CAN: Creative Adversarial Networks, Generating "Art" by Learning About Styles and Deviating from Style Norms (1706.07068v1)

Abstract: We propose a new system for generating art. The system generates art by looking at art and learning about style; and becomes creative by increasing the arousal potential of the generated art by deviating from the learned styles. We build over Generative Adversarial Networks (GAN), which have shown the ability to learn to generate novel images simulating a given distribution. We argue that such networks are limited in their ability to generate creative products in their original design. We propose modifications to its objective to make it capable of generating creative art by maximizing deviation from established styles and minimizing deviation from art distribution. We conducted experiments to compare the response of human subjects to the generated art with their response to art created by artists. The results show that human subjects could not distinguish art generated by the proposed system from art generated by contemporary artists and shown in top art fairs. Human subjects even rated the generated images higher on various scales.

• The paper presents CAN, which adapts GAN objectives to balance valid art production with intentional stylistic deviation.
• It employs dual discriminator pressures that enforce traditional art criteria while promoting novelty and ambiguity in creations.
• The study demonstrates CAN's potential in producing art comparable to human works, highlighting its implications for AI-driven creative industries.

A Critical Examination of Creative Adversarial Networks for Art Generation

The paper introduces the Creative Adversarial Network (CAN), a novel method for art generation utilizing machine learning techniques, particularly focusing on improving the creative aspects of the generated outputs. Building upon Generative Adversarial Networks (GANs), the CAN aims to move beyond merely replicating learned patterns from training data by intentionally fostering stylistic ambiguity within the generated artworks.

Methodological Insights

The CAN framework modifies traditional GAN objectives to achieve creativity by maximizing deviation from established styles while still conforming to the broader distribution of artistic data. The methodology leverages two conflicting forces within the discriminator: one that ensures the generator’s outputs are recognized as 'art' and another that penalizes adherence to familiar styles, thereby encouraging stylistic ambiguity. This dual-pressure system encourages the exploration of more creative spaces within the art generation process.

Martindale’s principle of "least effort" and Berlyne’s notions of arousal potential are foundational to this approach. The model capitalizes on the idea that creative works should increase arousal potential by introducing novelty, complexity, and ambiguities — aspects that are computationally fostered by the CAN’s design.

Results and Assertions

Experimental results are indicative of the CAN’s capacity to generate art that is indistinguishable from contemporary human art by human evaluators. Human subjects reportedly rated the images produced by the CAN to be as artistic as human-created artworks exhibited at prestigious art events like Art Basel. Furthermore, participants could not effectively differentiate between artworks produced by the CAN and those by recognized artists or art exhibitors, signifying the model's effectiveness.

Quantitative assessments demonstrated that, in a controlled survey, subjects frequently interpreted CAN-generated art to possess intentionality, visual structure, and communicative elements that rival those of human-produced art. While caution must be maintained when interpreting subjective results owing to inherent biases, the paper provides compelling evidence of the CAN’s ability to imitate the human creative process closely.

Theoretical and Practical Implications

Theoretically, the paper suggests a feasible computational pathway to model aspects of human creativity computationally, emphasizing the role of stylistic deviation as a creative catalyst. By breaking from conventional GAN methodologies, the CAN architecture proposes a innovative way to externally evaluate creativity through style ambiguity and distributional conformity.

Practically, CANs possess the potential for significant implications in the fields of AI and art. This includes broad applications in creative industries that require automated generation of novel content which still respects artistic integrity and aesthetics. Further, the continuous adaptability of the CAN — learning from new art without requiring semantic understanding of styles — offers a dynamic tool for artistic exploration.

Future Prospects

Future research could aim at enhancing the semantic understanding of generated artworks, thus providing deeper insights into not merely imitating but innovating art styles. Development of sophisticated evaluation metrics beyond human perceptual surveys may provide more granular control and insight into the creative processes underpinning these networks. Moreover, extending CAN frameworks to other artistic domains (e.g., music, architecture) could evaluate the generalizability of stylistic deviation as a universal component of computational creativity.

Through this research, the authors contribute an innovative approach to generating art that aligns closer with human creativity, thus expanding the horizons of what is achievable with machine learning in the arts. While not without its limitations, the CANs provide a compelling step forward in the machine-based creation of culturally significant artifacts.