acceptodds
Under review as a conference paper at ICLR 2027

Existence-Field Diffusion Model for Spatial Point Processes with Variable Cardinality

Abstract

We study generative modeling of spatial point processes (SPP), where both the number of points and their spatial configuration are governed by a joint distribution. While diffusion models have achieved strong performance in modeling complex distributions, extending them to variable-cardinality SPP remains challenging. Existing approaches either sample cardinality before generating locations conditionally, or introduce specialized discrete operations to change the number of points during generation. We propose the existence-field diffusion model (EFDM), which associates each potential point with a continuous variable representing its degree of presence. EFDM uses Gaussian diffusion to jointly model point locations and cardinality through a simple continuous representation. Existence probabilities can increase or decrease in both forward and reverse diffusion, without explicit point-addition or point-deletion steps. The same construction naturally accommodates categorical attributes, allowing locations, existence, and attributes to be modeled together. Experiments demonstrate improved molecular stability and validity over the compared baselines on the molecular dataset, alongside competitive performance on real-world spatial and synthetic datasets.

Primary area:
generative models
Keywords:
diffusion models, spatial point processes, set generative modeling, variable-cardinality sets, point processes, generative modeling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_iVCtOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes the existence-field diffusion model (EFDM), which associates each potential point with a continuous variable representing its degree of presence. EFDM uses Gaussian diffusion to jointly model point locations and cardinality through a simple continuous representation. Existence probabilities can increase or decrease in both forward and reverse diffusion, without explicit adding or deleting steps. The experiments demonstrate improved performance.

Critical strengths
  1. The proposed methodunifies coordinates, existence, and category attributes. By jointly diffusing existence variables and coordinates, and extending this to category attributes through shared existence probabilities, the model can simultaneously generate geometry, cardinality, and atomic types.
  2. The experiments are quite extensive, and the evaluation metrics are comprehensive.
Critical weaknesses
  1. EFDM applies Gaussian diffusion and MSE-based denoising to the existence logits, yet it ultimately derives discrete existence values ​​via a Bernoulli distribution or an argmax operation. The paper does not provide a variational lower bound (ELBO) or establish a rigorous connection to the diffusion of categorical existence variables. The design—weighting coordinate losses by σ(ui(t))\sigma(u_i(t)) and using standard MSE for the existence loss—appears heuristic. It remains to be clarified whether this continuous relaxation effectively optimizes the likelihood of a valid discrete set of points, and whether the final inferred distribution aligns with the training objective.
  2. Although EFDM allows for continuous variation in cardinality during the diffusion process, the actual modeling space is a bounded cardinality space defined by ∣x∣≤N|x| \le N. In experiments, NN is set to a value close to the maximum cardinality observed during training. This implies that the model cannot naturally extrapolate to cardinalities exceeding NN, and that both computational and memory requirements scale with NN.
  3. The proposed method is currently limited to spatial point processes. Whether it has the potential to be naturally extended to spatio temporal point processes, which is a more general and common scenario, remains to be discussed.
Questions for authors
  1. In Table.3, EFDM underperforms FDDM on the Count metric W1W_1, but performs better on JMMD2\text{JMMD}^2/ICMMD2\text{ICMMD}^2. How do the authors explain this trade-off? Does this suggest that while the model improves the distribution within the conditional space, it does so at the expense of the marginal count distribution?
  2. How should cardinality NN be chosen? If the actual cardinality during testing exceeds the maximum NN used during training, how will the model fail?
  3. Is there a theoretical connection between Gaussian diffusion in logit space and the final Bernoulli/argmax readout?
  4. Can the model be naturally extended to spatio temporal point processes?
  5. In a multi marked point process, is the assumption of a shared existence probability eie_i too restrictive? How can EFDM be extended to accommodate scenarios where different attributes exhibit distinct missingness patterns? Furthermore, how significant is the discrepancy between the deterministic argmax readout and probabilistic sampling-based readout when applied to the QM9 dataset?
3 comments
AuthorsOpenReview ↗

Response to Reviewer iVCt

Question 1. Is there a trade-off between marginal and conditional accuracy?

