Banner picture generated using the RAISE methodology in Dreamstudio. Prompt strategy is discussed here!
Learn how to create visual experimental stimuli quickly and effectively with GenAI.
This is a Web version of the Companion Appendix for RAISE: A New Method to Develop Experimental Stimuli for Advertising Research with Image Generative Artificial Intelligence, developed by César Zamudio, Jamie Grigsby, and Meg Michelsen, and published in the Journal of Advertising. Questions? Feel free to contact us!
THE RAISE METHODOLOGY
The goal of the RAISE (Rapid Artificial Intelligence Stimuli for Experiments) methodology is to develop rigorous image generative AI stimuli while maximizing equity and accessibility. As such, we developed RAISE assuming that the researcher will not require any programming skills. Furthermore, the methodology relies on software and tools that are commercially available and that are low-cost (either free, freemium, or requiring minimal funds).
The AI tool must also accommodate for image-to-image generation and image strength. The image generative AI tool chosen for this manuscript, as it satisfies these requirements while minimizing costs, is DreamStudio. However, other commercially available tools, such as Midjourney, BlueWillow/LimeWire, Nightcafe, or Dreamlike could potentially be used as well. Advanced users can also choose to install advanced image generative AI distributions on their computer – however, a graphics processing unit (GPU) of at least 8 – 12 VRAM is recommended (e.g., NVIDIA GeForce RTX 3060). If you do this, keep in mind that laptop GPUs with a similar name (e.g., NVIDIA GeForce RTX 3060) should be expected to be slower, since desktop GPUs can outperform laptop GPUs by 20% to 30% according to informal analyses.
Aside from a suitable image generative AI tool, two other resources might be relevant. Note that, while RAISE does not accommodate for the generation of text, this can be straightforwardly added after the stimuli are generated with easy-to-use software such as Microsoft PowerPoint, Photoshop, or GIMP, or with Web tools (an example involving using experimental stimuli generated with RAISE and then embedding them on a social media stimulus constructed online as shown in our example). As such, optionally, access to these resources might be necessary. Finally, for certain tasks, a smartphone or standalone camera might be useful, although not necessary.
The AI tool must also accommodate for image-to-image generation and image strength. The image generative AI tool chosen for this manuscript, as it satisfies these requirements while minimizing costs, is DreamStudio. However, other commercially available tools, such as Midjourney, BlueWillow/LimeWire, Nightcafe, or Dreamlike could potentially be used as well. Advanced users can also choose to install advanced image generative AI distributions on their computer – however, a graphics processing unit (GPU) of at least 8 – 12 VRAM is recommended (e.g., NVIDIA GeForce RTX 3060). If you do this, keep in mind that laptop GPUs with a similar name (e.g., NVIDIA GeForce RTX 3060) should be expected to be slower, since desktop GPUs can outperform laptop GPUs by 20% to 30% according to informal analyses.
Aside from a suitable image generative AI tool, two other resources might be relevant. Note that, while RAISE does not accommodate for the generation of text, this can be straightforwardly added after the stimuli are generated with easy-to-use software such as Microsoft PowerPoint, Photoshop, or GIMP, or with Web tools (an example involving using experimental stimuli generated with RAISE and then embedding them on a social media stimulus constructed online as shown in our example). As such, optionally, access to these resources might be necessary. Finally, for certain tasks, a smartphone or standalone camera might be useful, although not necessary.
Methodology overview
Our conditionally accepted Journal of Advertising article briefly introduced the RAISE methodology, a novel approach to construct experimental stimuli in Marketing using image generative AI. For your reference, an overview of RAISE is shown below. In this section, we'll discuss each step of the RAISE methodology.
Step 1: Determining the feasibility of AI for your research goal
Consistent with our research paper, the RAISE methodology is intended for experiments that will rely on visual stimuli, and in which the main visual element that experimental subjects will be exposed to must be manipulated along some attribute (e.g., gender, race, shape), but the rest of the stimulus must be held constant. For example, Liljedal, Berg, and Dahlen (2020) investigate the role that gender plays when portraying non-stereotypical occupations in ads and utilize stock photos of a male and female firefighters; while the pictures are not identical, the main visual element manipulated (firefighters of different genders) is clearly shown and systematically varied, while the rest of the stimuli details (e.g., the firefighters’ uniforms, background, and so forth) are reasonably similar. Still, as noted in the paper's general discussion ("When can AI be used? A precision continuum", and Figure 10), limitations in image generative AI technology and the researcher’s desired stimuli precision, will also determine how effective applying RAISE might be for a given research project.
Images purchased from Shutterstock.
As can be seen, Liljedal, Berg, and Dahlen (2020) used two Shutterstock images to represent a male and a female firefighter. RAISE can accomplish this using image generative AI quickly and cost-effectively.
