In organizational settings, employees’ affective states are highly dynamic. Affective Events Theory asserts that discrete daily work incidents trigger short-lived emotional reactions that subsequently color attitudes and behavior (Weiss and Cropanzano, 1996). Diary and experience-sampling research confirms this volatility: a substantial share of the total variance in momentary positive and negative affect resides within rather than between persons (Ilies et al., 2015). Such intra-individual fluctuations matter: ambulatory assessment studies show that transient shifts in affect forecast on-the-spot safety compliance, vigilance errors and interaction quality (Klumb et al., 2009). Therefore, organizations require assessment tools that can capture affect in situ and at cadences fast enough to support just-in-time interventions, for example, prompting micro-breaks when arousal drops or alerting supervisors when tension spikes.

Ambulatory assessment therefore seeks to record psychological states in situ through computer-assisted self-reports, behavioral logs or physiological sensors while people perform their normal duties (Bachmann et al., 2015). Likewise, ambulatory assessment is highly suitable for affective research, which are complex neurobiological and psychological phenomena characterized by high variability (Weiss and Cropanzano, 1996).

Three broad methodological streams can be distinguished: Physiological indicator measures, External behavioral observations, and Self-reports. (1) Physiological measurements monitor heart-rate variability, electromyography, skin conductance, and other autonomic or central nervous system signals. This approach yields comprehensive data but remains intrusive and can disrupt normal activity (Cowie and Douglas-Cowie, 1996). Even though recent research has introduced the feasibility of lightweight devices such as smartwatches (Toshnazarov et al., 2024), detailed physiological data continue to pose non-trivial privacy compliance risks, and the costs of both hardware and software still remain substantial (Bolpagni et al., 2024; Doherty et al., 2025). (2) External behavioral observations focus on visible cues like facial expressions, postures, and voice tone. Utilizing high-definition cameras and streaming video for real-time analysis, this approach provides a relatively objective measure of emotional fluctuations. However, this non-intrusive method could infringe on personal privacy and confidentiality, making it less viable in organizational settings (Betella and Verschure, 2016). (3) In the self-reporting method, emotion is typically assessed through textual scales. Unfortunately, although such methods collect extensive emotional data, it requires that the scales be simple and engaging, and may pose significant challenges for individuals with limited reading skills (Cranford et al., 2006; Ebner-Priemer and Sawitzki, 2007; Wilhelm and Schoebi, 2007).

Taken together, existing ambulatory techniques struggle to balance ecological validity, intrusiveness and scalability in organizational contexts. Physiological measures and external behavioral observations are costly and primarily limited to experimental scenarios. Textual self-report methods, though cheaper and more versatile, require well-designed scales and a certain level of participant literacy. Considering these limitations, pictorial scales may prove to be a solution.

Pictorial scales are a measurement method that utilizes visual elements to convey the meaning of items (Sauer et al., 2021). These scales offer several advantages, including simplicity (Wissmath et al., 2010), rapidity (Schreiber and Jenny, 2020), low dropout rates (Baumgartner et al., 2019), and low cognitive demands (Desmet et al., 2016; Obaid et al., 2015). These scales are particularly effective for individuals with lower educational attainment or limited reading skills. Consequently, pictorial scales can be better suited to organizational settings that include a diverse workforce, especially employees with limited literacy (Sauer et al., 2021).

Depending on different forms of presentation, pictorial scales can be divided into three categories: grid-based, humanoid-based, and realistic-picture-based, as shown in Table 1. As for grid-based scales, some typically use a two-dimensional affective space (valence * arousal), such as Affect Grid (Russell et al., 1989) and Feeltrace (Cowie et al., 2000). However, grid-based tools also have notable drawbacks. Because valence and arousal are reported with a single click inside a matrix, respondents must trade one dimension against the other, which lowers the effective resolution for each construct. Locating a precise cell further imposes a non-trivial cognitive load that slows response time and disadvantages participants with limited numeracy or literacy (Cowie and Douglas-Cowie, 1996; Russell et al., 1989).

Humanoid pictorial scales utilize emoticons or humanoid cartoons as anchors. For instance, the Self-Assessment Manikin (SAM) uses five progressively changing humanoid cartoons to measure valence, arousal, and dominance. Humanoid or manikin scales likewise present limitations. Interpretation of the stylized figures is filtered through culturally embedded gender and age stereotypes, leading to systematic bias across demographic groups. Moreover, the extreme postures used to depict very high arousal become visually ambiguous, reducing discriminability at the upper end of the arousal continuum (Betella and Verschure, 2016; Bradley and Lang, 1994).

