Immersive virtual reality (VR) technologies are becoming more and more promising for cognitive behavioral therapy and food science. Their evolution as well as the graphic computing capacities of Personal Computers (PC) now allow complex and rich 3D environments to be displayed. Integration and evaluation of food contents in VR have increased in the last few years, as evidenced in literature by recent reviews in the sensory and consumer science field (Wang et al., 2021) and for eating disorders rehabilitation (So et al., 2022). Studies involving food contents in virtual environments have different goals linked to the desire to eat elicited by the food stimuli (product selection, food evaluation, eating disorder diagnosis and therapy); the 3D food models considered in these studies are heterogeneous: they come from different creation process (e.g., manual design, photogrammetry) and thus exhibit different degrees of visual quality. This visual quality may have a significant influence on the outcomes of these studies, but this influence has never been demonstrated or isolated.

Therefore, in this paper, we present a study designed to evaluate the influence of the graphical quality of 3D food models on the desire to eat they arouse, in a virtual environment.

For this purpose, we conducted a user experiment in virtual reality; participants (N = 28) were asked to report their eating desire for a large number of 3D food stimuli of different types and different visual qualities. Results allow us to draw relevant conclusions and recommendations for the design of VR environments and experiments involving food stimuli.

The remainder of this paper is organized as follows. Section 2 reviews the related work about the use of food stimuli in VR. Section 3 presents our objectives and hypotheses. Section 4 describes the creation of the stimuli and the design of the VR environment as well as the materials. Section 5 presents the conducted experiment. Results and discussion are provided in Section 6 and Section 7, respectively. Finally, in Section 8, we conclude and propose several perspectives.

The study of food-related behaviors and feelings tends to develop more and more in VR immersive environments. Uses of food stimuli have been shown to be as effective as real ones, even stronger in some cases (van der Waal et al., 2021), and more effective than static images in generating emotional responses (Gorini et al., 2010). Multiple use cases of VR technologies for food science are identifiable and can be classified into different categories (Xu et al., 2021). Particularly, rehabilitation of eating disorders through the use of VR is considered as a promising strategy for assessments and treatments of these kinds of disorders (de Carvalho et al., 2017). So et al. (2022) published a survey that highlighted such rehabilitation studies in VR, despite their limited number. Virtual buffet in VR has also been used for evaluating non-conscious and uncontrollable aspects of food selection, or as a validated tool for studies on food preferences (Marcum et al., 2018; Persky et al., 2018; Ung et al., 2018; Cheah et al., 2020). Nutrition education also finds interest, especially for child feeding research (Persky et al., 2018). Finally, physiological measurements are also increasingly used in food-related VR experiments, especially gaze behavior (due to the increasing implementation of eye-trackers in VR head-mounted displays), e.g., for studying the effect of food presence on visual attention behavior (Hummel et al., 2018) or eye activity during a food choice experiment (Machín et al., 2019). Techniques used to create the 3D food stimuli used in these experiments are variable. Photogrammetry is frequently used to create and display realistic content in VR. Alba-Martínez et al. (2022) compared real cakes and 3D cakes made using a photogrammetric process for a visual characterization task. Chen et al. (2020) used scanned prepared foods with real-life proportions in a VR buffet to assess food selection processes. Gouton et al. (2021) captured real-life cookies by a photogrammetric process to reproduce conditions for validation of attribute characterization in VR. 360° video is also a method to reproduce a 3D environment in a realistic way. For example, Crofton et al. (2021) used 360° video to reproduce multiple contextual conditions for beef steaks and chocolate sensory analyzes. 3D food stimuli can also be created with 3D modeling software. Persky et al. (2018) used this modeling process to create a VR buffet for child feeding assessment. van der Waal et al. (2021) and Ammann et al. (2020) reproduced respectively chocolate, juices, and cakes using 3D modeling to study the color influence on flavor identification. This plurality of means used to create 3D food stimuli leads to disparate levels of realism and visual quality. Usually, food stimuli created using 3D modeling software are of lower visual quality and realism than those created by photogrammetry. The potential impact of this visual quality on the observed results has not been studied yet. Therefore, our study will focus on examining how the visual quality of food stimuli in virtual reality influences the elicited desire to eat.

The 3D environment (i.e., the visual context of the food stimuli) is also a factor that has been shown to influence the desire to eat. Specifically, the type of environment and the evoked context (i.e., conditions where the food is supposed to be consumed (Meiselman, 2006)) influence the appreciation scores of food contents in an immersive environment (Meiselman et al., 2000). An environment consistent with the product where it should be consumed generally makes the product more appealing (Picket and Dando, 2019) and situational appropriateness is more and more considered in recent research topics (Schöniger, 2022). Appreciation scores are lower under artificial eating conditions (e.g., laboratory) than from naturalistic eating conditions (Delarue and Boutrolle, 2010). Schöniger (2022) confirmed the aforementioned observation between immersive and laboratory environments. Since the aim of our study is to isolate and evaluate the effect of the visual quality of the food stimuli, we chose a neutral visual context (a sensory analysis booth) to avoid the effect of this context (which depends on the type of food), even if the induced response could be slightly reduced.

The purpose of our study is to characterize the variation of desire to eat felt by VR users when facing virtual food stimuli. As a main objective, we want to study the effect of visual quality on this feeling. Indeed, since we propose here a user experience in a controlled virtual environment, we can easily modify food stimuli, and thus the perception that users have of them, sight being a crucial sense for food and appetite in humans (Rolls et al., 1976; Rolls et al., 1983). In addition, we expect that some individual factors might have an effect on users’ responses to virtual food stimuli, as well as food type (Asp, 1999; Chen and Antonelli, 2020).

