RigAnything: Template-Free Autoregressive Rigging for Diverse 3D Assets

1. Template-Free Autoregressive Rigging. A transformer-based model that sequentially generates skeletons without predefined templates, enabling automatic rigging across diverse 3D assets through probabilistic joint prediction and skinning weight assignment.
2. Support Arbitrary Input Pose. Generates high-quality skeletons for shapes in any pose through online joint pose augmentation during training, eliminating the common rest-pose requirement of existing methods and enabling broader real-world applications.
3. Fast Rigging Speed. Achieves 20x faster performance than existing template-based methods, completing rigging in under 2 seconds per shape.

Pipeline. The input shape and previously predicted skeleton sequence are tokenized separately and processed through autoregressive transformer blocks with a hybrid attention mask. Afterward, a skinning module decodes skinning weights, a joint diffusion module samples the next joint position, and a connectivity module predicts the joint’s connection based on the sampled position.