# Prompting Image Generators for Training-free Primitive Shape Abstraction

arXiv: https://arxiv.org/abs/2607.05568 · HTML: https://mlsanity.com/paper/2607.05568/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 32 of its 51 references that are in the map.
- First posted on arXiv 6 Jul 2026.

## Builds on (most cited first, 30 of 32)

- [Language Models are Few-Shot Learners](https://mlsanity.com/paper/2005.14165/) (2020)
- [Segment Anything](https://mlsanity.com/paper/2304.02643/) (2023)
- [3D-R2N2: A Unified Approach for Single and Multi-view 3D Object Reconstruction](https://mlsanity.com/paper/1604.00449/) (2016)
- [PartNet: A Large-scale Benchmark for Fine-grained and Hierarchical Part-level 3D Object Understanding](https://mlsanity.com/paper/1812.02713/) (2018)
- [Learning Shape Abstractions by Assembling Volumetric Primitives](https://mlsanity.com/paper/1612.00404/) (2016)
- [Superquadrics Revisited: Learning 3D Shape Parsing beyond Cuboids](https://mlsanity.com/paper/1904.09970/) (2019)
- [Using Shape to Categorize: Low-Shot Learning with an Explicit Shape Bias](https://mlsanity.com/paper/2101.07296/) (2021)
- [PartSLIP: Low-Shot Part Segmentation for 3D Point Clouds via Pretrained Image-Language Models](https://mlsanity.com/paper/2212.01558/) (2022)
- [Neural Parts: Learning Expressive 3D Shape Abstractions with Invertible Neural Networks](https://mlsanity.com/paper/2103.10429/) (2021)
- [SAMPart3D: Segment Any Part in 3D Objects](https://mlsanity.com/paper/2411.07184/) (2024)
- [PARTFIELD: Learning 3D Feature Fields for Part Segmentation and Beyond](https://mlsanity.com/paper/2504.11451/) (2025)
- [Unsupervised Learning for Cuboid Shape Abstraction via Joint Segmentation from Point Clouds](https://mlsanity.com/paper/2106.03437/) (2021)
- [P3-SAM: Native 3D Part Segmentation](https://mlsanity.com/paper/2509.06784/) (2025)
- [Image Generators are Generalist Vision Learners](https://mlsanity.com/paper/2604.20329/) (2026)
- [Differentiable Blocks World: Qualitative 3D Decomposition by Rendering Primitives](https://mlsanity.com/paper/2307.05473/) (2023)
- [Img2CAD: Reverse Engineering 3D CAD Models from Images through VLM-Assisted Conditional Factorization](https://mlsanity.com/paper/2408.01437/) (2024)
- [Robust and Accurate Superquadric Recovery: a Probabilistic Approach](https://mlsanity.com/paper/2111.14517/) (2021)
- [MeshSegmenter: Zero-Shot Mesh Semantic Segmentation via Texture Synthesis](https://mlsanity.com/paper/2407.13675/) (2024)
- [SuperDec: 3D Scene Decomposition with Superquadric Primitives](https://mlsanity.com/paper/2504.00992/) (2025)
- [Marching-Primitives: Shape Abstraction from Signed Distance Function](https://mlsanity.com/paper/2303.13190/) (2023)
- [PartSAM: A Scalable Promptable Part Segmentation Model Trained on Native 3D Data](https://mlsanity.com/paper/2509.21965/) (2025)
- [PASTA: Controllable Part-Aware Shape Generation with Autoregressive Transformers](https://mlsanity.com/paper/2407.13677/) (2024)
- [3D Part Segmentation via Geometric Aggregation of 2D Visual Features](https://mlsanity.com/paper/2412.04247/) (2024)
- [BenchCAD: A Comprehensive, Industry-Standard Benchmark for Programmatic CAD](https://mlsanity.com/paper/2605.10865/) (2026)
- [Iterative Superquadric Recomposition of 3D Objects from Multiple Views](https://mlsanity.com/paper/2309.02102/) (2023)
- [PrimitiveAnything: Human-Crafted 3D Primitive Assembly Generation with Auto-Regressive Transformer](https://mlsanity.com/paper/2505.04622/) (2025)
- [SegviGen: Repurposing 3D Generative Model for Part Segmentation](https://mlsanity.com/paper/2603.16869/) (2026)
- [Light-SQ: Structure-aware Shape Abstraction with Superquadrics for Generated Meshes](https://mlsanity.com/paper/2509.24986/) (2025)
- [Exploiting GPT-4 Vision for Zero-shot Point Cloud Understanding](https://mlsanity.com/paper/2401.07572/) (2024)
- [SweepNet: Unsupervised Learning Shape Abstraction via Neural Sweepers](https://mlsanity.com/paper/2407.06305/) (2024)

## Abstract (arXiv)

Compact primitive abstractions represent 3D shapes with a few geometric primitives while preserving recognizable components. Learned methods depend on their training classes, and optimization-based methods split shapes geometrically rather than into parts. We instead reuse the visual part knowledge of pretrained models without task-specific training or fine-tuning. A vision-language model names parts in multi-view renders, and an unmodified image generator paints color-coded part masks. Reprojection and spatial clustering recover 3D instances, and a classical optimizer fits one tapered and bent superquadric per part. With five to eight primitives per object, the abstractions match the Chamfer distance of the strongest learned baseline on HumanPrim, improve on it by 10% on Toys4K, and have the lowest overlap among compact methods, while chair legs, backrest bars and wheels remain separate primitives. Our accuracy also transfers better than theirs to objects outside the learned methods' ShapeNet training classes. Replacing the generated masks with part labels from the 3D segmentation methods P3-SAM or PartField lowers IoU by 7 to 17 points under the same fitter. Further studies relate the remaining volumetric error to part granularity and to parts that the rendered views observe from one side only.