A comprehensive survey on facial expression generation: From GANs to LLM-guided multimodal models

Abstract

Facial expression generation (FEG) has emerged as a vital area in human–computer interaction, virtual avatars, and affective computing, aiming to synthesize natural and expressive facial behaviors across diverse interaction contexts. This survey presents a comprehensive analysis of recent advances in FEG, organized into six key paradigms: speech-driven expression generation, facial reaction generation, face video generation, facial animation, avatar-based generation, and text-driven expression generation. We review a wide range of model architectures, including VQ-VAEs, Generative Adversarial Networks (GANs), 3D Morphable Models (3DMMs), Transformers, and diffusion-based approaches, and compare their performance using commonly adopted evaluation metrics such as Fréchet Distance (FD), Peak Signal-to-Noise Ratio (PSNR), and expression realism measures. We further examine the emerging integration of Large Language Models (LLMs) into FEG pipelines. While recent studies suggest that LLMs can introduce high-level semantic reasoning and contextual awareness into facial expression generation, current approaches remain limited by the lack of standardized evaluation protocols, reliance on intermediate representations (e.g., motion tokens), and insufficient large-scale empirical validation. In addition, we provide a structured comparison of widely used datasets, highlighting variations in modality, annotation strategies, and interaction settings. Based on this analysis, we identify critical challenges in the field, including inconsistencies in evaluation metrics, limited personalization capabilities, and unresolved issues in cross-modal semantic alignment. This survey aims to serve as both a technical reference and a forward-looking roadmap for developing more robust, interpretable, and context-aware facial expression generation systems.