Image generated with Nightcafe. Prompt: "A group of many cute humorous robots trying to fix a machine."
Practice makes perfect - and, sometimes, restarting!
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!
As outlined in the RAISE methodology overview, using image generative AI to construct stimuli might not always be satisfactory. Hence, for any stimuli, we recommend to always generate multiple candidate images at a time. This will result in a set of candidate stimuli among which the best stimulus would be retained and possibly refined. We discuss three types of refinements, and how to address them.
Cosmetic issues
A candidate stimulus might be close to acceptable but have cosmetic issues –for example, incorrect fingers, colors, background, or other visual errors or imperfections. We recommend one of three solutions (or a combination of them) to this effect:
- One solution is to edit the original prompt by changing, removing, or adding words to more carefully represent the desired stimulus. For example, if a stimulus depicting a woman is acceptable, but shorter hair is required, one might add “She has short hair” to the prompt. In our applications, we find that, when generating Stimulus A, editing the original prompt is usually sufficient to alleviate cosmetic issues.
- A second solution is to generate variations: very similar versions of the same stimulus with relatively high image strength, such as in our body movement manipulation examples. Among the resulting stimuli, the latter, pertaining to the higher movement condition, shows differently colored sleeves and a design on the shorts. Generating such variations might potentially solve these minor cosmetic inconsistencies. Usually, image generative AI tools can automatically generate variations, with automatic settings that can be manually adjusted. Note that different AI tools refer to this approach in different ways: “variations”, “remixing”, "evolving," “generate similar”, and so forth.
- A third solution is to directly use the candidate stimulus’ seed. A seed is a random number used by the image generative AI tool as a starting value to generate images. This is usually provided once an image is generated, and using the same seed, while modifying a text prompt, will result in an image of a similar style, but with modified features as desired.
- An approach to be used in combination with any of the above three solutions is to attempt to correct imperfections through negative prompts. For example, suppose a candidate stimulus is acceptable, but for the presence of a tie. A researcher might then generate variations adding the negative prompt “tie” to generate updated stimuli that might be similar to the candidate but without the tie.
It must be noted that a final approach is to edit a candidate stimulus outside of the image generative AI tool. For instance, removing a logo or undesirable detail might be retouched with Microsoft Photos, which is freely available on any Windows computer; the Adobe Creative Suite; GIMP, an open-source photo editor; or affordable AI-powered tools like Photoroom. In our view, a simple photo editor such as Microsoft Photos, or Photoroom, are sufficient for most experimental stimuli development tasks with minor cosmetic issues.
Style issues
Correcting artistic style can be straightforwardly accomplished through our recommendations about manipulating artistic style within a text prompt, using an image generative AI tool’s preset styles, or both. Since this is a correction, the image generative AI tool must receive the candidate stimulus’ image as input as well, or, potentially, the candidate stimulus’ seed.
Major errors
Major errors involve generating an image that is not consistent with the original stimuli details desired – for instance, an incorrect aspect ratio, angle, background, and so forth. If this is the case, researchers should respecify the initial stimuli details, or reconstruct their text prompt and review the image generative AI tool’s settings.
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.