, prior exposure to the stimuli (i.e., learning) might reduce the experiences of unconformity and uncertainty to the unexpected (i.e., non-biological) stimuli in a PP sense, which may only trigger a ‘faint copy’ of uncertainty resolution (Van De Cruys et al., 2024). As such, the actual in-the-moment aesthetic experience may not be captured, but the stimuli would be labelled as aesthetically appealing owing to their past affordance of the aesthetic experience. Therefore, the aesthetic processes operated during the pre-rating and MEG measurements might not be quantitatively and qualitatively identical. Thus, our findings should be replicated in a future study using a simultaneous recording design. Fourth, the definitions of biological and non-biological stimuli were arbitrary. We operationally defined the biological and non-biological categories as extremes of the single continuous scale of ‘biologi-ness’ (section 1), but this idea needs further experimental support. Additionally, the stimuli sets were selected based on the relative scores within each participant (the top and bottom 20 images in the ‘biologi-ness’ rating). Another categorisation approach, such as the use of the absolute ‘biologi-ness’ score, may show different results; however, this is beyond the scope of the present study and should be investigated further in future psychological and neuroscientific studies.