FullPart: Generating each 3D Part at Full Resolution
Lihe Ding, Shaocong Dong, Yaokun Li, Chenjian Gao, Xiao Chen, Rui Han, Yihao Kuang, Hong Zhang, Bo Huang, Zhanpeng Huang, Zibin Wang, Dan Xu, Tianfan Xue
TL;DR
FullPart introduces a hybrid 3D part generation framework that combines implicit and explicit methods, generating detailed parts at full resolution within individual voxel grids, and leverages a large dataset for improved results.
Contribution
The paper proposes a novel FullPart framework that integrates implicit and explicit 3D generation paradigms, and introduces PartVerse-XL, the largest annotated 3D part dataset.
Findings
Achieves state-of-the-art results in 3D part generation.
Enables detailed synthesis of small parts at full resolution.
Demonstrates effectiveness of the combined implicit-explicit approach.
Abstract
Part-based 3D generation holds great potential for various applications. Previous part generators that represent parts using implicit vector-set tokens often suffer from insufficient geometric details. Another line of work adopts an explicit voxel representation but shares a global voxel grid among all parts; this often causes small parts to occupy too few voxels, leading to degraded quality. In this paper, we propose FullPart, a novel framework that combines both implicit and explicit paradigms. It first derives the bounding box layout through an implicit box vector-set diffusion process, a task that implicit diffusion handles effectively since box tokens contain little geometric detail. Then, it generates detailed parts, each within its own fixed full-resolution voxel grid. Instead of sharing a global low-resolution space, each part in our method - even small ones - is generated at…
Peer Reviews
Decision·ICLR 2026 Poster
1. The proposed per-part full-resolution generation paradigm is technically sound and yields satisfactory results. As shown in Figures 4 and 5, this method can improve the generation performance of both part-level and object-level. Besides, the proposed center-corner encoding mechanism utillizes a unified positional encoding strategy to preserve global spatial coherence across parts of different scales, effectively addressing the spatial discrepancy between different parts. 2. This work present
1. Generating each part within a full-resolution voxel grid substantially increases memory usage and computational costs, particularly for complex objects with numerous parts. The paper does not include an analysis of computational efficiency or stress testing in this regard. 2. The test set is limited to 100 manually selected untrained objects, which may not comprehensively evaluate the model's generalization capability and overall performance. 3. Quantitative ablation analysis results are mi
1. The combination of implicit and explicit representations is well-motivated and effectively leverages the strengths of both paradigms. 2. The construction of the PartVerse-XL dataset is a valuable contribution that can advance research in part-level 3D generation. 3. The proposed Center-Corner Encoding mechanism is conceptually simple yet empirically effective.
1. Section 2.2 discusses mainly recent (2025) part-level 3D generation works. The related work could be more comprehensive by including earlier implicit Tri-plane–based approaches, such as Frankenstein [Yan et al., 2024] and StdGen [He et al., 2024].
1. It identifies the problem that the resolution given to small parts is limited in 3D part generation by explicit voxel representation. And solves this problem by generating each part in an independent full-resolution voxel grid. 2. It creatively proposes a new positional embedding strategy to maintain global coherence in part generation with independent voxel grids. 3. The paper introduces PartVerse-XL, a large human-annotated 3D part dataset, which is a valuable resource for the research comm
1. It looks like generating each part in full-resolution voxel grids will significantly increase the computational cost of training and inference. However, the paper does not provide any analysis or discussion on the computational efficiency of FullPart compared to existing methods. 2. The evaluation is only conducted on the PartVerse-XL dataset, which is introduced by the authors themselves. It would be more convincing if the authors could also evaluate their method on other established 3D part
Code & Models
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.