Image-based AI for Industrial Design

How aligning semantics among desingers helps them use AI tools more effectively

Master thesis (2024)

Authors

B.H.P. Theijse Industrial Design Engineering

Contributors

Vera van der Burg Creative Processes - IDE (supervisor 1)
J.D. Lomas Design Aesthetics - IDE (supervisor 1)
Stef de Groot (coach)
Faculty
Industrial Design Engineering
AI Design Semantics Industrial Design Labelling LoRA Low-Rank Adaptation Image-based AI
More Info
expand_more

Published Date

29-04-2024

Language

English

Reuse Rights

Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.

Faculty

Industrial Design Engineering

Abstract

This research explores the opportunities and limitations of image-based AI tools for industrial design. In collaboration with designers at Royal Gazelle, several tools were tested. AI tools can broaden and speed up the creative processes but also lack control for the user. AI models and designers use different terms to articulate perceptions and control output. This misalignment stems from misalignment among humans which translated to the training datasets of AI.

This report proposes co-creative image labelling sessions to align perception and articulation of perception in design teams. The aligned perception creates a strong vibe. Collectively labelling an image dataset increased alignment of the perception of images by 30.8% when comparing post- to pre-session questionnaire results. Additionally, participants used more of the same vocabulary after the session.

Training a low-rank adaptation model on these labelled datasets could externalise tacit design knowledge and can be used for further development on AI models that align with designers.