Editing Discrete Latent Variables with Reinforcement Learning

Controlled and interpretable generation through minimal edits to discrete VQ-VAE codes.

This project develops a reinforcement-learning agent that transforms an input toward a specified target class by editing discrete VQ-VAE latent codes. The objective is to achieve the desired semantic change while minimizing the number of codebook edits.

The training reward combines classifier confidence with discriminator realism, encouraging goal-directed transformations that remain visually plausible. Multi-step editing trajectories provide a direct view of how the agent navigates the latent representation and reveal task-dependent structure in the codebook.

The broader goal is to make generative control more precise and interpretable by exposing a compact sequence of discrete decisions rather than relying on an opaque, unconstrained latent update.