Image generated with Nightcafe. Prompt: "A high-quality aerial photo of a colorful crowd of every race, size, and gender happily waving at the camera."
Our bodies are diverse, and image generative AI can effectively capture that.
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 can be used to manipulate the size of any manipulated visual stimuli. To illustrate, consider the work of Lou, Tse, and Lwin (2019), who feature an average and a plus-sized model as experimental stimuli. In line with this type of manipulation, Figure W4 (p. 32), which is originally shown in the manuscript and included here for convenience, illustrates the body size manipulation of a young woman across three sizes (thin, average, plus-sized).
As in the case of Lou et al. (2019) or other articles which investigate body size (Joo et al. 2021; Shoenberger et al. 2020), a full-body (or nearly so) picture is usually desirable. As noted in our Aspect, Style, and Details section, therefore, a somewhat elongated aspect ratio might be suitable to construct Stimulus A. Moreover, a pose that clearly illustrates the person’s body is also desirable. Hence, to construct Stimulus A, one possibility would be to construct the initial text prompt including terms related to body posing and use it as input in an image generative AI tool to produce a set of candidates for Stimulus A, refining them afterwards - however, without more advanced tools (e.g., ControlNet), refining a specific full-body pose can be a lengthy process.
To speed up this process, however, if a suitable picture illustrating a person and pose is available (e.g., a stock photo, or a picture taken by the researcher), then the picture can serve as an initial template (it is possible to find a suitable template with a different aspect ratio from the researcher’s desired aspect ratio. If this is the case, the correct aspect ratio can be induced by cropping, resizing, or taking a screenshot.). However, the template picture cannot be directly used as Stimulus A, because the characteristic “look” induced by the specific AI generative tool used must be applied consistently across stimuli. Hence, using RAISE, we construct Stimulus A using the initial text prompt, the template picture, and a relatively large image strength parameter (in this application, 80%), such that Stimulus A is almost identical to the template while allowing for its style to be recast consistent with the AI tool’s.
As in the case of Lou et al. (2019) or other articles which investigate body size (Joo et al. 2021; Shoenberger et al. 2020), a full-body (or nearly so) picture is usually desirable. As noted in our Aspect, Style, and Details section, therefore, a somewhat elongated aspect ratio might be suitable to construct Stimulus A. Moreover, a pose that clearly illustrates the person’s body is also desirable. Hence, to construct Stimulus A, one possibility would be to construct the initial text prompt including terms related to body posing and use it as input in an image generative AI tool to produce a set of candidates for Stimulus A, refining them afterwards - however, without more advanced tools (e.g., ControlNet), refining a specific full-body pose can be a lengthy process.
To speed up this process, however, if a suitable picture illustrating a person and pose is available (e.g., a stock photo, or a picture taken by the researcher), then the picture can serve as an initial template (it is possible to find a suitable template with a different aspect ratio from the researcher’s desired aspect ratio. If this is the case, the correct aspect ratio can be induced by cropping, resizing, or taking a screenshot.). However, the template picture cannot be directly used as Stimulus A, because the characteristic “look” induced by the specific AI generative tool used must be applied consistently across stimuli. Hence, using RAISE, we construct Stimulus A using the initial text prompt, the template picture, and a relatively large image strength parameter (in this application, 80%), such that Stimulus A is almost identical to the template while allowing for its style to be recast consistent with the AI tool’s.
The text prompt for Stimulus A (average size, which is shown above) is as follows:
A young woman wearing jeans and a white cropped tank top, her entire body shows, facing forward standing still. The background is white and blurred
After Stimulus A has been constructed, the process unfolds sequentially, as in other examples. However, in this case, manipulating body size relies as much on what the text prompt indicates the generated image should look like as much as on what it should not look like. To account for features to avoid in image generative AI, most tools easily allow to include negative text prompts, which the generative AI tool would then avoid. Hence, when conditionally generating a thinner-than-average model, the prompt might add words such as “thin” and “slender,” while negative prompts might include “plus-sized” and “chubby.” Consequently, the structure of text prompts and negative terms is as follows, building the Thin (B) stimulus first,
and followed by the Plus-sized (C) stimulus:
Below is a handy summary sheet illustrating the entire process you learned today:
It must also be noted that in the example discussed here, we defined the average model as Stimulus A, and the thinner and plus-sized models as Stimulus B and C, respectively. More generally, when using RAISE to manipulate body weight, Stimulus A should always serve as the basis from which all other subsequent body sizes are conditionally generated.
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.