acceptodds
Under review as a conference paper at ICLR 2027

Capacitor Chain Field Matching for Electrostatic Generative Modeling

Abstract

Electrostatic Field Matching (EFM) constructs data-to-data generative transport by placing source and target distributions on oppositely charged plates in an augmented space. However, its single-capacitor geometry produces backward-oriented and exterior field lines, requiring stochastic branching for exact transport and complicating practical ODE-based inference. We propose Capacitor Chain Field Matching (CFM), a framework that addresses these geometric limitations and revisits how electrostatic transport fields are learned. First, CFM replaces a single pair of charged plates with an infinite periodic chain of alternating source and target distributions. The resulting symmetry confines field lines between adjacent plates. We prove that the ideal construction transports the source distribution to the target along forward-oriented field lines. Second, we derive a training objective with single-sample targets and theoretically grounded perturbation kernels, replacing EFM's multi-sample field estimates and heuristic perturbation scheme. Together, these contributions preserve EFM's electrostatic interpretation while enabling practical transport and training without multi-sample field estimation.

Primary area:
generative models
Keywords:
generative modeling, electrostatic field matching, Poisson Flow generative models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_24NQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents a method for data-to-data transport using a physics inspired approach which views transport paths as electrostatic field lines. This idea was originally implemented in the electrostatic flow matching (EFM) paper, where target and source data in RD\mathbb{R}^D was embedded in RD+1\mathbb{R}^{D+1} along two hyperplanes P\mathbb{P} (positive) and Q\mathbb{Q} (negative), mimicking an electronic capacitor. Then, electrostatic field lines between P\mathbb{P} and Q\mathbb{Q} would correspond to transport paths for the data. In practice, EFM suffered from pathologies which limited its effectiveness in practice, most notably in that oftentimes the field lines would not lie entirely in [0,L][0,L] making ODE integration hard and necessitate heuristics to quell. This paper introduces capacitor chain field matching (C$^2$FM), which replaces the two plate approach with an repeating hyperplane pattern, with positively charged planes at ±L,±3L,±5L,...\pm L, \pm 3L, \pm 5L, ... and negatively charged planes at 0,±2L,±4L,...0, \pm 2L, \pm 4L, .... The reflection symmetry induced by the periodic structure allows forces the field lines to be forward oriented (Theorem 3.1), and thus forces the particles to stay in [0,L][0,L]. In practice, the authors approximate this with CC plates of each charge, and show how increasing CC gets closer and closer to the theoretical limit of totally forward ordered field lines. The authors first demonstrate the coherence of the method with simple experiments transporting gaussian mixtures and MNIST digits, before moving onto the learning experiments where they demonstrate learned data transport. This learning also introduces an objective function containing three terms: a denoising, axial, and mixing loss; importantly, this loss can be estimated with a single latent sample, another improvement over EFM which requires multiple per loss computation.

Critical strengths

-The authors propose a novel solution to fix the pathologies of the EFM approach by forming a capacitor chain of plates, which induces reflection symmetry which fixes several of the problems in the EFM method.

-The authors provide theoretical justification for their approach (Theorem 3.1, 3.2, 5.1), along with empirical justification for these theorems.

-The authors quantify their method with MNIST translation and generation compared to traditional flow matching.

Critical weaknesses

-The method is very computationally expensive. At large CC, the field becomes very expensive to compute and slow to train. This seemed to be why the only comparisons done were on a simple MNIST task and not more ambitious data-to-data transport targets.

-Much of the paper was spent comparing against EFM, but the evaluations only compared against the original Flow matching method, which seemed to defeat the purpose.

Questions for authors

-It seems possible that one could try to embed the data distribution in a naturally periodic domain (like a torus) instead of trying to approximate the capacitor chain with large CC. Would this be possible/was this attempted?

-More generally, do the authors see any way that this method could potentially enjoy greater computational success via further algorithmic improvements? As it stands now the method is mostly just theoretically pleasant, but lacking in real world applicability.

-It seems like there are several extra tricks employed needed to make the ODE integration work (in 5.2). How do these compare to the heuristics used by EFM?

