IDiffPose: Equilibrium Substitution for Lightweight Diffusion-Graph-Based Pose Estimation

Recent diffusion-graph frameworks such as DiffPose have advanced monocular 2D-to-3D pose lifting by coupling graph convolutional networks (GCNs) with denoising diffusion models. However, their depth is still bounded by the number of explicit attention-GCN blocks, limiting the receptive field under a fixed parameter budget. We propose IDiffPose, short for Implicit Diffusion Pose, an equilibrium substitution that replaces a selected deep block with a weight-tied operator whose representation is computed via a small number of unrolled fixed-point iterations. The training objective remains standard DDPM-style denoising, inference uses a deterministic DDIM sampler, and the forward process injects heteroscedastic Gaussian noise scaled by 2D heat-map uncertainty. Because the same weights are re-used across iterations, the model attains large effective depth with a compact parameter footprint; when non-equilibrium blocks are frozen, the number of trainable parameters drops substantially. On Human3.6M, substituting only the middle block improves MPJPE from 31.55 to 31.10 mm and P-MPJPE from 24.72 to 24.47 mm with 167,521 trainable parameters (vs. 1,025,674 explicit), while making all blocks implicit reaches 30.90/24.51 mm. These results show equilibrium substitution is a practical drop-in upgrade for diffusion-based pose lifting, yielding consistent accuracy with favorable efficiency controls.