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
How to generate and assess machine-made art using prior human art remains unclear. CAN modifies GANs to depart from established styles while staying within the art distribution; human subjects regularly confused its outputs with human art and sometimes rated them higher on high-level scales.
Problem
How human artists integrate knowledge of prior art into new forms remains largely unknown, and assessing machine-generated art’s creativity is difficult.
Source: §1 · p. 2, §4.2 · p. 12Method
CAN modifies GANs to maximize deviation from established styles while minimizing deviation from the art distribution, using the discriminator to encourage stylistic ambiguity.
Source: §1 · p. 2, §2.4 · p. 6Results
Human subjects regularly confused CAN-generated art with human art and sometimes rated it higher on high-level scales.
Source: §5 · pp. 20–21Takeaways & Limitations
Human subjects’ responses indicate that they saw CAN-generated images as art.
Source: Experiment III · p. 18The authors leave open whether higher ratings reflect style-backward or aesthetic-assessment bias and note that more experiments are needed.
Source: §5 · p. 21
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 proposes generating art without a human artist in the creative process, while learning from human-made art. It adapts GAN objectives to depart from established styles while remaining close to the art distribution.
1 Introduction
- Generating human-level creative products, including visual art, has been a longstanding goal in artificial intelligence research.The paper frames this ability as fundamental to demonstrating that AI algorithms are intelligent.Source: §1 · p. 1
- The paper investigates art generation without a human artist participating creatively, while using human creative products during learning.The authors argue that relating art-generation processes to human-made art matters because human artists draw on prior exposure to art.Source: §1 · p. 2
- Inspired by Martindale [15], the approach treats style breaks as a way to raise arousal potential, with increases kept minimal to avoid negative observer reactions.Martindale’s principle of least effort and explanation of style breaks inspire the proposed approach.Source: §1 · p. 2
- The proposed system modifies GAN objectives to maximize deviation from established styles while minimizing deviation from the art distribution, aiming to generate creative art.The paper argues that GANs in their original design are limited in generating creative products.Source: §1 · p. 2
2 Methodology
The methodology treats creativity as a controlled increase in arousal potential: CAN aims to generate work that remains recognizable as art while deviating from established styles. It contrasts this goal with GANs that emulate the art distribution without a creative objective.
2.1 Background
- Arousal-raising collative properties include novelty, surprisingness, complexity, ambiguity, and puzzlingness; habituation pressures art to change gradually rather than excessively.The account favors moderate arousal: too little is boring, while too much can provoke aversion and negative response.Source: §2.1 · p. 4
2.2 Art Generating Agent
- The agent stores exposed art and seeks to raise arousal through stylistic ambiguity while keeping generated work close enough to art to avoid excessive novelty.CAN is designed to generate work that does not follow established styles and confuses viewers about which style it belongs to.Source: §2.2 · pp. 4–5
2.3 GAN: Emulative and not Creative
- Ordinary GANs reward distribution-matching, so they can reproduce existing art or conditioned styles but provide no pressure to create novel work.Even conditioning generation on style labels can yield emulation rather than the novelty, surprise, ambiguity, or puzzlingness associated with creative art.Source: §2.3 · p. 6
2.4 From being Emulative to being Creative
- The art/not-art signal keeps outputs near the learned art distribution, while the style-classification signal penalizes easy assignment and drives stylistic ambiguity.Together, the signals encourage novel work without moving so far from art that it risks excessive arousal and negative response.Source: §2.4 · pp. 6–7
3 Technical Details
The section defines GANs as adversarial learners of image distributions, then presents CAN’s style-classification and style-ambiguity objectives for generating realistic art that deviates from learned style norms.
