acceptodds
Under review as a conference paper at ICLR 2027

Multi-Parameter Topological VAEs for Conditioning Image Generation

Abstract

Topological Data Analysis (TDA) is the field of data science whose aim is to detect and encode topological patterns in data sets. As such, it has become common in machine learning to complement traditional models with TDA-based features and losses to enforce geometrical constraints into the models behaviour. However, while several methods can ensure that topology can be preserved at a global level (eg, total number of components or cycles), constraining topology locally (eg, imposing the presence of topological features at specific areas in space) has been relatively unexplored, despite its usefulness in several application domains. In this article, we propose to fill this gap with multi-parameter persistent homology. More precisely, we leverage a recently proposed descriptor, called the Multi-parameter Module Approximation (MMA), to design new losses for generative models. Indeed, MMA descriptors are easily interpretable and are thus more amenable to the construction of losses than other standard descriptors. On the theoretical side, we prove that MMA descriptors are locally Lipschitz and definable (under suitable assumptions), so that stochastic subgradient descent converges, which allows us to use them for steering neural net behaviours at training time. On the practical side, we showcase the efficiency of our localized topological constraints for mask and image generation with variational autoencoders, by providing superior results over both single-parameter and non-topological baselines.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
multiparameter persistent homology, topological data analysis, topological loss functions, variational autoencoders, conditional image generation

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_1ghdOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a way to use MMA as a loss function and apply it for conditioning VAEs with topological features at spatially predefined positions. Its core contribution lies in making MMA for multi-dimensional persistence work for gradient-based optimization and proposing a second filtration parameter encoding proximity to the prescribed location.

Critical strengths
  • Making MMAs fit for gradient-based training and using them as loss functions is novel and an interesting research direction
  • Using 2-dimensional filtrations for encoding intensity and proximity of individual features to a target for solving the spatial awareness problem when no ground truth is available is a novel and suitable approach
  • The math seems to be sound, although my expertise in multi-parameter persistence is limited
Critical weaknesses

The critical weaknesses are threefold.

First, the motivation for spatially-aware and topologically correct mask generation is unclear to me. The presented use case seems constructed, and the used dataset is synthetic. The binary mask generation is already so restricted (by feature type and location) that a VAE might not be necessary anymore, and the mask can simply be generated deterministically, e.g., as in how the synthetic dataset itself was created. Furthermore, the paper uses artificially small networks to demonstrate the effectiveness of its solution, since larger networks can solve the task inherently (see line 365). I would like to see use cases where the paper's approach is actually the best solution.

Second, the experimental evidence is weak. One aspect of this weakness relates to the first point: only a single synthetic dataset is used. Furthermore, no significance test is reported, and standard deviations are heavily overlapping. The most important baseline, Scoccola et al., is mentioned but neither experimentally nor theoretically compared to the presented method. Other important baselines could include Oner et al. or Stucki et al.'s Betti Matching.

Third, the presentation is poor and the paper is hard to follow. Especially the intro lacks a clear storyline and jumps between seemingly unrelated concepts. Some statements are vague or unsupported. For example, line 73 states that some related work is "preliminary" without explaining what that means or why it is preliminary. Line 56 mentions that some theoretical aspect is "more involved". Line 68 mentions some challenges (that the paper potentially solves), but it does not explicitly mention what they are. In general, I believe the paper should be rewritten with a strong focus on clarity and cohesion. I can recommend following the principles from "Style: Lessons in Clarity and Grace" by Joseph Williams.

Questions for authors
  • Please provide statistical significance tests for your results
  • Please include a runtime analysis of your approach that includes the number of lines and the patch size and comparisons to other methods
  • Please compare your approach theoretically and experimentally to the approach in Scoccola et al.
  • How exactly is the 1PH baseline defined?
  • Please justify the relevance of your approach with realistic use-cases
  • In the abstract, intro, and related work, the spatial awareness problem of TDA is described generally, although non-generative tasks can solve this problem principled and efficiently via induced matchings. Please narrow down your framing throughout the paper and mention this line of work.
  • Your citations seem incorrect; potentially, you used citet instead of citep?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 1ghd

Use cases and motivation. We thank the reviewer for the detailed review. The use cases we envision are those of Gupta et al. (2025), such as an image with an exact number of animals or an aerial view of a road network with an exact number of regions, where our loss additionally prescribes where each object or region lies. The reviewer is right that our experiments use a single synthetic dataset. The method applies unchanged to other shapes and datasets: only the masks and the spatial profiles change. We are generating images from the masks of our models with ControlNet, as in Fig. 1, and we will report them in this thread.

