PaperlayerView exact revision on arXiv

Source-linked AI summary

BANG: Dividing 3D Assets via Generative Exploded Dynamics

Longwen Zhang, Qixuan Zhang, Haoran Jiang, Yinuo Bai, Wei Yang, Lan Xu, Jingyi Yu

arXiv:2507.21493v1cs.GR

TL;DR

Current 3D creation tools require substantial expertise and manual effort for part-level decomposition. BANG generates smooth exploded sequences that separate parts while maintaining coherence, and supports part-level generation, control, and analysis.

  • Problem

    Current 3D creation tools do not readily replicate intuitive part-level decomposition and reassembly, which requires substantial artistic expertise and manual effort.

    Source: §1 · p. 2

  • Method

    BANG uses generative exploded dynamics to progressively separate an input geometry into parts, with an adapter conditioning a diffusion model on geometry and timestamps and temporal attention supporting sequence continuity.

    Source: §1 · p. 2, §3 · p. 4, §3.1 · p. 6

  • Results

    BANG dynamically decomposes complex 3D assets into interpretable parts through a smooth, consistent process that preserves structure and coherence.

    Source: §1 · p. 2, §7 · p. 17

  • Takeaways & Limitations

    BANG supports part-level generation, decomposition, and reassembly, including interactive analysis and applications such as 3D printing.

    Source: §1 · p. 3

    Trained on 20k exploded dynamic data, BANG struggles with highly complex objects, particularly those with poorly defined structural components.

    Source: §7 · p. 17

Abstract

from arXiv · show

3D creation has always been a unique human strength, driven by our ability to deconstruct and reassemble objects using our eyes, mind and hand. However, current 3D design tools struggle to replicate this natural process, requiring considerable artistic expertise and manual labor. This paper introduces BANG, a novel generative approach that bridges 3D generation and reasoning, allowing for intuitive and flexible part-level decomposition of 3D objects. At the heart of BANG is "Generative Exploded Dynamics", which creates a smooth sequence of exploded states for an input geometry, progressively separating parts while preserving their geometric and semantic coherence. BANG utilizes a pre-trained large-scale latent diffusion model, fine-tuned for exploded dynamics with a lightweight exploded view adapter, allowing precise control over the decomposition process. It also incorporates a temporal attention module to ensure smooth transitions and consistency across time. BANG enhances control with spatial prompts, such as bounding boxes and surface regions, enabling users to specify which parts to decompose and how. This interaction can be extended with multimodal models like GPT-4, enabling 2D-to-3D manipulations for more intuitive and creative workflows. The capabilities of BANG extend to generating detailed part-level geometry, associating parts with functional descriptions, and facilitating component-aware 3D creation and manufacturing workflows. Additionally, BANG offers applications in 3D printing, where separable parts are generated for easy printing and reassembly. In essence, BANG enables seamless transformation from imaginative concepts to detailed 3D assets, offering a new perspective on creation that resonates with human intuition.

1 INTRODUCTION

Current 3D tools make part-level decomposition and adjustment labor-intensive, while mainstream generation and understanding approaches leave a gap between creating objects and reasoning about their parts. BANG addresses this with smooth generative exploded dynamics and tools for controlling, analyzing, and creating part-level 3D assets.

Motivation

  • BANG addresses the effort required for part-level 3D creation, which demands substantial artistic expertise and tedious manual labor with current tools.Source: §1 · p. 2

BANG

  • BANG dynamically decomposes complex 3D assets into interpretable parts through a smooth, consistent exploded process.Intermediate states progressively separate parts while preserving geometric and semantic consistency, culminating in a fully divided state.Source: §1 · p. 2

Generative Exploded Dynamics

  • The approach fine-tunes a large-scale latent diffusion model with an exploded view adapter and temporal attention for geometry- and timestamp-conditioned decomposition.Part-aware trajectory tracking associates components with the original mesh for reassembly.Source: §1 · p. 2

Controllable Generation

  • Bounding boxes and surface-region prompts enable users to specify parts, while aligned 3D and 2D features support selection from rendered views or sketches.The framework describes integration with multimodal models including DINOv2 and Florence-2.Source: §1 · p. 2

Applications

  • BANG supports part-level generation, decomposition, and reassembly from text or image inputs, as well as interactive analysis and 3D-printing applications.Source: §1 · p. 3

2 RELATED WORK

The related work reviews 3D structural understanding, 3D generation, and dynamic or exploded representations. Across these areas, it positions BANG as a generative approach to part-level decomposition that includes internal structure and coherent reassembly.