Accordingly, we propose the following hypotheses.

• H1: the desire to eat felt by a VR user is correlated to food stimulus visual quality, and decreases when this one is degraded.

The analysis of demographic data (gender, age, education level) did not reveal any particular effect on desire-to-eat scores, neither did the global desire to eat, the last time participants took a meal, their smoke habits, and their experience with immersive technologies.

Two of our twenty-eight participants were excluded from analyzes due to incongruous ratings. One was declared to be “unresponsive to VR environment” because of a desire-to-eat score of 0 for all stimuli during the first exposure session; therefore, we did not to conduct the second session. During the debriefing, the participant declared (I cannot project myself on the virtual environment; I cannot say “I can eat it.”) The other excluded participant presented outlier values in their score answers (values were considered as outliers when below the 1st quartile − 3 * interquartile range (IQR) or above the 3rd quartile + 3 * IQR, across all the participants’ results). We considered then twenty-six participants for the next statistical analyses.

Due to the sample size of participants used in our trial, and because the collected data are ordinal, we used non-parametric tests to perform analyzes. The level of significance was set to 5%.

First, the results from our experiment do support our Hypothesis H1: the desire to eat increases when the visual quality of virtual food increases too. More precisely, desire-to-eat scores are significantly lower for levels 1 and 2, and starting from level 3, they increase, even forming a “top” group for levels 6 and 7, with more moderated significant differences.

This observation suggests that a quality level of 3 seems to reflect the minimum required quality to trigger a significant desire to eat, with levels 4 and 5 triggering a similar desire. Significantly better results are obtained starting from level 6 (no significant improvement is brought by level 7). In terms of pseudo-MOS values, it means that a pseudo-MOS of 3 (on a scale from 1 to 5) is a minimum value to start to trigger significant desire-to-eat for VR food studies; the best choice, if possible regarding computing capacity, is a pseudo-MOS of 4.5 (without improvement after this value). Given the fact that the tool for computing pseudo-MOS is publicly available (Nehmé et al., 2023), this result is of great interest for optimizing 3D food models in food-related virtual reality experiments (e.g., for consumer research, eating disorder therapy, and so on).

Our results also support our Hypothesis H2. Desire-to-eat scores did vary according to the type of food, since we found significant differences of scores depending on the type of food presented, considering all visual quality levels together. Some comments given by the participants can help to understand some of these differences. For example, in the answers to our post-experiment debriefing, we noticed that the words “pain” and “chocolat,” which represent the pastry of the same name in French, were the most spoken words. They were used in combination with words regarding its appearance, using expressions such as “crushed in a bag,” “burnt” or “too orange.” Therefore, it tends to reveal that this food stimulus has been perceived as not appetizing as could be expected.

Hypothesis H3 was verified in our experiment since the participants’ eating habits (i.e., their taste for the corresponding reference food models) did influence significantly their desire-to-eat scores. When a food stimulus was not very appreciated, it produces a lower desire to eat than the other food. Similarly, the most preferred food produced the highest desire to eat, for all visual quality levels.

We also conducted an extra analysis of the desire-to-eat scores in regard with participants’ habits and the visual quality of the stimuli. We found that the difference between the extreme levels (1 and 7) significantly varied, with lowest differences when the food stimulus was “not at all or not really” liked, compared to the most preferred food that elicited highest contrasts regarding the desire-to-eat across the visual quality levels (see supplementary material file Section 4).

Nonetheless, we found interesting facts for some specific stimuli, despite our global results about hypotheses H2 and H3, from the debriefing interviews of the participants. Figure 11 shows a word cloud containing the most frequent words told by participants. Color differences represent word frequencies. This includes the findings about the “pain au chocolat” described above. This result could explain the low desire-to-eat scores elicited by this virtual stimulus despite its high food habit values. This example can be representative of a larger effect: despite being created by a high-quality photogrammetry process, the appearance of food models can seem not realistic for the participants. This observation suggests that, despite the high-quality brought by photogrammetry, there is still a need for more accurate processes that would be able to capture the physical properties of the objects (e.g., the way it reflects the light). It is even more important in the field of food science in virtual environments since food desire is also influenced by color contrasts (Paakki et al., 2019).

The proposed study evaluated the influence of the visual quality of food stimuli on the elicited desire to eat in virtual reality. Results show that the visual quality of virtual food stimuli significantly influences the desire to eat felt in virtual reality, where the desire decreases along with the visual quality. Moreover, our results pave the way to the optimization of 3D food models since we showed significant differences between some levels, as well as the absence of perceived desire to eat between other levels, e.g., between two very high quality levels. The desire to eat is also influenced by the type of food and the eating habits of the participants. Those results allowed us to make useful recommendations for designing virtual reality experiments involving food stimuli. Several perspectives may be considered. Firstly, the stimulation of other senses, in particular smell, should highly contribute to the intensity of the elicited responses (emotions, desire to eat). Flavián et al. (2021) showed that scent can strengthen the link between affection and conation; Javerliat et al. (2022) proposed an open-source reproducible olfactory device compatible with autonomous HMDs, that would allow an easy integration of this sensory cue in further experiments. Secondly, the use of reference food models representing a larger coverage of different diets and evaluated by a panel with a greater diversity could allow to confirm our results in a more general setting. Finally, as stated above, a perspective is to achieve a more realistic object appearance able to reproduce light interactions and thus improve food realism. For example, recent deep-learning techniques allow to reconstruct the accurate geometry and Spatially-Varying Bidirectional Reflectance Distribution Function (SVBRDF) (Bi et al., 2020) from a set of sparse images of a 3D object.