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

1 Introduction

• Part-aware generation supports controllable modeling, localized editing, and articulation by representing objects as structured assemblies rather than holistic shapes.The passage notes that structured, localizable representations also align with scalable 3D generation.Source: §1 · p. 2
• Existing part-aware diffusion pipelines rely on predefined segmentations or denoised layouts, and diffusion scalability limits hinder generation of objects with hundreds of parts.The cited image-generation studies report that naive diffusion models struggle as token counts increase.Source: §1 · p. 2
• MegaParts addresses the fidelity-efficiency tradeoff with a vector-quantized tokenizer that assigns adaptive-length part tokens according to geometric complexity.A coding-rate distortion objective determines each part’s token budget, concentrating geometric information in prefix tokens so simple parts use fewer tokens while intricate parts retain capacity.Source: §1 · p. 2
• The autoregressive model predicts an object bounding box, then each part’s bounding box and shape tokens, preserving explicit structure and enabling object- and part-level control.Efficient training strategies support long-context sequences.Source: §1 · p. 2
• MegaParts generates higher-quality meshes than conventional autoregressive baselines and diffusion methods while scaling to objects with 300 parts and 256k-token sequences.The framework generates object-level boxes, part boxes, and part geometry in a unified sequence.Source: §1 · pp. 2–3

2.1 Representations for Structured 3D Shape Modeling

• VecSet latent pools can lose intricate details [Chen et al. 2025d; Li et al. 2025b,c; Wu et al. 2024], while voxel grids allocate capacity by fixed spatial rules [Chen et al. 2025b; He et al. 2025; Hunyuan3D 2025b; Lai et al. 2025; Luo et al. 2026; Deng et al. 2025b].These limitations motivate part-aware representations that vary capacity by part complexity and separate global structure from part-level representation.Source: §2.1 · p. 3

2.2 High-Fidelity Component-Aware 3D Generation Pipeline

• Diffusion-based part-aware methods typically target fewer than 50 parts, as detailed geometry and coherent inter-part dependencies drive up token, memory, and computation costs.One stream generates geometry from prescribed part decompositions [Chen et al. 2025c; Lin et al. 2025; Tang et al. 2025b; Yan et al. 2025; Yang et al. 2025a; Zhang et al. 2025]; another jointly predicts layouts and geometry [Dai et al. 2025; Ding et al. 2025; Wang et al. 2025b].Source: §2.2 · p. 3

2.3 Autoregressive Models for 3D Asset Modeling

• Autoregressive 3D methods generate low-poly meshes directly or use structured discrete latents, including triplane and octree representations [Chen et al. 2024; Hao et al. 2024; Wei et al. 2025].Other examples include SAR3D [Chen et al. 2025a], Cube [Roblox 2025], OctGPT [Wei et al. 2025], and OAT [Deng et al. 2025a].Source: §2.3 · p. 3
• No prior autoregressive method explores long-context modeling for highly complex part-aware assets that need global coherence and fine-grained geometry across many components.Source: §2.3 · p. 4

3.1 Causal Autoencoder for Signed Distance Field Modeling

• Causal attention in the VQ-VAE encoder and decoder encourages early latent tokens to capture essential geometry, making reconstructions more robust to prefix truncation.The encoder updates learnable latent queries through cross-attention to input tokens or causal self-attention, then applies spherical vector quantization; the decoder uses causal self-attention and predicts signed distance values for mesh extraction.Source: §3.1 · p. 4

3.2 Token-Efficient Stable VQ-VAE with Phased Optimization

• Random-prefix reconstruction trains the VQ-VAE to concentrate geometry in short prefixes, while alternating encoder-only and decoder-only phases stabilize code assignments.The training scheme combines optimal-transport quantization and Stochastic Gradient Shortcut (SGS) to promote efficient codebook utilization.Source: §3.2 · p. 4
• The VQ-VAE objective combines reconstruction, near-surface eikonal regularization, surface-normal alignment, and vector-quantization losses.The eikonal constraint is applied within a near-surface band rather than throughout the volume.Source: §3.2 · p. 4

