Source-linked AI summary
CAN: Creative Adversarial Networks, Generating "Art" by Learning About Styles and Deviating from Style Norms
Ahmed Elgammal, Bingchen Liu, Mohamed Elhoseiny, Marian Mazzone
TL;DR
The paper addresses the limited creativity of GANs that emulate a training distribution by proposing CAN, which learns art styles while encouraging stylistic deviation within the art distribution. Human subjects regularly confused CAN-generated art with human art and sometimes rated it higher on high-level scales, though the authors leave these responses open to further interpretation.
Problem
GANs can learn to emulate an art distribution, but their original design lacks a mechanism encouraging creative deviation from established styles.
Method
CAN modifies GAN objectives to maximize stylistic ambiguity while keeping generated images within the art distribution, using art images labeled by style.
Results
Human subjects regularly confused CAN-generated art with human art and sometimes rated it higher on various high-level scales.
Takeaways & Limitations
The results indicate that subjects perceived CAN-generated images as art and as having creative characteristics under the reported human evaluations.
Takeaways & Limitations
The authors leave open whether higher CAN ratings reflect style-backward bias, aesthetic assessment bias, or limited creativity, motivating further experiments.
Abstract
from arXiv · showhide
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.
1 Introduction
The paper frames creative art generation as a longstanding AI goal and proposes modifying GANs to balance art-like outputs with deviation from established styles.
- Creative art generation is presented as a longstanding goal spanning poetry, stories, jokes, music, paintings, and creative problem solving.
- The paper seeks an art-generating system that learns from human art without involving a human artist in the creative process.
- Martindale’s theory motivates increasing art’s arousal potential through limited style breaks that avoid negative reactions from observers.
- GANs can emulate a training distribution, but their original objective lacks a force encouraging exploration of creative space.
- CAN modifies the GAN objective to maximize deviation from established styles while minimizing deviation from the art distribution.
2 Methodology
The methodology combines an arousal-based account of creativity with a GAN variant that generates art-like images while deviating from established styles. CAN uses art and style information to encourage stylistic ambiguity rather than straightforward emulation.
- 2.1 Background: Arousal potential denotes stimulus properties that raise arousal, including novelty, surprisingness, complexity, ambiguity, and puzzlingness.
- 2.1 Background: Moderate arousal is preferred because too little is boring, whereas too much can activate aversion and produce negative responses.
- 2.1 Background: Martindale’s model balances pressure to change art against a least-effort constraint that limits the pace of stylistic change.
- 2.2 Art Generating Agent: The proposed agent aims to generate art that is novel but not so novel that it leaves the accepted art distribution or enters a negative hedonic range.
- 2.2 Art Generating Agent: CAN is designed to avoid established art movements and styles by generating art that maximally confuses viewers about its style.
- 2.4 From being Emulative to being Creative: Standard GAN training can produce images resembling existing art without motivating the generator to explore creative space.
- 2.4 From being Emulative to being Creative: Style-conditioned GANs can emulate genres such as Renaissance, Impressionism, or Cubism, but this emulation alone does not constitute the paper’s creative objective.
- 2.4 From being Emulative to being Creative: Unlike a standard GAN, CAN receives signals for both art-versus-not-art classification and ambiguity across established styles.
3 Technical Details
The paper adapts GAN training so the discriminator learns both art realism and style, while the generator produces realistic images with stylistic ambiguity. Training uses alternating optimization, random noise, and an entropy-based ambiguity loss across artistic styles.
- Generative Adversarial Networks: GANs train a generator and discriminator in a two-player game to make generated samples match the training-data distribution.The generator produces samples without direct access to the training data, while the discriminator distinguishes generated from real images.
- Training Procedure: Training alternates discriminator and generator updates using mini-batch stochastic gradient descent, with random noise passed through the generator to create fake images.Algorithm 1 specifies mini-batch inputs, random-noise sampling, generator forward passes, and entropy loss computation.
- Creative Adversarial Networks: The style ambiguity loss is based on maximizing uncertainty over style classes, with the inverted cross entropy heavily penalizing confidently classified samples.Both entropy and the illustrated inverted cross entropy are maximized when classes are equiprobable, but inverted cross entropy tends to negative infinity at the boundaries.
- Creative Adversarial Networks: The modified objective adds style classification and style ambiguity losses to the standard real-versus-generated image discrimination objective.The discriminator learns both the art distribution and style categories, while the generator is optimized against these objectives.
- Creative Adversarial Networks: The generator is pushed toward realistic art while maximizing cross entropy between generated-image style posteriors and a uniform distribution.This objective encourages generated images to remain art-like while making their style classification ambiguous.
4 Results and Validation
The paper evaluates CAN against GAN-based baselines and human-created art using qualitative comparisons and human-subject experiments. CAN generated novel, aesthetically appealing images, and subjects often judged them as artist-created or rated them highly.
- Qualitative Validation: The 64x64 DCGAN baseline failed to emulate trained art, producing no recognizable figures or art genres.The baseline was trained on art data but did not generate images resembling its learned distribution.
- Qualitative Validation: The style-classification-CAN baseline better emulated the art distribution by learning style classes, but lacked the style ambiguity loss associated with creativity.Its generator used the GAN loss, while the discriminator additionally learned to classify styles.
- Qualitative Validation: CAN images appeared novel and aesthetically appealing while avoiding recognizable figures, genres, styles, and subject matter.The authors attribute this combination to the style ambiguity loss, which encourages novelty while remaining closer to the art distribution.
- Experiment I: 53% versus 35%: subjects judged CAN images artist-created more often than GAN images, while liking CAN more, 3.2 versus 2.8.These differences were statistically significant, with p-values of 1.9932e-15 for the artist/computer question and 9.3634e-06 for liking.
- Experiment II: 75% versus 65%: in Experiment II, subjects more often judged CAN images artist-created than higher-resolution GAN images.The difference was statistically significant with t-test p-value=1.0147e-05, consistent with Experiment I.
- Experiment IV: 59.47% of comparisons favored CAN as more novel, and 60% favored CAN as more aesthetically appealing over style-classification-CAN.The experiment used art-history students to assess the effect of adding the style ambiguity loss.
5 Discussion and Conclusion
CAN generates art by learning from historical art and optimizing stylistic ambiguity while remaining within the art distribution. Human evaluations found that viewers confused CAN images with human art and sometimes rated them higher, while the interpretation of those responses remains open.
- The proposed system learns from art images with style labels and generates art by maximizing stylistic ambiguity while staying within the art distribution.
- Human subjects regularly confused CAN-generated art with human art and sometimes rated it higher on high-level scales.
- The system’s novelty arises from interactions between generation signals that encourage deviation from established styles while remaining close enough to be recognized as art.
- The system does not semantically understand art, subject matter, or explicit artistic elements and principles; it learns through exposure to art and style concepts.
- The authors leave open whether subjects’ higher ratings reflect style-backward bias, aesthetic-assessment bias, or implications for creativity, motivating further experiments.