The realistic-picture-based scales are based on psychological projection techniques and use a multitude of real pictures as anchors, measuring emotions by having participants select the image that best matches their mood. For example, the Photographic Affect Meter (PAM) scale presents 16 images of different categories and themes in a single session, and determines participant’s valence and arousal based on affective labels corresponding to the selected image (Pollak et al., 2011). Similarly, the Highly Dynamic and Reusable Picture-based Scale (HDRPS) displays a quartet of thematically matched photographs (e.g., four animal images) that span an unpleasant-to-pleasant continuum. Participants click the single image that best matches their current affect, and its ordinal rank is recorded as the valence score for that trial (Liu et al., 2023).

Compared with instruments like grid and manikin, realistic photographs offer three distinct advantages for high-frequency organizational use. First, they deliver rich, multimodal cues that can be decoded in under a few seconds, matching the pace of diary sampling (Lang et al., 2005). Second, the concrete imagery eliminates the abstraction burden of coordinate mapping or icon interpretation, thereby reducing error variance and respondent fatigue (Dan-Glauser and Scherer, 2011). Third, photo sets can portray diverse contexts and actor identities, which helps minimize gender- or culture-specific stereotype bias and yields smaller social-desirability effects than stylized faces (Betella and Verschure, 2016). These properties make realistic-picture scales particularly suitable for continuous affect monitoring in the workplace.

As the main problem addressed in this study is the continuous measurement of individual affects in organizational contexts. On the one hand, in labor-intensive sectors such as construction, manufacturing, warehousing, and hospitality where many frontline employees have limited reading proficiency, measurement tools must remain simple and easy to understand. On the other hand, considering the need for repeated measurements, these tools should also be engaged to enhance participant motivation. Considering these factors, measurement tools that use real images as anchors are particularly suitable for this context.

However, there are relatively few realistic-picture-based scales available, and those that exist exhibit obvious limitations. For example, the PAM presents 16 images simultaneously to measure emotional states, raising unresolved questions about whether this multi-image display method might alter the subjects’ emotional states. Second, the image materials in PAM have not undergone rigorous rating, making the use of self-reported emotions as emotional labels for the images less precise. Third, PAM also mixes images of different types and themes, which does not account for the influence of personal preferences or visual appeal on the measurement results (Sauer et al., 2021). Finally, with only 100 images in its database and presenting 16 images at once, there is a high possibility of repetition, which can decrease response motivation.

Considering the limitations of existing realistic-picture-based scales, our team previously developed HDRPS. This scale initially collected 22,054 raw images, which were subjected to a three-step evaluation process: image usability testing, emotional type assessment, and emotional scoring experiments. Through more rigorous evaluations, images were carefully selected and categorized, resulting in the creation of a structured pictorial scale with 3,386 images, each tagged with detailed attribute labels.

Thus, HDRPS refines the selection and display of images to enhance measurement accuracy and reduce bias. By presenting fewer images at once and employing a rigorous vetting process for each image, it ensures more consistent and reliable emotional assessments. Additionally, HDRPS expands the image database, thereby reducing repetition and sustaining participant engagement throughout the study. This results in a more robust and effective tool for capturing nuanced emotional responses in research settings.

However, HDRPS still faces two significant limitations that necessitate further optimization. Firstly, in terms of measurable dimensions, while HDRPS images are labeled with both valence and arousal, the scarcity of high arousal images has led to a design focus primarily on valence. This limits the scale’s ability to accurately capture the precise emotions of the subjects. Secondly, regarding the presentation of images, although HDRPS reduces the influence of personal preferences and visual appeal by displaying four images of the same category (e.g., all animal images) at once, it still fails to eliminate the impact of content differences (e.g., images of a cat, dog, bird, and fish) on the measurement outcomes. These issues highlight the need for further refinements in HDRPS to enhance its accuracy and applicability.

