EI-Part: Explode for Completion and Implode for Refinement

Conference: CVPR 2026
Paper: CVF Open Access
Code: None open-source (Project page: https://cvhadessun.github.io/EI-Part/)
Area: 3D Vision
Keywords: Part-level 3D Generation, Explode-Implode Strategy, Sparse Voxels, Structured Diffusion, Structural Consistency

TL;DR

EI-Part proposes an "Explode-Implode" part-level 3D generation framework: incomplete segmented parts are exploded into a dispersed state to make room for structural completion, then imploded back to a compact state to dedicate full resolution to detail refinement. Self-attention is used in both states to maintain structural consistency among parts, ultimately outperforming SOTA models like HoloPart, X-Part, and OmniPart across Voxel IoU, CD, and F-Score.

Background & Motivation

Background: Open-world 3D generation (CLAY, TripoSG, HunYuan3D, TRELLIS, etc.) can now generate high-fidelity global geometry directly from images or text. However, most output monolithic meshes—lacking part-level decomposition, which makes downstream editing, rigging, and animation difficult. Consequently, "part-level generation" has become a critical branch: decomposing a 3D shape into several parts that are structurally coherent, geometrically plausible, detailed, and efficiently generated.

Limitations of Prior Work: The authors categorize failures of existing part generation methods into four criteria:

• Poor Structural Consistency: HoloPart treats each part as an independent set of latents and generates them separately, ignoring inter-part relationships; completion looks like "simple hole-filling," and assembly lacks coordination.
• Geometric Implausibility: OmniPart uses adaptive voxel resolution, but allocating voxels in the merged state causes voxel overlap between parts (creating ambiguity during completion) and limits completion to active voxels within bounding boxes, preventing expansion.
• Inaccurate Details: BANG stuffs all parts into a single latent set, and PartPacker uses dual latent sets; the representation capacity for individual parts is compressed, leading to insufficient geometric detail.
• Low Efficiency: PartCrafter and X-Part use cross-attention to enhance consistency but rely on fixed-length part tokens—large parts lack expressive power while small parts waste capacity, resulting in inefficient resource allocation.

Key Challenge: The spatial resolution requirements for part completion (needing "space" to grow plausible structures) and detail refinement (needing "high resolution" to depict surfaces) are in conflict. Satisfying both in the same merged or fixed layout leads to either restricted completion (within boxes, overlapping) or diluted details.

Goal: To provide sufficient space for completion expansion and enough resolution for refinement within the same framework while maintaining inter-part consistency and efficiency.

Key Insight: Since the two stages have different spatial requirements, the spatial layout of parts should switch according to the stage—exploding parts to occupy space during completion and imploding them back during refinement.

Core Idea: A dual-state sparse voxel diffusion using Explode for completion and Implode for refinement. The same set of voxels is reused across stages, but by changing the spatial arrangement of parts, resolution is "used where it matters most."

Method

Overall Architecture

Inputting a 3D shape \(O\), the output is a set of independent, structurally consistent, geometrically plausible, and finely detailed parts \(g_d\). The pipeline consists of three sequential stages: ① Diffusion-based part segmentation first partitions \(O\) into \(K\) semantic parts \(\{p^k_s\}\); ② Explode-based completion converts these incomplete parts into sparse voxels, explodes them radially, and performs conditional diffusion completion in the dispersed state to obtain coarse but complete structures \(e_c\); ③ Implode-based refinement implodes the completed voxels back to a compact state \(g_c\) for a second conditional diffusion to depict fine-grained geometry, outputting \(g_d\). Both diffusions occur in a structured latent space and use self-attention to allow parts to "see" each other for consistency.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input 3D Shape O"] --> B["Diffusion Multi-view Part Segmentation<br/>SAM→MVSegNet→InSegNet<br/>Obtain K Semantic Parts"]
    B --> C["Explode for Completion<br/>Parts to Sparse Voxels→Radial Explode<br/>Conditional Diffusion Completes Structure e_c"]
    C --> D["Implode for Refinement<br/>Iterative Implosion to Compact State g_c<br/>Conditional Diffusion Refines Details g_d"]
    C -.->|Dual-state Shared| E["Dual-state Self-Attention<br/>+ Cross-modal Fusion for Consistency"]
    D -.->|Dual-state Shared| E
    D --> F["Output: Structurally Consistent,<br/>Detailed Parts g_d"]

Key Designs

1. Diffusion-based Multi-view Part Segmentation: Clear Boundaries via 2D Texturing Logic

Direct 3D segmentation (e.g., PartField, P3-SAM) is limited by 3D resolution, often resulting in blurry boundaries; pure 2D-to-3D projection fails in occluded regions. The authors adopt a two-stage 3D texturing approach: first rendering normal maps \(\{n_i\}_{i=1}^6\) and Canonical Coordinate Maps (CCM) \(\{c_i\}_{i=1}^6\) from six orthogonal views. SAM provides front-view segmentation \(s_1\), which are fed into the multi-view segmentation diffusion model MVSegNet to generate consistent six-view segmentations. These are back-projected to 3D and passed to InSegNet to learn a continuous 3D semantic field, querying semantic colors \(\hat S(x)=\text{InSegNet}(x, f_{2D}, f_{3D})\) for each surface point \(x\), supervised by L1 loss \(L_{seg}=\frac{1}{n}\sum_i |\hat S(x_i)-S(x_i)|\). This inherits the generalization and sharp boundaries of 2D foundation models while filling occluded areas via the semantic field, providing \(K\) globally consistent, seamless parts \(\{p^k_s\}\).

