In this paper, we propose a feed-forward framework (Fig. 1) for human video compression based on 3D generative reconstruction. Recent surveys [1], [2] highlight the need for more efficient and semantically aligned solutions. Our approach disentangles video content into complementary structural and motion layers: the structural layer encodes regularized texture from a single frame, while the motion layer leverages the SMPL-X prior to represent complex dynamics with a compact set of pose and shape parameters. A hierarchical coding scheme exploits the heterogeneity of these representations for improved efficiency. After decoding, a feed-forward 3D reconstruction pipeline with facial feature extraction is employed, in which a multimodal transformer and a Gaussian head synthesize parametric cues that are fused with motion signals for accurate animation and high-fidelity rendering. Experiments show over$1000 \times$compression while preserving structural and semantic fidelity. The method consistently outperforms strong baselines (especially at$0.04-0.1 \text{kbpp})$, with significant gains in rate-distortion, FVD, and perceptual quality, as well as robust generalization across identities and scenes.

Neural network-based in-loop filtering significantly enhances video compression efficiency. However, high computational complexity hinders their deployment in real-time and ultra-high-definition scenarios. To address this, we propose a lightweight CNN-based in-loop filter for the luma component. In terms of model design, we utilize a U-Net-like architecture to learn the residual signal, incorporating depthwise separable 3 × 3 convolutions and 1 × 1 convolutions to reduce computational complexity, which results in a low complexity of only$37.707 \text{kMACs} /$pixel. For deployment optimization, we implement memory linearization to improve cache efficiency and combine blocked matrix multiplication with SIMD to maximize parallelism, ensuring cross-platform compatibility without third-party libraries. Experimental results on AVS4 EVM-0.9 (All-Intra) on a CPU platform show that the proposed method achieves BD-rate reductions of$1.36 \%, 0.35 \%$, and 0.34% for$\mathrm{Y}, \mathrm{U}$, and V components, respectively. Furthermore, the optimizations lead to a 91.5% reduction in decoding time, resulting in a decoding complexity of 7757% compared to the anchor.