In this article, we present the development and validation of HDRPS+, an affect measurement instrument designed for real organizational contexts. In response to the practical needs for dynamic and continuous measurement, we conducted 3 sub-studies to develop and validate HDRPS+, which can continuously measure changes in valence and arousal of organizational members. As an important optimization of HDRPS, developed by our team previously (Liu et al., 2023), HDRPS+ incorporates arousal dimension into measurement process, which can more accurately characterize individuals’ emotional states. Compared to traditional emotion research instruments, HDRPS+ is shown to have broader applicability, lower economic costs, simpler testing procedures, higher response rates, and lower susceptibility to social desirability effects. As a result, it can provide valuable support for continuous affect measurement in organizations.

Specifically, the three sub-studies were built upon the foundational designs of PAM and HDRPS to refine and test HDRPS+ thoroughly. First, we identified image acquisition ideas based on the results of semi-structured interviews, and constructed a picture pool through picture acquisition and supplemental acquisition (Study 1). Then, we screened and constructed picture attributes by evaluations on available, affective category, and affective score (Study 2). Finally, after comparing different picture presentation methods, we determined the optimal measurement and conducted a large-scale trial to verify the reliability and validity of the newly developed tool (Study 3).

Compared with prior measurement tools, the special contributions of this study are fourfold. First, the development and validation process of HDRPS+ is more rigorous, providing a beneficial reference for the development of pictorial scales. Second, the design of HDRPS+ draws on the principles of simple-drawing scales such as SAM, emphasizing the correlation of appearance or content among each set of images, thus effectively controls the influence of personal preferences and image attractiveness on the measurement results. Third, HDRPS+ was developed based on the valence-arousal emotion model, using 3 same-themed images to measure the subjects’ valence and arousal, which can draw individual affect portraits more quickly and accurately. Fourth, the database comprises 20 thematic image sets, including animals, plants, and scenes, offering richer material and a more balanced category distribution that can accommodate a wide range of subsequent research designs.

As realistic pictures are diverse and complex, clear clues for picture collection must be established when using them as measurement anchors. To this end, we recruited 57 students (29 males and 28 females, with an average age of 21.81) from Sichuan University who had participated in HDRPS picture ratings, and conducted semi-structured interviews in which lasted approximately 10 min each, aimed to address the following questions:

These interviews yielded 8 h and 43 min of audio recordings, which were transcribed into 150,600 words. Using ATLAS.ti 9.0 software, the data underwent three stages of coding, resulting in 496 primary codes, 26 secondary codes, and 3 core propositions (overall factors, emotional dimensions, and picture categories). The key findings are as follows: (1) Personal preferences would affect the evaluation of pictures in animals, plants and scenes. To avoid this, HDRPS+ should use group pictures with the same main subject matter. For example, in the case of images featuring people, all images within a set should feature the same individual, while for animal images, all images within a set should have similar appearances. (2) The factors influencing the evaluation of valence and arousal differ. Therefore, image collection should be dimension-specific to accurately convey emotional meanings. (3) The collection strategy for specific images should be adapted according to the circumstances. Due to the poor emotional discernibility of object-related images, we have temporarily restricted the image categories to people, animals, plants, and scenes. Additionally, considering that expressions and postures can affect the affective judgment of animal-related images, we have only selected mammalian animals with clear and visible facial expressions during picture acquisition.

We referenced the methods utilized by HDRPS and NAPS (Nencki Affective Picture System, Marchewka et al., 2014) for picture acquisition. With the exception of some self-taken photographs, the majority of the pictures were obtained from publicly available networks. A total of 60 sets of pictures (415 pictures in all) were collected, including 4 sets of people, 17 sets of animals, 19 sets of plants, and 20 sets of scenes. Since it was difficult to collect pictures of the same person in different affective states from open networks, we attempted to use facial expression databases (e.g., the Japanese Female Facial Expression Database, Lyons, 2021) or screenshots of film and TV works as picture materials. However, due to the following four considerations, we ultimately decided to remove images of people (26 in total). Firstly, the purpose of developing these facial expression databases was to provide standard materials for inducing specific emotions. Therefore, using such images as emotional anchors is highly likely to induce affective changes in the subjects. Secondly, the performers in film and TV works are often celebrities, and the subjects’ preferences for these performers could influence the evaluation results. Thirdly, the emotional suggestiveness of people figures can be strong, and subjects are more susceptible to be influenced by social desirability during testing, leading to measurement bias. Fourthly, there are potential legal risks related to issues such as portrait rights.