Step 2: Determining initial stimuli details
Once feasibility has been established, the details of the stimuli to develop must be decided.
- First is the context, which is determined by the research question and industry relevant to the researcher.
- Next are several stimuli details pertaining to the image generative AI method.
- First, aspect ratio, that is, the shape of the resulting stimuli, which can be square (1:1), landscape (7:4), portrait (4:7), among others.
- Additionally, the main visual element shown in the stimuli must be chosen. For instance, in studying the role of sex appeal in advertising, Liang, Wu, Su, and Jin (2023) present a model wearing either a bikini or fully dressed.
- The stimuli’s background. Researchers must determine if a plain background is appropriate (such as Liang, Wu, Su, and Jin (2023), who use a solid white background); otherwise, it must be decided what the background will display.
- In addition, the artistic style of the stimuli (e.g., Wang et al., 2023), which governs the overall look of the stimulus, must be determined.
- Finally, camera details might also be relevant – for example, lighting effects, zoom, type of lens, etc., which can be accomplished by including these details into the image generative AI’s text prompts. See Section 13, “Aspect ratio, artistic style, and camera details” (p. 19) for an in-depth discussion featuring various examples.
- Note that the background of the stimulus could be the visual element to manipulate, which is discussed in our Technical Details section.
Step 3: Drafting the initial text prompt
Initial text prompt template. The next step is to construct a text prompt that a generative AI tool can use to generate an array of potential stimuli, among which the highest-quality stimulus would be retained. While text prompts might vary in their effectiveness depending on the tool used and can also grow large and very detailed to accommodate a wide range of visual details, we propose a concise and standardized prompt template below. Elements to systematically manipulate in the researcher’s experiment are shown bolded, elements to hold constant are in parentheses, and optional elements are in brackets.
A (object characteristics) object characteristic to manipulate (situation characteristics) [artistic style]. [The background shows (background details)]. [The background is blurred]. [Camera details] [Aspect ratio].
Negative terms: [negative prompts
Negative terms: [negative prompts
To illustrate prompt construction, suppose the goal is to conduct a single-factor, two-condition experiment (Gender: Man, Woman) depicting young service workers, a white man and a white woman. Then, the only attribute to vary is gender, and the prompts might be written as follows –note that, here, artistic style and camera details were not used; additionally, since aspect ratio was also not specified, the default (1:1, i.e., square) would apply:
A (young white) man/woman (dressed in a white shirt and orange apron greeting someone at a restaurant). [The background shows (a busy restaurant). [The background is blurred].
Negative terms: None
Negative terms: None
In any prompt using this technique, the only elements to systematically manipulate are those related to the researcher’s experimental conditions. If, in a further study, the researcher would like to add the role of race, in a two-factor, two-condition experiment (Gender: Man, Woman; Race: White, Black), the prompt might be extended to:
A (young) white/black man/woman (dressed in a white shirt and orange apron greeting someone at a restaurant). The background shows (a busy restaurant)]. [The background is blurred].
Negative terms: None
Negative terms: None
We call this prompt the initial prompt. Note that, in this initial prompt, we do not use any negative terms. These terms, which the image generative AI tool would attempt to avoid instead of to encourage, are an important part of the RAISE methodology, and are discussed in further scenarios (e.g., body size, fine body movement).
Step 4: Generate Stimulus A
To streamline the discussion of stimuli generation, we will focus on the single-factor, two-condition (Gender: Man, Woman) experimental design example in the services context. As noted in the manuscript, we will refer to the experimental stimuli as A and B for convenience. However, stimuli for any number of experimental conditions can be constructed.
The initial stimuli details are the following:
The initial stimuli details are the following:
|
Experimental condition: Gender (Man/Woman)
Context: Services Marketing (restaurant) Object to manipulate: Waiter Aspect ratio: 1:1 (square) |
Background: Restaurant, blurred
Artistic style: None Camera details: None |
To generate Stimulus A, the initial prompt (excluding the illustrative parentheses and brackets used earlier) can be input directly in an image generative AI tool:
Prompt for Stimulus A: "A young white man dressed in a white shirt and orange apron greeting someone at a restaurant. The background shows a busy restaurant. The background is blurred."
The results are shown above. The tool will output several different candidate stimuli. Generally, these candidate stimuli will require several refinements, depending on whether the candidate stimuli exhibit cosmetic issues, style issues, or major issues. How to refine these candidate stimuli is discussed in the Technical Details section. Therefore, in what follows, we assume that the researcher has generated several candidate stimuli, and refined them, such that a final Stimulus A has been selected and retained.
The initial stimuli details were easily accounted for –the aspect ratio is the default for most image generative AI tools, and the blurred background detail was established directly in the initial prompt. Assuming refinements are complete, the male waiter shown above is retained as Stimulus A. The next step is to construct Stimulus B.