2.1 3D Structural Understanding

  • Surface-focused segmentation and multimodal analysis leave internal volumetric structure overlooked, whereas BANG displaces parts to model interiors and support precise manipulation.The review describes segmentation methods as relying on outer-surface point clouds or meshes and multimodal methods as focusing on surface geometry.Source: §2.1 · p. 3

2.2 3D Generation

  • The review covers 2D lifting, 3D-native, and part-aware generation; BANG natively decomposes objects into meaningful parts and ensures coherent reassembly.The overview groups its pipeline into data synthesis, dynamics generation, trajectory optimization, and controllable generation.Source: §2.2 · pp. 3–4, Figure 2 · p. 4

2.3 4D Generation and Exploded View

  • Dynamic 4D generation and traditional exploded views provide related context, but prior 3D exploded-view work is described as largely overlooked while BANG integrates 3D part decomposition in a generative framework.Source: §2.3 · p. 4

3 GENERATIVE EXPLODED DYNAMICS

Generative Exploded Dynamics turns an assembled 3D geometry into a smooth sequence of exploded states, preserving part-level geometric and semantic consistency. The section describes the model, its training data, and post-generation part tracking.

3 GENERATIVE EXPLODED DYNAMICS

  • BANG progressively separates complex 3D assets into interpretable parts while preserving geometric and semantic consistency across intermediate states.The conditioning-generative process simulates a smooth, radial transition from an assembled geometry to its constituent parts.Source: §3 · p. 4

3.1 Exploded Dynamics Generation Model

  • A frozen pretrained 3D diffusion model is adapted to generate geometry sequences using an input-conditioned Exploded View Adapter and temporal attention.The adapter conditions generation on input geometry and time; temporal attention models cross-frame context to support smooth transitions, using a lightweight module that can be trained independently.Source: §3.1 · pp. 5–6

3.2 Dataset Construction

  • The training dataset contains approximately 20k filtered Objaverse assets, with synthetic exploded sequences generated by optimizing component translations to reduce bounding-box collisions.The filtering prioritizes quality and excludes unsuitable assets; optimized translations are interpolated from assembled to fully exploded states, with overly close or excessively large results discarded.Source: §3.2 · pp. 6–7

3.3 Part Trajectory Tracking

  • After generation, connected components in the fully exploded mesh are tracked back toward the assembled state by optimizing per-part translations against SDF values.The tracking parametrization assumes linear translation, and optimization masks surface points inside another part because their negative SDF values are invalid for the loss.Source: §3.3 · pp. 7–8

4 CONTROLLABLE GENERATION

BANG provides spatially conditioned exploded views and connects 2D image regions to 3D geometry, giving users ways to guide which parts are decomposed. Its framework also supports feeding generated 3D assets into workflows for interactive, targeted manipulation.

4.1 Spatial Control

  • Bounding-box and surface-region prompts let users specify volumetric areas or finely detailed surface areas for decomposition.Bounding boxes apply even when the geometry lacks internal structure; training with varying numbers of prompts supports arbitrary prompt counts at inference.Source: §4.1 · p. 8, §4 · p. 8
  • A dedicated transformer branch encodes spatial prompts and integrates them with geometry features through interleaved cross-attention.Bounding-box corners receive positional embeddings, while sampled surface points are encoded as tokens; an auxiliary binary token indicates whether unselected regions should also be exploded.Source: §4.1 · p. 8, Figure 7 · p. 8

4.2 Cross-modal Creative Framework

  • The geometry decoder is adapted to produce features aligned with DINOv2, allowing 2D regions of interest to be mapped to corresponding 3D surface regions.Users can select regions from rendered images or concept images; feature similarity identifies corresponding surface points.Source: §4.2 · p. 9, Figure 9 · p. 9
  • The framework can take a newly generated 3D mesh and apply bounding boxes, surface regions, or 2D region selections to guide its decomposition.This connects concept generation from text or images with targeted part-level manipulation.Source: §4.2 · p. 9

5 APPLICATIONS

BANG’s applications use part-level decomposition for geometry refinement, multimodal structural understanding, and physical prototyping. The examples cover regenerating and reassembling detailed parts, dialogue-guided manipulation, and printing components for assembly.