FDDM/IDM's lower Count W1W_1 is consistent with its explicit modeling of the cardinality marginal. Its dedicated count model gives it direct control over this marginal but does not by itself guarantee better conditional spatial fidelity. The differing metric rankings do not establish an inherent trade-off between marginal and conditional accuracy. Count W1W_1, ICMMD$^2$, and JMMD$^2$ can all improve simultaneously.

Question 2. How should NN be chosen, and what if a test configuration exceeds it?

With NN slots, EFDM cannot generate more than NN points. However, when test configurations contain far more points than any training configuration, every method must extrapolate beyond the observed cardinality range. No method can guarantee accurate generation in that regime from the training data alone. Increasing NN raises this limit but does not by itself resolve the extrapolation problem. For our experiments, it suffices to choose NN to cover the largest training configuration, with room for a modest margin when the computational budget permits.

Question 3. How does Gaussian diffusion relate to the final readout?

Yes. Both readouts follow from how we encode the training data for Gaussian diffusion. For binary presence, we encode each slot as a logit uiu_i and keep it with probability σ(ui)\sigma(u_i). Even with perfect denoising, this Bernoulli draw flips the clean label with probability δ\delta.

On QM9, argmax recovers the correct atom type or absence from every clean ±1\pm1 code (Appendix A.3). The same encoding also supports sampling: since type kk is encoded as +1+1 when present and −1-1 otherwise, its probability is qk=(u^k+1)/2q_k=(\hat u_k+1)/2 when u^k\hat u_k is the conditional mean of that code. The sum over atom types is the probability of keeping the slot.

Question 4. Can EFDM model spatiotemporal point processes?

Yes. In additional experiments on the Earthquakes dataset, we added time as an extra coordinate and trained EFDM with three random seeds, following the PSD setup. EFDM achieved lower mean SL and MMD than the published PSD results.

ModelSL ↓\downarrowMMD ↓\downarrow
EFDM (mean ±\pm SD)0.03947±0.002670.03947\pm0.002670.01580±0.000770.01580\pm0.00077
PSD (Table 7)0.042±0.0030.042\pm0.0030.023±0.0030.023\pm0.003

Question 5. Can attributes be missing individually, and how does sampling compare with argmax on QM9?

Mark-specific missingness is an interesting question that we have not tested. For conditional prediction, we could model attribute values given that the point exists and use a separate indicator for whether each attribute is observed. This would let us predict a missing attribute from the observed ones.

On the same 10,000 saved QM9 terminal states, sampling from the probabilities derived from the signed codes came within 0.37 percentage points of argmax on all three metrics.

ReadoutAtom stability (%)Molecule stability (%)Validity (%)
Argmax (reproduced)99.0491.1996.38
Sampling with decoded probabilities98.9990.8296.04

Weakness 1. What does Eq. (4) optimize, and how does it relate to the generated point set?

Eq. (4) trains EFDM to denoise coordinates and existence jointly. For a given noisy input and time, the coordinate weight is a positive constant, so the optimal noise prediction under weighted MSE is still the conditional mean of the noise. The final readout defines a distribution over finite point sets. At the clean training codes, Bernoulli sampling recovers each binary label with probability 1−δ1-\delta, and argmax recovers each QM9 category exactly. We do not claim that Eq. (4) is an ELBO for the decoded point sets.

Weakness 2. See Question 2 for how we choose NN. Larger slot budgets increase computation and memory use.

DatasetFDDM (s/epoch)EFDM (s/epoch)EFDM overhead
Earthquake2.19712.37017.9%
Trip0.38750.442914.3%

Weakness 3. See Question 4 for our spatiotemporal Earthquakes experiment and comparison with PSD.

Area_Chair_2gocOpenReview ↗

Please respond

Hi iVCt,

Thanks for your work so far on this paper. Please respond to the authors before the end of the discussion period, even if to acknowledge their response.

I appreciate the authors for the effort and clarification. After carefully reading the author's response and other reviewers' comments, I decide to keep the rating.

