This manuscript adheres to the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) guidelines. This study replicates and extends a prior VR-based self-compassion therapy [3], which serves as the control condition for study 1. The original iVR experience included an onboarding interface where participants selected avatars, environments, and companion avatars with emotional behaviors (crying or upset). The therapy consisted of two stages: delivering compassion and receiving compassion.

Three between-subject studies were conducted:

Each study built on the previous, allowing for pairwise comparisons. Stylized avatars and new features were integrated by editing the original Unity project files.

Figure 1 compares the stylized avatar used in experimental conditions with the realistic avatar from the original system. Based on Weidner et al [36], stylized avatars feature simplified textures and nonhuman proportions, while realistic avatars maintain human morphology and detail. The avatars of the control study (Figure 1, right) had such realistic features and therefore aligned with generic realistic avatars as defined in Weidner et al [36].

For clarity, for the remainder of the paper, the term “realistic” refers to the generic realistic category [36].

Two key instruments were used:

From study 2 onward, the 8-item Patient Health Questionnaire (PHQ-8) [38], a validated measure for screening depression symptoms, was introduced to support downstream clinical effects of the iVR intervention. As this paper continues the work of Halim et al [3], it was considered prudent to include ongoing depression measurement.

Within study 3, presence was measured using the Slater-Usoh-Steed (SUS) questionnaire [8,39], a 6-item scale to support the analysis of UX. With respect to measurement, presence was prioritized over embodiment measures due to practical constraints; also, as introduced earlier, presence performance implies embodiment quality [13,14].

A qualitative questionnaire gathered open-ended feedback about the avatars (studies 1 and 2) and a virtual mirror (study 3). These questions targeted participants’ likes and dislikes to probe for system improvements and aimed to elicit uncanny valley (UV) responses via an open-ended question in studies 1 and 2:

In study 3, participants were given questions pertaining to their experience with the virtual mirror:

In the absence of eye tracking technology, these responses provided a means of validating mirror interactions. Thematic analysis was used to evaluate these responses [40].

Two comparisons were made:

For each quantitative data set, the Shapiro-Wilk normality test was first applied. Given the normality of the data, either a Mann-Whitney U test or a 2-tailed t test was performed. The null hypothesis assumed no significant differences between conditions.

Participants were recruited from the general community via posters, social media, university mailing lists, and word of mouth. Most were students or staff at The University of Queensland. Eligibility required being ≥18 years of age and capable of using immersive VR; those with severe motion sickness or conditions that might interfere with VR use were excluded.

Sessions were conducted one-on-one in a controlled room at The University of Queensland’s extended reality laboratory (Brisbane, Australia), with a trained researcher present. Participants were recruited and tested between 2023 and 2024. Studies 2 and 3 included a 2-week in-person follow-up. Some individuals who registered did not attend their session and were not awarded course credit; no reasons were collected.

Sample sizes were determined pragmatically based on prior VR studies, available resources, and recruitment feasibility in a university setting. No formal power analysis was conducted due to the exploratory nature of the research.

This study was approved by The University of Queensland Human Research Ethics Committee (approval #2023/HE000468). All participants provided written informed consent prior to participation. Data were de-identified before analysis to protect participant privacy and confidentiality. Participants received course credit through the university’s research participation system. No images or materials in the manuscript or supplementary files contain identifiable individuals.

A flow diagram showing participant progression and the assessment schedule is presented in Figure 5.

A total of 20 participants completed session 1 (n=12 female participants, n=6 male participants, and n=2 undisclosed).

A total of 49 participants completed session 1; 42 (86%) returned for session 2. The gender of the participants was as follows:

A total of 38 participants completed session 1; 35 (92%) returned for Session 2. Gender of the participants was as follows:

Open-ended feedback was collected from 100 participants across all three studies; participants completed their assigned iVR therapy exposures prior to providing responses. Deductive and inductive thematic analysis was conducted on participants’ open-ended responses. Participants were labeled P1-P20 (study 1), P21-P62 (study 2), and P63-P100 (study 3). The deductive thematic approach (Table S3 in Multimedia Appendix 1) includes codes corresponding to UEQ dimensions from Schrepp et al [46], which provided a structured lens through which to interpret qualitative feedback.

