Abstract
While Generative Adversarial Networks (GANs) have achieved high fidelity in face synthesis, generating images that consistently align with human aesthetic preferences remains a challenge. This research presents a framework that integrates Reinforcement Learning (RL) with EigenGAN to optimize the aesthetic quality of generated Caucasian faces. By utilizing a custom-trained aesthetics scoring model as a reward signal, an RL agent learns to navigate the latent space to enhance specific semantic attributes. Experimental results, implemented via the MindSpore framework, show that the adjusted images exhibit significantly higher aesthetic levels compared to the original EigenGAN outputs.
Introduction: The Aesthetic Gap in GANs
Modern GANs can produce incredibly realistic faces, yet they lack an inherent understanding of human "beauty." Previous attempts to "beautify" generated images often faced two primary hurdles: Identity Loss, where the subject's identity changes during adjustment, and a Lack of Granular Control, where beauty is treated as a single binary trait rather than a combination of specific attributes.
This paper addresses these issues by allowing users to choose specific semantic attributes—such as lighting, bangs, or smiles—and employing an RL agent to find the optimal values for those attributes to maximize aesthetic appeal.
Methodology: Aesthetics Driven Reinforcement Learning
The framework is built upon EigenGAN, an architecture that learns interpretable "eigen-dimensions" representing semantic attributes in an unsupervised manner. The core innovation is the Aesthetics Driven Reinforcement Learning loop.
The Reward Signal
The researchers developed an aesthetics scoring network using an EfficientNet-B4 backbone enhanced with Efficient Channel Attention (ECA). Trained on the SCUT-FBP5500 dataset, this model provides a "beauty score" that serves as the reward for the RL agent.
The RL Agent
To navigate the high-dimensional latent space effectively, the authors implemented a Soft Policy Gradient approach. This method incorporates "maximum entropy," which encourages the agent to explore a diverse range of attribute values, preventing the model from collapsing into a single "perfect" face and ensuring variety in the output.
Stability and Optimization
To stabilize training, a self-critical baseline was utilized. This allows the agent to compare its current attribute choices against its own average performance, eliminating the need for a complex separate value network. The agent observes a face, tweaks specific eigen-dimensions, and receives a reward based on the resulting change in the aesthetic score.
Experimental Results
The proposed aesthetics model achieved a Root-Mean-Square Error (RMSE) of 0.3058, outperforming previous state-of-the-art models on the SCUT-FBP5500 benchmark.
Quantitatively, the average aesthetic score across a test set of 1,000 images rose from 0.5652 (original EigenGAN) to 0.6179 after RL adjustment. Subjective user studies further confirmed these findings: RL-optimized images were preferred by 28% of participants, more than double the preference for the original baseline.
Key Insights
- Beauty as a Reward: Treating aesthetics as a goal for an RL agent to navigate toward, rather than a fixed latent direction, allows for better preservation of other facial features.
- Attribute Correlation: Using Spearman’s rank correlation, the authors identified that attributes like Bangs, Lighting, and Smiling have the strongest positive impact on perceived beauty.
Limitations and Future Work
Despite the improvements, the method is dependent on the disentanglement quality of the underlying EigenGAN. If attributes are entangled, artifacts may appear. Additionally, the training dataset (SCUT-FBP5500) is 72% Asian faces, which introduced certain biases in the RL agent's preferences, such as a tendency toward darker hair. Future research will focus on "inverting" real-world photos into the EigenGAN latent space to enable aesthetic editing of actual photographs.
Summary updated