FuguReport

LUNA: Learning Universal 3D Human Animation Beyond Skinning

Authors Peng Li, Rawal Khirodkar, Junxuan Li, Yuan Dong, Chen Cao, Yuan Liu, Wenhan Luo, Yike Guo, Shunsuke Saito
Affiliations Meta / The Hong Kong University of Science and Technology
Categories Method / 3D Animation / End-to-end 3D animatable model, Application / Human Modeling / 3D human animation beyond skinning, Evaluation / Visual Fidelity / Photorealistic visual fidelity comparison
License CC BY 4.0

Abstract Overview

This paper presents LUNA, an end-to-end 3D human animation framework that avoids Linear Blend Skinning (LBS) and instead maps diverse 2D driving inputs directly to deformations of a canonical 3D Gaussian avatar. The method reconstructs canonical Gaussians from a small set of identity images, then uses a transformer-based implicit neural animator that separates global rigid motion from local non-rigid deformation. To address the ambiguity of 2D-to-3D lifting, the training scheme combines photometric rendering losses with LBS-teacher distillation and supports both labeled data and large-scale unlabeled in-the-wild videos. The reported goal is to preserve photorealism while reducing pose-induced artifacts and enabling cross-identity, cross-modality animation from images, keypoints, sketches, and out-of-domain characters.

Novelty

The main novelty is an LBS-free, end-to-end 3D animatable model that supports implicit 2D driving across multiple modalities, rather than relying on explicit body fitting or parametric pose control. The paper also introduces a hybrid supervision strategy that uses LBS only as a soft structural prior while learning a direct 2D-to-3D deformation mapping.

Results

Experimentally, LUNA outperforms the compared methods on the non-rigid Cloth10K benchmark, reaching 22.072 PSNR, 0.027 L1, and 0.131 LPIPS, while remaining comparable to strong feed-forward baselines on NeuMan. It also improves temporal stability substantially, reducing MSJ to 0.0032 and MAE to 0.0225, including a reported 4.5x MSJ reduction over the strongest baseline (MV-LHM). In cross-identity driving on NeuMan, its performance stays close to self-reenactment and achieves the lowest reported normalized 2D keypoint error among compared methods.

Key Points

  1. LUNA directly animates canonical 3D Gaussians from heterogeneous 2D controls, including RGB images, 2D keypoints, sketches, and unseen character images, without explicit body fitting or preprocessing.
  2. The architecture disentangles global rotation/translation from local per-Gaussian residual deformation, aiming to capture both articulated motion and fine non-rigid effects such as loose clothing.
  3. Hybrid supervision is important to performance: LBS-guided distillation prevents geometric collapse, and multiview fine-tuning improves spatial compactness and visual quality in the ablations.

References

This page was created using generative AI such as GPT-5, Claude Opus 4, Gemini 3, Gemini 3.1 Flash Image, and their higher-end successor versions. No guarantee can be made regarding its contents.