In traditional machine learning, success is often binary: a classifier either identifies a cat or it does not. However, in the realm of computational creativity—spanning poetry, music, and visual design—there is no single "correct" answer. As Dinh-Viet-Toan Le et al. observe, the absence of standardized benchmarks for subjective outputs prevents consistent comparison across different generative architectures. This challenge, often termed the "aesthetic quantification challenge," stems from the inherent difficulty in aligning formal computational metrics with the nuances of human perception.
The Failure of Objective Ground Truth
The primary hurdle in evaluating creative systems is the lack of objective ground truth. In tasks like image captioning, metrics such as BLEU measure word-level overlap between a model's output and a reference. Yet, as Jianlong Fu et al. argue, these metrics are ineffective for art because they prioritize literal similarity over thematic spirit, novelty, and emotional impact. Similarly, Rui Yan notes that reference-based similarity frequently fails to capture the quality of creative outputs due to the open-ended nature of artistic expression.
This subjectivity extends to the data itself. When researchers like Alexey Tikhonov et al. build systems to generate "creative" narratives, the definition of "good" is often tethered to small, subjectively labeled datasets that reflect the specific tastes and biases of the creators rather than a universal standard.
Decomposing Beauty into Computable Pillars
To move beyond "I know it when I see it," researchers have attempted to decompose abstract aesthetic value into measurable components. Disha Shrivastava et al. propose four pillars to quantify creativity: Value, Novelty, Influence, and Unexpectedness. By breaking down artifacts (such as movies) into these dimensions, they found that computational metrics could align more closely with professional critic scores.
In visual domains, this decomposition involves translating artistic principles into geometry. Yan Ke et al. designed high-level features to mimic human perception of photographic quality, encoding concepts like "simplicity" and "realism" into mathematical formulas. More recently, LAN YANG introduced the Geometry-Aware Classification Layer (GACL), which uses the magnitude of feature vectors as a proxy for quality. The intuition is that "better" creative artifacts (like well-drawn sketches) are geometrically further from class decision boundaries, providing a learnable metric that requires no human-annotated quality scores.
The Observer and the Creative Process
A unique facet of computational aesthetics is that the perceived value of an artifact is often inseparable from the perceived intent of its creator. Amílcar Cardoso et al. highlight that evaluating creative systems is fundamentally different from task-oriented AI because creativity depends on the relationship between the creator and the observer.
This "story" of creation matters. Stephen J. McGregor et al. investigated whether transparency about an AI's statistical processes changes how humans value its poetry. They found that there is no established cultural consensus on the criteria for machine creativity, leading to high variability among human raters. This variability suggests that aesthetic evaluation is not just a measurement of the output, but an assessment of the process and the context in which it was generated.
Bridging the Semantic-Aesthetic Gap
Recent work has moved toward multimodal frameworks that attempt to explain why something feels aesthetic. Luwei Xiao et al. developed the Chimera framework to generate semantic and aesthetic rationales for sentiment. However, the authors found that generating rationales for abstract aesthetic impressions is significantly harder than for factual content, as the subjectivity leads to lower ground-truth quality and reduced factual clarity.
Similarly, in specialized domains like web design or fashion, the gap remains wide. Tao Zhang et al. demonstrated that while models can assess objective criteria like professionalism, they struggle with aesthetic design due to the high variability in human perception. As Wang-Cheng Kang et al. point out in the context of fashion, while a model can optimize for a user's predicted preference, directly assessing the "desirability" of a novel design remains an open challenge.
Ultimately, the field is shifting from seeking a single "beauty" formula to developing specialized models—such as Nir Diamant et al.’s Beholder-GAN or Xin Jin et al.’s ILGNet—that learn to approximate specific human-centric values within constrained domains. The goal is not to solve subjectivity, but to build systems that can navigate it.