In parallel, an inductive analysis was performed to identify emergent themes not captured by the UEQ framework. These data-driven themes reflected patterns in participant experience that appeared to be consistent with experimental changes. The results of the inductive analysis are summarized in Table S4 in Multimedia Appendix 1.

This study explored how stylized avatars, avatar customization, and virtual mirrors influenced UX (RQ1) and self-compassion outcomes (RQ2) across three iVR therapy conditions. In study 1, qualitative feedback indicated general acceptance of the stylized avatars. Study 2 showed a significant improvement in self-compassion outcomes compared to study 1. Study 3 produced significant improvements in four of the six UEQ dimensions, which suggests improved UX.

For cross-study comparisons, see Table S2 in Multimedia Appendix 1. Study 3 yielded the highest attractiveness scores, with a statistically significant improvement compared to study 2 (P=.01). Qualitative data indicated this may be attributed to the enjoyment of avatar customization, which was introduced in study 2 and praised by participants in study 3. However, immersion-related feedback also emerged, such as “some of the environment designs feel too rigid…which breaks the immersion” (P78). Since Study 3 introduced virtual mirrors—known to support embodiment—this addition may explain the rise in UX.

On the perspicuity scale, study 3 again scored highest and significantly outperformed study 2 (P=.01). Participants highlighted the ease of learning, but across all studies, a lack of narrative context was noted. For example, participants reported difficulty understanding the avatars’ emotional states, affecting their ability to engage in role-play scenarios.

Study 3 also significantly improved dependability scores (P<.001). The stylized avatars, rather than triggering discomfort, were perceived as appealing—“they’re all cute” (P37)—and met participant expectations for a supportive environment. Thematic analysis suggested that the participants felt the system supported therapeutic outcomes: “I think this can make me feel relaxed” (P83).

Stimulation was significantly higher in study 3 than in study 2 (P=.04). Participants reported increased interest when viewing themselves delivering compassion, suggesting that the mirror feature enhanced engagement.

With respect to novelty, all studies received positive feedback—“innovative and unique” (P20)—but only study 1 reached an “above average” benchmark. Some participants reported reduced novelty due to restricted avatar movement, for example, “better if the users got a little more freedom during the simulation” (P56). This may indicate a design trade-off between structured therapeutic focus and exploratory freedom.

The UEQ served as the primary UX tool due to its alignment with iVR design goals for mental health. Though no significant differences were observed between studies 1 and 2, stylized avatars performed well compared to the UEQ benchmark, validating their continued exploration in therapeutic contexts. However, despite enhanced customization in study 2, UEQ scores did not improve significantly. One possible explanation may lie in elevated participant expectations (eg, limited skin tone options), which may have led to reduced satisfaction. This supports the idea that “less is more”; a limited but well-executed customization system may better serve therapeutic VR experiences.

Feedback across studies emphasized the importance of avatar individualization for user acceptance. Study 3’s improvement in multiple UEQ dimensions suggests that mirrors—without being explicitly highlighted—supported user presence (and likely embodiment). “Attractiveness is a pure valence dimension” [46], that is, representative of the user’s general experience. Attractiveness rose significantly in study 3, likely due to parallel improvements in perspicuity, dependability, and stimulation.

This connection is further supported by SUS findings. Significant improvements on items related to perceived realness—questions 2 and 3 of the SUS questionnaire [8]—and memory of the experience indicate that participants recalled the virtual mirror scenario as immersive and believable. Given the high scores (Table S5 in Multimedia Appendix 1), this kind of recall implies a strong sense of presence [39].

Taken together, study 3 suggests that the inclusion of virtual mirrors was a key turning point in enhancing iVR UX. While prior studies did not yield significant gains, mirrors may have “unlocked” the potential of the design. Whether these effects are solely due to mirrors or the cumulative impact of prior changes remains unclear, but no prior condition showed similar improvements.

In study 2, participants exposed to customizable stylized avatars over two sessions showed a significant increase in self-compassion (SCS). This supports the notion that stylistic, customized avatars in iVR interactions may enhance therapeutic benefit. However, these gains were not observed in study 3, despite the enhanced UX. This suggests that while mirrors improve UX and presence, they may not directly translate to better therapeutic outcomes.

Nevertheless, qualitative feedback highlighted that participants valued the ability to create avatars resembling themselves. It was noted that this customization improved their emotional connection to the experience in stage 2 of the experiment. For example, “customizing it to look like me did make the difference of how I perceived the audio replay” (P60). This statement implies that appearance alone was insufficient; users needed a visual feedback loop (eg, mirrors) to fully benefit from avatar customization.

