Core Concepts
Human-LRM presents a template-free large reconstruction model for feed-forward 3D human digitalization from a single image, guided by diffusion models.
Abstract
The article introduces Human-LRM, a novel approach for reconstructing 3D humans from a single image. It addresses the limitations of existing methods in capturing fine geometry and appearance details, achieving generalization across datasets. The model leverages dense novel views generated by a conditional diffusion model to enhance the fidelity of full-body human reconstructions. By training on extensive datasets and using a three-stage approach, Human-LRM outperforms previous methods significantly in terms of geometry and appearance quality.
Stats
Trained on more than 10K identities.
Utilizes multi-view RGB data and 3D scans.
Achieves enhanced generalizability across various scenarios.
Quotes
"Our method does not rely on a human mesh template such as SMPL and thus does not suffer from this problem."
"Our results demonstrate exceptional generalizability to challenging cases such as people in difficult poses."