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

The paper proposes an existence-field diffusion model (EFDM) for generating spatial point patterns. The method uses a prespecified fixed number of potential points, each of which is associated with a spatial configuration and a continuous existence variable. Then the method applies Gaussian diffusion jointly to the coordinates and existence variables. The final point configuration is obtained by sampling the existence indicators. EFDM avoids explicit point insertion and deletion in the diffusion model. The paper further extends the framework to categorical point attributes and estimates them jointly with the spatial configurations. The empirical performance is competitive compared with several existing diffusion-based methods including FDDM and TDDM on several real datasets and synthetic spatial point patterns.

Critical strengths

The EFDM framework uses a joint modeling strategy for the spatial configuration and the existence variables, which avoids fixing cardinality before denoising or introducing trans-dimensional moves. The method is simple yet seems highly effective for the range of sample sizes under consideration.

The empirical performance of EFDM is demonstrated competitive, compared with the existing diffusion model methods, such as FDDM, TDDM, and PSD. Under most of the metrics, EFDM seems to perform similarly or slightly better than FDDM, and under some metrics, significantly better than TDDM, IDM, and PSD. The choices of evaluation metrics are comprehensive as they assess different aspects of spatial point patterns.

Critical weaknesses
  1. EFDM requires the specification of size NN, which should be larger than the observed cardinality nn, and the diffusion model is trained on NN potential points. This makes it unclear whether EFDM remains computationally scalable to massive spatial point patterns. For Transformer-based architectures with full attention, the computational cost of attention typically scales quadratically with the number of points. Therefore, if N≫nN\gg n, I wonder if the training cost of EFDM is significantly higher than that of FDDM.

  2. The reason for choosing the point-set representation and decoding strategies needs more explanation. For decoding, the paper uses the Bernoulli sampling distribution as in equation (5). For the point-set representation (on page 15), it is said that `` inactive coordinates are sampled from the observed points in the same set." Even though δ=0.003\delta=0.003 is a small probability for the inactive slots in the training representation, this cannot rule out the possibility that for a larger nn, some of the inactive slots could be retained when they coincide with some active slots. I expect that such scenario may possibly occur when N≫nN\gg n with a positive probability, and I wonder if this could lead to learning a slightly distorted spatial configuration in the diffusion models.

Questions for authors
  1. Can the authors provide more explanation on why the weights in equation (4) can be set the same as the existence probabilities? I believe this is crucial for training the diffusion model.

  2. I wonder if in the experiment on QM9, the authors have used training, validation, and testing sets to evaluate the performance with the percentages reported in Table 1, similar to those for the trip and earthquake datasets? If not, can the authors show similar empirical evidence for the QM9 data?

2 comments
AuthorsOpenReview ↗

Response to Reviewer rrpZ

Question 1. Why weight the coordinate loss by existence in Eq. (4)?

Only retained points contribute coordinates to the output set. We therefore weight each slot's coordinate loss by its current existence value wi=σ(ui(t))w_i=\sigma(u_i^{(t)}), giving more weight to the geometry of slots with larger current existence values. Every slot receives equal weight in the existence loss.

Given the denoiser input Z=(y(t),t)Z=(y^{(t)},t), wiw_i is a positive constant with respect to the prediction. Over unrestricted predictors, the weighted and unweighted MSE therefore have the same optimal denoising target at each input,

arg⁡min⁡a∈Rd E[wi∥ϵx,i−a∥2∣Z]=E[ϵx,i∣Z]\underset{a\in\mathbb{R}^d}{\arg\min}\,\mathbb{E}[w_i\lVert\epsilon_{x,i}-a\rVert^2\mid Z]=\mathbb{E}[\epsilon_{x,i}\mid Z].

This argument requires only a positive weight determined by the noisy input. It applies to our finite-δ\delta encoding without assuming that wiw_i is a calibrated posterior probability of clean presence.

Question 2. QM9 data splits and checkpoint selection

We follow the QM9 data split used in TDDM. The 130,831 molecules are divided into disjoint training (100,000, 76.43%), validation (17,748, 13.57%), and test (13,083, 10.00%) sets. We fit the model only on the training split. Atom stability, molecule stability, and validity measure the chemical quality of generated molecules directly, without a reference set.