The sentence at the opening of Sec. 4 on the U-Net diffusion model of Gupta et al. (2025) is too general, and we will rewrite it. We designed the conditioning of this paper so that the topological losses and the metrics that evaluate them are easy to define and interpret, for MMA and single-parameter PH alike: it prescribes the position, the type and the number of the objects through one spatial profile per object. This conditioning is so informative that the U-Net solves the task without topological supervision. With a less informative conditioning, such as the number of objects of each type and their coarse position in the image (e.g., top right), the U-Net produces masks whose topology does not match the prescribed one, and we can provide these results on request. The CVAE fails already under the informative conditioning, so this setting measures the contribution of the topological loss where the loss and the metric are easiest to define and interpret, which in our view makes it the right setting for a first application of MMA-based losses.

Significance. We have implemented a paired Wilcoxon signed-rank test between MMA and 1PH on the same 100 test images for each object count NN. On the submitted runs, the difference is significant for N=2,…,8N = 2, \dots, 8 (p≤0.012p \leq 0.012), and not significant for N=1N = 1 (p=0.06p = 0.06), N=9N = 9 (p=0.12p = 0.12) and N=10N = 10 (p=0.82p = 0.82). The standard deviations of Table 1 measure how much the score varies from one test image to another, and all models are evaluated on the same images, so the paired test is the appropriate comparison. We will add these tests to the revised paper.

Scoccola et al. We are training the CVAE with a loss on the Hilbert decomposition signed measure of Scoccola et al. in place of MMA, on the same bifiltrations, and we will report the results in this thread. The signed measure and MMA are different descriptors, and MMA is the more global of the two. With MMA every term of our loss has a direct interpretation: each summand is one topological feature, so the loss rewards the largest summand in each degree and penalizes the others. On the signed measure, the same loss raises three problems.

  1. The positive and negative points of the signed measure do not necessarily correspond to single topological features, so a loss of the form "one component, and one hole for hollow objects, at each prescribed location" has no direct counterpart.
  2. A loss that reads individual points of the signed measure, for instance one that rewards the nn points of largest norm, loses regularity exactly where features are born or die. A positive and a negative point that fall in the same cell of the grid sum up, and if their net weight is zero the point is discarded, so the rewarded set of points changes abruptly and the loss jumps.
  3. The losses covered by the differentiability theory of Scoccola et al. read the measure as a whole, such as optimal transport to a target measure (their Example 3.22). Our baseline therefore uses optimal transport to a fixed target measure for each object and degree.

Oner et al. and Betti Matching. Both are segmentation losses that compare the prediction with a ground-truth mask, while our loss also supervises samples from the prior, which have none. We will discuss both in the related work.

Runtime. We are implementing a runtime analysis of the topological losses on our cluster, and we will report it in this thread. Preliminary results confirm that training with MMA takes longer than with 1PH, which in turn takes longer than training without topological term.

The 1PH baseline. It uses the same CVAE, losses and coefficients as MMA (Table 2), with a one-parameter persistence loss in the style of Gupta et al. (2025) whose target is the pair of Betti numbers of the ground-truth mask. We will describe it in Appendix E.

Framing and presentation. We will restrict the framing to generative models, cite the work on induced matchings (Bauer and Lesnick, 2015; Stucki et al., 2023), rewrite the introduction around a single line of argument, make lines 56, 68 and 73 precise, and correct the citation commands.

General: I thank the authors for their response, which addresses some of my concerns. I appreciate the work's theoretical aspect, and I can imagine useful applications. However, I'm not convinced that the paper's setting is such an application. My main hesitation in increasing my score is the method's (and the paper's entire setting's) usefulness. My main two questions are: (1) Why do I need a model that can create masks with a predefined topology and location of topological features? Can SOTA image generation models still not solve this task? Maybe you could include tests on a real downstream task. (2) Under the assumption that I need the output of such models, would it not be much easier to use alternatives (existing methods that can be applied in this setting, see below; or a deterministic algorithm, such as the one used for creating the synthetic dataset)? The authors would need to (empirically) demonstrate the benefit of their method over alternative approaches in real-world settings and/or in a downstream task. Also, please upload the updated manuscript. Other points below.

