Information and affective valence influence judgments of complexity, liking and understanding

Abstract

Visual complexity is a key factor in perceptual and evaluative judgments. People's representation of visual complexity is constructed from quantitative and structural image features, but it is also influenced by familiarity and expertise. We examined how people represent visual complexity and its impact on perception and evaluation, focusing on information about paintings and their affective valence on judgments of visual complexity, liking and understanding. Seventy‐six participants rated 60 representational artworks of negative, neutral, and positive valence on complexity, beauty and understanding. Half of the participants received written information about each artwork. Results showed that negative artworks were judged as more complex than neutral artworks and positive ones, but this effect was attenuated by the provided information. Liking judgments increased with judged complexity, were higher for positive artworks than neutral ones, for neutral than negative ones, and were higher when information was provided. Understanding judgments were higher for positive artworks than neutral ones, and higher for neutral artworks than negatively valenced ones. Information increased understanding only for negative artworks, and judged complexity did not affect these judgments. In sum, the representation of the visual complexity of an image is influenced by its valence and the available information, modulating judgments of complexity and liking, but not of understanding.

BACKGROUND

Evaluative judgments are assessments people make about the value, quality or significance of something. They often involve determining whether an action, situation, object, person or idea is appropriate or inappropriate, good or bad, pleasing or displeasing, beautiful or ugly, liked or disliked, and so forth. Judgments of beauty, liking or pleasingness are often referred to as aesthetic judgments because they involve the appraisal of the cognitive and affective impact of certain sensory features, such as colour, symmetry, balance, order or complexity (Berlyne, 1972; Chatterjee, 2003; Chatterjee & Vartanian, 2014; Leder et al., 2004; Russell, 1994; Skov & Nadal, 2021). While these judgments are often explicit and consciously made, much of the aesthetic experience may occur unconsciously. Viewers may respond to artworks intuitively, where emotional and cognitive processes shape their perceptions without the need for deliberate reflection on factors, such as complexity. Explicit evaluative judgments are often used in psychological studies, as they provide a measurable means to access these internal states (Nadal & Vartanian, 2022), but the role of unconscious evaluations should not be overlooked, as both conscious and unconscious processes contribute to aesthetic judgments.

There is a long research tradition showing that complexity is one of the main factors correlated with evaluative judgments (Marin, 2022; Skov & Nadal, 2021). Today, research on the impact of complexity on judgments of beauty, liking and pleasingness continues to thrive. Recent work has explored, among other issues, the dimensions underlying complexity and their measurement (Van Geert & Wagemans, 2020), the processing of visual complexity in the brain (Zhou et al., 2023), the representation of complexity across different sensory modalities (Clemente et al., 2021), the implementation of automated computational metrics of visual complexity (Carballal et al., 2020) and the application of this knowledge to architecture (Lee & Ostwald, 2023), educational materials (Stoesz et al., 2020), usability of websites (Lazard & King, 2020; Michailidou et al., 2021), and generative and evolutionary art (McCormack & Cruz Gambardella, 2022). The present study aims to contribute to this research by increasing our understanding of the kinds of information people base their representation of visual complexity on and of how this representation influences different kinds of judgments.

The idea that complexity influences evaluative judgments can be traced back to classical philosophical aesthetics, which posited that beauty arises from the proportion and arrangement of an object's parts—such as their size, uniformity, number and interrelationships (Tatarkiewicz, 1972). This general aesthetic principle found empirical expression in Fechner's (1876) scientific aesthetics, particularly in his principle of coherent connection of variety. According to Fechner, people require a certain degree of variation in their activities or sensory experiences to feel comfortable and engaged. An object must offer diverse points of engagement to sustain attention, yet those elements must also be coherently connected in time and space. Without sufficient variety, an object may be perceived as monotonous, boring and dull; without coherence, it may seem fragmented, dispersed or incoherent. Thus, according to Fechner, pleasurable engagement with an object depends on the presence of a coherently connected variety of elements.

Birkhoff (1933) was among the first to translate this theory into a mathematical formulation. In his formula, features contributing to order (O) enhance the aesthetic measure (M), while elements contributing to complexity (C) diminish it. Therefore, the aesthetic measure is best estimated by the ratio of order to complexity, expressed as M = O/C. Subsequent tests of his formula yielded mixed results (Beebe‐Center & Pratt, 1927; Davis, 1936; Eysenck, 1942; Eysenck & Castle, 1970), revealing that it was a poor predictor of evaluative judgments. This led Eysenck (1941) to develop a new empirical aesthetic formula, where both order and complexity positively influence aesthetic preference: M = O × C (Eysenck, 1942).

Complexity also played a central role in Daniel Berlyne's motivational theory. Berlyne (1971) posited that the hedonic tone, or the pleasure and preference elicited by a stimulus, depended on the level of arousal it induced relative to the organism's current arousal level. Collative properties, Berlyne suggested, influence arousal by how they compare with previously experienced objects. In the context of aesthetics, Berlyne suggested that interest and preference for an image are largely influenced by its perceived complexity (Berlyne, 1963; Berlyne et al., 1968). Factors contributing to perceived complexity include pattern regularity, the number of elements, their heterogeneity and form irregularity (Berlyne, 1970). Under typical conditions, with an intermediate arousal level, people generally prefer stimuli of moderate complexity over those that are highly complex or very simple. However, preferences may shift towards more complex or simpler objects depending on the individual's current arousal level.

Berlyne's (1971) conception of complexity as a collative variable departed from most previous work, which had conceived visual complexity as a property of objects, defined in terms of physical and spatial characteristics. As a collative variable, complexity involves contrast and comparison, meaning that responses to an object are influenced by the degree and kind of similarity or difference with its background, with objects experienced previously, or with anticipated objects (Berlyne, 1971). Berlyne stated:

‘Of course, motivational, including aesthetic, phenomena must depend directly, not on the usual objective information‐theoretic measures, but rather on their subjective equivalents (…). Objective uncertainty, amount of information and so on, are determined by the objective sample space (i.e., the alternative events that could actually occur) and by objective probabilities (the relative frequency with which events belonging to the sample actually occur). But preferences and emotional reactions will surely be governed by the membership of a subjective sample space (i.e., the events that the subject represents to himself as possible alternatives) and by subjective probabilities (i.e., how likely the subject considers each of the alternative possibilities to be)’ (Berlyne, 1971, p. 46).

In sum, collative variables are inherently subjective, as they are influenced by the relationship between the physical and statistical characteristics of stimuli and the processes within an organism. This means that a pattern may appear more novel, complex or ambiguous to one person compared with another, or may vary for the same person at different times.

