Multi-Distribution Knowledge Distillation With Wasserstein Distance for Efficient 3D Human Reconstruction

Recently, transformer-based Human Mesh Recovery (HMR) from monocular images has achieved remarkable progress by effectively modeling long-range dependencies among body parts. However, the large-scale transformer architectures that drive these state-of-the-art results impose prohibitive computational and storage demands, severely limiting their deployment in real-time or resource-constrained scenarios. Existing knowledge distillation methods offer a potential solution but typically rely on a single distribution, either final outputs or intermediate features, thus failing to capture the diverse and complementary representations inherent in transformer-based HMR models, such as attention patterns. To address these limitations, we propose Multi-dIstribution kNowledge Distillation (MIND), a framework tailored for transformer-based HMR that transfers knowledge from multiple complementary distributions: final outputs for high-level task knowledge, intermediate features for visual representation learning, and internal attention heatmaps to preserve spatial focus patterns. Furthermore, motivated by the inherent conceptual similarity between Generative Adversarial Networks (GANs) and knowledge distillation, we introduce a Wasserstein distillation strategy that leverages the Wasserstein-GAN framework to robustly align these distributions between the teacher and student models. Extensive experiments on Human3.6M, 3DPW, and COCO demonstrate that MIND not only surpasses existing distillation baselines, reducing PA-MPJPE by 3.1mmon 3DPW, but also matches or exceeds the state-of-the-art teacher model HMR2.0 [1] while using only 17.8% of its parameters and 13.7% of its computation cost.