KungFuCap: monocular martial 3D reconstruction via Gaussian splatting and diffusion-guided view synthesis

Martial arts are a treasured but fragile cultural heritage. To safeguard their intricate motions with the simplest possible setup, we introduce KungFuCap, a monocular-video framework that yields high-fidelity, fully animatable 3D human models of martial arts sequences. For each frame from a single-view recording, a pose-conditioned Stable Diffusion model, fine-tuned with SMPL-driven control signals, synthesizes a geometrically consistent 180° complementary novel view - ensuring temporal stability via frame-to-frame diffusion guidance, pose fidelity through joint-level keypoint alignment, and clean background isolation via inpainting masks to mitigate style drifts or occlusion artifacts. The resulting dual-view clip is then processed by a SMPL-guided 3D Gaussian Splatting optimizer: SMPL initializes Gaussian points on its mesh vertices for robust starting geometry, enforces body topology constraints during optimization to prune extraneous voxels and redistribute points, and leverages inter-frame deformation fields to mitigate motion-induced discontinuities in rapid martial arts sequences, thereby enhancing temporal coherence. Extensive evaluations, including generation quality metrics, ablation studies on pose errors and SMPL constraints for geometric/temporal fidelity, on our newly collected KungFuCap dataset and the public DNA-Rendering benchmark demonstrate that the proposed method consistently outperforms prior single-view approaches across multiple perceptual and geometric metrics, providing an effective foundation for the digital preservation of martial arts.