3.1 Generative Adversarial Networks
- The GAN generator produces samples intended to match the training-data distribution, while its discriminator distinguishes real images from generated ones; the networks are optimized alternately.The generator is trained to maximize log D(G(z)) for stronger gradients than minimizing −log(1 − D(G(z))).Source: §3.1 · pp. 7–8
3.2 Creative Adversarial Networks
- CAN minimizes cross-entropy between generated-image style posteriors and a uniform target, encouraging style ambiguity while sharply penalizing images assigned high probability to one style.Unlike entropy, the cross-entropy loss rises sharply as a class posterior approaches 1, producing large losses and gradients for confidently classified generated images.Source: §3.2 · p. 8, Figure 3 · p. 8
- CAN’s discriminator learns real-versus-generated status and style labels, while its unlabeled generator aims to produce realistic images with ambiguous style posteriors.The discriminator uses real images with style labels; the generator’s objective combines realism with cross-entropy against a uniform style distribution.Source: §3.2 · p. 9
- CAN’s generator maps a 100-dimensional noise vector through convolutional upsampling to a 256 × 256 image, while a shared discriminator body feeds real/fake and K-style classification heads.The discriminator body has six stride-2 convolution layers; its real/fake head is fully connected, and its style head uses three fully connected layers.Source: §3.2 · p. 9
- CAN training used 128-image mini-batches, Adam at learning rate 0.0001 for 100 epochs, Batch Normalization, and five 90%-size crops per painting.Weights were initialized from a zero-centered Normal distribution with standard deviation 0.02, and the LeakyReLU slope was 0.2.Source: §3.2 · p. 9
4 Results and Validation
Human-subject experiments favored CAN on novelty, aesthetic appeal, and art-related ratings, while qualitative comparisons suggest its style ambiguity loss helps generate novel yet aesthetically appealing images. Results also show that participants often attributed CAN images to artists.
4.1 Training the model
- The networks were trained on 81,449 paintings from 1,119 artists in the WikiArt collection downloaded in 2015.The paintings spanned the Fifteenth through Twentieth centuries.Source: §4.1 · p. 10
4.2 Qualitative Validation
- CAN produced novel-looking, aesthetically appealing images without typical figures, genres, or subject matter, while style-classification-CAN generated recognizable art elements.Removing the style ambiguity loss yielded the style-classification variant, which better emulated the art distribution; the authors argue the ambiguity loss encourages novelty while remaining close to that distribution.Source: §4.2 · pp. 12–13
4.3 Quantitative Validation
- The quantitative comparison included Art Basel 2016 paintings chosen at random after excluding clear figures and obvious brush strokes to reduce potential subject bias.The paper describes inclusion in this contemporary art fair as an indication of work at the frontier of human creativity, as judged by experts and the art market.Source: §4.3 · p. 13
4.4 Experiment II:
- In Experiment II, 75% of responses attributed CAN images to an artist, versus 65% for 256x256 DCGAN images (p-value=1.0147e −05).The artist-attribution question followed other image ratings; the authors report that this ordering produced a higher CAN attribution rate than in Experiment I.Source: §4.4 · pp. 16–17, Table 3 · p. 17
Experiment III:
- CAN images scored higher than the Abstract Expressionist and Art Basel sets on intentionality, visual structure, communication, and inspiration.The authors cautioned that the meaning and statistical significance of higher scores are debatable, but interpreted similar levels to human art as indicating that subjects saw CAN images as art.Source: Table 4 · p. 18, Experiment III · p. 18
Experiment IV:
- Art-history students selected CAN images as more novel 59.47% of the time and more aesthetically appealing 60% of the time than sc-CAN images.Experiment IV compared paired images from CAN and the style-classification variant, using novelty as a proxy for creativity.Source: Experiment IV · p. 18
5 Discussion and Conclusion
The CAN generates art by maximizing stylistic ambiguity while remaining within the art distribution; human-subject experiments found that people regularly confused its images with human art and sometimes rated them higher. Its style-based learning enables novelty and self-assessment, but it lacks semantic understanding of art.
5 Discussion and Conclusion
- The system learns from art images spanning the 15th to 21st centuries and their style labels, but it has no semantic understanding of subject matter or explicit artistic elements and principles.The authors say it can continue learning from new art and adapt its generation accordingly.Source: §5 · pp. 18, 21
- Human subjects regularly confused CAN images with human art and sometimes rated them higher on high-level scales.The authors leave open why CAN art was ranked above Art Basel samples and call for more experiments to interpret those responses.Source: §5 · pp. 20–21
- By maximizing stylistic ambiguity while staying within the art distribution, the system explores beyond established styles while remaining recognizable as art.This interaction is presented as enabling novel artifacts and self-assessment; human-subject experiments provide evidence of artifact quality, aligning with Colton’s criteria.Source: §5 · pp. 20–21
- The authors leave open whether CAN’s higher ratings than Art Basel samples reflect style bias, aesthetic-assessment bias, or lower creativity, and call for more experiments.The Art Basel comparison set contains 25 paintings selected from the 2017 fair.Source: Figure 8 · p. 20, §5 · p. 21