Abstract
From Paleolithic cave paintings to Impressionism, human painting has evolved to depict increasingly complex and detailed scenes, conveying more nuanced messages. This paper attempts to emerge this artistic capability by simulating the evolutionary pressures that enhance visual communication efficiency. Specifically, we present a model with a stroke branch and a palette branch that together simulate human-like painting. The palette branch learns a limited colour palette, while the stroke branch parameterises each stroke using B'ezier curves to render an image, subsequently evaluated by a high-level recognition module. We quantify the efficiency of visual communication by measuring the recognition accuracy achieved with machine vision. The model then optimises the control points and colour choices for each stroke to maximise recognition accuracy with minimal strokes and colours. Experimental results show that our model achieves superior performance in high-level recognition tasks, delivering artistic expression and aesthetic appeal, especially in abstract sketches. Additionally, our approach shows promise as an efficient bit-level image compression technique, outperforming traditional methods.
Introduction
Why did human painting evolve the way it did? Most modern AI art generators act as mimics, replicating styles from vast datasets. However, this research explores a more fundamental question: can painting ability emerge autonomously from the pressure of "visual communication efficiency"? Historically, painting served as a tool for information transfer—identifying prey or recording events.
By treating painting as an optimization problem where the goal is to be recognizable with minimal resources (strokes and colors), the authors demonstrate that AI can independently develop techniques that mirror the trajectory of human art history.
Methodology: Recognition-Driven Evolution
The proposed model utilizes a dual-branch architecture to simulate the painting process:
- Stroke Branch: Employs differentiable rasterization to parameterize strokes as B'ezier curves. The model optimizes control points, thickness, and placement.
- Palette Branch: Learns a restricted, representative color palette for each image by localizing coordinates within the RGB space.
- Evolutionary Pressure: The "fitness" of a painting is determined by a high-level recognition module (CLIP). The model is rewarded when a pre-trained vision model correctly identifies the subject.
- Complexity Estimation: Images are categorized by complexity (object density, size, overlap) to study evolution across different scenarios.
The optimization balances semantic recognition, structural geometry, and color perception.
Convergence with History
A striking finding of this research is that the model, without being shown historical art, produced paintings stylistically similar to ancient Chinese and Medieval European works. This suggests that historical styles may have been optimal solutions for communication under physical constraints like limited pigments and manual labor.
The model independently "discovered" the Five-Value System, a fundamental art principle that divides light and shadow into five zones to convey 3D volume on a 2D surface. Furthermore, the researchers observed that structural "shorthand" (black-and-white sketches) often yielded higher recognition accuracy than colored versions at low stroke counts, emphasizing that structure is more vital for communication than decoration.
Key Results and Efficiency
The model demonstrates significant improvements over existing methods:
- Recognition: With only 32 strokes, it achieved a Top-3 recognition accuracy of 50.53%, surpassing CLIPasso.
- Human-Level Skill: Professional artists estimated the model's mastery of shading allows it to outperform approximately 30% of untrained humans with two weeks of practice.
- Image Compression: The model functions as an extreme compression technique, achieving 74.51% recognition accuracy at just 0.059 bits per pixel (bpp), significantly more efficient than standard JPEG formats for recognition tasks.
Limitations and Future Outlook
Despite its successes, the model currently lacks a sophisticated understanding of linear perspective, often rendering 3D structures in a flattened, pre-Renaissance manner. It also struggles with specific Eastern techniques like varying line density to imply water depth. Future work may apply this evolutionary approach to other creative domains or explore its potential in privacy-preserving vision, where sketches convey essential information without revealing sensitive details.
Summary updated