Part-aware 3D generation

MEGAPARTS

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

Manwen Liao*Xinyu Lian*Jian Mao*Kaixu ChenLi LuoJinghao YanWanshui GanQiao YuWeitian ZhangChunhua ShenGuang Chen†Bo Dai†Xudong XuZhaoyang Lyu†

Research team

* Equal contribution † Corresponding author

MegaParts is the base model for MeshCoder 1.5.

The University of Hong Kong

Shanghai Artificial Intelligence Laboratory

Fudan University

Tongji University

University of Science and Technology of China

Shanghai Jiao Tong University

Zhejiang University

Shanghai Innovation Institute

The idea

From monolith to assembly

Complex objects are assembliesNot single shapes

Part-aware generation makes 3D assets controllable, editable, and ready to articulate.

Existing methods lose efficiency and detail as the number of components grows. MegaParts represents an object as one structured sequence: a global bounding box, localized part boxes, and compact shape tokens for every component.

Its adaptive tokenizer assigns more capacity only where geometric complexity demands it, allowing long-context generation to scale without abandoning fine structure.

300: parts in a single generated object
256K: tokens in the unified generation sequence
Adaptive: token budgets matched to part complexity

System architecture

A factory line for geometry

From global boundsTo complete geometry

The system learns compact discrete geometry tokens, then predicts global layout, part layout, and local shape as one coherent sequence.

MegaParts method diagram showing token-efficient vector-quantized representation learning and part-aware autoregressive generation — Pipeline / overview Predict object bounds, part bounds, and local shape in one structured autoregressive sequence.

Token-efficient representation

A causal vector-quantized autoencoder concentrates essential geometry in prefix tokens, so each part can be reconstructed at an appropriate token length.

Part-aware generation

A long-context transformer predicts object bounds, part bounds, and local shape tokens in order, preserving explicit structure throughout generation.

Adaptive representation

Rate-distortion aware

Spend tokensWhere geometry gets difficult

Simple components need only short sequences. Intricate parts receive more tokens to preserve shape, normals, and bounding-box fit. This tradeoff makes hundreds of individually modeled parts practical in one object.

Compact partsDetailed partsHighest complexity

Generated robot with its parts colored by token count and reconstructions at different token budgets — Token allocation16 → 4096

Evaluations

Structure you can direct

Structured 3D generationWith explicit part-level control

Generate from language, reshape the global layout, specify individual part bounds, or use the explicit decomposition for localized edits and articulation.

Language conditioned generation

MegaParts follows detailed prompts while retaining explicit component structure, producing coherent geometry across objects ranging from cameras to buildings.

Text-conditioned generation comparison between MegaParts and four baseline systems

Controllable layouts

Global and part-level boxes act as spatial instructions. Changing those bounds alters the assembled asset while preserving semantic and geometric coherence.

Objects generated under different global bounding-box proportions

Camera part localization comparison showing ground truth, MegaParts, XPart, and FullPart

Geometry with working parts

Explicit decomposition exposes small components, articulation, and internal structure that holistic representations routinely flatten or omit.

Generated piano and house with highlighted articulated and editable parts — 01 / ArticulationEditable components stay structurally connected.

Three generated dishwashers with doors open and detailed interior racks and dishes — 02 / Interior geometryNested racks, dishes, and doors remain explicit.

Two complex generated structures composed of many colored parts — 03 / Composed scenesMany local pieces resolve into coherent structures.

Citation

BibTeX

@misc{liao2026megapartsscalingpartaware3d,
  title={MegaParts: Scaling Part-Aware 3D Object Generation to 300 Parts via Token-Efficient Autoregressive Modeling},
  author={Manwen Liao and Xinyu Lian and Jian Mao and Kaixu Chen and Li Luo and Jinghao Yan and Wanshui Gan and Qiao Yu and Weitian Zhang and Chunhua Shen and Guang Chen and Bo Dai and Xudong Xu and Zhaoyang Lyu},
  year={2026},
  eprint={2608.14783},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2608.14783},
}