We screened 41 EMA checkpoints from epochs 4,000–8,000 at 100-epoch intervals, using 128 generated molecules per checkpoint. We then evaluated three candidates from this screening and three additional late checkpoints, all within epochs 6,000–8,000, using 10,000 generated molecules each. Given their comparable performance, we report the final EMA checkpoint at epoch 8,000 in Table 1, with all three metrics computed from a single 10,000-molecule evaluation. The performance ranges across the six evaluated checkpoints are reported in Appendix A.2.

Weakness 1. Computational cost of the fixed slot budget

EFDM does more attention work than FDDM. However, when test configurations contain far more points than any training configuration, both EFDM and FDDM face a cardinality extrapolation problem. Increasing NN alone does not resolve this problem. For our experiments, it therefore suffices to choose NN to cover the largest training configuration, with room for a modest margin when the computational budget permits. The table reports the measured training time per epoch and the overhead relative to FDDM.

DatasetFDDM (s/epoch)EFDM (s/epoch)EFDM overhead
Earthquake2.19712.37017.9%
Trip0.38750.442914.3%

Weakness 2. Effect of Bernoulli decoding on generated point sets

We agree that Bernoulli decoding can bias the generated count with our choice of δ=0.003\delta=0.003. We therefore evaluated a deterministic final readout that retains slots with existence probability above 0.50.5 on Earthquake, Trip, and the synthetic dataset.

DatasetSeedsBernoulli (original paper): Count W1W_1 ↓Threshold (p>0.5p>0.5, new): Count W1W_1 ↓
Earthquake36.2±1.56.2\pm1.55.7±1.45.7\pm1.4
Trip313.81±7.3613.81\pm7.3613.52±7.2913.52\pm7.29
Synthetic12.0782.0782.1522.152

Earthquake and Trip values are means ±\pm sample standard deviations across seeds.

On QM9, we tested the reverse change, replacing the reported argmax readout with sampling from the probabilities derived from the ±1\pm1 atom-type codes. On the same 10,000 terminal states, all three metrics were within 0.37 percentage points of argmax.

ReadoutAtom stability (%)Molecule stability (%)Validity (%)
Argmax (reproduced)99.0491.1996.38
Sampling with decoded probabilities98.9990.8296.04

In the revised paper, we will specify the decoding rule used for each dataset and report how stochastic and deterministic decoding affect the results.

I would like to thank the authors for their response and for clarifying my concerns. I will maintain my current score.

Reviewer_EHKNOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes a diffusion model for variable cardinality data within an upper bound N. It does so by introducing an latent existence variable (coined existence field) which is modelled by the existence probability per node/element. They show experiments on QM9 and some SPP datasets.

Critical strengths

The general idea and motivation is simple and well established.

Critical weaknesses

Misrepresenting method as a SPP model The paper claims to model SPPs but in fact does not. A spatial point process assigns mass to configurations of unbounded cardinality, and the existence-field representation is defined only relative to a fixed slot budget N.

The papers idea of modelling a existence state is not novel. Published, for exactly this purpose, in: DrugFlow (virtual “no atom” node type, n_virt ~ U{0,...,N_max} added per training sample), FlowMol3 (fake atom type, count ~ U(0, 0.3N)), MiAD (mirage atoms as type 0, padded to N_m), UniMat (empty periodic-table cells), CrystalDiT (null atom for crystals under 20 atoms) and many more. Further, modelling the discrete state via its hyperparmeter has a long standing history: e.g. Variational Flow Matching for Graph Generation, Dirichlet Flow Matching with Applications to DNA Sequence Design and many more.

Missing Baselines. DrugFlow and FlowMol3 and MiAD, which are the same mechanism and the direct comparators; Branching Flows and Edit Flows and FlexMDM, the current variable-cardinality methods. There is no matched fixed-length oracle: same backbone, same training, given the true cardinality. Without it, EFDM’s QM9 numbers cannot be attributed to the existence field rather than to the EGNN+Transformer backbone and 8,000 epochs, especially since the baselines use different architectures entirely.

Weak evaluation Single seed comparison.