Scoccola et al.: Having this as a baseline is a necessary improvement, but not yet sufficient to provide solid empirical evidence of the approach's usefulness. However, please also provide a detailed description of the specific design, as this will probably determine if it is a strong or weak baseline. Specifically, please specify exactly how the target measure is defined, where the masses are placed, and with what signs. Also report the tuning framework for the baseline.

Oner et al. and Betti Matching: It is true that both are segmentation losses, but they can be applied almost directly to your setting, and both aim to solve the same problem of spatial correspondence among topological features. You train in a reconstruction setting with ground-truth masks, to which both methods apply.

Runtime: Please report it here and put it in the updated manuscript. Also compare to new baselines. Include the number of lines and the patch sizes.

AuthorsOpenReview ↗

Updated results and answers to the follow-up

Revised paper. We have uploaded the revised paper. Table 1 now compares MMA with 1PH, windowed 1PH, Betti matching and the ablation, on reruns in which MMA and the loss are computed on the perturbed bifiltration. Paired two-sided Wilcoxon signed-rank tests show that MMA improves on 1PH and windowed 1PH for N=2,…,9N = 2, \dots, 9 (p≤3.2⋅10−3p \leq 3.2 \cdot 10^{-3}), and on Betti matching and the ablation at every NN (Appendix F).

Usefulness of the setting. For geometric masks such as those of our benchmark, the reviewer is right: a deterministic algorithm can produce them. For more complex images, such as the animals of Gupta et al. (2025), no simple deterministic algorithm exists, to the best of our knowledge. This is the use case we target, and our loss applies to it unchanged, exactly as it applies to the masks generated here.

Scoccola et al. We built the baseline with the loss of Scoccola et al., the optimal-transport distance to a fixed target measure (their Example 3.22), on the Hilbert decomposition signed measure of the same bifiltrations as MMA. For each object, the target is a unit positive mass at the lower corner of the box in degree 0. In degree 1, the target of a hollow object is the signed measure of the box, with positive masses at the lower and upper corners and negative masses at the other two, and the target of a filled object is the zero measure. The outside term has a unit mass at the upper corner. We reused the coefficients and the schedule of MMA, without a separate tuning. The mean score of this baseline lies between 0.96 and 1.00 at every object count. This is the difficulty raised in our first response: the points of the signed measure do not correspond to single topological features, so designing a loss that prescribes one feature at each location is not straightforward, and we did not find a satisfactory one in the time of the rebuttal.

Oner et al. and Betti matching. Following the reviewer's suggestion, Betti matching is now a baseline in Table 1. It needs the ground-truth mask, so it acts on the posterior term only, and its scores (0.45 to 0.69) lie well above those of MMA, mainly through spurious features and features of the wrong degree. We did not run Oner et al.

Runtime. Appendix E reports the time per optimization step on one H100 GPU with 32 CPU cores, batch size 16 and 64×6464 \times 64 images: 13.8 ms without topological loss, 61.6 ms with 1PH and 3.64 s with MMA. The MMA loss uses ∣L∣=25|L| = 25 lines on the full 64×6464 \times 64 grid, with no patch: it computes one bifiltration per object plus one for the outside term, on 4 of the 16 images of each minibatch, so its cost grows with the number of objects. 1PH runs on all 16 images at resolution 48×4848 \times 48.

Reviewer_WPbuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks how to prescribe both the topology (filled vs. hollow) and the location of objects in generated binary masks. On the theory side it studies the Multi-parameter Module Approximation (MMA) descriptor of Loiseaux et al. (2025): it partitions the space of filtrations into finitely many semiaffine "blocks" on which the number of summands and corners is constant (Props. 2–5), proves that MMA is locally 1-Lipschitz on the generic set Ω\Omega of filtrations without ties (Thm. 10) and along segments that cross block boundaries one adjacent transposition at a time (Thm. 11), and deduces definability and local Lipschitz continuity of MMA-based losses (Thm. 12, Cor. 13), which places them in the scope of stochastic subgradient convergence results. On the practical side it builds, for every prescribed object jj, a bifiltration (1−x^, 1−gj)(1-\hat x,\,1-g_j) from the mask and a spatial profile, and a loss (Eqs. 5–7) that rewards one dominant degree-0 summand (plus one degree-1 summand for hollow objects) near each location and penalizes everything else. The loss supervises posterior reconstructions and prior samples of a SPADE-conditioned CVAE (Eq. 8). On synthetic 64×64 masks with 1–10 objects, the MMA loss gives the lowest score on a custom benchmark for 1–8 objects; single-parameter PH (1PH) is slightly better at 9–10 objects, and both are far better than no topological loss (Table 1).