Despite this subjectivity, collative properties and subjective informational variables tend to change in tandem with objective measures from classical information theory (Attneave, 1957; Attneave & Arnoult, 1956; Berlyne, 1971, 1974; Hall, 1969). These observations led to two main lines of research: one that aims to determine the statistical image properties that approximate human perceptual (e.g. complexity) and evaluative (e.g. beauty) judgments, and another that aims to determine the kinds of object features that people use to construct their perceptual representations of complexity and how these impact evaluative judgments (Nath et al., 2024; Van Geert & Wagemans, 2020).

Along the first of these research lines, several different computational and statistical methods have been used to approximate human visual complexity judgments (Van Geert & Wagemans, 2020). Quantification of different image components has shown that complexity ratings increase with the number of lines and distinct figures (García et al., 1994; McDougall et al., 1999). Edge detection algorithms have shown that complexity ratings increase with the number of edges in an image (Forsythe et al., 2003, 2011). The pyramid of histograms of orientation gradients (PHOG) methods have shown that complexity ratings increase with HOG complexity and fractal dimension (Van Geert & Wagemans, 2021). Algorithmic information theory‐based image compression techniques have shown that complexity ratings increase with compressed file lengths (Donderi & McFadden, 2005; Forsythe et al., 2011; Güçlütürk et al., 2016; Jakesch & Leder, 2015; Marin & Leder, 2013; Palumbo et al., 2014). Information from several of these algorithms can be combined through machine learning and neural networks to produce better estimations of complexity ratings (Gartus & Leder, 2017; Machado et al., 2015), but this requires a large number of stimuli in the learning set. Some of these measures have also been used to approximate human evaluative judgments of beauty (Hayn‐Leichsenring et al., 2017; Mallon et al., 2014), appeal (Bies et al., 2016), pleasantness (Marin & Leder, 2013) and liking (Güçlütürk et al., 2016; Iigaya et al., 2021).

The second line of research has examined the kind of image features people take into account to construct their representation of visual complexity. Berlyne (1958, 1971) showed that people's judgments of visual complexity were influenced by different aspects of an image's elements and their arrangement: the number, heterogeneity, incongruence and irregularity of the elements, as well as the irregularity of their arrangement and the degree to which they are perceived as a unit. Berlyne et al. (1968) confirmed these two factors that contribute to judged complexity: the number of elements, which explained 67.4% of complexity ratings, and disorganization, which accounted for 23.6%. Chipman's (1977) study on complexity judgments of simple matrix patterns identified two components influencing ratings of complexity: one quantitative, the number of corners and the other structural, including symmetry, rotation and repetition. Chipman (1977) suggested that the quantitative component sets an upper bound for visual complexity, whereas the structural component can reduce perceived complexity. Ichikawa (1985) found that with short presentations, complexity ratings reflected mainly quantitative aspects. With longer presentations, quantitative and structural aspects influenced the representation of complexity. Later work using different materials and methods has confirmed that judgments of visual complexity take into account quantitative and structural aspects of images (Gartus & Leder, 2017; Jakesch & Leder, 2015; Nadal et al., 2010). Moreover, several studies have found that both influence evaluative judgments of beauty (Nadal et al., 2010; Nath et al., 2024) and preference (Van Geert & Wagemans, 2021): Beauty and preference judgments increase with quantitative complexity and decrease with structural complexity (Nadal et al., 2010; Nath et al., 2024; Van Geert & Wagemans, 2021). Thus, both aspects influence evaluative judgments, though in different directions. But it is the representation of complexity, rather than the depicted complexity, that influences evaluative judgments (Nath et al., 2024).

There is some evidence that people's representation of visual complexity is influenced by aspects that are not strictly properties of the image. Among these aspects, research has focused mostly on familiarity. Snodgrass and Vanderwart (1980) collected ratings of familiarity (‘according to how usual or unusual the object is in your realm of experience’) and complexity (‘the amount of detail or intricacy of line in the picture’) for 260 images. They found a significant correlation of r _s  = −.466 (p < .01) between the complexity and familiarity ratings, indicating that images that looked less familiar looked more complex and images that looked more familiar looked less complex. An analysis of similar data collected on different sorts of images (Rossion & Pourtois, 2004) showed significant correlations between familiarity and complexity, ranging from r = −.41 to r = −.50 (all p < .01) (Forsythe et al., 2008). Furthermore, the experimental manipulation of familiarity has shown that participants who have acquainted themselves with a set of images judge them as less complex than participants who had no familiarity with them (Forsythe et al., 2008). Tinio and Leder (2009) also found that massive familiarization changed the effects of complexity, in that contrasting levels of complexity were liked better. For example, after massive familiarization with highly complex patterns, simple ones gained in aesthetic appeal. Marin and Leder (2013) found that in both environmental scenes and representational paintings, familiar images were consistently rated as less complex, and this effect became more pronounced with longer viewing durations. Marin et al. (2016) extended these findings by showing that familiarity influences complexity ratings differently across visual stimuli, such as environmental scenes, cartoons and paintings. In paintings, the negative correlation between familiarity and complexity was particularly strong, suggesting that familiarity plays a significant role in shaping how complex an image appears. Another 2016 study by Marin and Leder revealed that familiarity also interacted with emotional responses, where familiar images led to lower arousal and lower perceived complexity.

Other research has shown that the effects of knowledge on representations of complexity are not limited to familiarity. Li and Choi (2023) demonstrated that prior knowledge about the image can, in some cases, modulate perceived complexity, particularly for less complex images. Expert art historians rate the complexity of artworks lower than non‐experts, showing that domain‐specific knowledge and skills acquired with years of training also influence complexity judgments (Commare et al., 2018). Additionally, they highlighted that the semantic content of images, beyond their structural features, plays a significant role in how complexity is perceived. Semantic complexity refers to the narrative or conceptual elements within the artwork, which can heighten perceived complexity even if the structural details remain simple. This further supports the idea that complexity judgments are multifaceted and influenced by both structural and semantic aspects of the artwork.

Putting all the research reviewed above together allows us to synthesize the process of constructing the representation of an image's complexity and its role in perceptual and evaluative judgments. Both computer and human perceptual systems extract several relevant features from an image. Computers can use such features, alone or in combination, to quantify the complexity contained in the image. Human perceptual systems transform these features into representations of objects and their interrelations, which contribute to the quantitative and structural aspects of the representation of complexity. However, this representation of visual complexity is susceptible to modulation by other factors, such as familiarity with the image, domain‐specific expertise, and viewing time. The representation of complexity can be translated into judgments of complexity and has a substantial influence on evaluative judgments.

