Attribute Controllable Beautiful Caucasian Face Generation by Aesthetics Driven Reinforcement Learning

Xin Jin, Shu Zhao, Le Zhang, Xin Zhao, Qiang Deng, Chaoen Xiao

TL;DR

This paper presents a novel framework that combines EigenGAN with Reinforcement Learning (RL) to generate Caucasian faces that are both high-fidelity and aesthetically optimized. By using a custom-trained aesthetics scoring model as a reward signal, the RL agent learns to navigate the latent space to "beautify" specific facial attributes (like lighting, bangs, and smiles) while maintaining precise control over the image generation process.

What Problem Does This Paper Address?

While modern Generative Adversarial Networks (GANs) can produce incredibly realistic faces, they don't inherently understand human "beauty" or aesthetics. Previous attempts to "beautify" generated images often suffered from two main issues:

  1. Identity Loss: Methods like Beholder-GAN struggle to keep the person looking like the same individual after the aesthetic adjustment.
  2. Lack of Control: Many models treat "beauty" as a single binary attribute (beautiful vs. not), which is reductive and offers no granular control over why a face looks better.

The authors aim to bridge this gap by allowing a user to choose specific semantic attributes (e.g., "make the lighting better" or "adjust the bangs") and having an RL agent find the optimal values for those attributes to maximize aesthetic appeal.

Methods

The researchers built their system on top of EigenGAN, a GAN architecture that unsupervisedly learns "eigen-dimensions" representing interpretable semantic attributes (like pose, gender, or hair color) in different layers.

The core innovation lies in the Aesthetics Driven Reinforcement Learning loop:

  • The Reward Function: They trained a dedicated aesthetics scoring network using an EfficientNet-B4 backbone with Efficient Channel Attention (ECA). This model predicts a "beauty score" for any generated face.
  • The RL Agent: Instead of using standard RL, which often fails in high-dimensional latent spaces, they implemented a Soft Policy Gradient approach. This incorporates "maximum entropy," which encourages the agent to explore diverse attribute values rather than collapsing into a single "perfect" face.
  • Self-Critical Baseline: To stabilize training, they used a "self-critical" baseline (borrowed from image captioning research). This allows the agent to compare its current choices against its own average performance, removing the need for a complex separate "value network."
  • Workflow: The agent observes a generated face, chooses how to tweak specific "eigen-dimensions," and receives a reward based on the change in the aesthetic score.

Datasets

  • SCUT-FBP5500: A benchmark dataset containing 5,500 frontal faces with diverse ethnicities and beauty scores. This was used to train the aesthetics scoring model.
  • ImageNet: Used for pre-training the feature extraction modules (ResNet18 and EfficientNet).
  • EigenGAN Pre-trained Weights: Used as the base generator for the face synthesis.

Key Insights and Ideas

  • Beauty is a Reward, Not an Attribute: Instead of trying to find a "beauty direction" in the latent space (which is often entangled with other features), the authors treat beauty as a goal for an RL agent to navigate toward.
  • Entropy is Essential: By using Soft Policy Gradients, the model maintains high entropy during training. This prevents the "quick diminishing of entropy" (where the model gets stuck on one look) and ensures the generator can still produce a wide variety of attractive faces.
  • Attribute Correlation: The authors used Spearman’s rank correlation to identify which latent dimensions actually impact aesthetics. They found that attributes like Bangs ($z_{3,4}$), Lighting ($z_{4,4}$), and Smiling ($z_{4,5}$) had the strongest positive correlations with beauty scores, while others (like background) were irrelevant.

Key Results

  • Superior Scoring: Their aesthetics model achieved a Root-Mean-Square Error (RMSE) of 0.3058, outperforming the previous state-of-the-art on the SCUT-FBP5500 dataset (0.3325).
  • Quantitative Improvement: The average aesthetic score of 1,000 test images increased from 0.5652 (original EigenGAN) to 0.6179 when using the multi-attribute RL adjustment.
  • Subjective Preference: In a user study, participants were asked to pick the "best-looking" image from a group. The RL-optimized images ($M_{aes}$) received the highest percentage of votes (28%), more than doubling the preference for the original EigenGAN images (12%).
  • MindSpore Implementation: The authors successfully utilized the MindSpore framework to implement their multi-attribute model.

Limitations and Future Work

  • Disentanglement Dependency: The method's success depends entirely on how well the underlying EigenGAN can separate attributes. If attributes are "entangled" (e.g., changing hair color also changes gender), the RL agent will produce artifacts.
  • Manual Filtering: The authors still had to manually traverse latent dimensions to identify which ones corresponded to specific human-understandable traits.
  • Dataset Bias: The SCUT-FBP5500 dataset is 72% Asian faces. The authors noted this biased the RL agent toward certain traits, such as preferring darker hair colors.
  • Gender Blindness: The generator sometimes lacks "gender perception," leading to odd results like male faces being assigned reddish lipstick.
  • Future Direction: The authors plan to extend this to real-world image editing by first "inverting" real photos into the EigenGAN latent space.

Summary updated

Abstract

In recent years, image generation has made great strides in improving the quality of images, producing high-fidelity ones. Also, quite recently, there are architecture designs, which enable GAN to unsupervisedly learn the semantic attributes represented in different layers. However, there is still a lack of research on generating face images more consistent with human aesthetics. Based on EigenGAN [He et al., ICCV 2021], we build the techniques of reinforcement learning into the generator of EigenGAN. The agent tries to figure out how to alter the semantic attributes of the generated human faces towards more preferable ones. To accomplish this, we trained an aesthetics scoring model that can conduct facial beauty prediction. We also can utilize this scoring model to analyze the correlation between face attributes and aesthetics scores. Empirically, using off-the-shelf techniques from reinforcement learning would not work well. So instead, we present a new variant incorporating the ingredients emerging in the reinforcement learning communities in recent years. Compared to the original generated images, the adjusted ones show clear distinctions concerning various attributes. Experimental results using the MindSpore, show the effectiveness of the proposed method. Altered facial images are commonly more attractive, with significantly improved aesthetic levels.
BibTeX
@inproceedings{jin2022attribute,
  author = {Xin Jin and Shu Zhao and Le Zhang and Xin Zhao and Qiang Deng and Chaoen Xiao},
  title = {{Attribute Controllable Beautiful Caucasian Face Generation by Aesthetics Driven Reinforcement Learning}},
  booktitle = {ACM Multimedia},
  year = {2022},
  url = {https://openreview.net/forum?id=SJzTjIsaToe},
}

Loading related content...

Figure