-Possibly crazy idea: An attractive property of having a physically inspired solution is that one could try to build a physical device that produces an electric field that yields a 2 dimensional transport path. Have the authors considered if this is possible or if this would be useful for anything?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 24NQ (Part 1/2)

Q1: Periodic-domain formulation

Yes, a periodic-domain formulation is possible. The infinite capacitor chain has period 2L2L along the auxiliary coordinate zz. We can therefore identify zz with z+2Lz+2L and represent the chain by one source distribution and one target distribution on a circle of circumference 2L2L. Repeating this cell recovers the infinite chain.

We use this periodic structure in practice through the Poisson-summed Fourier representation in Appendix F.2. Its coefficients account for the infinite repetition of the charges. This allows us to evaluate the field using Fourier modes without explicitly summing a large number of image copies.

In practice, we combine this Fourier representation with direct image sums. We describe the switching rule and report its cost and numerical accuracy in W1 below and Appendix H.3.

Thus, our implementation already exploits periodicity along the auxiliary axis. The data coordinates remain Euclidean. Embedding the data coordinates themselves on a torus would introduce a different geometry, which we have not explored experimentally.

Q2: Further algorithmic improvements

Yes. In the revision, we introduce an algorithmic extension and evaluate its practical effect on more challenging image tasks.

We extend C$^2$FM to multiple auxiliary dimensions. As in PFGM++, increasing dd yields lighter-tailed training perturbations. Our ablation shows that the d=4096d=4096 configuration achieves substantially lower FID than d=1d=1 at the same training budget on both Shoes→Bags and CIFAR-10 (Appendix H.2). Section 3.1 establishes confinement and exact transport for the generalized infinite-chain construction.

Using d=4096d=4096, we add unpaired Shoes→Bags translation and unconditional CIFAR-10 generation at 32×3232\times32, achieving FIDs of 5.2 and 8.0, respectively. Section 4.3 and Figure 4 report the comparison with Flow Matching and show samples. These results demonstrate that the same electrostatic formulation can produce realistic images for both types of tasks.

We also combine image sums with a Fourier representation to evaluate the chain efficiently. As detailed in our response to W1 and Appendix H.3, the combined evaluator adds only 3.6% to full training iteration time on the Shoes→Bags benchmark relative to the matched C=0C=0 baseline.

Further directions include improving the analytic mixing-logit approximation and extending the method to higher resolutions; we discuss these remaining limitations in the revised manuscript.

Q3 + W2: Integration choices and comparison with EFM

Our main integration change is arc-length parametrization. In C$^2$FM, contributions from the image charges can nearly cancel in the axial direction. Using the axial coordinate as time then requires dividing by a very small axial component, making integration poorly conditioned. Arc-length parametrization avoids this division by normalizing the full field instead (Section 4.2).

Arc-length parametrization addresses numerical conditioning in C$^2$FM, but it does not resolve EFM’s transport limitations. It changes the speed along a field line, not the line itself or where it crosses the target plane. Applying it to EFM therefore preserves the exterior branches: trajectories that first cross the target plane far from the target data still do so, even with exact integration.

In our learned-field comparison (Appendix H.1), EFM and C$^2$FM use matched training and inference protocols: the same backbone, denoising objective, training budget, source inputs, arc-length solver, and stopping rule. Under this shared protocol, C$^2$FM achieves FID 1.0, compared with 21.7 for EFM. Some EFM outputs remain saturated noise, while C$^2$FM produces recognizable threes from the same inputs. This reproduces the pattern observed with the exact fields in Figure 3, despite both learned fields receiving the same numerical treatment.

AuthorsOpenReview ↗

Response to Reviewer 24NQ (Part 2/2)

Q4: Could the construction be realized as a physical device?

This is an interesting physical question. A literal realization with ordinary conducting capacitor plates would not reproduce the construction, because free charges on a conductor redistribute in electrostatic equilibrium rather than remaining in an arbitrarily prescribed nonuniform surface-charge profile. A more plausible analogue would use insulating layers with deliberately deposited or trapped charge, so that simple spatial charge distributions, for example, Gaussian-like profiles or mixtures, could be imposed approximately.