The method in the main text isn’t the method used for QM9. Main text (Eq. 5, Fig. 1): each slot gets an existence variable u_i, diffused, then ẑ_i ~ Bernoulli(σ(û_i)) decides which slots are real. QM9 (A.2, A.3): no u_i at all — just type logits with the inactive category pinned at U_i0 = 0 (Eq. 13), decoded by argmax (Eq. 17), trained on ±1 targets. A.2 says the δ = 0.003 existence encoding doesn’t apply to QM9. Eq. 18 admits argmax can disagree with the probabilistic readout. So the existence field + Bernoulli readout runs on earthquake/trip/synthetic, while QM9 uses a plain (K+1)-way categorical field in continuous logits. The headline result comes from the configuration with the distinctive machinery stripped out and the two are never compared on the same dataset.

Questions for authors

See weaknesses

2 comments
AuthorsOpenReview ↗

Response to Reviewer EHKN

1. Definition of a spatial point process

Section 3.1 follows the general definition in the Handbook of Spatial Statistics (van Lieshout, 2010, Def. 16.1, p. 265). The definition requires that for each bounded Borel set A⊆DA\subseteq D, “the number N(A)=NX(A)N(A)=N_{\mathcal{X}}(A) of points falling in AA is a (finite) random variable.” This condition only requires the point count in each bounded Borel set to be finite almost surely; it does not require the count distribution to have unbounded support.

2. Methodological contribution

The cited molecular and crystal models encode absence using special atom types or coordinates, whereas EFDM jointly diffuses a separate continuous existence variable and the coordinates of each potential point. We use the same formulation to model earthquake locations in Japan and e-scooter pickups in Louisville. We will cite these methods.

3. Baseline comparisons

We chose FDDM, TDDM, and PSD as baselines because our focus is on models for variable-cardinality spatial point processes rather than on all methods evaluated on QM9. Branching Flows models spatial point sets of variable cardinality, but its published QM9 results omit both atom stability and molecule stability, making a direct comparison on these metrics difficult. Edit Flows and FlexMDM generate variable-length discrete sequences and therefore do not directly apply to our SPP setting.

On QM9, EFDM and TDDM use EGNN backbones of the same depth and width, providing a fair comparison in this respect.

4. Number of training seeds

On Earthquake and Trip, we trained and evaluated every method with three seeds per dataset. Each seed varied both the data split and the training randomness. Tables 2 and 3 report the mean and sample standard deviation across the three runs.

5. Existence modeling and decoding on QM9

We disagree with the conclusion that QM9 removes existence modeling. In the QM9 representation described in Appendix A.3, the dummy category represents absence. The same checkpoint can also be evaluated with an alternative stochastic readout. For an ideal MSE noise predictor, the clean-code estimate at a noisy input estimates the conditional mean of 21{C=k}−12\mathbf{1}\{C=k\}-1, motivating the transformation qk=(u^k+1)/2q_k=(\hat u_k+1)/2. For generated terminal codes, we project these values onto qk≥0, ∑kqk≤1q_k\geq 0,\ \sum_k q_k\leq 1 and use them as categorical sampling probabilities. This is algebraically equivalent to sampling presence with probability p=∑kqkp=\sum_k q_k and, if present, sampling type kk with probability qk/pq_k/p.

ReadoutAtom stability (%)Molecule stability (%)Validity (%)
Argmax (reproduced)99.0491.1996.38
Sampling with decoded probabilities98.9990.8296.04

On the same 10,000 saved QM9 terminal states, stochastic decoding came within 0.37 percentage points of argmax on all three metrics without retraining or temperature tuning. Compared with Bernoulli sampling, retaining points with final existence probability above 0.50.5 reduced count W1W_1 on Earthquake and Trip and increased it on the synthetic dataset.

DatasetSeedsBernoulli (original paper): Count W1W_1 ↓Threshold (p>0.5p>0.5, new): Count W1W_1 ↓
Earthquake36.2±1.56.2\pm1.55.7±1.45.7\pm1.4
Trip313.81±7.3613.81\pm7.3613.52±7.2913.52\pm7.29
Synthetic12.0782.0782.1522.152

Earthquake and Trip values are means ± sample standard deviations across seeds. We will include the derivation and readout comparisons in the revision.

