MegaParts: Scaling Part-Aware 3D Object Generation to 300 Parts via Token-Efficient Autoregressive Modeling

MegaParts is a new 3D generation framework aimed at a problem that matters immediately to asset-heavy pipelines: part-aware modeling stops scaling once objects get too complex. The system is built to generate coherent assemblies of semantic parts, which is useful for controllable modeling, editing, and articulation in games and tools workflows.

The key change is a token-efficient vector-quantized tokenizer for part geometry. Instead of spending huge token budgets on detailed meshes, it learns discrete latent representations that adapt token count to geometric complexity while still reconstructing shapes faithfully. That compact representation then feeds a large language model that generates object bounding boxes, part bounding boxes, and part shape tokens in one structured sequence.

The headline numbers are the ones developers will notice: support for objects with up to 300 parts and sequence lengths up to 256k tokens. That scale is a big deal for anyone trying to generate modular props, mechanical assets, or articulated characters without the usual memory blowups that make long-context generation impractical.

The broader implication is that compressed discrete part tokens may be a viable alternative to diffusion for large-scale part-aware 3D generation. The system reports higher mesh quality than baseline autoregressive and diffusion approaches, suggesting that token efficiency is not just a cost-saving trick but a path to better fidelity and more usable control for downstream content creation.