Creative Adversarial Network (CAN)

• Creative Adversarial Network (CAN) is a generative model that promotes creative deviation by integrating a style ambiguity loss into the traditional GAN framework.
• CAN enhances the standard GAN architecture with a dual-headed discriminator that assesses both authenticity and stylistic classification, ensuring outputs are plausible yet artistically novel.
• Extensions like conditional and unrolled variants improve diversity and quality in creative outputs, successfully applying CAN principles to both visual art and music generation.

A Creative Adversarial Network (CAN) is a generative model architecture that extends the standard Generative Adversarial Network (GAN) by explicitly encouraging creative deviation from established style norms while preserving the underlying structure of learned data distributions. Originating in visual art generation, CANs augment the adversarial framework with a mechanism for style ambiguity, operationalized through a discriminator that simultaneously detects authenticity and classifies samples into stylistic categories. By introducing a tailored loss term—maximizing the entropy of the predicted style distribution for generated samples—CANs are designed to produce content that is simultaneously plausible and novel within a learned domain. Subsequent work generalizes this approach to conditional generation (CCAN) and adapts it to symbolic music (Elgammal et al., 2017, Hereu et al., 2024, Nag, 2024).

1. Architectural Foundations

The canonical CAN architecture builds directly upon the DCGAN backbone. It consists of a generator $G$ and a discriminator $D$, with the following key enhancements:

• Generator ($G$): Receives a noise vector $z \sim \mathcal{N}(0, I)$, processed through a stack of transposed convolutional layers (e.g., $5$–$6$ layers, with channel reductions such as $512 \to 3$ and upsampling from $4 \times 4$ to $64\times64$ or $256\times256$) to synthesize an image or, for symbolic domains, a piano-roll.
• Discriminator ($D$): Processes input images through a CNN trunk, with two output heads:
  • Real/Fake Head ($D_r$): Outputs a scalar probability via sigmoid, indicating whether the instance is real or generated.
  • Style-Classification Head ($D_c$): Outputs a categorical (softmax) distribution over $K$ pre-defined style classes. Typically, deep FC layers process the final conv features to produce the $K$-way vector.

For Conditional CAN (CCAN), both $G$ and $D$ integrate learned embeddings of style labels: - $G$ receives $[z; E_G(c)]$ as input, concatenating a style embedding to the noise vector. - $D$ incorporates a style embedding $E_D(c)$ by tiling or concatenation with the input.

In music generation, the CAN is adapted by treating symbolic scores as image-like tensors (e.g., $128\times128$ piano-rolls) (Nag, 2024). The same basic structure applies, with convolutional generators and discriminators, and an analogous style-classification head for composer identity.

2. Objective Functions and Style-Ambiguity Mechanism

CAN redefines the standard GAN objective by introducing a style confusion term designed to maximize entropy in the discriminator’s style predictions for generated data. The min–max game becomes:

$$\min_{G}\;\max_{D}\;V(D,G) = \mathbb{E}_{(x,\hat c)\sim p_{\mathrm{data}}}\left[ \log D_r(x) + \log D_c(\hat{c}|x) \right] + \mathbb{E}_{z\sim p_z}\left[ \log(1 - D_r(G(z))) - \sum_{k=1}^K \left( \frac{1}{K}\log D_c(c_k|G(z)) + \left(1-\frac{1}{K}\right)\log(1 - D_c(c_k|G(z))) \right) \right].$$

• For Discriminator ($D$):
  • Standard GAN loss: pushes $D_r(x)$ to $1$ (real), $D_r(G(z))$ to $0$ (fake).
  • Style-classification loss: cross-entropy for classifying real images into their true style $\hat c$.
  • Creativity loss: encourages uniformity over styles for generated samples (i.e., style confusion).
• For Generator ($G$):
  • Fool $D_r$: maximize $D_r(G(z))$ (“look real”).
  • Maximize style confusion: push $D_c(\cdot|G(z))$ toward uniform over all classes, so generated images cannot be easily assigned to a single style.

The generator’s overall loss comprises the adversarial term plus a “creativity” loss, which, in information-theoretic terms, regularizes toward maximal style-classification entropy (Elgammal et al., 2017, Hereu et al., 2024).

