PaperlayerView exact revision on arXiv

Source-linked AI summary

Generative Modeling with Orbit-Space Particle Flow Matching

Sinan Wang, Jinjin He, Shenyifan Lu, Ruicheng Wang, Greg Turk, Bo Zhu

arXiv:2605.02222v1cs.GRcs.CV

TL;DR

Particle generators must handle permutation symmetry and can exploit physical-space geometry that grid-based flows leave unused. OGPP addresses both through terminal canonicalization, identity embeddings, and geometric probability paths, achieving strong results across minimal-surface, ShapeNet, and single-shape encoding tasks. The framework remains limited by fixed particle counts, quadratic full attention, non-Wasserstein paths, and more moderate canonicalization gains on highly variable data.

  • Problem

    Particle systems are permutation-symmetric physical-space representations, but standard grid-based generative models do not efficiently transfer to them.

  • Method

    OGPP combines terminal orbit-space canonicalization, particle index embeddings, and geometric probability paths whose terminal velocities encode surface normals.

  • Results

    OGPP performs strongly across minimal-surface generation, ShapeNet generation, and single-shape encoding, including up to two orders of magnitude lower minimal-surface metric error in one step.

  • Takeaways & Limitations

    The results support particle-native flow matching with structured identities, symmetry-aware endpoints, and geometric attributes carried by terminal velocity.

  • Takeaways & Limitations

    The framework uses a fixed particle count and full attention, while its geometric paths are not Wasserstein-2 optimal transport and canonicalization helps less on highly variable data.

Abstract

from arXiv · show

We present Orbit-Space Geometric Probability Paths (OGPP), a particle-native flow-matching framework for generative modeling of particle systems. OGPP is motivated by two insights: (i) particles are defined up to permutation symmetries, so anonymous indexing inflates per-index target variance and yields curved, hard-to-learn flows; and (ii) particles live in physical space, so the flow terminal velocity has physical meaning and can encode geometric attributes, e.g., surface normals. OGPP instantiates three key components: (1) orbit-space canonicalization of the probability-path terminal endpoint, (2) particle index embeddings for role specialization, and (3) geometric probability paths with arc-length-aware terminal velocities that generate normals as a byproduct of the flow. We evaluate OGPP on minimal-surface benchmarks, where it reduces metric error by up to two orders of magnitude in a single inference step; on ShapeNet, where it matches the state of the art with 5x fewer steps and reaches airplane EMD comparable to DiT-3D with 26x fewer parameters and 5x fewer steps; and on single-shape encoding, where it produces normals and reconstructions competitive with 6D generators while operating entirely in 3D.

1 Introduction

OGPP adapts flow matching to particle systems by addressing permutation symmetry and using geometric probability paths to encode particle attributes. Its evaluations span particle geometry, shape generation, and physics-oriented tasks.

  • Particle systems appear throughout graphics, including sampling, point-set geometry, physical simulation, and agent-based animation.
  • Permutation symmetry makes standard grid-based generative models ill-suited to particles because reindexing changes vectorized representations without changing configurations.
  • OGPP combines terminal orbit-space canonicalization, particle index embeddings, and geometric probability paths into one particle-native flow-matching framework.
  • Canonicalization orders terminal particles by a geometric criterion, while identity embeddings let indices specialize to stable canonical roles and reduce target variability.
  • OGPP reduces minimal-surface metric error by up to two orders of magnitude in one step, matches ShapeNet state of the art with 5× fewer steps, and remains competitive with 6D generators for encoding.
  • Geometric probability paths encode per-particle surface normals in terminal tangents, allowing positions and attributes to be generated by one flow.

2 Related Work

Related work covers continuous-time generative models, alternative probability paths, point-cloud generators, permutation handling, symmetry modeling, and physics-aware particle evaluation.

  • Continuous-time generative models use SDE or ODE formulations, including diffusion and flow-based models.
  • Alternative probability paths modify linear interpolation through blue-noise injection, generalized VP schedules, or latent flow-matching designs.
  • Point-cloud generation has progressed from GANs, VAEs, and CNF models toward scalable two-stage latent representations.
  • Permutation ambiguity is often handled through deterministic spatial orderings such as Morton or Hilbert curves, including in recent Transformer-based models.
  • Symmetry-aware flow models use OT couplings or related constraints, but reported training complexity can reach O(B^2N^3), limiting scalability.
  • Methods operating on voxel grids without orbit-space canonicalization retain an Eulerian formulation associated with ill-conditioned regression and highly curved flows.
  • Physics-grounded particle systems provide intrinsic evaluation criteria such as spectra, fractal dimensions, residual forces, and surface deviation beyond distribution-matching metrics.

