EMERGE: Resolution-Agnostic Point Cloud Generation with Equivariant Graph-Based Diffusion

1 Introduction

• The paper identifies a gap: graph-based architectures are largely absent from 3D generative modeling, and continuous 3D equivariance remains underexplored.Existing generative approaches also overlook non-grid structures and continuous spatial symmetries.Source: §1 · pp. 1–2
• EMERGE is presented as the first fully SE(3)-equivariant graph-based diffusion model designed for point-cloud generation.The model is designed to preserve continuous spatial symmetries while modeling point-cloud topologies.Source: §1 · p. 2
• A parameter-free distribution-alignment technique addresses density-induced shifts, enabling inference at arbitrary resolutions without retraining, fine-tuning, or architectural changes.The contribution list describes this as zero-shot high-fidelity 3D super-resolution.Source: §1 · p. 2
• EMERGE achieves state-of-the-art generation quality while requiring up to 10x fewer training epochs than existing methods.The stated contributions also report evaluation against prior methods quantitatively and qualitatively.Source: §1 · p. 2

GNNs in 3D Vision

• GNNs suit unordered 3D point clouds and have been used in discriminative tasks, but graph-based architectures have remained absent from 3D generative modeling.Examples include segmentation [38], detection [34], and classification [19].Source: §2 · p. 2

Point Cloud Generation

• Point-cloud generation methods evolved from continuous normalizing flows and VAEs to DDPMs, trajectory optimization, hybrid point-voxel models, latent diffusion, and time-varying denoising.The cited methods include PVD [42], LION [41], and TIGER.Source: §2 · p. 2
• Prior generative approaches largely ignore explicit non-grid topology and continuous spatial symmetry, which EMERGE addresses with a graph-based, SE(3)-equivariant backbone.The paper describes zero-shot super-resolution at arbitrary inference resolutions as another distinction.Source: §2 · p. 2

2D/3D Equivariant Models

• Equivariant GNNs have supported predictive and generative chemical-compound tasks [33] [3] [17] [36]; point-cloud models used limited discrete rotation groups for downstream analysis [8] [43], while current generative models lacked 3D equivariance.This absence motivates a universal, continuous geometric inductive bias for point-cloud synthesis.Source: §2 · p. 3

Diffusion Models

• Diffusion models progressively add Gaussian noise to a point cloud over discrete timesteps and learn to reverse the process to synthesize samples.The forward chain ends with a standard normal sample; DDPM training approximates the true posterior q(Xt−1|Xt).Source: §3 · p. 3
• In the stated DDPM objective, the network predicts the added noise ε under a variance schedule βt.The passage identifies εθ as the neural network trained to predict ε.Source: §3 · p. 3

SE(3)-Equivariance and Invariance

• SE(3)-equivariance means rotating or translating an input point cloud transforms the output by the same spatial transformation.The text defines SE(3) transformations using a rotation matrix and translation vector.Source: §3 · p. 3
• SE(3)-invariance means a function’s output is unchanged when its input is rotated or translated.The paper uses “equivariance” and “invariance” to mean SE(3) versions unless stated otherwise.Source: §3 · p. 3

4.1 EMERGE Architecture

• Each EMERGE block alternates multi-scale local processing with global attention and ultimately predicts denoised coordinates while preserving SE(3) symmetries.The local phase uses Multi-Scale EGNN and Canonical Frame Voxel Pooling; the global phase injects shape context through Global Invariant Feature Attention.Source: §4.1 · p. 3