To advance understanding of how visual complexity is represented and its effects in judgment tasks, we set out to answer the following research questions:

1. How does the information participants have about an image modulate their representation of its complexity? Extensive research has shown that various types of information can significantly impact aesthetic preferences and evaluations of artworks. For instance, Temme (1992) found that 79% of participants' appreciation improved after reading descriptive labels of paintings at exhibitions, particularly when assessing ‘aesthetically ambiguous’ artworks. Additional information appeared to make these artworks more favourable. Cupchik et al. (1994) conducted a comprehensive study on the effects of three types of information on artwork appreciation: physical descriptions, formalistic information related to emotional arousal and contextual information introducing socio‐cultural aspects. Their findings showed that descriptive information decreased both affective and cognitive ratings, while contextual information increased both. Formalistic information, which emphasizes visual attributes, raised cognitive ratings but lowered emotional ones. Millis (2001) further supported this by showing that both descriptive and elaborative titles improved viewers' comprehension of illustrations and photographs, with elaborative titles being more effective in enhancing the hedonic value of stimuli. Similarly, Russell (2003) found that providing descriptive information increased judgments about the meaningfulness and pleasingness of abstract and semi‐abstract artworks. Leder et al. (2006) extended these insights by exploring how different types of titles influence the appreciation and understanding of artworks, discovering that descriptive titles affect early stages of aesthetic processing, while elaborative titles play a larger role in comprehension and interpretation. These findings underscore the critical role of information in shaping aesthetic experiences, motivating us to explore how information affects the perception of complexity in aesthetic judgments.

2. How does an image's affective valence modulate participants' representation of its complexity? Studies show that affective valence significantly influences the perception of visual complexity. Ochsner (2000) identified a relationship between the affective valence of images and their perceived complexity, demonstrating that emotional responses to stimuli from the International Affective Picture System (IAPS) were linked to the complexity of the visual content. In line with this, Silvia (2005) emphasized that emotional experiences are not merely the product of objective stimulus features but are moderated by subjective experiences, meaning affective valence impacts how complex a stimulus feels. Further research, such as Marin and Leder (2013), supports this finding, showing that environmental scenes and representational paintings were rated as more or less complex depending on the affective valence they evoked, with positive stimuli often perceived as less complex. Marin et al. (2016) expanded on these findings by showing that affective valence and emotional arousal significantly modulate perceived complexity across different types of visual stimuli, including paintings and cartoons, with negative or more arousing stimuli often rated as more complex. These findings reinforce the idea that complexity judgments cannot be fully separated from the affective context in which the stimuli are presented. Thus, affective valence seems to impact the perception of complexity.

3. How does the representation of an image's complexity influence participants' judgments of their liking and understanding of it? The paragraphs above have shown how complexity affects aesthetic judgments, but research suggests that complexity can also affect understanding, as more complex stimuli often require greater cognitive resources to be understood (Leder et al., 2004). Theories of cognitive fluency indicate that simpler stimuli are easier to process and thus often lead to higher judgments of understanding (Reber et al., 2004). However, in more complex artworks, participants may feel less certain about their understanding due to the higher cognitive demands.

In the present study, we specifically examined whether the information participants received and the affective valence of the images modulate the representation of complexity, and whether this representation in turn influences judgments of complexity, liking and understanding. To this effect, and to answer the questions above, we designed an experiment in which participants judged negative, neutral and positive artistic paintings on complexity, liking and understanding scales, with half of the participants receiving information about the artworks. We expected to find that participants who received information about the artworks would judge them as less complex, more likable and more understandable than participants who did not receive the information, and that negatively valenced artworks would be judged as more complex, less likable and less understandable than positively valenced artworks.

METHOD

Participants

The study comprised 76 participants aged between 18 and 55 (M = 26.9, SD = 8.7), including 40 women (M _age = 27.3, SD = 9.1) and 36 men (M _age = 26.4, SD = 8.1). To determine the required sample size, we conducted an a priori power analysis using GPower (version 3.1.9.7). The study employed a mixed design with one between‐subject factor (information: with/without) and one within‐subject factor (valence: negative, neutral and positive). Based on previous findings suggesting a moderate effect of contextual and semantic information on complexity perception (Forsythe et al., 2008; Marin & Leder, 2013), we set the expected effect size to f = 0.25. With α = .05, desired power = 0.75, and assuming a correlation of .5 among repeated measures, the analysis indicated that a total of 76 participants would be sufficient. Accordingly, our final sample consisted of 76 individuals. Participants for the study were enlisted via the volunteer participant management system, an online recruitment platform. The cohort predominantly comprised students from the Psychology Department at the University of Vienna. Participation in the study was voluntary, with individuals choosing to engage without receiving any monetary compensation, though they earned academic credits. None of the participants had specific training or education in art. The study adhered to the ethical guidelines outlined in the World Medical Association Declaration of Helsinki. All participants possessed normal colour and image perception capabilities; those with visual impairments utilized corrective eyewear, such as glasses or contact lenses, to facilitate their participation.

Stimuli

The stimulus set consisted of 60 representational artworks, each categorized as negative, neutral or positive in valence. These categories were established based on a preliminary pre‐test involving a sample of 32 participants (26 women; M _age = 28.1, SD = 4.7) taken from the same population, who rated 240 reproductions of representational artworks spanning from the 17th to the 21st century on 7‐point Likert scales of valence and arousal. All participants were tested together in a classroom with each image projected on a screen for 8 s while participants used a pen to mark the corresponding valence and arousal ratings on an individual paper response sheet. Each image used in the study was approximately 1000 × 768 pixels and was originally stored in JPEG format. The size of the images was kept consistent across all stimuli to ensure uniformity in terms of visual presentation. The images were selected from diverse art history books and museum exhibition catalogues aiming to represent diverse art traditions, including classical realism, impressionism, cubism and expressionism, and to potentially cover a broad range of visual complexity by including images with fewer elements and less intricacy to images with more elements and intricacy. Based on valence ratings, the images were divided into three groups: 80 negatively valenced, 80 neutral and 80 positively valenced images. The most negative images depicted themes, such as nature in decay, sorrow or death, while the most positive images showed nature in prosperity and joyful individuals. We excluded the 30 most and the 30 least arousing images in each of the three valence categories. This resulted in a final set of 20 artworks of intermediate arousal in each of the three valence categories. Additionally, four neutral‐valence artworks were chosen for practice trials. Title, author and year information for these stimuli can be found at https://osf.io/zphjg/.

