Organizations: 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
Part-aware 3D object generation is essential for graphics applications such as controllable modeling, editing, and articulation, where objects are represented as coherent assemblies of semantic parts. However, existing part-aware generation methods, do not scale well to highly complex objects. As the number of parts increases, generating detailed geometry becomes prohibitively expensive in token length and memory. We introduce MegaParts, a scalable autoregressive 3D generation framework to address this challenge by combining structured sequence modeling with a token-efficient vector-quantized shape tokenizer. Our tokenizer learns discrete latent representations for part-level geometry by minimizing token usage subject to high-fidelity reconstruction, enabling adaptive-length tokenization based on geometric complexity. On top of this compact representation, we train a large language model to generate object bounding boxes, part bounding boxes, and part shape tokens within a unified structured sequence. Combined with efficient long-context training strategy, our token-efficient formulation scales to objects with up to 300 parts and sequence lengths up to 256k tokens. This substantially extends the scale of part-aware 3D generation while preserving compositional structure and enabling fine-grained part-level control. Our method achieves higher mesh quality than baseline autoregressive and diffusion models, showing that compressed discrete part tokens improve not only scalability but also the achievable fidelity of generated geometry. These results suggest that LLM native token-efficient autoregressive modeling is a compelling alternative to diffusion for large-scale part-aware 3D generation. The project page is available at https://expmaster.github.io/megaparts_webpage.
Figures & tables
Figure 1. MegaParts can generate complex objects with up to 300 parts. Like factory assembly lines, it autoregressively manufactures part meshes and then assembles them into complete 3D objects, substantially extending the scope and capability of part-aware 3D generation while maintaining geometry quality. figure description
Figure 2. Overview of MegaParts . Our framework consists of two stages. Top: we learn a token-efficient vector-quantized representation for part geometry. Bottom: we train a part-aware autoregressive generator that represents a 3D object as a structured sequence conditioned on a text prompt and optional object- and part-level bounding boxes. If bounding boxes are absent, the model first predicts the object box and part boxes, followed by autoregressive generation of local shape tokens for each part. The decoded part geometries are then assembled into the final part-aware 3D asset.
Figure 3. Given a complex object consisting of multiple parts, we adaptively assign each part a suitable token budget. During inference of the VQ-VAE we reconstruct part meshes under different token budgets and determine the optimal token number based on a carefully designed rate distortion metric. Our method enables flexible representation volume for part meshes with different complexity while introducing minimum visual degradation.
Part-level
Object-level
Method
CD( ×10−3 ) ↓
NC ↑
CD( ×10−3 ) ↓
NC ↑
Cube (1024)
3.95
0.89
1.89
0.85
Ours (1024)
0.12
0.93
1.52
0.89
Ours (512)
0.35
0.90
1.92
0.80
Ours (2048)
0.07
0.97
1.32
0.93
Ours (4096)
0.06
0.98
1.25
0.94
Table 1. Quantitative reconstruction results of VQ-VAEs at both the part and object levels. At the matched budget of 1024 tokens, our model consistently outperforms Cube. Moreover, increasing the token budget from 512 to 4096 yields progressively better reconstruction quality, indicating a clear coarse-to-fine trend in the learned discrete representation.
SAR3D
Cube
TRELLIS-text
ShapeLLM-Omni
Ours
FID ↓
94.69
55.58
54.81
70.99
43.40
CLIP ↑
0.21
0.26
0.25
0.23
0.27
Table 2. Generation evaluation on PartObjaverse-Tiny with normal-map–based semantic alignment. Our method achieves better performance over text-conditioned generation baselines.
Metric
FullPart
XPart
Ours
Part CD( ×10−2 ) ↓
8.59
8.01
3.01
Part IoU ↑
0.41
0.52
0.63
BBox IoU ↑
0.76
0.59
0.94
Table 3. Results for part-bounding-box-conditioned shape generation. Our method achieves the best performance among the compared baselines.
Setting
Scope
512
1024
2048
4096
CD( ×10−3 ) ↓
NC ↑
CD( ×10−3 ) ↓
NC ↑
CD( ×10−3 ) ↓
NC ↑
CD( ×10−3 ) ↓
NC ↑
Ours (full)
Part
0.35
0.90
0.12
0.93
0.07
0.97
0.06
0.98
Object
1.92
0.80
1.52
0.89
1.32
0.93
1.25
0.94
w/o phased VQ-VAE training
Part
0.42
0.77
0.22
0.82
0.18
0.83
0.18
0.88
Object
3.05
0.76
2.88
0.81
2.51
0.82
2.17
0.87
w/o causal attn.