Definition of a spatial point process

We agree that a bounded-cardinality process is a SPP but not the reverse, i.e., that a bounded-cardinality model can capture abitrary Point Processes.

  • Model class vs. stated target. Just have a look at your own definition: Sec. 4.1 states the goal as approximating p∗p^{\ast} on X={x⊂Rd:∣x∣<∞}\mathcal{X} = \lbrace x \subset \mathbb{R}^d : |x| < \infty \rbrace, but EFDM assigns zero mass to every configuration with ∣x∣>N|x| > N (Eq. 2). Standard SPPs with unbounded count support (Poisson, Cox, cluster processes) are therefore excluded by construction. This is a support restriction. A.2 sets NN to the maximum training cardinality (for QM9, exactly N=29N = 29). Any held-out configuration above that receives probability exactly zero by design, rather than merely being hard to learn.
  • NN is not a free hyperparameter. Every sample is processed with all NN slots, so attention always costs O(N2)O(N^2), no matter how many points the set actually contains. A set with n≪Nn \ll N points pays the same cost as the largest set in the data. The overhead reported in the rebuttal (7.9% / 14.3% over FDDM) is measured with NN close to the training maximum, which is the cheapest possible setting. Choosing NN larger to cover larger configurations, as the paper suggests is possible, raises this cost quadratically for every sample. In practice this ties NN to the training maximum, which brings back the support restriction above.

Methodological contribution

  • The paper itself contradicts the stated distinction. The rebuttal says the cited methods encode absence via special atom types, whereas EFDM diffuses a separate existence variable. However, on QM9, EFDM uses a (K+1)(K+1)-way categorical with the inactive logit fixed at Ui0=0U_{i0} = 0 (Eq. 13, A.3); the rebuttal itself calls it "the dummy category". This is the same mechanism as the virtual / fake / mirage atom types in DrugFlow, FlowMol3 and MiAD.
  • The two representations are equivalent by the paper's own derivation. Eqs. 14–16 establish a bijection between the shared-existence and dummy-category parameterizations. In the binary case (K=1K = 1, v1=0v_1 = 0), Eq. 14 gives U1=uU_1 = u: the binary existence field is exactly a two-category logit with the null category fixed at zero.
  • What remains is applying the established virtual-node / padding approach to non-molecular SPP data. That can be a contribution, but the paper should then be positioned as such, with these methods discussed as prior work and included as baselines. Citing them alone does not resolve the novelty concern.
  • Not addressed: the second part of our comment, on prior work modelling discrete states through continuous relaxations (e.g., Variational Flow Matching for graphs, Dirichlet Flow Matching, and EDM's own continuous one-hot atom types). Please position EFDM relative to these.

Baseline comparisons

  • The matched fixed-length oracle is not addressed.
  • DrugFlow, FlowMol3 and MiAD, the direct comparators using the same mechanism, are not discussed in the response.
  • Branching Flows is a variable-cardinality point-set model. Even if its published QM9 results omit stability, it can be evaluated on Earthquake, Trip and the synthetic benchmark, or on QM9 using its public code. We accept that Edit Flows and FlexMDM are more complicated empirical comparisons.

Existence modeling and decoding on QM9

  • The readout experiment answers a different question. Our concern is that the method presented in Sec. 4 (separate uiu_i, the existence-weighted loss in Eq. 4, and the Bernoulli readout in Eq. 5) is not the method that produces Table 1. Argmax and stochastic decoding giving near-identical results shows that the readout does not drive QM9 performance. I.e. it is precisely the dummy-category categorical, which brings us back to our point 2.
  • The Bernoulli readout does not seem to be the right one. On Earthquake and Trip, the new deterministic threshold readout improves Count W1W_1 over the Bernoulli readout of Eq. 5, which further moves the method towards a plain padding model.

Minor

  • The "reproduced" argmax molecule stability (91.19%) differs from Table 1 (91.18%).
  • Since inactive coordinates are copied from real points, please report the fraction of generated points with a near-duplicate neighbour.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.