Quantification of depicted complexity

To quantify the depicted complexity, we used one of the data compression methods described by Marin and Leder (2013) and Forsythe et al. (2011): the ratio between the original file size and the compressed GIF file size for each image. This information theory‐based method can be thought of as an approximation to Kolmogorov complexity, understood as the length of the shortest algorithm in any programming language that computes a particular binary string (Aksentijevic & Gibson, 2012; Donderi, 2006b; Nath et al., 2024). From this perspective, simple images contain more redundant information that can be represented by shorter strings of bits, so image compression algorithms produce smaller file sizes than more complex images. The compressibility of an image, thus, reflects its complexity, with less compressibility implying higher complexity. The ratio between original and compressed file sizes correlates highly with complexity ratings of marine charts and radar images (Donderi, 2006a, 2006b), simple icons (Forsythe et al., 2003), abstract patterns (Palumbo et al., 2014), designs, photographs of natural scenes and artworks (Forsythe et al., 2011; Marin & Leder, 2013). We, therefore, converted the set of images, initially stored in JPEG format, into lossless GIF using Adobe Photoshop with the same settings as Marin and Leder (2013): palette local selective, colours 256, forced black‐white colours, no transparency, 75% dither diffusion, exact colours and normal order of lines. We thereafter calculated the ratio between the compressed file size and the original file size, and the result was considered as each image's depicted complexity.

Statistical checks

Given our research question 2 about the impact of affective valence on the representation of complexity, we had aimed to produce three sets of stimuli that were clearly distinct in affective valence, while avoiding extremes of arousal in each valence level. The reason for avoiding extreme values of arousal is that arousal is positively associated with perceived complexity (Madan et al., 2018). Arousal itself may stem from different sources, including not only the emotional significance of the image but also its low‐level visual features, such as colour and shape (Van Paasschen et al., 2014). Also, according to Berlyne's (1971) arousal theory, greater complexity elicits higher arousal, which in turn shapes aesthetic judgments. The following tables present the descriptive statistics of the 20 stimuli selected for each level of the affective valence variable, including valence and arousal ratings (Table 1).

TABLE 1.

Descriptive statistics for valence and arousal ratings across affective valence levels.

	Affective valence level	Min	Max	Mean	SD
Valence	Negative	1.84	2.97	2.40	0.32
	Neutral	3.5	4.5	4.06	0.36
	Positive	4.97	5.59	5.15	0.17
Arousal	Negative	3.25	4.53	3.93	0.33
	Neutral	2.84	4.63	3.51	0.45
	Positive	3.03	4.09	3.48	0.32

A Kruskal–Wallis test revealed significant differences in valence among the levels of the affective valence variable ($\chi$ ² = 52.5, p < .0001). A follow‐up Wilcoxon rank‐sum test, Benjamini–Hochberg adjusted for multiple comparisons, revealed that the valence ratings of images in the negative level were significantly lower than those of images in the neutral level, (W = 0, p < .0001), and that the valence ratings of images in the positive level were significantly higher than those of images in the neutral level, (W = 0, p < .0001), indicating that we had successfully created three subsets of images that differed significantly in affective valence. A Kruskal–Wallis test revealed significant differences in arousal among the levels of the affective valence variable ($\chi$ ² = 15.52, p < .001). A follow‐up Wilcoxon rank‐sum test, Benjamini–Hochberg adjusted for multiple comparisons, revealed that the arousal ratings of images in the negative affect level were significantly higher than those of images in the neutral affect level (W = 316, p = .002) and than those of images in the positive affect level (W = 334, p < .001), but that arousal ratings for images in the neutral and positive levels were not statistically different (W = 194, p = .892).

We also thought it prudent to check depicted complexity values across levels of the affective variable. Table 2 shows these values:

TABLE 2.

Descriptive statistics for depicted visual complexity across valence levels.

	Affective valence level	Min	Max	Mean	SD
Depicted complexity	Negative	0.14	0.34	0.22	0.05
	Neutral	0.16	0.43	0.24	0.07
	Positive	0.19	0.41	0.29	0.06

A Kruskal–Wallis test revealed significant differences in depicted among the levels of the affective valence variable ($\chi$ ² = 12.91, p < .01). A follow‐up Wilcoxon rank‐sum test, Benjamini–Hochberg adjusted for multiple comparisons, revealed that the depicted complexity of images in the positive level were significantly higher than those of images in the neutral level (W = 98, p < .01) and of images in the negative level (W = 75, p < .01). There was no statistical difference in the depicted complexity levels of images in the neutral and negative levels of the variable affective valence (W = 184, p > .05). These differences in depicted complexity will be taken into account when interpreting the results.

Information texts

In the main experiment, half of the participants viewed each artwork together with a brief text (371–571 characters, M = 449.53, SD = 38.25) containing information about its title, author, year of creation, historical context, iconography and composition, compiled from the corresponding art history book or museum exhibition catalogue. Here are two examples of the information texts. The first is information about a positively valenced image and the second is about a negatively valenced one:

John William Waterhouse

Gossip (1885)

The Victorian British painter John William Waterhouse (1849–1917) was well known for his careful and expressive treatment of face and hands, as well as costume and foliage. In this painting, he portrays a group of women, one of his most popular themes, during a break in their domestic activities, providing a quiet, languorous, and relaxed mood. The communicative intent among the characters is achieved through the intense gazes among the women.

George Bellows

Return of the Useless (1918)

Here, George Bellows (1882–1925) depicted German soldiers unloading sick and disabled Belgian citizens who were being returned home, as they were no longer fit to work for the Germans. He expressively depicted the anguish of the prisoners and cleverly used the same colour for the rusty boxcar as he used for the flushed face of the German soldier who is venting his anger on the fallen and cowering man and the bloodied skin of some of the prisoners.

Procedure

Upon arrival at the laboratory, participants were welcomed and provided with a general overview of the tasks they would be performing. They were then asked to fill out the German version of the ‘Positive and Negative Affect Schedule’ (PANAS; Krohne et al., 1996; Watson et al., 1988), designed to assess mood through self‐reported positive and negative affect. Instructions were given to the participants to answer the questionnaire based on their current feelings, with an emphasis on the necessity for honesty and precision in their answers.