4.2 Local Processing: Multi-Scale EGNN

• Local EGNN updates use dynamically constructed k-nearest-neighbor graphs, residual coordinate and feature updates, and diffusion-timestep conditioning in edge messages.The network injects local invariant geometric features into its update functions.Source: §4.2 · pp. 4–5
• A 13-dimensional invariant feature vector describes neighborhood geometry, including properties such as curvature, linearity, and planarity, without breaking equivariance.The vector is derived from neighborhood covariance eigenvalues and eigenvectors; invariant scalar weights preserve equivariant coordinate updates.Source: §4.2 · p. 5
• Canonical-frame voxel pooling aligns points to an object-centered frame before pooling, avoiding the rotation-dependent cell assignments of a fixed axis-aligned grid.The canonical frame is built from centered coordinates and covariance eigenvectors; pooled coordinates are projected back to world space.Source: §4.2 · p. 5
• Voxel point counts condition pooled features through FiLM, and those counts remain invariant because pooling occurs in canonical space.The counts represent differing local densities across pooled nodes.Source: §4.2 · p. 5
• Timestep-conditioned coefficients combine coordinate and feature predictions across voxel scales and the unpooled base resolution.The scale coefficients are invariant scalars, preserving the respective SE(3) symmetries of coordinate updates and invariant features.Source: §4.2 · p. 6

4.3 Global Processing: Invariant Feature Attention

• Global Inducing Point Attention compresses node features into learnable inducing points, processes global context, then broadcasts it back to nodes.The module avoids full attention across all points and operates on invariant node features, retaining their spatial invariance.Source: §4.3 · p. 6

5 Experiments

• Training and evaluation use ShapeNetv2 airplane, chair, and car categories with 2,048 points per shape; the training sets contain 2,832, 4,612, and 2,458 shapes, respectively.Source: §5 · p. 6
• EMERGE exceeds all baseline metrics under standard splitting and leads 5 of 6 under LION; EMD 1-NNA is 50.71% for cars, 53.33% for airplanes, and 52.42% for chairs.Table 1 reports 1-NNA using Chamfer Distance and Earth Mover’s Distance, with 50% as the theoretical optimum.Source: Table 1 · p. 7, §5 · p. 7
• Generation quality is assessed with 1-NNA using Chamfer Distance and Earth Mover’s Distance, where 50% accuracy is the ideal distribution match.Source: §5 · p. 7, Table 1 · p. 7
• EMERGE converges in 3,100 epochs for airplanes, 2,600 for chairs, and 7,000 for cars, compared with LION’s 32,000 total epochs.The LION regime comprises 24,000 epochs for its latent DPM and 8,000 for its VAE.Source: §5 · p. 8
• Figure 3 compares 2,048-point airplane, chair, and car samples; EMERGE is reported to produce clear local surfaces and globally coherent shapes relative to other methods.Source: Figure 3 · p. 7, §5 · p. 8
• Across all 55 ShapeNet categories, EMERGE generates diverse, structurally complex point clouds without explicit class conditioning.Source: §5 · p. 8, Figure 4 · p. 8

6 Zero-Shot Super-Resolution

• EMERGE supports zero-shot super-resolution at arbitrary inference densities above the training density; without alignment, higher-resolution inference can distort geometry and collapse structure.The aligned pipeline preserves global topology and fine-grained details.Source: §6 · p. 8, Figure 5 · p. 8, §6.1 · p. 9
• At higher density, voxel counts rise with the density ratio and shift the count-conditioning FiLM inputs beyond their training distribution.Source: §6 · p. 8
• Point-count alignment scales inference voxel counts by Ntrain/Ninf before normalization to preserve the count-conditioning input distribution.Source: §6 · p. 8
• Under an approximately uniform locally smooth 2D surface assumption, k-NN receptive-field radius shrinks with higher point density; dilated k-NN restores rtrain while retaining k neighbors.The dilation factor is d = floor(Ninf/Ntrain), querying k·d neighbors and subsampling every d-th neighbor.Source: §6 · p. 9

6.1 Super-Resolution Inference Pipeline

• The inference pipeline upsamples a base cloud by duplication, injects forward noise to separate overlapping points, then denoises from an intermediate timestep using both alignment modifications.The reported qualitative examples denoise from tres = 150 with 30 DDIM steps.Source: §6.1 · p. 9

