Research map / 3D vision / 3D generation
Efficient Part-level 3D Object Generation via Dual Volume Packing
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 PartPacker, an end-to-end single-image-to-3D model that generates an arbitrary number of complete parts by packing them into two complementary volumes via bipartite-graph contraction, avoiding 2D segmentation priors.
Problem
Part-level 3D object generation from a single image is hard because objects have a variable number of parts, existing latent denoising models output fused meshes, and segmentation-then-completion pipelines rely on 2D/3D segmentation priors and process parts sequentially.
Approach
Pack disjoint parts into two shared volumes by converting the part-connectivity graph into a bipartite graph via heuristic greedy odd-cycle edge contraction (weighted by penetration depth), so a fixed-length dual-latent flow model can generate all parts jointly and in parallel.
Key results
- Single-volume pretrained model (Ours†) gets ULIP 0.1729 and Uni3D 0.3941, comparable to TripoSG and Hi3DGen; the dual-volume part-level model scores slightly lower (ULIP 0.1715, Uni3D 0.3906).
- Generates part-level meshes from a single image in about 30 seconds regardless of part count, versus HoloPart's multi-minute preprocessing plus per-part completion that scales linearly with parts.
- Approximately 254K of 386K multi-part meshes remain after balance filtering.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use dual volumes instead of the fused-shape single-volume model | Ours image-to-3D model · Image-to-3D test set (40 images) | ULIP cosine similarity: 0.1729 -> 0.1715 | Hurt |
| Use dual volumes instead of the fused-shape single-volume model | Ours image-to-3D model · Image-to-3D test set (40 images) | ULIP-2 cosine similarity: 0.4022 -> 0.3986 | Hurt |
| Use dual volumes instead of the fused-shape single-volume model | Ours image-to-3D model · Image-to-3D test set (40 images) | Uni3D cosine similarity: 0.3941 -> 0.3906 | Hurt |
| Add part-level VAE finetuning (ablation) | Ours part-level generation model · Part-level shape reconstruction | reconstruction robustness to spatial variations | Helped |
| Filter unbalanced training examples (ablation) | Ours part-level generation model · Part-level generation | segmentation quality and diversity | Helped |
| Add a learnable part embedding to the second latent code (ablation) | Ours dual-volume flow model · Part-level generation | differentiation between dual volumes | Helped |
Surprising findings
- Without data filtering the model tends to produce overly simplistic segmentations, often only two parts.
Limitations
- Limited control over part granularity because parts come from noisy scene-graph annotations.
- Two-volume bipartite packing cannot represent three mutually contacting parts, giving suboptimal results.
- Failure cases include incorrect internal structures and unstable segmentation across random seeds.
Experimental setup
Datasets: Trellis500k (Objaverse-XL subset)
Benchmarks: 40-image in-house test set
Metrics: ULIP cosine similarity ULIP-2 cosine similarity Uni3D cosine similarity inference time
Baselines: TripoSG Hi3DGen Hunyuan3D-2 HoloPart
Builds on: Dora Trellis 3DShape2Vecset TripoSG HoloPart
Closest papers by meaning
- PartGen: Part-level 3D Generation and Reconstruction with Multi-view Diffusion Models (CVPR 2025 · similarity 0.90)
- PartCrafter: Structured 3D Mesh Generation via Compositional Latent Diffusion Transformers (NeurIPS 2025 · similarity 0.89)
- MeshArt: Generating Articulated Meshes with Structure-Guided Transformers (CVPR 2025 · similarity 0.89)
- AutoPartGen: Autoregressive 3D Part Generation and Discovery (NeurIPS 2025 · similarity 0.87)
- Interactive3D: Create What You Want by Interactive 3D Generation (CVPR 2024 · similarity 0.87)
- SINGAPO: Single Image Controlled Generation of Articulated Parts in Objects (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.