Participants were randomly divided into two groups for the study. The first group was shown images without additional context and tasked with rating each on a scale from 1 (least) to 7 (most) based on their liking (‘How much do you like this image?’), perceived complexity (‘How complex is this image?’) and understanding (‘How much do you understand this image?’). Participants were told that complexity referred to the number and variety of elements and the degree of detail and intricacy, liking referred to the degree to which they favoured the images and understanding referred to their ability to grasp the artistic objective or intention of the painter. The second group viewed the same images, each paired with information in German about the artwork: title, painter and a brief descriptive text about its style, content and pictorial resources (i.e. directions of gazes or use of certain colours). Participants in this group rated the images using the same scales and with the same instructions as the first group. For both groups, the rating was self‐paced. After viewing each image in full screen and pressing the ‘Enter’ key, the same image was displayed again in a smaller size alongside the rating scales, and they first rated the liking scale, then the complexity scale and then the understanding scale. Each scale was presented separately, and participants were required to provide a response before proceeding to the next. After rating the three scales, there was a 1500 ms pause before displaying the next image. The order of stimulus presentation was randomized individually for each participant.

Given that it has been shown that domain‐specific expertise influences complexity ratings (Commare et al., 2018), the Kunstwissen und ‐interesse Fragebogen (KIF), a German questionnaire designed to gauge interest and knowledge in the visual arts domain, was distributed. This questionnaire aimed to quantify participants' interest in and knowledge of visual art, encompassing aspects, such as frequency of art exposure, familiarity with art terminology and personal interest in art activities. This measure was included to identify the potential influence of art knowledge and interest on the image evaluation task.

Data analyses

The raw data can be found at https://osf.io/zphjg/. We used generalized linear mixed‐effects models (GLMMs) to analyse the effects of provided information and image valence on participants' complexity, liking and understanding ratings (Hox, 2010; Snijders & Bosker, 2012). These models efficiently handle both between‐subjects and within‐subjects effects of the independent variables (Baayen et al., 2008), providing a significant advantage over ANOVAs. Unlike ANOVAs, which typically require averaging across stimuli, linear mixed effects models do not inflate the empirical Type I error rate, thereby avoiding the replication issues often seen with significant effects in different samples (Judd et al., 2012, 2017). These models are particularly adept at analysing hierarchically structured data where responses are dependent on or nested within participants (Nezlek, 2001). These models are known for their robust handling of outliers, flexibility and ease of use (Judd et al., 2012). An important feature of these models is their ability to produce generalizable conclusions beyond the sample providing the data (Judd et al., 2017; Nezlek, 2001). This makes them ideal for analysing preference responses that often vary significantly between individuals and objects (Silvia, 2007).

In this study, the models were structured to capture the effect of the main predictors on participants' responses, adhering to Barr et al.'s (2013) recommendation to include the maximal random effects structure justified by the design. This approach minimizes power loss, reduces Type I error and enhances the generalizability of results to other participants and stimuli. Analyses were conducted in the R statistical environment, version 4.3.3 (R Core Team, 2024), using the ‘lmer()’ function from the ‘lme4’ package, version 1.1‐35.1 (Bates et al., 2015), with REML estimation. The ‘lmerTest’ package, version 3.0‐3 (Kuznetsova et al., 2017), provided p‐values for t‐tests using the Satterthwaite approximation for degrees of freedom, known to maintain acceptable Type I error rates (Luke, 2017). Cook's distance was calculated using the base R function ‘cooks.distance()’ to assess influential data points.

The model predicting complexity judgments included interactions among depicted complexity, information (with, without) and valence (negative, neutral and positive) as fixed effects, with intercepts and slopes of depicted complexity, valence and their interaction treated as random effects within participants, and the intercept and slope of information treated as a random effect within stimulus. The models also included participants' scores for art knowledge, art interest, PANAS positive and PANAS negative. The models predicting liking and understanding judgments were identical, except that we modelled the effect of judged complexity instead of depicted complexity, in line with previous research (Nath et al., 2024), Berlyne's (1971) conception of complexity as a collative variable, and the framework of visual complexity representation and its impact on evaluative and understanding judgments outlined above. In all cases, the categorical predictors were treatment coded, and the continuous predictors were scaled to a mean of 0 and a standard deviation of 1. Reference levels for the categorical variables were set as ‘without information’ and ‘neutral valence’. When necessary, follow‐up paired comparisons were conducted using the ‘test()’ and ‘emtrends()’ of the R package ‘emmeans’ (Lenth, 2024), applying the Holm method to control for the family‐wise error rate and obtain the adjusted p‐values. To quantify explained variance in our linear mixed‐effects model, we computed both marginal and conditional R ² values using the method proposed by Nakagawa and Schielzeth (2013), extended by Johnson (2014) to accommodate models with random slopes. This approach partitions variance into components attributable to fixed effects (marginal R ²) and to both fixed and random effects (conditional R ²), while properly accounting for the additional complexity introduced by random slope terms.

RESULTS

Complexity judgments

The model was set up to predict complexity ratings from the triple interaction of depicted complexity, image valence and group, and participants' art knowledge, art interest, and positive PANAS and negative PANAS scores. We included the interaction of depicted complexity and image valence as random effects within participants and group as a random effect within the stimulus. This model, however, did not converge. Following Barr et al.'s (2013) recommendations, we simplified the random effects structure by removing the interaction from the random effects within participants and left depicted complexity and image valence as main effects within participants. The final model retained random intercepts for both participants as well as random slopes for the main effects of depicted complexity and valence within participants. Before exploring the model, we conducted an analysis of influential cases based on Cook's distance, revealing three highly influential participants. We removed these participants from the dataset and performed the same analysis.

The model excluding the highly influential participants revealed three significant effects. First, image valence influenced complexity ratings: negatively valenced images were rated as more complex (M = 0.30, 95% CI [0.10, 0.50]) than neutral ones (M = 0.04, [−0.15, 0.23]) [β = .56, t(67.34) = 2.818, p = .006], and positively valenced images were rated as less complex (M = −0.34, 95% CI [−0.56, −0.11]) than neutral ones [β = .65, t(73.96) = 3.029, p < .005]. Second, this effect of valence was moderated by a significant two‐way interaction between valence and group [β = .29, t(89.78) = 2.189, p = .031, Figure 1]. The group of participants who received no information rated negatively valenced images as more complex (M = 0.33, 95% CI [0.11, 0.56]) than neutral ones (M = −0.02, [−0.23, 0.20]) [z = 2.818, p = .015] and positively valenced images as less complex (M = −0.42, [−0.69, −0.16]) than neutral ones (z = 3.029, p = .010). In the case of participants who received information, the difference between complexity ratings for negatively valenced images (M = 0.27, 95% CI [0.33, 0.50]) and neutral ones (M = 0.10, [−0.13, 0.32]) was non‐significant (z = 1.257, p = .210), although they did rate positively valenced images (M = −0.25, [−0.52, 0.02]) as significantly less complex than neutral ones (z = 2.486, p = .026).