5.1 Per-part Geometric Detail Enhancement

  • Per-part regeneration produces more detailed geometry than single-mesh pipelines while supporting artist-friendly topologies and part-specific animations.The process scales exploded parts into normalized space, regenerates them from coarse geometry and image regions, then reassembles them; Fig. 10 illustrates the workflow.Source: §5.1 · p. 10, Figure 10 · p. 10

5.2 Multi-modal Integration for Structural Understanding and Control

  • Multimodal integration supports automatic part descriptions after explosion and natural-language guidance for targeted decomposition.GPT-4 can describe visually marked parts; user instructions can be mapped through ROI selection, segmentation, and geometric feature matching before applying spatial prompts.Source: Figure 11 · p. 10, §5.2 · p. 11

5.3 Expedite Combinable Structure 3D Printing

  • BANG generates separated, aligned components for individual 3D printing and post-print assembly, including printed robot parts with programmatically generated interlocking structures.The described workflow supports optimized print orientations, tailored material choices, and movable joints between components.Source: §5.3 · p. 11, Figure 12 · p. 11

6 EXPERIMENTS

BANG is evaluated as a pipeline for generating and refining exploded 3D parts, with spatial control and interactive applications. Comparisons and ablations report part-quality, temporal-consistency, and control findings, while the user study favors BANG over two segmentation methods.

6.1 Experimental Setup and Implementation Details

  • The training pipeline adds a four-layer exploded-view adapter to a pretrained latent diffusion model, then trains temporal attention for sequences of up to five frames.The base model is pretrained on about 500,000 geometries; the adapter is trained on 20,000 exploded-view examples.Source: §6.1 · pp. 11–12

6.2 Visualization of Generated Exploded Dynamics

  • BANG demonstrates concept-image-to-part generation, prompt-controlled decomposition, part-wise refinement, interactive exploration, and individually printable parts for reassembly.Its steampunk workshop example proceeds from a concept image to a 3D mesh, exploded components, and enhanced parts.Source: §6.2 · p. 13

6.3 Structural Segmentation Comparison

  • Across tested geometries, BANG generates volumetric part decompositions, while SAMesh and SAMPart3D produce surface-based regions and struggle more with complex or non-artist-created meshes.The comparisons include artist-crafted meshes, remeshed assets, and watertight assets generated by BANG’s base model; baselines are manually separated for exploded-view display.Source: §6.3 · p. 14, Figure 15 · p. 14
  • Across 10 generated assets, users evaluated which method best aligned with intuitive part decomposition and visual appeal, favoring BANG’s exploded views (65.5%) over SAMesh (26.2%) and SAMPart3D (8.3%).BANG averaged 45 seconds per asset, compared with 386 seconds for SAMesh and 940 seconds for SAMPart3D.Source: §6.3 · p. 15

6.4 Evaluations

  • The evaluation measures part tracking on 50 held-out PartObjaverse-Tiny objects using weighted IoU for localization and an SDF objective for surface alignment.Exploded parts are transformed back to their original positions before evaluation.Source: §6.4 · p. 15
  • Temporal attention improves weighted IoU by 18.8% and reduces the SDF objective by 31.5% in the ablation study.The table caption associates temporal attention with improved temporal consistency and stopping gradients for overlapping points with more accurate part trajectory tracking.Source: §6.4 · p. 16, Table 1 · p. 16
  • For generated sequences, trajectory-tracking metrics improve through five frames; further increases face training and inference constraints and rising computational cost.Ground-truth sequences stabilize after three frames, while the model’s training was limited to five frames by GPU memory.Source: §6.4 · p. 16, Figure 19 · p. 16
  • Part-count conditioning offers coarse rather than precise control, changing decomposition granularity while retaining semantic consistency.Fewer requested parts can merge functionally related components; higher counts produce more detailed structural decompositions.Source: Figure 20 · p. 16, §6.4 · p. 17

7 DISCUSSIONS AND CONCLUSIONS

BANG decomposes complex 3D assets into interpretable parts through smooth exploded views, using an Exploded View Adapter and Temporal Attention Module. Its current scope is bounded by difficulty with complex geometry, fidelity losses, and an artistic rather than physically constrained pipeline.

7 DISCUSSIONS AND CONCLUSIONS

  • BANG combines an Exploded View Adapter conditioned on geometry and timestamps with a Temporal Attention Module for smooth transitions across exploded states.The framework supports 3D decomposition, generation, and enhancement.Source: §7 · p. 17
  • Trained on 20k exploded-dynamics data, BANG struggles with highly complex objects, especially those with poorly defined structural components.The paper identifies broader training data, particularly real-world mechanisms, as a direction for improving coverage.Source: §7 · p. 17
  • Highly detailed regions can lose local detail and drift from the original geometry because training lacks explicit per-part supervision and the latent representation has limited token length.Fig. 21 illustrates these failure cases.Source: §7 · p. 17, Figure 21 · p. 17
  • BANG follows an artistic, geometry-focused pipeline, so it may not meet physical-assembly needs in manufacturing or robotics and currently omits material and appearance properties.The paper points to physical simulation and integrating material and appearance attributes as future directions.Source: §7 · p. 17

Related papers

Loading 2507.21493v1…