MegaParts: Scaling Part-Aware 3D Object Generation to 300 Parts via Token-Efficient Autoregressive Modeling
MegaParts claims part-aware 3D generation can scale to objects with up to 300 parts and 256k-token sequences.
The system compresses part-level geometry into adaptive discrete tokens, then uses an autoregressive language model to generate object boxes, part boxes, and part shape tokens in one structured sequence. The paper says this cuts token and memory pressure while preserving fine-grained part control. Its reported results beat baseline autoregressive and diffusion models on mesh quality. HF Daily Papers' note
The system compresses part-level geometry into adaptive discrete tokens, then uses an autoregressive language model to generate object boxes, part boxes, and part shape tokens in one structured sequence. The paper says this cuts token and memory pressure while preserving fine-grained part control. Its reported results beat baseline autoregressive and diffusion models on mesh quality. HF Daily Papers' note
score 5