7 Conclusion

• Canonical voxel pooling and global invariant feature attention support modeling complex hierarchical geometries while preserving continuous spatial symmetries.Source: §7 · pp. 9–10
• The conclusion reports that EMERGE achieves state-of-the-art generation fidelity and converges in significantly fewer training epochs than existing methods.Source: §7 · p. 10
• The authors state that density-induced distribution shifts can be isolated and negated, enabling zero-shot, high-fidelity generation at arbitrary inference resolutions.Source: §7 · p. 10

A SE(3) Equivariance of Voxel Pooling

• The canonical frame resolves eigenvector sign ambiguity using spatial extents and enforces a right-handed frame; degenerate covariance eigenvalues can make strict equivariance ill-defined.The authors report degeneracy as statistically insignificant for finite point clouds and observe it with pure Gaussian noise near t ≈ T, where alignment does not negatively affect downstream generation.Source: A SE(3) Equivariance of Voxel Pooling · p. 15
• Canonical-frame voxel pooling is exactly SE(3)-equivariant: its canonical representation and pooled nodes are invariant, while unprojected outputs transform with the input.The output satisfies P′ = PR + 1_Mt^T.Source: A SE(3) Equivariance of Voxel Pooling · p. 16

B.1 Definition of Invariant Features

• The local 13-dimensional invariant feature vector combines seven eigenvalue-based shape features with six absolute spatial moments along principal axes.The features are constructed from the covariance of each point’s k-nearest-neighbor relative vectors.Source: §B.1 · pp. 16–17

1. Translation Invariance; 2. Covariance Rotation; 3. Invariance of Eigenvalues; 4. Invariance of Spatial Moments

• The first seven features derive entirely from covariance eigenvalues and are strictly SE(3)-invariant; identical scalar projections make absolute-moment features 8–13 invariant, so all 13 features are SE(3)-invariant.Multiplying equivariant relative vectors by a function of these invariant features preserves the coordinate update mechanism’s SE(3)-equivariance.Source: §3 · p. 17, §4 · p. 18

C Feature-wise Linear Modulation (FiLM)

• FiLM layers inject conditioning information into intermediate features without disrupting coordinate-update equivariance, using a residual scale factor of 1 + γ for stable early training.Conditioning can be a global vector per shape or a value per node.Source: C Feature-wise Linear Modulation (FiLM) · p. 18

D Voxel Counts-Conditioning Normalization

• Voxel counts provide structural-density context, but their wide variation can destabilize features; the method log-transforms counts and min-max normalizes them per shape before FiLM conditioning.The bounded conditioning scalar is in [0, 1], and ε = 10^-8 prevents division by zero.Source: D Voxel Counts-Conditioning… · p. 18

E.1 Proof of Scale-Count Expectation Alignment

• For zero-shot super-resolution, raw inference voxel counts are scaled by N_train/N_inf before normalization to preserve the expected input distribution of count-conditioning FiLM layers.The expectation-alignment proof assumes uniform point sampling over a locally smooth 2D surface manifold.Source: §E.1 · p. 19, E Proofs for Zero-Shot Distribution… · p. 19
• Linear scaling aligns the expected FiLM-layer input distribution across all voxel scales with the training distribution, neutralizing density shift.The scaled inference count is introduced to align distributions before logarithmic normalization and FiLM conditioning.Source: §E.1 · p. 19

E.2 Proof of Dilated k-NN Receptive Field Preservation

• At inference density d times training density, querying k_ext = k·d neighbors and subsampling every d-th preserves the training receptive-field radius and selects exactly k neighbors.The proof assumes a locally smooth manifold and uniform point density.Source: §E.2 · p. 20
• Preserving exactly k neighbors keeps message-passing aggregation magnitude and variance in-distribution.The EGNN aggregates exactly k terms.Source: §E.2 · p. 20
• Dilated k-NN is used only for the base point-cloud EGNN; voxel-pooled EGNNs retain standard k-NN because their voxel locations and densities do not change with inference resolution.Source: §E.2 · p. 20