The resulting field lines could also be visualized experimentally. A standard electrostatics demonstration uses fine polarizable grains suspended in an insulating viscous medium; the grains align with the local electric-field direction and reveal the field-line pattern. A similar setup could therefore provide a physical visualization of the transport geometry between two prescribed charge distributions. We view this mainly as an analogue or educational realization rather than as a practical generative-computing device.

W1: Computational cost at large CC

Increasing CC adds terms to the analytical sums used for sampling and target construction. For finite chains, we evaluate the 2C+12C+1 image terms per side in parallel in FP64 on the GPU. The number of neural network evaluations per training iteration does not depend on CC.

The table below reports the mean time per full training iteration over 36 timed iterations per configuration on a single A100, including sampling, target construction, and model updates. Times are in milliseconds; parentheses show overhead relative to the matched C=0C=0 denoiser baseline.

ConfigurationCMNIST 16×1616\times16, batch 128Shoes→Bags 32×3232\times32, batch 256
C=0C=0152.98638.45
C=4C=4162.19 (+6.0%)647.41 (+1.4%)
C=16C=16161.80 (+5.8%)648.07 (+1.5%)
C=64C=64163.06 (+6.6%)648.77 (+1.6%)
C=32C=32 + Poisson175.69 (+14.8%)661.23 (+3.6%)

For the tested finite chains, the overhead remains below 7% on CMNIST and 2% on Shoes→Bags.

To approximate the infinite chain, we combine the image and Fourier representations described in Appendix F.2. In this benchmark, we use the C=32C=32 image sum for R/L<20R/L<20, where R=∥x−x′∥R=\lVert\mathbf{x}-\mathbf{x}'\rVert is the data-space distance between the query and the data point, and LL is the plate spacing. For R/L≥20R/L\ge20, we use the Poisson-summed Fourier series to compute the small axial contribution directly, avoiding cancellation between image terms. This calculation uses five nonzero harmonics, plus the zero-frequency term for the kernel sum, in FP64 on the GPU.

The infinite-chain approximation adds 14.8% to iteration time on CMNIST and 3.6% on Shoes→Bags, as reported in the last row. Appendix H.3 also evaluates its numerical accuracy.

Concluding remarks

We hope these clarifications have fully addressed your concerns. Should any points require further elaboration, we are happy to provide it. Based on these comprehensive explanations, we are hopeful that you will consider raising your rating.

Reviewer_F91dOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper builds on Electrostatic Field Matching (EFM), which transports a source distribution P to a target Q. EFM places P and Q on oppositely charged hyperplanes in an augmented space and follows the electric field lines between them.

The paper makes two contributions.

  1. Geometric construction. C²FM replaces the single capacitor with an infinite periodic chain of alternating copies of P and Q. Reflection symmetry across every plane then gives a strictly positive axial field inside the slab, so a deterministic field-line map transports P to Q. Explicit bounds control the finite-truncation error.

  2. Learning procedure. The field direction is rewritten exactly in terms of four Coulomb-weighted conditional moments and a mixing logit. These are learned with single-sample, denoising-style objectives, with query points sampled from the Coulomb kernel itself. Inference integrates field lines in arc length with an adaptive step rule and returns the network’s final denoised estimate as the output.

Critical strengths
  1. The geometric idea is simple and the core theory is clean. The periodic chain enforces reflection symmetry, which forces the normal field to vanish on the plates off the data support. The maximum-principle argument for E_z > 0 is short and convincing.

  2. The finite-chain analysis. Theorems B.1 and B.2 give explicit truncation rates and an endpoint-stability bound. Most physics-inspired generative models lack this kind of analysis.

  3. The decomposition of the field into Coulomb-weighted conditional moments is a useful insight. It connects electrostatic models to denoising and yields consistent single-sample objectives. Sampling queries from the Coulomb kernel removes the need for EFM’s hand-designed training-volume distribution.