3 Background

The background defines the paper’s notation, flow-matching probability paths, particle permutation symmetry, and canonicalization framework. It also explains how canonicalization concentrates terminal endpoints to simplify velocity regression.

  • 3.1 Naming Conventions: Bold symbols denote vectors or vector fields, capital letters denote random variables, and lowercase bold letters denote realizations or fixed values.
  • 3.2 Flow Matching: Flow matching trains a neural velocity field that transports p_init to p_data by integrating an ODE.
  • 3.2 Flow Matching: A conditional probability path interpolates from the initial noise distribution to a point mass at each target data point at t=1.
  • Independent and OT-based couplings spread possible terminal endpoints, whereas orbit-space canonicalization concentrates them into a smaller region.
  • This concentration is expected to reduce conditional covariance and simplify the velocity regression.
  • The relevant particle symmetry is G=S_N, acting by permuting particle indices after global pose normalization.
  • A canonicalization map selects one representative from each symmetry orbit using invariance and orbit-membership conditions, commonly through deterministic particle ordering.

4 Orbit-Space Probability Paths

This section combines orbit-space canonicalization with particle index embeddings to reduce permutation-induced target variability and simplify velocity learning. It further uses orbit-continuous canonicalization and one-sided endpoint treatment to encourage smoother, straighter flows, while a plain Transformer supports conditional particle generation.

  • Model Architecture: The model uses a plain Transformer encoder with particle index and global time embeddings, plus optional cross-attention for conditional generation.Variable anchor counts are handled by padding condition tokens with learnable missing embeddings and masking padded positions.
  • Conditional Generation: 3 inference steps produce smooth and accurate minimal-surface boundaries across anchor counts from 3 to 8.A single conditional model is trained to generalize across varying numbers and positions of anchors.
  • Orbit-Space Canonicalization and Index Embeddings: Canonicalizing the terminal endpoint X1 removes the additional covariance caused by random permutations in the regression target.The remaining covariance reflects intrinsic variability rather than variability from randomly permuted particle roles.
  • Orbit-Space Canonicalization and Index Embeddings: Orbit-space canonicalization and particle index embeddings work together to untangle mixed particle roles and produce easier-to-learn velocity targets.Canonicalization removes permutation-induced role ambiguity, while identity embeddings let indices specialize to distinct velocity-field roles.
  • Orbit-Continuous Canonicalization: A canonicalization map that sends nearby orbits to nearby representatives yields a continuous Bayes-optimal velocity field and encourages straighter particle trajectories.The velocity-field regularity is controlled by the orbit-continuity of the canonicalization map and the smoothness of the canonical means.
  • One-Sided Canonicalization: Canonicalizing only X1 avoids the directional cancellation and inflated local Lipschitz ratios associated with additionally canonicalizing the noise endpoint X0.The one-sided X1 strategy achieves the smallest local Lipschitz ratios and lowest prevalence of high-cancellation edges among the four tested regimes.

5 Geometric Probability Paths for Attribute Encoding

OGPP replaces linear probability paths with quadratic Hermite curves whose terminal tangents encode surface normals, while arc-length-aware magnitudes improve trajectory sampling. Joint endpoint canonicalization further reduces regression noise and improves normal-aware generation.

  • 5 Geometric Probability Paths for Attribute Encoding: Standard linear interpolation leaves terminal velocity geometrically unused, whereas OGPP turns it into a carrier of per-particle attributes.The geometric path uses this additional degree of freedom to encode surface normals.
  • 5 Geometric Probability Paths for Attribute Encoding: Quadratic Hermite paths bend trajectories so terminal velocity aligns with each particle’s surface normal.The path starts at noise, ends at the target position, and uses the terminal tangent to encode the normal.
  • 5.2 Arc-Length Terminal Velocity (ATV): Arc-length-aware terminal velocity uses chord length and chord-normal alignment to produce approximately uniform speed profiles under uniform time sampling.When the chord and normal align, the path is nearly straight; greater misalignment increases terminal speed to compensate for bending.
  • 5.3 Joint Canonicalization for Attribute-Encoded Paths: Joint canonicalization reduces conditional covariance for geometric-path regression because the target depends on the joint position-terminal-velocity endpoint.The method applies Hilbert ordering to concatenated position-normal vectors in R6.
  • 5.3 Joint Canonicalization for Attribute-Encoded Paths: 0.92 average cosine similarity and 0.61 1-NNA accuracy are achieved with joint canonicalization, compared with 0.91 and 0.78 under position-only ordering.Joint canonicalization improves both normal estimation and generation quality in the reported ablation.
  • 5.4 Marginal Velocity at Terminal Time: At terminal time, the marginal velocity equals the expected surface normal conditioned on position and supplies normals without additional computational cost.The generated point cloud obtains surface normals by normalizing the learned terminal velocity.

