Research map / 3D vision / 3D generation

MoCA: Mixture-of-Components Attention for Scalable Compositional 3D Generation

NeurIPS 2026 · 3D vision › 3D generation · approach: Text-to-3D scene generation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MoCA, a compositional 3D DiT whose Mixture-of-Components attention routes each component to full tokens of top-k important components and compressed tokens of the rest, enabling scalable part-level object and instance-level scene generation with up to 32 components.

Problem

Global attention across components in compositional 3D generation scales quadratically with the number of parts or instances, limiting how many components can be generated and wasting capacity on weak interactions.

Approach

Importance-based top-k routing via sigmoid anchor-token attention selects which components get full-token global attention, while the remaining components are compressed into a few tokens (ratio 8) so global context is preserved at reduced cost; load balancing uses stochastic routing instead of an auxiliary loss.

Key results

• On PartObjaverse-Tiny, MoCA achieves CD 0.1010 and Fscore-0.1 0.8708 versus best baseline PartPacker CD 0.1105 and Fscore-0.1 0.8484.
• On 3D-FRONT scene generation, MoCA reaches CD-Scene 0.1201 and Fscore-Scene 0.8637, beating PartCrafter (0.1366, 0.8379) and MIDI (0.1548, 0.8238).
• Ablation shows removing routing to important components raises CD from 0.1180 to 0.1969, and softmax router activation raises CD to 0.1451.

Surprising findings

• MoCA's routing overhead is substantial at N=32 (36 ms routing vs 79 ms global attention), and total runtime is still higher than PartCrafter at N=4 and N=8 despite lower global attention cost.

Limitations

• Vecset VAE is frozen, which may bottleneck reconstruction when many components have small volumes.
• Routing implemented in pure PyTorch adds substantial runtime overhead.
• Evaluation on PartObjaverse-Tiny lacks part correspondence with ground truth, so metrics are overall geometry only.

Experimental setup

Datasets: PartObjaverse-Tiny ABO 3D-FRONT Trellis-500K MIDI processed 3D-FRONT

Benchmarks: PartObjaverse-Tiny ABO 3D-FRONT

Metrics: Self-IoU Chamfer Distance F-score F-score-Scene CD-Scene CD-Obj

Baselines: HoloPart PartPacker PartCrafter MIDI

Builds on: 3DShape2VecSet Trellis PartPacker DeepSeek-V3 LoRA

Closest papers by meaning

• PartCrafter: Structured 3D Mesh Generation via Compositional Latent Diffusion Transformers (NeurIPS 2025 · similarity 0.87)
• MIDI: Multi-Instance Diffusion for Single Image to 3D Scene Generation (CVPR 2025 · similarity 0.86)
• DecompDreamer: A Composition-Aware Curriculum for Structured 3D Asset Generation (NeurIPS 2026 · similarity 0.86)
• AutoPartGen: Autoregressive 3D Part Generation and Discovery (NeurIPS 2025 · similarity 0.86)
• Structured 3D Latents for Scalable and Versatile 3D Generation (CVPR 2025 · similarity 0.85)
• Direct3D-S2: Gigascale 3D Generation Made Easy with Spatial Sparse Attention (NeurIPS 2025 · similarity 0.85)