Multi-view consistent 3D light-field-driven display method for dynamic digital humans

This paper proposes a multi-view consistent digital human reconstruction framework for three-dimensional (3D) light-field displays. Existing digital human reconstruction methods often emphasize single-view rendering quality, which may lead to geometric drift and cross-view inconsistency under dense-view display conditions. To address these limitations, an explicit facial mesh anchors Gaussian primitives through bounded residual offsets to improve geometric stability, while a spatial-angular attention network is embedded into 3D Gaussian splatting (3DGS) to jointly model fine-grained spatial details and view-dependent appearance. Furthermore, RGB, perceptual, and SSIM errors between rasterized images and ground-truth observations at calibrated camera viewpoints are back-propagated through the differentiable rasterizer to optimize Gaussian parameters; the physical light-field display is used for downstream validation. On temporally held-out test splits of three self-captured avatars, the proposed method achieves macro-average PSNR, SSIM, LPIPS, and FID values of 27.70 dB, 0.880, 0.097, and 18.71, respectively. The full model further achieves a warped cross-view error (WCVE) of 0.0213, compared with 0.0387 for the baseline. The generated digital humans exhibit improved geometric stability and view-dependent appearance consistency, enabling coherent 3D visualization on a physical light-field display.