6 Algorithm Overview

OGPP training canonicalizes terminal endpoints, specializes particle indices with identity embeddings, and trains on geometric paths with arc-length-aware terminal velocities. Inference integrates the learned velocity field from noise to generate particle positions and surface normals.

  • 6 Algorithm Overview: Training canonicalizes terminal position-normal endpoints to reduce conditional covariance and straighten the learned flows.Joint canonicalization is applied before constructing training trajectories.
  • 6 Algorithm Overview: Particle index embeddings let each index specialize to its canonical role, separating mixed regression targets into easier-to-learn trajectory families.The embeddings are provided to the velocity network during training.
  • 6 Algorithm Overview: Geometric probability paths replace linear paths with quadratic Hermite paths whose terminal tangents encode surface normals using arc-length-aware velocity.This component is implemented during per-particle trajectory construction.
  • 6 Algorithm Overview: Inference draws noise and integrates the learned velocity field forward from t=0 to t=1 with a standard ODE solver.The final positions are generated particle locations, while terminal velocity yields normalized surface normals.

7 Experiments

OGPP performs strongly across energy-driven particle generation, minimal-surface synthesis, ShapeNet generation, and single-shape reconstruction. Its advantages include early-step accuracy, competitive distribution quality with fewer inference steps, and jointly generated normals that support reconstruction.

  • Energy-driven Particle Generation: OGPP achieves better metric values with fewer inference steps and typically converges earlier than the baselines on energy-driven tasks.Its error remains low from early steps, while baselines converge more slowly and plateau at higher error levels.
  • Blue-Noise Generation: Pearson correlation reaches 0.999 and L2 error 0.014 for the 26M model on uniform blue-noise generation.
  • Minimal Surfaces: OGPP reduces minimal-surface metric error by up to two orders of magnitude in one inference step and produces accurate boundaries across varying anchor counts.With three anchors, the method achieves area fraction error 0.004, angle smoothness 0.33, and uniformity CV 0.34 in one step; with 3–8 anchors, it produces smooth boundaries in three steps.
  • Fractal Generation: OGPP produces realistic fractal morphology at 10 inference steps and maintains the closest resemblance to the dataset sample at 200 steps.
  • Multilayer Thomson Problem: At 20 steps, OGPP obtains tangential force RMS 4.99 versus 102.4 for Original FM and generates configurations with uniform shell spacing and proper inter-shell separation.The lower force RMS is reported as suggesting configurations closer to energy minima.
  • ShapeNet Generation: OGPP matches NSOT with approximately 5× fewer inference steps and reaches airplane EMD comparable to DiT-3D with roughly 26× fewer parameters and 5× fewer steps.It also achieves the lowest FID and KID among the flow-matching baselines on ShapeNet airplanes.
  • Single-shape Encoding: OGPP produces accurate, consistently oriented normals that enable high-quality Screened Poisson reconstructions across thin structures, solid objects, and complex topologies.PCA-based normals suffer orientation failures and reconstruction artifacts at thin structures such as wings and tail fins.
  • Ablations: The ablation results show that combining orbit-space canonicalization with identity-conditioned trajectories is necessary for smooth, well-formed minimal surfaces in one ODE step.Identity embeddings alone fail to produce high-quality surfaces, while canonicalization without identities is only comparable to vanilla flow matching.

8 Discussion

The discussion identifies when canonicalization helps most and how OGPP can adapt its ordering strategy to domain structure. Benefits are largest for simple, structurally consistent configurations and more moderate for diverse real-world shapes.

