︎ Motivation

Artificial intelligence (AI) is advancing, offering personalized services to cater to individual preferences. Despite people's strong desire for consistent context, industries like fashion, interior, and graphics currently utilize AI for basic product analysis and recommendations.
Our goal is to evolve from recommendations to creation and tackle the challenge of constructing more effective feature-matching algorithms in the visual domain.

︎ Task Domain & Goal

We aim to develop a system generating a visual core element suitable for visual context.
In our project, the model leverages user interior design information to craft a collection of graphic artworks that blend seamlessly with the overall interior context.

To prove our concept, we conducted a controllability test.

Based on our model, we provide 9 pictures of interfaces according to the color and selection of artwork described earlier. The participants were required to score on a 10-point Likert scale how satisfying each stimulus was. The order was randomized. Subsequently, participants saw all stimuli and were asked to rank each. Finally, we ask why they rated applications low or high and comments on the application.

47 millennials joined the survey. Their average age was 25.40 years (SD = 3.73).

The appropriate level of controllability is important.
But, if it is impossible, it is better for AI to select everything rather than give vaguely. In other words, when only a few options must be displayed, giving users no choice may provide greater satisfaction.

Creating results based on user preferences leads to high satisfaction.
To meet personal preferences, use AI to collect data, allow users to adjust some parameters, or provide a wide range of results.

Should offer Preview 
Preview should be included; Users would like to see applied.

︎ Scalability of Concept

As a new way to create harmony in a visual-to-visual domain, we propose our method for generating visual results with extracted parameters from visual inputs.

• This concept requires much less time than AI-enabled tasks and can result in a more affordable core element. 
• It can be used without pre-stacked data is a significant benefit. 
• It may also be useful in cases other than interior design. We predicted that this concept would be potentially applicable in the following examples:

︎Take-Away Messages

Simpler solution based on user preference
We can provide aesthetic satisfaction by meeting users' needs to display artworks that match the interior with spatial and economic efficiency. Of course, dealing with this issue was not just a matter of harmony. In design, especially decorating the personal spaces, the most significant variable was users' preference, which was also clearly revealed in user tests. In this respect, even during our team's initial concept development stage, machine learning models with user data were naturally presented as the most appropriate solution. However, we used a simpler method based on the premise that the interior will basically reflect users' tastes and preferences.

Importance of interaction
Generating beautiful artwork is important, but interaction is the most important aspect of creating an application using this model. Giving users suitable controllability aids in the generation of results that users can be satisfied with, as well as changing the user's contentment with the model. To deliver the most pleasant user experience, additional user research and trials about controllability with more details should be conducted prior to producing actual apps with the model.