This paper provides a much-needed conceptual cleanup of "Computational Aesthetics," a field that has grown increasingly messy as AI-generated content (AIGC) and deep learning have exploded. The authors propose a new hierarchical framework to distinguish between simply predicting a "beauty score" and truly understanding or generating art.
TL;DR
The authors propose a two-level framework for the field: Special Computational Aesthetics (SCA), which focuses on measuring and modeling the human aesthetic process, and General Computational Aesthetics (GCA), which extends this to include model-guided artistic generation. Crucially, they argue that "AI Aesthetics" is a methodological subset (a way of doing research) rather than a distinct category of aesthetic tasks.
What Problem Does This Paper Address?
As AI's involvement in art has deepened, the term "computational aesthetics" has become an overstretched umbrella. It is often used interchangeably with AI art, AIGC, or neuroaesthetics, leading to confusion about research goals.
Without a clear taxonomy, the field risks being narrowed down to just "image aesthetic quality assessment" (predicting if a photo is "good") or expanded so broadly that it loses its scientific rigor. The authors tackle this by defining the internal structure of the field and its boundaries with neighboring disciplines like affective computing and computational creativity.
Methods
This is primarily a theoretical and review paper that employs conceptual analysis. The authors synthesize decades of research—from Birkhoff’s 1928 mathematical formulas for "aesthetic measure" to modern deep learning agents—to build a new taxonomy.
Key methodological contributions include:
- The SCA/GCA Hierarchy: Distinguishing the "core" tasks of perception and understanding from the "extension" of generation.
- The Seven-Layer Task Model: A functional expansion of image aesthetic assessment, moving from simple numerical classification to complex linguistic guidance.
- Boundary Mapping: Using "research objects" and "objectives" to differentiate computational aesthetics from "aesthetic computing" (which is about making algorithms beautiful, rather than using algorithms to study beauty).
Datasets
While the paper is a review, it highlights several foundational and specialized datasets that have driven the field's evolution:
- AVA (2012): A large-scale database for aesthetic visual analysis that moved the field toward deep learning.
- APDD: Focused on paintings and drawings.
- VADB: A large-scale video aesthetic database with professional annotations.
- MADB: A music aesthetics dataset.
Key Insights and Ideas
- Perception vs. Understanding: The authors distinguish between "Aesthetic Perception" (identifying attributes and predicting scores) and "Aesthetic Understanding" (explaining how those judgments are formed). A model that predicts a high score accurately doesn't necessarily "understand" aesthetics if its internal logic doesn't map to human cognitive processes.
- The "Substantive Guidance" Rule: Generation (AIGC/AI Art) only falls under General Computational Aesthetics if an aesthetic model substantively constrains or optimizes the output. Simply using a GAN to make an image isn't "computational aesthetics" unless the system is explicitly guided by aesthetic principles.
- AI is a Method, Not a Task: They clarify that "AI Aesthetics" is simply computational aesthetics that happens to use AI. This prevents the field from being fragmented by technology rather than by research questions.
- The Computability of Emotion: Drawing on philosopher Feng Peng, they argue that the computability of aesthetic experience is essentially the computability of emotion, linking the field directly to affective computing.
Key Results
The authors successfully map the evolution of the field into a Seven-Layer Task Model for image aesthetics:
- Aesthetic Classification: (e.g., beautiful vs. ugly)
- Aesthetic Rating: (continuous numerical values)
- Aesthetic Distribution: (capturing group agreement/disagreement)
- Aesthetic Attributes: (composition, color, lighting)
- Aesthetic Commentary: (linguistic "art criticism")
- Aesthetic Question Answering: (interactive communication)
- Aesthetic Guidance: (providing specific suggestions for improvement)
Limitations and Future Work
- The "Black Box" Problem: The authors acknowledge that while deep learning models are great at predicting human scores, their internal processes often remain a black box. Future work must focus on "Aesthetic Process Modeling" to ensure computational models actually reflect human psychological and neural mechanisms.
- Cross-Media Challenges: Most research is currently siloed in images. Integrating "aesthetic intelligence" across photography, painting, video, and music remains a significant hurdle.
- Personalization: Moving beyond "group mean scores" to understand individual and cultural aesthetic preferences is a critical next step.
- Terminology: The authors note that cross-language (English/Chinese) terminology often muddies the waters, requiring researchers to be more precise about whether they are discussing a "style" or a "scientific field."
Summary updated