3.3 Adaptive-Length Representation via Rate–Distortion Optimization

• Each semantic or structural part is encoded in a normalized local frame, and rate–distortion optimization selects a prefix length balancing reconstruction fidelity against token use.The component bounding box is stored separately to recover its global position and scale; the distortion objective includes a bounding-box IoU term to avoid small floaters.Source: §3.3 · p. 5

3.4 Component Aware Generation With Structured Autoregressive Modeling

• The global representation serializes all component layouts before their geometry tokens, making the decomposition explicit and suitable for part-level control.Components are ordered by bounding-box-center position, while each component’s prefix tokens progress from coarse geometry to finer detail.Source: §3.4 · pp. 5–6

3.2 Token-Efficient Stable VQ-VAE with Phased Optimization

• At 1024 tokens, MegaParts outperforms Cube, and reconstruction quality improves progressively as the VQ-VAE budget rises from 512 to 4096 tokens.Source: Table 1 · p. 6

4.1 Evaluation Protocol

• The pipeline is trained on curated public and private data, and evaluated for reconstruction and generation on PartObjaverse-Tiny [Yang et al. 2024], with high-part-count results in Supplementary Material C1.The dataset is described in Supplementary Material A2.Source: §4.1 · p. 6

4.2 Evaluation of VQ-VAE

• VQ-VAE reconstruction is assessed with Chamfer Distance and Normal Consistency after meshes are normalized to [−1, 1]3 and 100,000 points are sampled; Cube [Roblox 2025] is the primary baseline.The main comparison uses baseline training token counts, with additional evaluation across token-count settings; reconstruction metrics are presented in Table 1.Source: §4.2 · pp. 6–7

4.3 Evaluation of the Generation Model

• Text-conditioned generation is evaluated with FID and CLIP score on rendered normal maps; layout-conditioned generation reports part-level IoU, part-level Chamfer Distance, and BBox IoU.For layout-conditioned comparisons, FullPart [Ding et al. 2025] and XPart [Yan et al. 2025] receive ground-truth part boxes to control for structural layout; qualitative text-conditioned comparisons appear in Figure 5.Source: §4.3 · p. 7
• Part-level metrics penalize fused meshes and primarily measure adherence to prescribed part decomposition and layout, so Table 3 should be read alongside qualitative comparisons.The authors caution that Table 3 is not a fully controlled cross-modality appearance benchmark.Source: §4.3 · p. 7
• Image-conditioned methods [Lin et al. 2025; Tang et al. 2025b; Yang et al. 2025d] are compared qualitatively because their different input modality precludes a straightforward, fully fair numerical protocol.Representative comparisons are shown in Figure 8.Source: §4.3 · p. 7

4.4 Ablation Study

• Dropping phased VQ-VAE optimization hurts reconstruction metrics, while removing causal attention also reduces performance, supporting staged optimization and a causal decoder for adaptive token budgets.The authors attribute the optimization effect to more stable optimization over discrete latents; generation-model ablations are in Supplementary Material B.Source: §4.4 · p. 7

5.1 Part-Level Editing and Secondary Creation

• Fine-grained part structure supports localized editing, recombination of parts across assets, and component-level refinement without rebuilding the whole shape.These structured outputs provide artists with a part-level prior for flexible secondary creation.Source: §5.1 · p. 7

5.2 Fine-Grained Articulated Assets Creation

• Separating small functional components lets each receive motion parameters, interaction logic, or physical constraints for animation, simulation, and digital interaction.Examples include individual piano keys and control-panel buttons and switches.Source: §5.2 · p. 7

6 Conclusions and Discussions

• The full VQ-VAE performs best across reconstruction settings; removing phased training degrades quality, while removing causal decoder attention weakens gains from larger token budgets.Table 4 reports part-level and object-level reconstruction metrics under different token budgets.Source: Table 4 · p. 8
• MegaParts generates objects with up to 300 parts while preserving geometric fidelity and structural coherence, outperforming prior autoregressive baselines and remaining competitive with diffusion-based methods.The framework combines a token-efficient vector-quantized tokenizer with adaptive-length token selection and a long-context autoregressive model.Source: §6 · pp. 7–8
• Long-context generation incurs substantial training and inference costs, especially for highly complex objects.The current framework models geometry only, omitting textures, materials, semantic annotations, and physical or functional relationships among parts.Source: §6 · p. 8
• Future work will explore more efficient sequence modeling, hierarchical representations, and unified modeling of geometry, appearance, and semantics.Fine-grained part-aware generation also supports articulated object creation and secondary creation guided by localized part priors.Source: Figure 4 · p. 7, §6 · p. 8