Critical weaknesses
  1. The writing is hard to understand. Key concepts are introduced in diversified terms (“axial clock”, “atoms”, “launch law”, “screening”, “readout”).These are not defined precisely in the main text. Heavy reliance on the appendix. The main text depends on equations that only appear there (Eqs. 64, 66, 68, 72, 75, 78, 88). Sections 5.1–5.2 cannot be read on their own. Disjointed structure. Experiments are split between Sec. 3.3 and Sec. 5.3, with related work placed between them.

  2. The experiments are far too limited to support the claims. Beyond 2D toy data, the only dataset is colored MNIST at 16×16 resolution (D = 768), with a single translation pair (2→3). There are no standard generation benchmarks (e.g., CIFAR-10, used by PFGM/PFGM++).

  3. Flow Matching is the sole comparison, and it “received no separate tuning”. Missing are EFM, IFM, PFGM++ (for generation), and translation methods such as rectified flow or OT-CFM. The Flow Matching translation FID of 16.85 is surprisingly poor for 2→3 on 16×16 MNIST and needs explanation.

Questions for authors
  1. Direct comparison with EFM. The paper’s motivation is fixing four limitations of EFM (Sec. 2.2), yet no experiment compares against a trained EFM. Could you compare C²FM with EFM, and ideally with IFM, on the same tasks using the same backbone, training budget and evaluation protocol?

2.Realistic data. Colored MNIST at 16×16 is the only non-toy dataset. Could you evaluate on standard benchmarks, such as CIFAR-10 generation against PFGM++ and Flow Matching, and a realistic unpaired translation task at higher resolution (e.g., CelebA attributes or AFHQ cat↔dog)? This would support the claim of practical transport.

  1. Measuring each limitation for the learned model compared with EFM. The exact-field figures (Figs 2, 3) are qualitative and do not involve the trained network. For the learned model, could you provide more detailed explanation and proof on how C²FM solves their limitations?
2 comments
AuthorsOpenReview ↗

Response to Reviewer F91d (Part 1/2)

Q3: Addressing EFM’s limitations in the learned model

EFM provides theoretical guarantees for transporting one distribution to another. However, implementing its exact transport requires handling the following difficulties.

Transport geometry

  1. Some trajectories travel outside the plates. The network therefore needs to learn the field both inside the slab and along these exterior trajectories.
  2. At the source plate, one must probabilistically choose whether to start toward the target plate or in the opposite direction.
  3. When a trajectory crosses the target plate, one must decide probabilistically whether to stop there or continue along the field.

The practical EFM implementation uses forward trajectories and stops at the first crossing. This simplification generally loses the exact-transport guarantee.

C$^2$FM resolves these difficulties through the reflection symmetry of the infinite chain. The exact infinite-chain field confines forward trajectories between the plates and provides exact transport (Section 3.1). Consequently, training can focus on the slab, and trajectories terminate at the target without a stochastic continuation decision.

Learning the field

  1. EFM estimates the field using several data samples and then normalizes this estimate. Normalization makes the resulting direction target generally biased.
  2. The distribution of training points must be chosen manually.
  3. In high dimensions, multi-sample estimates can be dominated by a few samples, limiting the benefit of using more samples. This behavior is observed for related estimators in [1].

Our learning procedure samples training queries from the Coulomb kernel and uses single-sample denoising targets. We prove that the population objectives recover the quantities needed to assemble the field direction (Section 4.1). This avoids normalizing a Monte Carlo field estimate and determines how to sample data-space queries once the auxiliary coordinates are chosen. The auxiliary-coordinate schedules remain explicit choices.

In the revision, we clarify which difficulties are resolved by the geometry and which by the learning procedure. The newly added comparison in Q1 below and Appendix H.1 tests whether the geometric advantage persists after learning, using the same denoising formulation for both fields.

