Research Goal
In this study, we wanted to understand how expert users use AI image-generation tools. We focused on their creative process and observing techniques that made them more effective at getting desirable results from AI image-generation tools.
Participant Profile
We recruited participants who had self-reported extensive experience using AI image-generation tools. We wanted them to have in-depth knowledge of techniques that could produce high-quality images. We purposely targeted participants from multiple sources to recruit the required level of expertise:
- Midjourney-community members: Midjourney is the most popular AI image-generation tool today. The Midjourney community has experienced users who serve as guides on the Midjourney Discord server. We reached out to them to take part in the study.
- AI-art enthusiasts on LinkedIn: Many AI-art creators advertise their work and experience on LinkedIn. We recruited creative professionals who showed high levels of expertise.
- Open-recruitment survey on social media: We shared our screener survey on NN/g’s social media and invited participation. The screener used critical incident questions like Tell me about a time you learned a new trick to improve the images you create with AI tools. This method helped us filter out inexperienced users.
We recruited 9 participants for the study.
Method
Contextual inquiry is well suited to understanding users in their context and sheds light on their techniques, strategies, and thought processes behind them. We asked participants to bring some projects they wanted to work on and then observed them attempt these during the session.
Session Structure
Each session was 2 hours long and was divided into two parts:
- Background interview (10-15 min)
- Observation and contextual interview (105-110 min)
Background Interview
We asked participants questions about their experience with AI image-generation tools and the tasks that they do day-to-day using them.
Observation and Contextual Interview
Most of each session involved observing participants generate images using the AI tool for a project of their choice. Participants brought a variety of image-generation tasks, based on their professions or hobbies:
- Creating art to put on the board of a board game the participant was designing
- Creating new furniture designs to be used as inspiration for designing 3D models of furniture for Second Life
- Creating Lego-tractor posters for a kid’s bedroom
- Creating a storyboard for a film
- Getting visual inspiration for a logo-design assignment
We also had a few backup tasks available if participants finished their tasks before the end of the two-hour session. We used a chatbot to help give us secondary-task ideas that were contextually appropriate and realistic based on what we knew about each participant’s role, goal, and experience (supplied in the screener).
Participants were asked to think aloud as they performed the actions. We asked probing questions when we observed a behavior we wanted to know more about. At the end of the interview, we asked remaining clarifying questions to ensure we understood their behaviors.
Conclusion
Overall, this approach allowed us to deep-dive into participants' typical AI image-generation workflows. We observed the different strategies they used to overcome the challenges they encountered in their day-to-day use of AI tools for image generation.