Part
0.66
0.89
0.22
0.90
0.13
0.93
0.10
0.95
Table 4. Ablation study of phased training and causal attention in our VQ-VAE. We report part-level and object-level reconstruction metrics under different token budgets. The full model performs best across all settings, while removing phased training degrades reconstruction quality and removing causal decoder attention weakens the benefit of larger token budgets.
Figure 4. Fine-grained part-aware generation supports diverse downstream applications. Left: articulated object creation with detailed part decomposition. Right: secondary creation guided by localized part priors.
Figure 5. Qualitative comparison between our text conditioned generation results against baselines. Our results show superior alignment with given text prompt, demonstrating our model’s strong prompt following ability.
Figure 6. Results of global bounding-box-conditioned generation. As the input bounding-box configuration changes, our model automatically synthesizes corresponding part layouts and meshes. The generated results align closely with the prescribed bounding boxes, demonstrating effective structural control.
Figure 7. Results of part bboxes conditioned generation. Our model produces semantically reasonable part meshes with high quality geometry.
Figure 8. Comparison between our results with image conditioned part-aware generation methods. We use Nano Banana to refine rendered images of our text conditioned generation results and use them as input to baselines. Our resulted geometry surpass baselines and our part segmentations are much more structured and detailed than other methods.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 9. Example of the text processed by Qwen3 .
Figure 10. Distribution of the number of parts for meshes in the dataset.
Figure 11. Distribution of the number of parts for meshes with more than 50 parts in the dataset.
Setting
BBox IoU ↑
w/o object bbox prediction
0.89
w/ object bbox prediction
0.93
Appendix
Table 5. Effect of object-level bounding-box prediction supervision on global controllability.
Ordering
CLIP ↑
FID ↓
z – x – y
0.27
43.40
Connectivity-graph BFS
0.25
41.68
Appendix
Table 6. Sensitivity to part ordering on the test set.
CLIP ↑
FID ↓
Method
[1,50)
[50,100)
[100,200)
[200,300]
[1,50)
[50,100)
[100,200)
[200,300]
SAR3D
0.24
0.23
0.22
0.23
124.0
118.9
119.9
116.4
Cube
0.26
0.26
0.27
0.26
93.5
99.5
95.0
100.9
TRELLIS-text
0.26
0.26
0.26
0.26
99.6
98.7
95.7
103.4
ShapeLLM-Omni
0.24
0.25
0.24
0.24
103.2
100.4
99.8
105.7
Ours
0.26
0.26
0.27
0.27
46.3
55.0
64.9
73.2
Appendix
Table 7. Text-to-3D generation on held-out sets stratified by ground-truth part count.
Method
Part range
Part CD ↓
Part IoU ↑
BBox IoU ↑
Success rate ↑
FullPart
[1,50)
8.3
0.38
0.66
0.89
[50,100)
10.8
0.26
0.37
0.17
[100,200)
27.2
0.13
0.33
0.05
[200,300]
–
–
–
0.00
XPart
[1,50)
7.7
0.45
0.42
0.82
[50,100)
–
–
–
0.00
Appendix
Table 8. Bounding-box-conditioned generation stratified by part count. Part CD is reported in units of 10−2 , matching Table 3 of the main paper. A dash "-" indicates that no valid result is available due to OOM. Metrics other than success rate are averaged over successful outputs.
Part range
[1,50)
[50,100)
[100,200)
[200,300]
AR generation
335
2066
3746
6735
Decoding + marching cubes
51
413
502
1076
Appendix
Table 9. Inference-time breakdown in seconds across part-count ranges.
Method
Text alignment ↑
Geometry quality ↑
SAR3D
0.00
0.00
Cube
0.18
0.16
TRELLIS-text
0.16
0.13
ShapeLLM-Omni
0.02
0.02
Ours
0.64
0.69
Appendix
Table 10. User studies for text alignment and geometry quality.
Figure 12. Three dishwashers with high quality interior details generated by our model.
Figure 13. More visual comparisons between our model and baselines on text conditioned 3D generation.