The initial stimuli details were easily accounted for –the aspect ratio is the default for most image generative AI tools, and the blurred background detail was established directly in the initial prompt. Assuming refinements are complete, the male waiter shown above is retained as Stimulus A. The next step is to construct Stimulus B.
Step 5: Generate Stimulus B as a function of Stimulus A
Before proceeding, it is crucial to discuss how the formation of further stimuli is conducted in the RAISE methodology. The key idea is to hold Stimulus A constant to produce a Stimulus B in which visual features except for the visual element to be manipulated are held virtually constant. This is necessary to maintain experimental rigor and prevent confounds (Geuens and De Pelsmacker, 2017). As noted, since we rely on image generative AI tools which take text as well as images as input, our approach relies on this ability by generating Stimulus A, and then generating candidates for Stimulus B using both the retained image of Stimulus A and the new prompt for Stimulus B as input. In this way, the generative AI tool can construct Stimulus B conditional on the image of Stimulus A, and the additional text prompt for Stimulus B will then alter Stimulus A accordingly.
Crucially, the tool must accommodate two important image generative AI features.
Crucially, the tool must accommodate two important image generative AI features.
- The first is image-to-image generation. This means that, to generate a stimulus, the AI tool can take as input an existing image in addition to a text prompt, to generate a new image as a function of both. This allows for Stimulus A to serve as a reference to construct Stimulus B, alongside an updated text prompt reflecting experimental condition B.
- The second is image strength adjustment. This means that, when generating a stimulus, the AI tool user can specify how “similar” an output image will be relative to an input image. This feature is necessary to construct Stimulus B, such that it is different enough from Stimulus B to capture the new experimental condition, but similar enough such that the major difference is the object manipulated. The RAISE methodology depends on both features – the chosen tool to develop the examples in this article, DreamStudio, includes both.
The picture above shows the Dreamstudio window that will be used to generate Stimulus B. The starred controls, Image and Image strength, indicate the fields in which to add the image of Stimulus A, and select a suitable image strength. Note that while manipulating image strength is a feature present in multiple tools, several offer only coarse control – for instance, instead of a continuum between 0% and 100%, a tool might instead simply allow to specify if the image or the prompt has “priority” or “importance.” In our experience, the latter degree of specificity is not enough to produce successful manipulations, and we do not recommend using tools that provide coarse image strength control.
To continue, the text prompt for Stimulus B would be:
To continue, the text prompt for Stimulus B would be:
Prompt for Stimulus B: "A young white woman dressed in a white shirt and orange apron greeting someone at a restaurant. The background shows a busy restaurant. The background is blurred."
The image file of Stimulus A must be uploaded into the AI tool, and a suitable image strength needs to be specified, as discussed earlier. Image strength lies between 0% (the image is ignored) or 100% (the image is perfectly replicated). In our multiple applications of the RAISE method, our experience indicates that there is not a consistent rule regarding image strength: it must be determined through experimentation in an ad-hoc manner, based on the researcher’s stimuli and context.
In this example, an adequate image strength was 35%, allowing for a Stimulus B that is sufficiently different from A to be recognizable as a woman, but with their major physical features being nearly identical (e.g., apron details, shirt sleeves, bowtie, background, and even the part in their hair). We generated a satisfactory Stimulus B – by comparing both stimuli, it is evident that gender has been manipulated, but the stimuli are otherwise almost identical.
Below is a handy summary sheet detailing what you learned today:
In this example, an adequate image strength was 35%, allowing for a Stimulus B that is sufficiently different from A to be recognizable as a woman, but with their major physical features being nearly identical (e.g., apron details, shirt sleeves, bowtie, background, and even the part in their hair). We generated a satisfactory Stimulus B – by comparing both stimuli, it is evident that gender has been manipulated, but the stimuli are otherwise almost identical.
Below is a handy summary sheet detailing what you learned today:
Summary card to create two waiters of different genders: man and woman. Feel free to save and share!
Once a satisfactory Stimulus B is generated, any number of Stimuli can be created based on Stimulus A. Therefore, RAISE can be used to create any number of subsequent visual experimental stimuli as needed. As stimuli are created, we recommend keeping track of as many details used when generating each stimulus, such as text prompts, image strength values, AI seeds, negative prompt words, the use of any image to assist the AI tool, etc. – these details are important if replication is needed and are discussed in the Technical Details section.
Now you have the know-how to quickly and cost-effectively generate experimental stimuli as in this page's banner!
Now you have the know-how to quickly and cost-effectively generate experimental stimuli as in this page's banner!
All images generated using the RAISE method with Dreamstudio.
Ready to learn more?
Feel free to browse our gallery of examples with full tutorials, and the technical details section to learn the finer points of generating visual experimental stimuli using the RAISE methodology.