To ensure consistent evaluation baselines for each HDRPS+ image set (i.e., having representative “neutral/medium” images for each dimension), we individually evaluated the valence (positive, neutral, negative) and arousal (high, medium, low) dimensions of all 389 images. This approach enables the identification of the valence baseline image (with a neutral valence) and arousal baseline image (with a neutral arousal) for each set of images.

At the operational level, we utilized OpenCV and Visual Studio 2015 to adjust 389 original images to 512*512 pixel JPG images. In addition, to facilitate comparison of images in the same set, images of the same evaluation dimension within the set were presented simultaneously using slides. During the process, 5 master’s and doctoral students with more than 2 years of emotion research experience successively evaluated the valence and arousal dimension of each set of pictures and were given a 5-min break between the two modules.

The agreement percentage (AP) refers to the percentage of evaluators who perceive a particular image to belong to a specific emotional category out of all evaluators (Wang and Chu, 2013; as shown in equation 1). In this study, an 80% agreement percentage was used as the threshold for judgement. Eventually, a total of 23 sets of pictures, comprising 10 sets of animals, 5 sets of plants, and 8 sets of scenes, were identified to possess both the valence and arousal baselines.

Note: AP refers to the agreement percentage; i represents the emotional category of the image (including positive valence, neutral valence, negative valence, high arousal, medium arousal, and low arousal); NEi denotes the number of evaluators who consider a certain picture to belong to the emotional category i; N represents the total number of evaluators.

As not every set among the 23 sets in Step 3 contains all categories under the arousal dimension (e.g., Scenes-1 lacks highly arousing images), and the number of pictures varied greatly between groups, we conducted supplementary picture acquisition. Except for two sets of images that were difficult to collect, the remaining 21 sets of images were moderately supplemented (10 images per set).

Following the supplemental acquisition, we obtained the initial pool for HDRPS+, which contains 9 sets of animals, 5 sets of plants, and 7 sets of scenes.

In this paper, we set up an available evaluation session of picture materials inspired by the development of textual scales. Twenty MBA students from Sichuan University participated in this experimental session, including 6 males and 14 females who confirmed that they were currently mentally and cognitively normal, without psychiatric diagnoses or use of psychotropic medications. As this phase was a small-scale professional evaluation rather than an effort to establish a population statistical sample, perfect gender balance was not strictly controlled.

To address potential dimensional contamination issues of evaluating valence and arousal simultaneously (Kurdi et al., 2017), we divided the experiment into two modules: valence evaluation and arousal evaluation. Prior to the experiment, the experimenter provided informed consent forms, instructions (including an introduction to the evaluation indexes and task descriptions), and scoring sheets. During the evaluation, participants judged the “availability” of each of the 21 picture sets—that is, whether the set met our predefined criteria of relevance, clarity, and specificity for both valence and arousal (see Table 2)—and then recorded a Yes/No response on the scoring sheet. Considering individual differences in evaluation speed, we did not impose restrictions on the playback speed of the images, but only required each subject to respond as quickly and accurately as possible based on their intuition.

The agreement percentage of 6 available evaluation indexes for each set of pictures was calculated according to Equation 1. Following the criteria proposed by Li (2014), sets of images were screened with a threshold of 60% agreement ratio. The results showed that a total of 15 sets of pictures had an agreement ratio higher than 60% for all 6 available indexes, including 8 sets of animals, 5 sets of plants, and 2 sets of scenes.

Based on Equations 6, 7, we finalized the valence and arousal labels of each image, and images with consistent affective category in step 2 and step 3 were retained. A total of 115 images remained for the valence dimension, and 116 images remained for the arousal dimension. In addition, we also analyzed the data by group (see Appendix 1), and found that the remaining all 14 groups of pictures, except for Animal-8, were able to reflect the complete affective space.

Note: Tv is the valence category of a certain image, T_A is the arousal category of a certain image.

Further analysis showed that the 14 newly collected sets were unevenly distributed—seven animal, five plant, and only two scene sets. To enrich the material and balance category counts, we supplemented the database with images from the original HDRPS database, whose pictures had already been judged using the same procedure but on a 9-point valence/arousal scale. To keep the granularity consistent across all materials, we recoded the 9-point HDRPS valence scores into five ordered categories (1–2 = very low, 3–4 = low, 5 = neutral, 6–7 = high, 8–9 = very high; see Table 3). After supplementation, the final database contained 20 thematic picture sets.