A.1 Implementation Details of Models

• The pipeline pairs an adaptive part-level 3D VQ-VAE with a fine-tuned LLM; its VQ-VAE expands to 4096 query tokens and a 65,536-entry codebook.It follows Cube v0.5 [Roblox 2025], copying pretrained weights with small random perturbations for the expanded queries and codebook.Source: §A.1 · p. 13
• The generator extends pretrained Qwen3-8B [Yang et al. 2025c] with shape, part, planning, and bounding-box tokens to serialize geometry and structure in one sequence.Part records and shape spans have explicit boundary tokens; object and part boxes encode three effective digits.Source: §A.1 · p. 13

A.2 Data Collection and Processing

• Training uses nearly 10M part meshes for VQ-VAE and approximately 440K assets for generation, including 340k public and 100k private assets.The generation corpus draws on public datasets and a high-quality private dataset.Source: §A.2 · p. 13
• Connected-component decomposition and iterative overlap-based merging produce meshes with at most 300 parts, followed by multi-stage hole filling and watertightening.Meshes over 300 components are reduced by merging the smallest component into its adjacent component with greatest voxel overlap; repair selects the first result below a Chamfer-distance threshold before watertightening and SDF extraction.Source: §A.2 · p. 13
• VQ-VAE training mixes object-level and part-level meshes 1:1, while generation training filters overlength or unreconstructable samples and balances four part-count ranges uniformly.The generation sequence limit is 256k tokens; part-count bins are [1, 50], (50, 100], (100, 200], and (200, 300], sampled at 1:1:1:1.Source: §A.2 · pp. 13–14
• Caption annotation renders eight spherical-camera views per asset and uses Kimi-K2.5 [Kimi 2026] to generate hierarchical categories, mesh condition, materials, and ten descriptions at varying granularity.Samples with incomplete or damaged geometry are filtered out.Source: §A.2 · p. 14

A.3 Training Recipes

• Generation training proceeds through token alignment, short-context pretraining, and long-context fine-tuning, with the VQ codebook initialized from the trained VQ-VAE and frozen throughout.The stages run for 1,000 iterations, 100,000 iterations, then 20k, 20k, and 50k iterations at maximum contexts of 32K, 128K, and 256K tokens, respectively; the final stages apply part-count balancing.Source: §A.3 · pp. 14–15

B.1 Object-Level Control

• BBox IoU rises from 0.89 to 0.93 with object-box prediction supervision, indicating stronger adherence to user-specified global constraints.The study compares explicit object-box supervision with prompting the box directly for controlled generation.Source: §B.1 · p. 15
• The object-control study compares explicit object-box prediction supervision with directly prompting the box for controlled generation.Adherence is evaluated by computing IoU between the user-provided object box and the generated mesh's actual bounding box.Source: §B.1 · p. 15

B.2 Sensitivity to Part Ordering

• Part-ordering sensitivity is tested by comparing the main paper's z–x–y spatial ordering with connectivity-graph breadth-first-search ordering.All other training and evaluation settings are unchanged.Source: §B.2 · p. 15
• For z–x–y ordering, part-box centers are quantized onto a 323 voxel grid and sorted lexicographically, with ties broken by bounding-box volume from large to small.This is the spatial-ordering procedure used in the main paper.Source: §B.2 · p. 15
• The alternative BFS ordering traverses a part-connectivity graph built by voxelizing components and computing their overlap.The traversal proceeds in breadth-first order.Source: §B.2 · p. 15
• Performance is comparable between z–x–y and BFS ordering, and z–x–y is retained because it avoids graph construction and speeds large-scale preprocessing.The authors interpret the comparable performance as limited sensitivity to these two reasonable serialization strategies.Source: §B.2 · p. 15

Held-Out Evaluation Sets

• Four mutually exclusive held-out sets contain 400 objects across part-count ranges [1, 50), [50, 100), [100, 200), and [200, 300], with no test objects in training.Text-to-3D uses normal-map rendering and CLIP/FID; bounding-box-conditioned evaluation reports Part CD, part IoU, BBox IoU, and success rate, defined as valid-mesh inference completion.Source: §C.1 · p. 15