2. Explode for Completion: Creating Space for Unambiguous Completion

Part segmentation provides only incomplete shapes. OmniPart's lesson was that allocating voxels in a merged state causes overlap and restricts growth to bounding boxes. The authors reverse this: \(\{p^k_s\}\) are converted to explicit sparse voxels \(\{v^k_s\}\) (large parts get more voxels, small parts fewer, achieving size-adaptive allocation more efficiently than fixed-length tokens). Following the BANG "explode" logic, explosion vector optimization is performed: axis-aligned bounding boxes are calculated for each part, and a translation vector is optimized to push voxels radially from a converged state to a dispersed state \(e_s=\text{Explode}(\{v^k_s\})\). The translation direction \(\{u_k\}\) and distance \(\{d_k\}\) for each part are recorded. Post-explosion, there is no voxel overlap ambiguity, and completion can extend beyond original boxes, gaining higher usable resolution. Completion is modeled as conditional structured diffusion \(p_\theta(e_c \mid e_s, n_1)\): conditioned on the front normal map \(n_1\) and exploded voxels \(e_s\), using rectified-flow + DiT trained with Conditional Flow Matching:

\[\mathcal{L}_{CFM}(\theta)=\mathbb{E}_{x_0,t,\epsilon}\,\lVert v_\theta(x,t)-(\epsilon-x_0)\rVert_2^2 .\]

Normal maps are encoded via DINOv2 into 2D tokens \(D_n\), and incomplete parts via SS-VAE into voxel condition features \(E_s\), generating structured latents \(E_c\) for complete parts, decoded into coarse complete voxels \(e_c\).

3. Implode for Refinement: Focusing Resolution on Details

The completion stage sacrifices detail for structural plausibility. The completed exploded voxels are then imploded back to a compact state, re-focusing limited resolution on surface details. Implosion is not a simple reversal via \(m'_k = m_k - d_k\cdot u_k\), but an iterative center-approach: parts are sorted by distance to the center and moved incrementally by step \(\alpha\) in the opposite direction:

\[m^{j+1}_k = m^j_k - \alpha\cdot u_k,\]

stopping if a collision occurs. This squeezes parts as tightly as possible without intersection, resulting in compact complete voxels \(g_c\). Refinement is also a conditional diffusion \(p_\theta(g_d \mid g_c, e_c, n_1)\), conditioning on imploded voxels \(g_c\), exploded complete voxels \(e_c\), and normal tokens \(D_n\), trained with CFM loss. A key difference is the use of a stronger Sparc3D VAE (Sparse Deformable Marching Cubes representation) to extract features and decode final geometry \(g_d\), producing ultimate parts with fine details.

4. Dual-state Shared Self-Attention + Multi-modal Fusion: Maintaining Structural Consistency

Processing parts independently loses inter-part relationships. The authors insert self-attention in both exploded and imploded states, allowing part latents to perceive each other before cross-modal fusion with normal conditions. The completion stage uses:

\[F_E = \text{CrossAttn}\big(D_n,\ \text{SelfAttn}(\text{Concat}(E_s, Z_t))\big),\]

And the refinement stage uses:

\[F_I = \text{CrossAttn}\big(D_n,\ \text{SelfAttn}(\text{Concat}(G_c, E_c, Z_t))\big),\]

where \(Z_t\) is the latent noise at step \(t\). Self-attention performs internal information perception on concatenated multi-part tokens (the source of structural consistency), while cross-attention injects geometric priors from normal maps (the source of fidelity). This mechanism serves as the "glue" across both stages.

Loss & Training

• Segmentation: MVSegNet and InSegNet use Adam, learning rate \(3\times10^{-4}\), trained for 4 days on 32 GPUs; InSegNet uses L1 semantic field loss \(L_{seg}\).
• Explode Completion Model: Adam, learning rate \(1\times10^{-4}\), gradient clipping max-norm 1.0, trained for 2 weeks on 64 GPUs; zero-conditioning probability of 0.3 used for robustness.
• Implode Refinement Model: Fine-tuned on Sparc3D checkpoints, Adam, learning rate \(1\times10^{-4}\), trained for 6 days on 64 GPUs.
• Both diffusions utilize Conditional Flow Matching under rectified-flow (Eq. 3).

Key Experimental Results

Data sourced from Objaverse / Objaverse-XL / ABO / 3D-FUTURE / HSSD, filtered for parts \(\leq 20\), and processed via mesh sub-grid splitting and merging to construct final GLBs. Evaluation on PartVerse benchmark using 100k sampled points for CD / IoU / F-Score calculations.

