Part-aware 3D object generation, where objects are represented as assemblies of semantic parts, is crucial for applications like controllable modeling, editing, and articulation. However, existing methods struggle to scale to highly complex objects, as increasing part counts lead to prohibitive token lengths and memory usage.
To address this, researchers from Hugging Face introduce MegaParts, a scalable autoregressive framework that combines structured sequence modeling with a token-efficient vector-quantized shape tokenizer. The tokenizer learns discrete latent representations for part-level geometry, minimizing token usage while maintaining high-fidelity reconstruction, enabling adaptive-length tokenization based on geometric complexity.
On top of this compact representation, a large language model is trained to generate object bounding boxes, part bounding boxes, and part shape tokens within a unified structured sequence. Combined with an efficient long-context training strategy, the formulation scales to objects with up to 300 parts and sequence lengths up to 256k tokens, substantially extending the scale of part-aware 3D generation while preserving compositional structure and enabling fine-grained part-level control.
The method achieves higher mesh quality than baseline autoregressive and diffusion models, demonstrating that compressed discrete part tokens improve both scalability and fidelity. The results suggest that LLM-native token-efficient autoregressive modeling is a compelling alternative to diffusion for large-scale part-aware 3D generation.
For more details, visit the project page or the paper on arXiv.