Text-to-3D Generation

• CLIP score stays at 0.26–0.27 across four part-count ranges, while FID rises from 46.3 to 73.2 yet remains below all compared baselines in every range.The ranges are [1, 50), [50, 100), [100, 200), and [200, 300].Source: §C.1 · p. 15

Bounding-Box-Conditioned Generation

• The bounding-box comparison uses identical prescribed part boxes, but differs in auxiliary modality: MegaParts uses text, FullPart [Ding et al. 2025] images, and XPart [Yan et al. 2025] point clouds.It therefore evaluates geometry adherence to matched layouts, not modality-identical unrestricted appearance quality.Source: §C.1 · p. 15
• Success rates are 0.99, 0.95, 0.85, and 0.77 as part counts rise; diffusion baselines sharply decline above 50 parts from CUDA out-of-memory errors.MegaParts also achieves consistently stronger part and box adherence; its decreasing Part CD at higher counts is attributed to simpler components receiving compact adaptive token budgets.Source: §C.1 · p. 15

Evaluation Limitation

• On a single NVIDIA A800 80 GB GPU, exceptionally long sequences can still fail; feasibility depends primarily on sequence length, and more memory or distributed inference can extend context.Part count alone does not determine sequence length: some objects below 300 parts require more tokens than some 300-part objects because of part complexity.Source: §C.1 · p. 15

C.2 Inference Efficiency and Memory

• The inference profile measures single-object generation on one NVIDIA A800 80 GB GPU in bfloat16, separating autoregressive generation, VQ-VAE decoding, and marching cubes.Shape extraction decodes each part sequentially and applies marching cubes at resolution 10243.Source: §C.2 · p. 15, Table 9 · p. 16
• 42.5 tokens/s with vLLM versus 17.0 tokens/s with naive Hugging Face generation shows the throughput gain from using standard LLM serving for the Qwen3-based generator.Quantization, speculative or multi-token decoding, and distributed inference are compatible directions for further acceleration, but no additional throughput is reported for them.Source: §C.2 · p. 15
• Sequentially decoding every part and running marching cubes at resolution 10243 adds extraction overhead even for simple geometry; lowering extraction resolution can reduce this cost.Source: §C.2 · p. 15
• Batching part decoding or overlapping shape decoding with autoregressive generation are proposed ways to reduce extraction overhead.Source: §C.2 · p. 16
• 71.24 GB of peak GPU memory suffices for one 300-part object with a context length of 188k tokens on a single 80 GB GPU.This is an example, not a fixed cost for all 300-part objects: memory depends mainly on token length, and part count is not strictly monotonic with sequence length.Source: §C.2 · p. 16

C.3 Rate–Distortion Selection Cost

• Rate–distortion selection sweeps nine candidate prefix lengths per part using GPU parallelism, taking 1.8 seconds per part on average.This is an offline preprocessing step for constructing adaptive-length LLM training sequences.Source: §C.3 · p. 16
• The rate–distortion sweep adds no generation-time overhead because the autoregressive model directly predicts the number of shape tokens used for each part.Source: §C.3 · p. 16

C.4 Human Perceptual Evaluation

• The human perceptual study assessed text-prompt alignment and overall geometry quality.Source: §C.4 · p. 16
• The study sampled 20 prompts and generated one result per method.Source: §C.4 · p. 16
• Each result was rendered from six fixed viewpoints with identical camera and lighting settings.Source: §C.4 · p. 16
• Method order was randomly shuffled and method identities were hidden from evaluators.Source: §C.4 · p. 16
• 64% of votes for text alignment and 69% for geometry quality went to our model, substantially exceeding the compared methods.The results are reported in Table 10.Source: Table 10 · p. 16, §C.4 · p. 16

D.1 Additional Emergent Generation Behaviors

• The model occasionally generated objects with plausible internal structures despite not being explicitly trained for interior-structure modeling.Representative examples appear in Figure 12.Source: §D.1 · p. 17

• Figure 13 presents more visual comparisons between MegaParts and baselines on text-conditioned 3D generation.Source: §D.2 · p. 17, Figure 13 · p. 18