Main Results

Quantitative comparison of part generation:

Method	Voxel IoU↑	CD↓	Voxel [email protected]↑	[email protected]↑	[email protected]↑	[email protected]↑
PartPacker	0.2586	0.1273	0.3768	0.8199	0.6428	0.2435
PartCrafter	0.0742	0.3474	0.1316	0.4429	0.2801	0.0749
HoloPart	0.6106	0.0431	0.7374	0.9557	0.9402	0.6400
X-Part	0.7478	0.0599	0.8413	0.9256	0.9087	0.7923
OmniPart	0.2861	0.1431	0.4007	0.7911	0.6516	0.2644
EI-Part (Ours)	0.7981	0.0194	0.8452	0.9910	0.9742	0.8129

Ours achieves the best performance across all 6 metrics. Most notably, CD is reduced to 0.0194 (less than half of HoloPart's 0.0431), and the strict [email protected] reaches 0.8129. While X-Part is competitive in Voxel IoU, its significantly worse CD confirms the limitations of fixed-length tokens in precision.

Ablation Study

Configuration	Observation	Explanation
Full EI-Part	Plausible structure + fine details	Both explode completion and implode refinement present
w/o Explode	Implausible part geometry	Lacks spatial expansion for completion; restricted growth
w/o Implode	Lacks fine-grained details	No resolution refocusing; surfaces remain blurry
Segmentation: Replaced with PartField / P3-SAM	Lower accuracy, blurry boundaries	MVSegNet+InSegNet provides more accurate, meaningful segments

Key Findings

• Explode and Implode are Indispensable: Removing Explode leads to geometric implausibility (no space to complete); removing Implode leads to lack of detail (resolution not focused).
• Segmentation is the Upstream Bottleneck: Even with shared segmentation, EI-Part outperforms baselines, but the proprietary segmentation method is a prerequisite for end-to-end quality.
• Sparse Voxels + Adaptive Allocation is more efficient than fixed tokens: prevents the "large parts insufficient, small parts wasteful" imbalance seen in X-Part.

Highlights & Insights

• Spatial Layout Switching as Resource Scheduling: Treating resolution as a scarce resource—exploding for completion and imploding for refinement—is a clever scheduling perspective. This "dual-state reuse" can be generalized to other generation tasks with conflicting structural vs. detail demands.
• Iterative Implosion: Recording \(\{u_k\}, \{d_k\}\) makes implosion traceable, and the iterative "move until collision" approach maximizes resolution utilization for details—a practical design choice.
• Unified Consistency via Self-Attention: Unlike HoloPart (independent) or X-Part (fixed tokens), self-attention on variable-length concatenated tokens balances consistency with size flexibility.

Limitations & Future Work

• High Training Cost: Multi-stage training requires 32/64/64 GPUs for 4/14/6 days respectively; high barrier for reproduction.
• Dependency on Upstream Segmentation: Success relies on correct segmentation; part count is hard-capped at \(\leq 20\).
• Lack of Physical Constraints: Generation is primarily geometry-driven; future work needs to incorporate physical principles for complex functional assemblies.

Related Work & Insights

• vs HoloPart: HoloPart generates parts independently, leading to inconsistency. EI-Part uses shared self-attention for mutual perception.
• vs X-Part / PartCrafter: They use fixed-length tokens, which are inefficient for varying part sizes. EI-Part uses sparse voxels with adaptive allocation, leading to better CD/F-Scores.
• vs OmniPart: OmniPart allocates voxels in a merged state, causing overlap and restricting completion to bounding boxes. EI-Part explodes parts to eliminate overlap and expand completion space.

Rating

• Novelty: ⭐⭐⭐⭐⭐ The "Explode-Implode" resource scheduling perspective is novel and directly addresses the spatial conflict between completion and refinement.
• Experimental Thoroughness: ⭐⭐⭐⭐ Comprehensively leads across 6 metrics with fair comparisons, though core ablations are primarily qualitative in the main text.
• Writing Quality: ⭐⭐⭐⭐⭐ Pain points are clearly mapped to baselines; motivations and mechanisms are well-explained.
• Value: ⭐⭐⭐⭐ High performance on critical part-level generation tasks, though high training costs and closed source code limit immediate accessibility.

Related Papers

• [CVPR 2026] Depth Hypothesis Guided Iterative Refinement for Event-Image Monocular Depth Estimation
• [CVPR 2026] Dehallu3D: Hallucination-Mitigated 3D Generation from a Single Image via Cyclic View Consistency Refinement
• [CVPR 2026] Learning Hierarchical Hyperbolic Mixture Model for Part-aware 3D Generation
• [CVPR 2026] Learning Spatial-Temporal Consistency for 3D Semantic Scene Completion
• [CVPR 2026] LaS-Comp: Zero-shot 3D Completion with Latent-Spatial Consistency