F Super-Resolution Inference Pipeline Pseudo-code

• The super-resolution pseudocode specifies base initialization, deterministic upsampling, forward noise injection, and parameter-free distribution alignment during reverse denoising.Source: F Super-Resolution Inference Pipeline… · p. 20

• The paper states that the architecture’s remaining operations are invariant to total point count, with no hidden distribution shifts when inference uses more points than training.Source: G Higher-Resolution Distribution-Shift… · p. 20
• The inference procedure initializes a base cloud, upsamples it deterministically, injects forward noise, then applies distribution alignment during reverse diffusion transitions.The denoising loop may use DDPM or DDIM sampling.Source: G Higher-Resolution Distribution-Shift… · p. 21
• At higher inference resolution, scale voxel counts by N_train/N_inf and use mean pooling, whose expected feature magnitude is independent of the number of points in a voxel.Mean pooling acts as a denser Monte Carlo estimate of the local continuous expectation.Source: G Higher-Resolution Distribution-Shift… · p. 21
• Dilated k-NN is constructed only at the base point-cloud scale; matching neighborhood cardinality and spatial bounds keeps coordinate and feature aggregation statistics unshifted.Source: G Higher-Resolution Distribution-Shift… · p. 21
• Global inducing-point attention uses fixed C inducing points with C much smaller than N; normalized weights yield a convex combination whose expected magnitude and variance remain aligned as N increases.The model does not use positional embeddings for point features.Source: G Higher-Resolution Distribution-Shift… · p. 22

H Extended Experimental Setup and Implementation Details

• The six-block architecture interleaves multi-scale EGNNs and global inducing-point attention; evaluation applies PCA to generated and ground-truth shapes before pose-dependent CD and EMD metrics.The setup uses seven resolution branches, 128 inducing points, and eight attention heads; generation takes about 41 seconds for 100-step DDIM, while 15,000-point super-resolution takes about 34 seconds with 30 steps.Source: H Extended Experimental Setup and… · p. 22

I Min-SNR Weighting Strategy Formulation

• The training objective uses Min-SNR-γ weighting to mitigate instability and slow convergence from conflicting optimization directions across diffusion timesteps, with γ set to 5.Source: I Min-SNR Weighting Strategy Formulation · pp. 22–23

J Extended Qualitative Results

• The extended qualitative results section presents additional examples to demonstrate EMERGE’s generative capacity, diversity, and scalability.Source: J Extended Qualitative Results · p. 23

J.1 Standard Benchmark Categories

• For Airplane, Car, and Chair, Figure 7 samples are described as high-fidelity geometries with smooth surfaces and sharp structural details.Source: §J.1 · p. 23, Figure 7 · p. 24

J.2 55-Class Generation

• Figure 8 presents samples across all 55 ShapeNet classes, spanning dense volumetric objects and intricate thin structures.Source: §J.2 · p. 23, Figure 8 · p. 24

J.3 Zero-Shot Super-Resolution

• Without distribution alignment, Figure 9 super-resolution attempts collapse structurally; Figure 11 reports geometric stability at 40,000 points.Source: §J.3 · p. 24

K Limitations

• Dynamic k-nearest-neighbor graph construction and analytical eigenvector computations introduce noticeable latency.The eigenvector computations support canonical-frame extraction and invariant feature representations.Source: K Limitations · p. 24
• Wall-clock time per training step and inference pass is higher than for standard diffusion baselines.This is the trade-off despite EMERGE requiring vastly fewer total epochs to converge.Source: K Limitations · p. 25
• Supplementary samples cover varying resolutions across all 55 ShapeNet classes, Airplane, Car, and Chair alignment comparisons, and zero-shot super-resolution up to 40,000 points.Figure 9 compares samples with and without distribution alignment for Airplane, Car, and Chair.Source: Figure 10 · p. 25, Figure 9 · p. 25, Figure 11 · p. 26