An Abstraction Technique-Based Approach to Creativity Generation for Inspiring Painting Sketches.

Many novice artists struggle to generate ideas when facing a blank canvas, and existing generative systems often offer inadequate support. Although data-driven models can generate visually striking images, they are frequently complex, opaque, or misaligned with users' intentions and thoughts. This thesis addresses this gap by investigating how abstraction techniques can support creativity in the early stages of artistic sketching.The research proposes a rule-based framework that structures the creative process into three phases: syntactic abstraction for selecting basic visual elements, semantic abstraction for arranging these elements into meaningful sequences, and pragmatic abstraction for configuring the spatial hierarchy to convey atmosphere. Each phase operates through a process of abstraction and instantiation, which involves broad concepts such as intentions, themes, and atmosphere being abstracted into formal attributes and then instantiated into specific combinations. A prototype system was implemented to demonstrate how rules based on geometric combinations and conceptual associations can generate sketch-like outputs that stimulate thought, rather than replacing human creativity. In comparison to deep learning approaches, the rule-based method offers transparency, interpretability, and adaptability, making the process and outcomes adjustable and predictable.The main contributions comprise three aspects: i) the formulation of a novel three-phase framework that formalises abstraction-instantiation cycles for creative computing; ii) the development of an operational set of rules—mapping intentions to polygon combinations, themes to horizontal layout heuristics, and atmospheres to vertical hierarchy via an interpretable score—formally specified and demonstrated for feasibility and operability through a working prototype; and iii) a reflective discussion of the framework’s implications, reporting preliminary validation and outlining directions for future empirical study.In summary, the thesis demonstrates that abstraction and instantiation can serve as an effective mechanism for inspiring artistic creativity, especially for novice artists. While validation is currently confined to self-testing, the findings indicate promising opportunities for future multi-user and cross-cultural studies. Furthermore, future work could also explore combining the proposed rule-based framework with existing deep learning approaches to develop controllable generation technologies. By using deep learning outputs as benchmarks and comparing them with controllable generation results, it would be possible to gain deeper insights into the distinctive value of the rules introduced in this thesis.