FIGURE 1.

Complexity ratings as a function of image valence and information condition. Raw participant ratings are shown as jittered semi‐transparent dots, overlaid on violin plots representing the full distribution of responses. Model‐estimated means from the linear mixed‐effects model are depicted as black squares, with vertical error bars indicating 95% confidence intervals, and asterisks indicate significant pairwise differences (p < .05). Facets reflect the two information conditions: ‘Without information’ and ‘With information’. Valence levels are ordered from negative to positive.

Third, the model also showed that there was a two‐way interaction between depicted complexity and group [β = .17, t(253.3) = 2.374, p = .018, Figure 2], indicating that the effect of depicted complexity on complexity ratings was stronger for participants who had received information about the images (α = .045, [−0.063, 0.154]) than for participants who had not (α = −.005, [−0.107, 0.097]). None of the other factors or interactions had significant effects on complexity judgments (all ps > .1).

FIGURE 2.

Relationship between perceived complexity and depicted complexity with and without information. Relationship between depicted complexity and perceived complexity ratings. The Y‐axis indicates the perceived complexity ratings (1–7 rating scale), while the X‐axis is the scaled depicted complexity (scaled to have a mean of 0 and a standard deviation of 1). The shaded area represents the 95% confidence interval. The left panel shows the condition without information and the right panel shows the condition with information.

Regarding random effects, the variance for the participants' intercept was 0.482 and the variance for the stimulus intercept was 0.273. The variance for the slope of depicted complexity was 0.019, the variance for the slope of the negative to neutral valence within participants was 0.170, the variance for the slope of the positive to neutral valence within participants was 0.291, and the variance of the slope for information within stimulus was 0.001. The residual variance was 1.42. The estimated marginal R ² was .071, indicating that the fixed effects alone accounted for approximately 7.1% of the variance. In contrast, the estimated conditional R ² was .411, suggesting that the combination of fixed and random effects explained approximately 41.1% of the total variance. This difference underscores the substantial contribution of the random effects structure—particularly participant‐level variability in responses to depicted complexity and valence—to the overall model fit, and highlights the large inter‐individual differences in the way depicted complexity and valence influence the judgment of complexity.

Liking judgments

The model was set up to predict liking ratings from the triple interaction of judged complexity, image valence and group, participants' art knowledge and art interest scores, and participants' positive and negative PANAS scores. We included the interaction of judged complexity and image valence as random effects within participants and group as a random effect within stimulus. Before exploring the model, we conducted an analysis of influential cases based on Cook's distance, revealing two highly influential participants. We removed these participants from the dataset and performed the same analysis.

The model revealed that participants' liking for the images could be predicted by their own complexity ratings, whether they had received information about the images, the image valence, and their positive and negative PANAS scores. First, the model indicated a significant effect of complexity ratings on liking [β = .30, t(73.41) = 3.201, p = .002]: the higher participants judged an image's complexity, the more they liked it. Second, participants who received information about the images liked them more (M = 4.09, 95% CI [3.87, 4.32]) than participants who did not (M = 3.72, [3.50, 3.94]) [β = .37, t(67.15) = 2.544, p = .013]. Third, participants liked negative images (M = 3.30, 95% CI [3.02, 3.59]) less than neutral images (M = 4.01, [3.79, 4.23]) [β = .68, t(96.90) = 3.848, p < .001] and liked positive images (M = 4.41, [4.14, 4.68]) more than neutral ones [β = .36, t(95.38) = 1.995, p = .049]. Fourth, the higher the participants' positive PANAS score, the more they liked the images [β = .033, t(68.52) = 2.672, p = .009], and the higher their negative PANAS score, the less they liked the images [β = −.037, t(68.81) = 2.352, p = .022]. None of the other factors or interactions had significant effects on liking ratings (all ps > .092).

Regarding random effects, the variance for the participants' intercept was .262 and the variance for the stimulus intercept was 0.142. The variance for the slope of judged complexity was 0.224, the variance for the slope of the negative to neutral valence within participants was 0.441, the variance for the slope of the positive to neutral valence within participants was 0.468, the variance for the slope of the interaction between judged complexity and negative to neutral valence within participants was 0.070, the variance for the slope of the interaction between judged complexity and positive to neutral valence within participants was 0.035, and the variance of the slope for information within stimulus was 0.003. The residual variance was 1.598. The marginal R ² was .165, indicating that the fixed effects explained approximately 16.5% of the variance in liking ratings. The conditional R ² was .311, indicating that the full model, including both fixed and random effects, explained approximately 31.1% of the variance. These results suggest that the fixed predictors captured a moderate amount of the overall variance, while the random effects—reflecting individual and stimulus‐level variability—contributed additional explanatory power, and highlighted the substantial inter‐individual differences in liking for the stimuli in general and in the impact of judged complexity and image valence on liking ratings.

Understanding judgments

The model was set up to predict understanding ratings from the triple interaction of judged complexity, image valence and group, participants' art knowledge and art interest scores, and participants' positive and negative PANAS scores. We included the interaction of judged complexity and image valence as random effects within participants and group as a random effect within stimulus. Before exploring the model, we conducted an analysis of influential cases based on Cook's distance, revealing three highly influential participants. We removed these participants from the dataset and performed the same analysis.