Interestingly, mirror placement was not emphasized to participants yet still yielded a significant impact. However, incorrect or unnatural mirror placement disrupted immersion. For example, “[mirror] could be weird [to see] in the park environment” (P64) or “I see the mirror and then walk through it. I’m [a] ghost and that makes me a little bit scar[ed]” (P76). This underscores the importance of environmental congruence in design-embodied iVR therapy.

The appearance of stylized avatars was generally well received, supporting the idea that less realistic avatars can be effectively used in emotionally sensitive and immersive environments such as iVR self-compassion therapy, where avatar-owner resemblance is important. This finding aligns with prior research suggesting that stylized avatars may reduce the risk of UV effects while still fostering user connection [42,43,48]. Participants responded positively to avatars with cartoonish or humanoid qualities. “I don’t recommend realistic avatars as much. I think the more cartoonish appeal is better” (P41).

The UV model suggests that entities close to human-likeness [44], but imperfect, can cause discomfort. The use of stylized avatars in this study—located near “Humanoid Robot” on the affinity line—appears to have mitigated this risk. Mirrors further supported predictability and safety (dependability), suggesting that this combination may serve as a UX-enhancing strategy to avoid UV interference.

That said, UV was not the primary aim of this study but a theoretical background for design rationale. However, these reflections are offered as considerations for future iVR therapy design. It is important to note that violations of physical logic (eg, walking through mirrors, unnatural placement) also negatively impacted immersion and should be carefully addressed in future iterations.

This study shows that stylized avatars, when paired with suitable UX enhancements such as virtual mirrors and avatar customization (along with other individualization concepts, environment selection, companion avatars, emotional behavior, etc), can offer stronger UX without losing therapeutic benefit, which might be an assumed consequence of selecting a lesser fidelity avatar. Pursuing hyperrealistic avatars without considering context may be counterproductive, as this ignores the Sisyphus-like limitation that avatars will risk falling back into the UV when realism is applied without careful design intent. This study gathers evidence to suggest that lower-fidelity avatars can be more appropriate in therapeutic settings but also that context matters when selecting avatar designs.

This study used the SCS—a validated tool—but it has known psychometric limitations [49]. Additionally, participants were not filtered based on PHQ-8 scores; however, 74% had mild or higher range depression results. Future study iterations should include codesign and testing with clinical populations.

While the fidelity comparison was carefully controlled across studies, data were collected at different time points, which introduces the possibility of unmeasured societal or contextual differences. Future studies should aim to run control and experimental conditions concurrently to further minimize potential time-related confounds. Also, avatar fidelity was classified using a binary framework [36]. Future research may benefit from applying more continuous fidelity measures—such as user-rated realism scales or quantitative visual complexity metrics—to better capture subtle design differences.

Mirror engagement was inferred rather than directly measured, as the Meta Quest 2 headset lacks built-in eye tracking. However, mirrors were centrally positioned, and qualitative responses strongly suggest that participants did engage with them.

Finally, some tests were underpowered, which limited the ability to detect small-to-moderate effects. To help interpret these findings and provide additional insight, qualitative methods were included to support quantitative results.

If compassion means “to suffer together” [50], then does engaging in self-compassion mean suffering alone? Perhaps not with these avatars. In this iVR intervention, participants were comforted by their own stylized, customized avatars within immersive environments enhanced by virtual mirrors. Participants experienced self-compassion in a space tailored to their preferences, offering new insights into how avatar design and environmental cues shape therapeutic experiences.

This study sequentially investigated whether stylized avatars, avatar customization, and virtual mirrors influence UX and therapeutic outcomes in self-compassion VR therapy. Across three studies, results demonstrated that stylized avatars were well accepted and did not diminish self-compassion outcomes. Avatar customization improved self-compassion scores in one study, while virtual mirrors significantly improved UX across multiple dimensions. However, enhanced UX alone did not guarantee stronger therapeutic outcomes.

These findings suggest that individualized avatars and virtual mirrors can meaningfully enhance iVR experiences without triggering UV effects. More broadly, this study supports the feasibility of lower-fidelity avatars in mental health VR applications. Future research should explore these design strategies in clinical populations to better understand avatar-based therapy interventions.