🍐 PEAR :Pixel-aligned Expressive humAn mesh Recovery

Reconstructing detailed 3D human meshes from a single in-the-wild image remains a fundamental challenge in computer vision. Existing methods often suffer from slow inference, produce only coarse body poses, and exhibit misalignments or unnatural artifacts in fine-grained regions such as the face and hands. These issues make current approaches difficult to apply to downstream tasks. To address these challenges, we propose PEAR—a unified framework for human mesh recovery and rendering. PEAR explicitly tackles three major limitations of existing methods: slow inference, inaccurate localization of fine-grained human pose details, and insufficient photometric supervision during self-reconstruction. Specifically, we train a clean, unified ViT-based model capable of recovering expressive 3D human geometry (SMPLX + FLAME) from a single image without cropping any specific body parts. This preprocessing-free design enables real-time inference at over 100 FPS. In addition, we integrate the model with a neural renderer to jointly optimize geometry and appearance, significantly improving the reconstruction accuracy of fine-grained human geometry and producing higher-quality rendering results. Finally, we propose a modular data annotation strategy that enriches the training data and enhances the robustness of the model. Extensive experiments on multiple benchmark datasets demonstrate that our approach achieves substantial improvements in both geometric reconstruction accuracy and rendering quality.

We propose PEAR, a unified framework for real-time expressive 3D human mesh recovery. It is the first method capable of simultaneously predicting SMPL-X and FLAME parameters at 100 FPS, and it exhibits strong robustness to diverse human-part image inputs, accurately handling face-only, upper-body, and full-body images.