Q1: Direct comparison with EFM

Please, note that Figure 3 already compares the exact EFM and C$^2$FM fields on CMNIST. Some EFM trajectories reach the target plane far from the target digits, producing saturated noise, while C$^2$FM produces recognizable threes from the same inputs. This experiment isolates the effect of the field geometry, without neural approximation error. Learning may alter individual trajectories, but it provides no guarantee of eliminating the exterior branches present in the exact EFM field.

Following your question, to further test whether this behavior persists after learning, we added a comparison on CMNIST 2→32\to3 (Appendix H.1). We train the EFM and C$^2$FM fields using the same backbone, a shared denoising formulation, and the same training budget. Both are trained on points inside the slab and evaluated from the same inputs with the same solver and stopping rule.

After 30,000 training steps, C$^2$FM achieves an FID of 1.0, compared with 21.7 for EFM. The samples reproduce the pattern observed with the exact fields, with some EFM outputs remaining saturated noise while C$^2$FM produces recognizable threes.

AuthorsOpenReview ↗

Response to Reviewer F91d (Part 2/2)

Q2 + W2 + W3: Experimental scope and high-dimensional extension

We have added experiments on unpaired Shoes→Bags translation and unconditional CIFAR-10 generation at 32×3232\times32 resolution (Section 4.3 and Figure 4). These evaluate the same electrostatic formulation on both data-to-data transport and noise-to-data generation.

For these experiments, we use our extension to multiple auxiliary dimensions with d=4096d=4096. This configuration achieves substantially lower FID than d=1d=1 at the same training budget (Appendix H.2). The additional note below explains the extension.

We compare with Flow Matching under matched training budgets and evaluation protocols. On Shoes→Bags, C$^2$FM achieves an FID of 5.2, compared with 81.4 for FM. On CIFAR-10, the corresponding scores are 8.0 and 5.1. Figure 4 shows samples from both methods. These results demonstrate that C$^2$FM can produce realistic images for both translation and generation within the same electrostatic formulation.

Our central goal is to address the geometric and learning limitations of EFM while retaining its ability to transport between two prescribed data distributions. This makes unpaired translation a central application and EFM the primary comparison, which we additionally evaluate with learned fields in Appendix H.1. Flow Matching provides a reference for sample quality across both tasks.

Regarding FM’s performance on the original CMNIST task and the newly added Shoes→Bags task, similar difficulties have been reported in prior work. The EFM paper notes that FM does not always correctly translate the shape and color of CMNIST digits [3]. Su et al. [4] identify sparse supervision as a challenge for data-to-data flow matching. With finite source and target datasets, training points lie along interpolation paths that sparsely cover high-dimensional space, which can lead to poor generalization of the learned velocity field. Our observations are consistent with this explanation.

W1: Writing and structure

We thank the reviewer for pointing this out. We revised the presentation in three ways. First, we removed or replaced informal and metaphorical terminology such as “axial clock”, “atoms”, “launch law”, “screening”, and “readout”, and now use consistent technical terminology throughout the paper.

Second, we rewrote Sections 5.1–5.2 so that they can be read independently of the appendices. The main text now explains the complete training and inference procedures at the level needed to understand the method, rather than referring the reader equation-by-equation to appendix material. We retain only a single pointer to the appendix for technical derivations and implementation details.

Third, we removed the standalone “Relation to Other Works” section and integrated the relevant comparisons and discussion into the corresponding parts of the paper. This removes the previous interruption between the theoretical and experimental sections and makes the overall narrative more continuous.

Additional note on auxiliary dimensions

With one auxiliary dimension, Coulomb-based training perturbations have heavy tails, so training must handle large deviations from the data. As observed in PFGM++ [2], increasing the auxiliary dimension makes these tails lighter. This motivates extending C$^2$FM beyond d=1d=1.

We generalize the construction to any d≥1d\ge1 and establish confinement and exact transport for the infinite chain (Section 3.1). Our ablation compares the d=1d=1 and d=4096d=4096 configurations at the same training budget. FID decreases from 84.4 to 5.2 on Shoes→Bags and from 25.4 to 8.0 on CIFAR-10 (Appendix H.2). These results motivate using d=4096d=4096 in the main experiments.

