Make-It-Animatable: An Efficient Framework for Authoring Animation-Ready 3D Characters

TL;DR

Make-It-Animatable presents a data-driven, template-free framework that converts any 3D humanoid input (mesh or 3D Gaussian splats) into an animation-ready model in under a second. It fuses a particle-based shape autoencoder with a coarse-to-fine representation and a structure-aware, next-child-bone transformer to predict high-quality rigging (blend weights), bones (head/tail positions), and pose-to-rest transformations, robust to non-standard shapes and poses. The approach leverages low-rank dynamics, geometry-aware attention, and body-prior losses to ensure accurate, stable deformations across diverse characters, including extras bones like ears or tails. Extensive experiments on Mixamo and VRoid datasets show superior accuracy and speed compared with auto-rigging and template-based methods, with strong ablations validating each component’s contribution and showing practical potential for real-time animation pipelines.

Abstract

3D characters are essential to modern creative industries, but making them animatable often demands extensive manual work in tasks like rigging and skinning. Existing automatic rigging tools face several limitations, including the necessity for manual annotations, rigid skeleton topologies, and limited generalization across diverse shapes and poses. An alternative approach is to generate animatable avatars pre-bound to a rigged template mesh. However, this method often lacks flexibility and is typically limited to realistic human shapes. To address these issues, we present Make-It-Animatable, a novel data-driven method to make any 3D humanoid model ready for character animation in less than one second, regardless of its shapes and poses. Our unified framework generates high-quality blend weights, bones, and pose transformations. By incorporating a particle-based shape autoencoder, our approach supports various 3D representations, including meshes and 3D Gaussian splats. Additionally, we employ a coarse-to-fine representation and a structure-aware modeling strategy to ensure both accuracy and robustness, even for characters with non-standard skeleton structures. We conducted extensive experiments to validate our framework's effectiveness. Compared to existing methods, our approach demonstrates significant improvements in both quality and speed. More demos and code are available at https://jasongzy.github.io/Make-It-Animatable/.

Figures (21)

• Figure 1: Given a 3D character represented by mesh or 3D Gaussian Splats with arbitrary pose and shape, our framework can produce high-quality results of rigging, skinning, and pose resetting for it within one second. The output 3D model is fully animatable with a fine-grained skeleton and optional bone topology of extra body structures.
• Figure 2: Pipeline of the proposed framework. Given an input 3D character, we produce high-quality blend weights, bones, and pose-to-rest transformation for it, so that any animation is within easy reach. First, we coarsely localize the joints with a pre-trained lite version of this framework, which helps enable a finer shape representation. Then the shape is encoded into a neural field with a particle-based autoencoder. The decoding process involves spatial and learnable queries for different animation assets. Finally, the structure-aware modeling of bones is proposed to better align the predictions with skeleton topology priors.
• Figure 3: Pipeline of the proposed structure-aware transformer. The per-bone shape-aware embedding is first added with its parent bone's latent, which is encoded from the autoregressive outputs (in inference) or the ground-truth values (in training). The summation is then fused with the ancestral bones' features via the masked causal attention. Eventually, bone attributes are decoded from the output shape- and structure-aware embeddings. In inference, the whole process follows the paradigm of next-child-bone prediction.
• Figure 4: Comparison with Meshy meshy and Tripo tripo. We feed them the same image as reference and compare the performance based on their generated 3D models respectively. The blend weights of two joints, i.e., Left Shoulder and Right Leg, are visualized. Given that these baselines can only apply preset motions and their rest-pose models cannot be exported, we apply a similar "running" sequence to all the methods for fair comparison. The T-pose models predicted by our method are also included as the front-view animating results.
• Figure 5: Comparison with RigNet xu2020rignet. We visualize the blend weights of selected joints and manually deform them to assess the impact of rigging quality on skinning results.