LARA-Gen: Enabling Continuous Emotion Control for Music Generation Models via Latent Affective Representation Alignment

Recent text-to-music models enable coherent music generation from text prompts, yet fine-grained emotional control remains unresolved. We introduce LARA-Gen, a framework for continuous emotion control trained with a Latent Affective Representation Alignment (LARA) loss that explicitly aligns the model’s hidden states with representations from an external music understanding model. In addition, we design an emotion control module based on a continuous valence–arousal space, disentangling emotional attributes from textual content and bypassing the bottlenecks of text-based prompting. Furthermore, we establish a benchmark with a held-out test set and a robust Emotion Predictor for objective evaluation of emotional controllability. Experiments demonstrate that LARA-Gen achieves continuous, fine-grained emotion control and a favorable balance between emotion adherence and music quality.

(a) LARA-Gen framework. A Proxy Network P_θ aligns the internal hidden states H of the backbone model with target features M̄ from a frozen MERT encoder.
(b) The architecture of Emotion Predictor. It uses a sliding window over MERT features and an Emotion Regression Head R_φ to produce a final valence-arousal prediction from given music.