3. Training Procedures and Implementation Details

CAN training replicates DCGAN protocols:

• Data: Visual domain experiments use the WikiArt dataset, filtered to specific classes (e.g., 14,245 portraits, 24 styles) (Hereu et al., 2024); music domain uses MIDI converted to 2D piano-rolls.
• Preprocessing: Uniform resizing (e.g., $64\times64$ or $256\times256$), normalization to $[-1,1]$, and five-crop augmentation.
• Optimization: Adam optimizer ($\text{lr}=1\times10^{-4}$, $\beta_1=0.5$, $\beta_2=0.999$); batch size 128; $>$100 epochs; typical training time is 24h per model on V100/A100 GPUs.
• Stabilization: BatchNorm in all layers except output, LeakyReLU in $D$, ReLU in $G$, strided convolutions replace pooling.

For Unrolled CAN in music (Nag, 2024), the generator update is computed with respect to a $k$-step-unrolled $D$, mitigating mode collapse and promoting greater diversity:

for each training iteration: D_temp ← D for i in 1..k: D_temp ← D_temp − η ∇_{D_temp} V(D_temp, G) G ← G − η ∇_G V(D_temp, G) D ← D − η ∇_D V(D, G)

This anticipates discriminator adaptation, enabling $G$ to seek genuine creative deviation rather than exploiting ephemeral $D$ weaknesses.

4. Novelty, Arousal Potential, and Evaluation Metrics

CAN formalizes “creative deviation” using the concept of arousal potential drawn from Berlyne and Martindale’s psychological theory: maximizing entropy in the style-classifier output is seen as enhancing novelty, surprise, and ambiguity.

• Arousal Potential Metric: For a given image $x$, arousal potential is quantified as $H_{\text{uniform}}(p(c|x))$, the entropy of the style posterior, with $p(c|x)$ given by $D_c$ (Elgammal et al., 2017).
• Music generation novelty: Assessed via an auto-encoder trained on the real data manifold; higher reconstruction MSE for generated samples indicates greater creative divergence (Nag, 2024).

Empirical results confirm that CANs produce samples rated as more creative (higher arousal potential, greater style ambiguity) than DCGAN or pure style-classification variants.

5. Conditional and Domain-General Extensions

Conditional CAN (CCAN) introduces control over the generative process:

• Conditional Embeddings: Style labels embedded and incorporated into both $G$ and $D$, steering samples toward target stylistic regions.
• Loss Structure: Retains the creativity term, ensuring that even when conditioned on a style, generated samples must display ambiguity within that family—emulating the human process of creating art that is rooted in tradition but breaks new ground (Hereu et al., 2024).

In music, CAN generalizes by learning stylistic classes corresponding to composer identity. The architectural and loss function principles persist, supporting general domain transfer of the style-ambiguity principle. Unrolled optimization further increases generative diversity (Nag, 2024).

6. Experimental Outcomes and Comparative Findings

Qualitative results in portraiture (Hereu et al., 2024):

• DCGANs yield plausible, but style-homogeneous, outputs.
• CANs generate greater stylistic variety—unusual facial features, palettes, and attire—occupying the “edges” of the training distribution without falling off the manifold.
• CCAN outputs, when conditioned on specific styles (e.g., Realism, Rococo), exhibit broad hallmarks of the designated style but incorporate unexpected and novel deviations.

Quantitative and perceptual studies (Elgammal et al., 2017):

• Human subjects rate CAN-generated works as more “artist-like” and aesthetically appealing than standard GAN outputs, and in certain respects comparable to contemporary art.
• Statistical significance is robust (e.g., $p < 1 \times 10^{-5}$ for ability to distinguish CAN and DCGAN works).
• In music, the Expert Gate novelty score demonstrates that unrolled CANs exceed GANs and basic CANs in deviation from data manifold while preserving semantic integrity (Nag, 2024).

7. Limitations and Prospects

CAN mechanisms do not endow models with semantic understanding; their “creativity” is strictly a function of style ambiguity, not content innovation (Elgammal et al., 2017). The approach currently relies on well-defined, annotation-rich style/class labels and cannot generalize to unstructured creativity or multi-modal context without further modification.

Prospective research avenues include incorporating higher-level compositional constraints, extending CANs to multi-modal generation (e.g., aligning music and visual art), and exploring long-term style evolution under adversarial creativity pressure (Elgammal et al., 2017, Hereu et al., 2024). A plausible implication is that coupling CAN frameworks with deep semantic modeling could yield systems with a broader conception of creativity, beyond stylistic ambiguity alone.