Concluding remarks

We hope these clarifications have fully addressed your concerns. Should any points require further elaboration, we are happy to provide it. Based on these comprehensive explanations, we are hopeful that you will consider raising your rating.

[1] Unlocking the Duality between Flow and Field Matching
[2] PFGM++: Unlocking the Potential of Physics-Inspired Generative Models
[3] Field Matching: an Electrostatic Paradigm to Generate and Transfer Data
[4] Three Forms of Stochastic Injection for Improved Distribution-to-Distribution Generative Modeling

Reviewer_tUdjOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes a new approach for electrostatic field matching, which stacks infinite periodic sequences of charged plates to resolve transport issues in single-capacitor geometry in data-to-data generative modeling. The resulting symmetry confines field lines between adjacent plates. The authors prove that the ideal construction transports the source distribution to the target along forward-oriented field lines. The authors construct finite-sample training objectives and empirically show improvement over the baselines on 2d toy problem as well as the high-dimensional MNIST dataset.

Critical strengths
  1. I am not familiar with electrostatic generative modeling, but I appreciate the authors' efforts toward clearly describing the relevant concepts.
  2. The idea is intuitive and well presented with theory.
  3. The application shows improvement over the baseline EFM approach.
Critical weaknesses

The usability of the approach in ultra-high-dimensional datasets such as CIFAR10 and ImageNet remains unexplored.

Questions for authors

None

1 comment
AuthorsOpenReview ↗

Response to Reviewer tUdj

We thank the reviewer for their positive assessment of our work and for highlighting the importance of evaluating the method beyond low-dimensional benchmarks.

We extend C$^2$FM from one auxiliary dimension to any d≥1d\ge1, preserving confinement and exact transport for the infinite-chain construction (Section 3.1). Increasing dd yields lighter-tailed training perturbations. At the same training budget, our d=4096d=4096 configuration achieves substantially lower FID than d=1d=1 on both Shoes→Bags and CIFAR-10 (Appendix H.2).

Using this extension, we add experiments on unpaired Shoes→Bags translation and unconditional CIFAR-10 generation at 32×3232\times32 resolution (Section 4.3 and Figure 4). C$^2$FM achieves FID 5.2 on Shoes→Bags and 8.0 on CIFAR-10. Under matched training budgets and evaluation protocols, the corresponding Flow Matching scores are 81.4 and 5.1. These results show that the same electrostatic formulation can produce realistic images for both data-to-data transport and noise-to-data generation.

The revision therefore directly addresses applicability to CIFAR-10 and adds a realistic unpaired translation task. Scaling to ImageNet and higher resolutions remains an open direction, which we acknowledge in Discussion and Limitations.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their constructive feedback. We have uploaded a revised manuscript with the changes highlighted in blue. The main changes are:

  • [tUdj,F91d,24NQ] We extend C$^2$FM from one auxiliary dimension to any d≥1d\ge 1. At the same training budget, the d=4096d=4096 configuration yields substantially lower FID than d=1d=1. See Section 3.1 and Appendix H.2 for details.

  • [tUdj,F91d,24NQ] We add 32×3232\times32 experiments on unpaired Shoes→Bags translation and unconditional CIFAR-10 generation. C$^2$FM (d=4096d=4096) performs well on both tasks using the same electrostatic formulation. See Section 4.3 and Figure 4 for details.

  • [F91d,24NQ] We compare learned C$^2$FM and EFM fields on CMNIST 2→32\to3 under matched training and evaluation protocols. As in the exact-field experiment (Figure 3), some inputs yield saturated noise with EFM and recognizable threes with C$^2$FM. See Appendix H.1 for details.

  • [24NQ] We measure full training iteration times for several values of CC. On Shoes→Bags at 32×3232\times32, our infinite-chain approximation adds 3.6% to iteration time relative to the matched C=0C=0 baseline. See Appendix H.3 for details.

  • [F91d] To make the paper easier to follow, we rewrite the descriptions of training and inference in the main text and integrate related work into the relevant sections.

Please consider the revised manuscript. We are happy to address any remaining points during the discussion phase.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.