  • Why canonicalization helps: Canonicalization concentrates each particle index in a stable spatial region by assigning consistent roles to structurally corresponding particles.
  • Conditions for reduced benefit: Canonicalization provides smaller gains when samples share less common structure or when the chosen ordering captures that structure less effectively.On complex real-world shapes such as diverse ShapeNet categories, the gains are more moderate than on synthetic minimal-surface benchmarks.
  • Domain-specific alternatives: OGPP can replace Euclidean space-filling curves with domain-specific orderings, including counterclockwise polygon ordering for minimal surfaces and potential spectral sorting for articulated shapes.

9 Conclusion

OGPP is a particle-native flow-matching framework that respects permutation symmetry and uses physical-space dynamics to generate particle geometry. Its limitations concern scalability, path optimality, canonicalization design, and dependence on dataset regularity.

  • 9 Conclusion: OGPP treats particles as persistent entities with identities, trajectories, and geometry-aware dynamics rather than as grid-based representations.The framework investigates whether respecting permutation symmetry and physical semantics improves particle-generation learning.
  • 9 Conclusion: OGPP combines terminal canonicalization, particle index embeddings, and geometric probability paths whose terminal velocities can carry per-particle attributes such as surface normals.Canonicalization addresses permutation ambiguity, embeddings support identity-aware conditioning, and geometric paths encode attributes in the flow.
  • 9 Conclusion: Fixed particle counts and full attention limit scalability because attention has quadratic cost in the number of particles.The current framework does not yet support variable particle counts.
  • 9 Conclusion: The geometric probability paths are not Wasserstein-2 optimal-transport geodesics and may therefore induce slightly more curved flows.The limitation follows from lacking the geodesic property of W2 displacement interpolation.
  • 9 Conclusion: Canonicalization introduces a design degree of freedom whose induced locality or ordered semantics are not explicitly exploited for additional information.The choice of canonicalization remains underutilized in the current framework.
  • 9 Conclusion: Canonicalization helps less on highly variable datasets than on structurally regular benchmarks.The benefit depends on how much common structure the data exhibits.

A Flow Matching Details

Flow matching constructs conditional paths from an initial distribution to data points and trains a velocity field to follow them. Marginalization makes the target field well-defined, while conditional flow matching provides a tractable training objective.

  • A Flow Matching Details: Each conditional probability path starts from p_init at t = 0 and collapses to a point mass at x_1 at t = 1.The path describes transport of noise samples toward the terminal data location.
  • A Flow Matching Details: Averaging conditional paths over data samples produces a marginal path interpolating between p_init and p_data.The induced ODE follows the corresponding conditional velocity field.
  • A Flow Matching Details: The marginal velocity field is a posterior-weighted average of conditional velocities, with weights given by the posterior of x_1 given X_t = x.This marginalization identity yields an ODE that transports p_init along the marginal path.
  • A Flow Matching Details: Flow matching avoids directly evaluating the intractable marginal field by regressing the model velocity onto conditional reference velocities.The conditional flow-matching loss is equivalent up to a constant to a marginal regression loss.

B Group Theory Details

The group-theoretic formulation represents particle permutations as symmetries acting on configurations. Orbits collect equivalent configurations, while canonicalization selects one invariant representative for each orbit after pose normalization.

  • B Group Theory Details: A group is a set with an associative binary operation, an identity element, and inverses, and it acts on a set through a structure-preserving map.The action satisfies identity and composition consistency.
  • B Group Theory Details: Particle configurations are first normalized for global translations and rotations, leaving permutations as the remaining symmetry.PCA-based alignment may remain ambiguous under axis sign flips or degenerate inertia tensors.
  • B Group Theory Details: An orthogonal representation maps each group element to an orthogonal matrix acting on Euclidean configuration space.Orthogonal matrices preserve inner products and satisfy R^T R = I.
  • B Group Theory Details: An orbit is the set of configurations reachable through group transformations, and configurations in one orbit represent the same object under the symmetry.A G-invariant function is constant on each orbit.
  • B Group Theory Details: A canonicalization map selects an orbit representative invariantly, with its output remaining inside the input configuration’s orbit.These conditions establish a correspondence between orbits and canonical representatives.

C Conditional Covariance Reduction via Orbit-Space Canonicalization: Detailed Derivation

The derivation shows that orbit-space canonicalization removes permutation-induced variability from endpoint targets, reducing conditional covariance and irreducible regression error. It also explains why canonicalizing the initial endpoint can instead worsen local velocity smoothness through directional cancellation.

