MegaParts: Scaling Part-Aware 3D Object Generation to 300 Parts via Token-Efficient Autoregressive Modeling
Manwen Liao, Xinyu Lian, Jian Mao, Kaixu Chen, Li Luo, Jinghao Yan, Wanshui Gan, Qiao Yu, Weitian Zhang, Chunhua Shen, Guang Chen, Bo Dai, Xudong Xu, Zhaoyang Lyu
cs.CV, cs.GR
2026-08-15
A HKU and Shanghai AI Lab team assigns each 3D part an adaptive-length discrete token budget via rate-distortion optimization, then autoregressively generates objects with up to 300 parts and 256k-token sequences, halving part-level Chamfer distance versus FullPart and XPart.
Editable 3D assets need semantic parts, not one fused mesh: a chair is four legs, a seat, a backrest, and only parts can be swapped or joint-bound in a game engine. Existing part-aware generation hits a hard ceiling: token length and memory explode as parts multiply, so a few dozen parts is the practical limit, and objects like vehicles or machinery with hundreds of parts are out of reach.
The root cause is representational. Diffusion models denoise all parts jointly in a flat continuous latent space, where both detail fidelity and cross-part consistency grow costlier with every added part. The autoregressive route fits the domain naturally (parts are assembled in sequence anyway), but only if per-part token counts compress, or a 300-part sequence becomes untrainably long.
MegaParts gives every part an adaptive token budget through three interlocking pieces:
Training data is self-curated: nearly 10M part meshes for the VQ-VAE and roughly 440K part-annotated objects for the generator (340K public, 100K private), with objects beyond 300 connected components iteratively merged down to 300. The tokenizer initializes from Roblox's Cube v0.5 with query tokens and codebook quadrupled (4,096 queries, 65,536 entries).
| Method | Part CD (×10⁻²)↓ | Part IoU↑ | BBox IoU↑ |
| FullPart | 8.59 | 0.41 | 0.76 |
| XPart | 8.01 | 0.52 | 0.59 |
| MegaParts | 3.01 | 0.63 | 0.94 |
This is the part-bounding-box-conditioned setting where all methods receive the same ground-truth boxes. On text-conditioned generation, FID lands at 43.40 versus SAR3D 94.69, Cube 55.58, TRELLIS-text 54.81.
Reconstruction leads as well: at a matched 1024-token budget, part-level CD is 0.12×10⁻³ against Cube's 3.95×10⁻³, a 33× gap, and quality climbs monotonically from 512 to 4096 tokens, confirming a coarse-to-fine structure. The framework handles 300 parts and 256k-token sequences, while baselines like like FullPart and XPart were built for objects an order of magnitude simpler.
For 3D content production this is the first part-aware generator operating at real industrial complexity, namely vehicles and machinery rather than demo chairs. The adaptive-token-budget idea travels beyond 3D to any generation task where sample complexity varies wildly, video blocks or audio segments included. And it is real evidence in the autoregressive-versus-diffusion debate: compressed discrete tokens don't just save compute, they raise the achievable fidelity ceiling.
In fairness, the text-conditioned baselines are not specialized part-aware methods; the harder comparison is the box-conditioned table, and the two read together make the case.
The authors name two: long-context autoregressive generation is expensive at both training and inference, especially for highly complex objects; and the system models geometry only, with no textures, materials, semantic annotations, or physical and functional relations between parts (which joint rotates, what bears load). The last gap leaves real distance to drop-in engine use.
Two more from a close read. The main benchmark, PartObjaverse-Tiny, skews low-part-count, so the 300-part claim rests on the appendix's high-part-count study rather than the main tables. And the paper itself cautions that Table 3's metrics heavily penalize baselines whose parts fuse into one mesh, a setup structurally friendly to part-by-part modeling, so those numbers measure adherence to the prescribed decomposition, not cross-modality appearance quality.