Introducing the EG3D Generative Model to Build a High-Fidelity Dynamic Face Actuation System for Gamified Film and Television Content

With the increasing demand for immersion and interactivity in gamified film and television content, maintaining high-fidelity dynamic facial expressions remains challenging due to insufficient detail preservation and temporal inconsistencies across consecutive frames. To address these issues, this paper proposes EG3D-TA, a framework that integrates EG3D implicit field modeling with a Temporal Awareness (TA) actuation network. The proposed method constructs a three-dimensional generative representation based on EG3D, where Neural Radiance Fields (NeRF) implicitly model high-resolution facial geometry, while a keypoint-driven mapping network transforms two-dimensional facial expressions into continuous W+ latent encodings. An optical flow-guided latent interpolation strategy based on PWC-Net is further introduced to optimize inter-frame trajectories and improve temporal coherence. In addition, model pruning and feature caching mechanisms enable low-latency real-time inference. Experimental results demonstrate that the proposed approach achieves high visual fidelity (LPIPS: 0.12– 0.19; PSNR: 36.0–38.2 dB) and excellent motion consistency (TVD: 0.08–0.12; FVD: 78.3–88.9), while maintaining an interactive performance of 21.6 FPS at 5122 resolution with an end-to-end latency of 46.3 ms. Beyond high-fidelity digital character generation, the proposed geometry-aware and temporally consistent representation framework provides valuable methodological support for three-dimensional signal reconstruction, immersive visual communication, and intelligent sensing systems in advanced electromagnetic information processing and future integrated communication–perception applications.