The model revealed that participants' understanding of the images could be predicted by whether they had received information about the images, the image valence, the interaction between information and valence, and their art interest score. First, participants understood negative images (M = 4.48, 95% CI [4.15, 4.80]) less than neutral images (M = 4.75, [4.44, 5.06]) [β = .49, t(66.84) = 2.176, p = .033] and understood positive images (M = 5.58, [5.27, 5.89]) more than neutral ones [β = .92, t(75.86) = 3.934, p < .001]. Second, participants who received information about the images understood them more (M = 5.17, 95% CI [4.91, 5.43]) than participants who did not (M = 4.70, [4.43, 4.97]) [β = .38, t(78.76) = 2.021, p = .047]. Third, the interaction between information and valence was significant [β = .44, t(74.83) = 2.788, p = .007], indicating that information increased the understanding especially for negative images (Figure 3): participants who received information understood negative images better (M = 4.89, [4.51, 5.26]) than participants who did not receive information (M = 4.07, [3.67, 4.46], z = 3.836, p < .001), but the impact of information on understanding was non‐significant for neutral images (M _Without = 4.56, [4.19, 4.93]; M _With = 4.94, [4.59, 5.29]; z = 2.006, p = .090) and positive images (M _Without = 5.48, [5.10, 5.86]; M _With = 5.68, [5.33, 6.04]; z = 1.049, p = .294). Fourth, the greater the participants' interest in art, the more they understood the images [β = .018, t(66.45) = 2.437, p = .018]. None of the other factors or interactions had significant effects on understanding ratings (all ps > .120), including the main effect of judged complexity [β = −.097, t(72.51) = 1.071, p = .288], indicating that judged complexity did not affect ratings of understanding.

FIGURE 3.

Understanding ratings as a function of information condition and image valence. Raw participant ratings are displayed as jittered semi‐transparent dots overlaid on violin plots showing the full distribution of responses. Black squares represent model‐estimated means from a linear mixed‐effects model, with vertical lines indicating 95% confidence intervals, and the asterisk indicates significant pairwise difference (p < .05). Facets correspond to the three valence categories.

Regarding random effects, the variance for the participants' intercept was 0.435 and the variance for the stimulus intercept was 0.420. The variance for the slope of judged complexity was 0.212, the variance for the slope of the negative to neutral valence within participants was 0.148, the variance for the slope of the positive to neutral valence within participants was 0.300, the variance for the slope of the interaction between judged complexity and negative to neutral valence within participants was 0.080, the variance for the slope of the interaction between judged complexity and positive to neutral valence within participants was 0.030, and the variance of the slope for information within stimulus was 0.082. The residual variance was 1.333. The marginal R ² was .171, indicating that the fixed effects accounted for approximately 17.1% of the variance in understanding. The conditional R ² was .382, suggesting that the full model, including both fixed and random effects, explained approximately 38.2% of the total variance. These results indicate that while the fixed predictors captured a meaningful share of the variance, a considerable proportion was additionally explained by subject‐ and stimulus‐level random effects, reflecting differences among individuals and stimuli in understanding ratings in general, and differences in the way judged complexity and valence influenced judgments of understanding.

DISCUSSION

The primary goal of our experiment was to advance our understanding of how visual complexity is represented and how these representations influence judgments of liking and understanding. Specifically, we sought to determine whether additional information about the artworks and their affective valence modulates judgments of complexity, liking and understanding. We asked our participants to rate the complexity, liking and understanding of 20 negatively valenced, 20 affectively neutral and 20 positively valenced artworks. Half of the participants received information about each of the artworks; half received no information. We hypothesized that participants who received information about the artworks would perceive them as less complex, more likable and more understandable, consistent with prior research showing that contextual information can enhance appreciation and reduce perceived complexity (Leder et al., 2006; Temme, 1992). We also hypothesized that negatively valenced artworks would be judged as more complex, less likable and less understandable compared with positively valenced artworks, in line with studies indicating that negative emotional valence increases perceived complexity and decreases liking (Marin & Leder, 2018; Silvia, 2005).

The first of our goals, thus, was to ascertain whether images' valence and information influence people's representation of their visual complexity. Our results showed that, indeed, affective valence and information affected the way participants rated the artworks' complexity. On the one hand, the more negative the valence, the higher the complexity ratings: negatively valenced artworks were rated as more complex than neutral ones (even though, as noted in the methods section, the depicted complexity of the former was significantly lower than that of the latter) and positively valenced artworks were rated as less complex, which aligns with the idea that negative affect can enhance perceived complexity (Reber et al., 2004). The strength of the effect of valence, however, was reduced by having information about the artworks: valence had a smaller impact on the complexity ratings of participants who received the information, suggesting that contextual information can moderate emotional influences on perception (Cupchik et al., 1994). On the other hand, we also found that information moderated the relation between depicted complexity and complexity ratings: Depicted complexity had a stronger effect on complexity ratings in participants who had received information about the images than in participants who had not, supporting the notion that cognitive engagement enhances sensitivity to visual details (Leder et al., 2006). Participants' art interest and initial affective state had no appreciable effect on their complexity ratings.

One possible reason why positive valence reduced complexity ratings and negative valence increased them could be that participants interpreted the affective valence of the image content, in part, as visual comfort and discomfort, associated with simpler and more complex images, respectively (Fernandez & Wilkins, 2008; Juricevic et al., 2010). Negatively valenced images might have been perceived as more visually discomforting and, therefore, interpreted as more visually complex than positive images, which might have felt more visually comforting and, therefore, interpreted as visually less complex. These findings may be attributed to emotional affect acting as a priming mechanism that modulates participants' cognitive processing. For example, according to Forgas' (1995) Affect Infusion Model, negative emotions tend to promote analytical processing and detail orientation, while positive emotions encourage a more holistic approach, considering the overall situation. Besides, positive emotions simplify information processing by reducing the cognitive load (Shah & Oppenheimer, 2008), which may also increase the sense of familiarity with the stimuli. This heightened familiarity can further reduce perceived complexity, as familiar stimuli are processed more fluently and are often judged as simpler. Therefore, these emotional states directly influence how individuals perceive and interpret the complexity of an image. Contrary to our initial hypotheses, although providing information modulated the effects of other variables, it did not straightforwardly reduce complexity ratings. It might be the case that the representation of visual complexity is affected by implicit knowledge and skills, acquired by familiarization through exposure (Forsythe et al., 2008) and domain‐specific expertise (Commare et al., 2018), but not declarative knowledge. Both of these tentative explanations deserve further consideration in future studies.

The second of our goals was to determine whether participants' representation of images' visual complexity influenced their liking and understanding ratings. Regarding the liking judgments, our results showed that participants' liking for the artworks could be predicted by their own complexity ratings (liking increased with judged complexity), image valence (liking was greater for positive than neutral images and for neutral than negative images), whether they had received information or not (participants who received the information gave higher liking ratings), and by their positive and negative affective states at the beginning of the experiment (participants with higher positive affect gave higher liking ratings, and participants with higher negative affect gave lower liking ratings). Our results are, therefore, in line with the literature that has shown that evaluative judgments increase with judged complexity (Nath et al., 2024; Van Geert & Wagemans, 2021) and with the literature showing that aesthetic preferences are higher for materials of positive valence than for materials of negative valence (Clemente et al., 2022; Leder et al., 2011, 2014), that they increase with author, title, content, style and contextual information about the work (Cleeremans et al., 2016; Darda & Chatterjee, 2023; Millis, 2001; Swami, 2013; Temme, 1992), and are higher with positive affective states than with negative affective states (Belke et al., 2006; Flexas et al., 2013). Together, these results highlight the interplay of perceptual, affective and cognitive processes in aesthetic evaluative judgments (Brielmann & Dayan, 2022; Chatterjee & Vartanian, 2014; Leder et al., 2004; Skov & Nadal, 2021).