  • C Conditional Covariance Reduction via Orbit-Space Canonicalization: Detailed Derivation: The total-covariance decomposition separates intrinsic fixed-permutation variability from an additional positive semidefinite covariance caused by random permutations.The additional term vanishes when the conditional expectation is permutation-independent.
  • C Conditional Covariance Reduction via Orbit-Space Canonicalization: Detailed Derivation: A smaller conditional covariance directly lowers the irreducible mean-squared error of the regression target.The covariance is scaled by 1/(1 − t)^2 in the stated target-covariance relation.
  • C Conditional Covariance Reduction via Orbit-Space Canonicalization: Detailed Derivation: Canonicalizing the terminal endpoint makes the conditional law of the canonicalized endpoint independent of the random permutation.This removes permutation variability from the endpoint distribution conditioned on the intermediate state.
  • C Conditional Covariance Reduction via Orbit-Space Canonicalization: Detailed Derivation: Permutation actions preserve covariance trace because they are orthogonal transformations, allowing the covariance reduction result to be compared through traces.Taking traces and combining the stated equations yields the main result.
  • C Conditional Covariance Reduction via Orbit-Space Canonicalization: Detailed Derivation: Canonicalizing both endpoints can make displacement vectors similar in scale, increasing near-cancellation in interpolated differences.Nearest-neighbor selection then favors configurations with small denominators but non-tiny numerators.
  • C Conditional Covariance Reduction via Orbit-Space Canonicalization: Detailed Derivation: Canonicalizing the terminal endpoint contracts pairwise dispersion by a factor α_0, with α_0 < 1 for nontrivial group symmetry.The contraction is used in the analysis of local velocity-field behavior.
  • C Conditional Covariance Reduction via Orbit-Space Canonicalization: Detailed Derivation: Two-sided canonicalization therefore yields large local Lipschitz ratios and a less smooth velocity field, whereas leaving X_0 uncanonicalized avoids most such cancellation events.The contrast is between one-sided terminal canonicalization and canonicalization at both endpoints.

E Orbit-Continuous Canonicalization and Straight Flows: Detailed Derivation

The analysis assumes smooth endpoint distributions over orbit space and requires a well-behaved canonicalization to preserve that regularity. This yields local Lipschitz control of the velocity field in both orbit and Euclidean configuration spaces.

  • Smoothness assumptions: The endpoint distribution is assumed to vary smoothly over nearby particle orbits.A poorly behaved canonicalization can introduce abrupt representative changes and destroy this smoothness.
  • Canonicalization regularity: Canonicalization must preserve orbit-space regularity rather than introduce discontinuous representative changes.The choice of canonicalization controls the orbit-Lipschitz constant used in the analysis.
  • Velocity regularity: Eq. (11) provides a local Lipschitz bound for the velocity field from orbit-Lipschitz regularity and canonical-mean smoothness.The time-dependent constant L_vel(t) depends on the orbit-Lipschitz constant and intrinsic smoothness of the canonical means.
  • Velocity regularity: The orbit-space bound also controls velocity-field regularity in the Euclidean configuration space used by the network.This follows when the standard orbit metric is used.

F Arc-Length Terminal Velocity: Detailed Discussion

The arc-length terminal velocity design addresses the nonuniform progress induced by normalized terminal velocities. It selects or approximates terminal speeds that minimize speed variation while retaining the prescribed terminal normal direction.

  • Motivation: Normalized terminal velocity assigns equal terminal speed to paths with different chord lengths, producing nonuniform speed profiles over time.Under NTV, distant and nearby points must compensate through early acceleration or deceleration, making t a poor progress surrogate.
  • Optimal speed selection: For the quadratic Hermite path, terminal-speed scaling is optimized by minimizing the variance of speed over t ∈ [0, 1].The terminal velocity is written as v_1 = α n̂_1, with α selected through a one-dimensional optimization.
  • Optimal speed selection: The optimized terminal velocity preserves the normal direction while making each Hermite trajectory’s speed as uniform as possible.The scalar speed factor is chosen separately for each initial point, endpoint, and terminal normal.
  • Practical approximation: The ATV approximation uses only chord length and chord-normal alignment to efficiently approximate the speed-variance optimum.It requires only norms and dot products and empirically produces more uniform speed profiles than NTV.
Loading 2605.02222v1…