semantic part decomposition

Existing part-level 3D shape decomposition methods suffer from limited controllability and insufficient semantic plausibility, hindering downstream applications such as retopology, UV mapping, and 3D printing. To address this, we propose a novel interactive part generation paradigm guided by axis-aligned bounding boxes (AABBs): leveraging AABBs as spatial prompting cues, integrating point-wise semantic feature encoding, and incorporating structural consistency optimization to achieve fine-grained decomposition that is semantically coherent, geometrically faithful, and topologically consistent. Our method unifies prompt-driven generative modeling with editable user interaction, enabling intuitive, precise control over part geometry and semantics. Evaluated on standard benchmarks, it achieves state-of-the-art performance in part-level shape generation, with outputs demonstrating production-grade quality and practical usability. To foster reproducibility and further research, we will publicly release the source code.

High-fidelity, part-level controllable 3D content generation and editing remain key challenges in computer graphics. To address this, this work proposes CompoSE, a method that leverages a diffusion Transformer architecture with alternating local-global attention to automatically infer semantic structure and symmetry from user-provided coarse part layouts—such as bounding boxes—without requiring part-level textual prompts. A novel layout-conditioning mechanism ensures strict alignment with input constraints. CompoSE enables fine-grained, context-aware editing operations, including part replacement, addition or removal, and style-preserving scaling. Experiments demonstrate that CompoSE significantly outperforms existing approaches in guided 3D synthesis, with both quantitative metrics and large language model–based evaluations confirming its superiority.

Existing 3D generative methods predominantly produce monolithic shapes, limiting interactive editing capabilities. To address this, we propose a two-stage collaborative framework for part-level controllable 3D asset generation. In the first stage, unlabeled 3D part layout planning is guided by 2D part masks to ensure structural consistency. In the second stage, an autoregressive structural planning module generates a sequence of part bounding boxes, while a pretrained spatially conditioned correction flow model jointly synthesizes geometry and texture. Our method supports user-defined part granularity and precise spatial placement, enabling unified modeling of semantic disentanglement and structural coherence. Extensive evaluations on multiple benchmarks demonstrate state-of-the-art performance in both fidelity and diversity, with significant improvements in editability and downstream applicability—e.g., for scene composition and iterative design.

This work addresses the problem of localized 3D mesh editing guided by a single edited image. We propose an interactive editing framework based on mask-conditioned reconstruction: user-specified 3D regions serve as geometric masks, and the edited image acts as a conditioning signal to guide a Large Reconstruction Model (LRM) to reconstruct only the masked regions while preserving high fidelity in the unmasked areas. To our knowledge, this is the first method to adapt LRM for real-time, mask-conditioned mesh editing. Our approach integrates multi-view-consistent mask rendering, stochastic 3D occlusion synthesis, and single-view conditional injection, enabling high-quality geometric updates in a single forward pass. The framework supports diverse semantic edits—including deformation, part replacement, and detail sculpting—achieving state-of-the-art reconstruction quality while accelerating inference by 10× over the best prior baseline.

This work addresses the limitations of existing 3D part decomposition methods, which predominantly rely on functional semantics while neglecting material boundaries and suffer from linearly increasing inference costs with part count. To overcome these issues, we propose an efficient material-aware decomposition framework that encodes the geometry of multi-material parts using a single global latent variable and decodes all parts in parallel within a single forward pass. Our approach integrates a diffusion model to generate initial material assignments, employs reinforcement learning to optimize part layout and suppress overlaps, and introduces a sparse voxel flow-matching model with part-wise attention to recover fine geometric details. This method achieves, for the first time, inference complexity decoupled from the number of parts, significantly improving both material decomposition accuracy and computational efficiency while preserving high geometric fidelity.

为解决3D对象生成中缺乏部件级精确控制的问题，MultiCube通过两阶段扩散过程和独立部件条件编码方法实现对各部分语义及空间布局的精细控制。

This study addresses the limitation of existing 3D shape decomposition methods, which frequently produce overlapping parts or gaps that hinder complete coverage. To overcome this, we propose a unified 3D segmentation model that enables controllable decomposition via point prompts. Building upon pretrained 3D generative models, our approach integrates a prompt encoder with a novel part decoder to jointly model all components within a shared latent space. By performing coarse-to-fine decoding of the complete partition, it ensures non-overlapping parts with full geometric coverage. Experimental results demonstrate that the proposed model surpasses state-of-the-art methods across multiple tasks, improving part compatibility by an order of magnitude.

Existing 3D editing methods are hindered by the scarcity of high-quality paired supervision data, leading to suboptimal performance in geometry preservation, multi-view consistency, and local controllability. This work introduces semantic parts as the fundamental editing units and proposes a part-based transformation supervision paradigm. The authors construct Pxform, a high-quality dataset comprising 100,000 paired samples, and design PartFlow, a feed-forward network that leverages a source-aware latent space to enable high-fidelity 3D editing without requiring explicit edit masks. By integrating part-level data generation, pretrained 3D priors, mask-aware velocity preservation, and rendering-space consistency constraints, the method significantly enhances boundary sharpness, semantic coherence, and source structure retention in both geometric and appearance editing tasks, achieving state-of-the-art performance.

Existing methods for 3D part generation struggle to simultaneously achieve global geometric consistency and high-quality, editable part decomposition: segmentation-based approaches fix the whole shape before partitioning, while additive strategies often produce discontinuous boundaries. This work proposes SCULPT, a novel subtractive part generation framework that operates in a structured 3D latent space, iteratively co-generating parts and the remaining object through joint denoising and sparse overlapping voxel support sets. The approach adaptively determines part counts and effectively eliminates inter-part gaps and material discontinuities. Evaluated on PartObjaverse, SCULPT achieves state-of-the-art geometric quality and excels in assembly-based reconstruction, further enabling fine-grained textured part decomposition from diverse inputs, including real-world images.

This study addresses the challenges of occlusion-induced structural ambiguity and insufficient inter-part coherence in single-image part-level 3D generation. To overcome these limitations, this work proposes a controllable 3D mesh generation framework that leverages segmentation as an explicit anchor. Methodologically, it pioneers the use of segmentation as an explicit localization signal for generation and introduces global context exchange alongside structured cross-part attention mechanisms. Combined with a newly constructed large-scale dataset, PartObjectNet, the proposed approach achieves high-quality part assembly without requiring post-processing alignment. Experimental results demonstrate that this method significantly outperforms existing approaches in geometric quality, part coherence, and controllable decomposition.