Regarding understanding judgments, our results showed that, as could be expected, the higher the participants' scores on the art interest scale, the higher their understanding ratings. More interestingly, participants' understanding ratings were influenced by our experimental manipulations of valence and information. We found that participants felt they did not understand negative artworks as well as neutral images and that they understood positive artworks better than neutral ones. We also found that participants who received information understood the artworks better than those who did not, but this effect was moderated by the valence variable: information increased understanding mainly for negative artworks.

One plausible explanation for participants' lower understanding of negative images is that non‐experts, such as our art‐historically naïve participants, tend to prefer artworks that are representational and of positive valence (Bimler et al., 2019; Leder et al., 2014; Winston, 1995; Winston & Cupchik, 1992). This is because when artistically untrained people look at art, they search for recognizable objects they can associate with stored knowledge, under the belief that understanding the artwork equates to understanding the depicted scene (Cupchik & Gebotys, 1988). Content‐related features, such as familiarity, personal feelings and affective valence, thus, are the most determinant of untrained people's evaluative judgments (Augustin & Leder, 2006; Cupchik & Gebotys, 1988; Nodine et al., 1993; Winston & Cupchik, 1992). They, therefore, become more familiar with positively valenced artworks and understand them better. Different kinds of information, such as titles or descriptive texts, increase people's self‐rated understanding of artworks (Leder et al., 2006; Millis, 2001; Temme, 1992), most especially for the sort of artworks they are less familiar with or understand less well to begin with, such as non‐representational or contemporary art (Bordens, 2010; Leder et al., 2006, 2012; Russell & Milne, 1997). Our results are in line with these studies: Information increased participants' understanding of the artworks, and this was especially true for the negatively valenced ones, the ones participants understood less well.

Interestingly, and contrary to our original hypothesis, participants' representation of visual complexity had no appreciable impact on judgments of understanding. This is an informative result, nonetheless, because it shows that representations of visual complexity impact certain judgments, such as perceptual and evaluative judgments, and not any judgment that could be made about images. This allows us to discard the notion that the impact of complexity representation on perceptual and evaluative judgments is due to general processes, such as translating the output of judgment computations into verbal or numerical scales or to anchoring effects (using the complexity rating as an anchor for the liking or understanding judgment) (Furnham & Boo, 2011; Tversky & Kahneman, 1974).

Some issues need to be taken into consideration when interpreting these results. First, our procedure to create three sets of images that differed in affective valence succeeded in creating three distinct and non‐overlapping levels of the variable affective valence. However, our procedure to exclude images with extremely low or high arousal did not produce full homogeneity in arousal across the three categories: the arousal ratings of images in the negative valence level were significantly higher than those of images in the neutral and positive valence levels. On the one hand, one could argue that this is what is commonly observed in such affective classifications of images, where negative stimuli tend to evoke higher arousal due to their greater emotional salience or motivational relevance (Marin & Leder, 2013; Marin et al., 2016). On the other hand, we aimed to examine the effect of affective valence, not arousal, which means that our results might have been influenced to a certain extent by this uncontrolled higher arousal in the category of negative affective valence. However, given that the difference in mean valence among the image categories (2.40–4.06–5.15) is several times larger than the mean difference in arousal (3.39–3.51–3.48), we believe that the variable affective valence indeed captures mainly differences in valence and not arousal. We consider this imbalance to be minimal and not detrimental to the main interpretation of our findings. Other procedures to assemble sets of artworks that vary independently in valence and arousal would be required to test their separate contribution to the effects of affect observed here. Future studies might consider constructing stimulus sets in which valence and arousal vary independently, allowing for a more precise examination of their respective contributions to the perception of visual complexity.

Second, given that the texts available to one of the groups included information about style, content and pictorial resources used by the artist, some texts, especially those accompanying stimuli in the positive and negative affective valence category, are not affectively neutral. Such texts included words like ‘light’, ‘leisure’ and ‘joy’, or ‘conflict’ and ‘anger’. It is, therefore, conceivable that they set an emotional frame that added to the image's own affective content and contributed to the interaction of the information and valence category observed for the complexity and understanding ratings. Maybe what we defined as ‘information’ partly overlapped with affective content. Future research could therefore employ affectively neutral descriptions and compare them with affectively charged ones to dissociate strictly informational from affective effects.

Finally, we selected liking and understanding ratings as instances of judgments more reliant on affect and cognition, respectively. Further research would be required to determine the impact of complexity, valence and information on other sorts of judgments that seem to rely on more complex interrelations of affective and cognitive processes, such as judgments of beauty.

CONCLUSIONS

Our results support previous research that has shown that people's representation of visual complexity does not rely solely on image features that have to do with the amount, variety of elements and their disorganization (Berlyne et al., 1968; Nadal et al., 2010; Nath et al., 2024; Van Geert & Wagemans, 2021), but also relies on knowledge acquired through familiarity and domain‐specific expertise (Commare et al., 2018; Forsythe et al., 2008). Specifically, our study has shown that people's representation of visual complexity also relies on an image's affective valence and the available information about it, and that these aspects interact with each other. Forsythe et al. (2008) referred to the influence of these variables as confounds that bias participants' judgment of visual complexity, and suggested that computer‐based measures of complexity are preferable because they are unaffected by them. From a psychological perspective, however, we believe that such variables should not be regarded as confounds, but as sources of information people use to construct a meaningful representation of visual complexity. Finally, our results have also shown that this complexity representation constructed upon image features, affect and knowledge acquired in different ways, influences judgments of complexity and liking, but not of understanding.

AUTHOR CONTRIBUTIONS

Xiaolei Sun: Writing – original draft; formal analysis. Jiajia Che: Writing – original draft; formal analysis. Marcos Nadal: Conceptualization; methodology; writing – review and editing. Helmut Leder: Conceptualization; writing – review and editing; funding acquisition; methodology.

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are openly available in OSF at https://osf.io/zphjg/.