Critical strengths
  • A meaningful gap is addressed with a well-designed loss. Topological losses used so far for generation control how many components or holes a sample has, not where they are. The per-object bifiltration with a spatial second parameter (Sec. 4.1, Eqs. 5–7) encodes "one component / one hole at this location" without prescribing the object's boundary, and the outside term handles spurious components with one extra evaluation per image. Fig. 1 shows the mechanism concretely: swapping two objects empties the degree-1 descriptor.

  • The differentiability analysis is structural rather than hand-waved. Expressing MMA as a piecewise-affine map into a corner-configuration space on each block (Props. 3–5), and proving a Lipschitz estimate on Ω\Omega and across single-transposition boundaries (Thms. 10–11, Prop. 30), is more than prior multi-parameter work offered for an interpretable descriptor. Together with the Table 1 improvements at N=1,…,8N=1,\dots,8 and the anonymous code/data release, this makes the approach credible in principle.

Critical weaknesses
  • The evaluation does not isolate the claimed benefit, and the scope is narrower than the title. (a) The only quantitative result is a composite score (Eq. 19) whose quality-deficit term vjv_j is itself computed from MMA on the generated mask (Eqs. 20–21); a model trained to maximize MMA summand weight is thus scored by a metric that rewards MMA summand weight. (b) No decomposition is reported — rate of correct local topology, missing/spurious counts, wrong-degree counts, displacement — so it is unclear which behaviour improves. (c) The 1PH baseline is never specified: which filtration(s), per object or global, whether it receives any spatial information, which coefficients and posterior/prior schedule. Since 1PH already closes most of the gap (ablation 0.41–0.62 → 1PH 0.09–0.12 → MMA 0.04–0.11) and MMA's advantage reverses at N=9,10N=9,10, the claim that the multi-parameter structure, rather than localized supervision per se, is responsible is not established. (d) The title, abstract and Fig. 1 promise image generation (via ControlNet), but all experiments are on synthetic masks; no image is generated or evaluated. The authors also state (Sec. 4) that they deliberately use a generator that underperforms, so whether the loss helps a competent generator is open.

  • Gaussian (Sec. 4.1, Fig. 1) vs. Lorentzian (E.4, Eq. 18) for gjg_j; please make the text consistent and state which one was used.

  • Sec. 4.1: "These labels are the respective of the two conditioning channels of the model" is garbled.

  • Abstract: "the models behaviour" → "the model's behaviour"; "(eg, …)" → "(e.g., …)". Sec. 4.2: "MMA-based models achieves" → "achieve".

  • The training loss uses ∣L∣=25|L|=25 lines while the metric uses the multipers default (557 lines); worth stating that this is intentional.

Questions for authors
  • Theory (clarification; decisive for Weakness 1). Which regularity statement covers the structural ties introduced by the coordinatewise-max extension? If a symbolic or numerical perturbation is used in the implementation, what is it, and why do the loss and its gradients inherit the Lipschitz bound of Thm. 11?

  • Baseline (clarification; decisive for Weakness 2). Please specify the 1PH objective exactly (filtration, spatial treatment, reward/penalty coefficients, posterior/prior terms, schedule). Does MMA retain its advantage against a 1PH loss evaluated within the same per-object windows, i.e. a localized single-parameter baseline?

  • Evaluation (request; affects Weakness 2). Can Table 1 be decomposed into the terms of Eq. 19 (unmatched, spurious, wrong degree, displacement, quality deficit), and can the score be reported with vjv_j removed or replaced by a non-MMA quality measure? Are there any image-generation results, or should the title and abstract be scoped to masks? Does the gain survive with a stronger generator (e.g., the U-Net diffusion model of Gupta et al., 2025)?

2 comments
AuthorsOpenReview ↗

Response to Reviewer WPbu

Assumption 6 and the cubical bifiltration. We thank the reviewer for this question, and the reviewer is right. Before submission, we planned to perturb the bifiltration so that no two simplices take the same value on a line of LL, but an unresolved bug prevented this, so the submitted code admits ties, including the face/coface ties of the coordinatewise-max extension. We have since fixed the bug: each simplex σ\sigma receives a fixed offset cσ=10−6r(σ)/Nc_\sigma = 10^{-6} r(\sigma)/N on both coordinates, where r(σ)r(\sigma) is its rank in an order of the NN simplices that lists faces before cofaces. The offsets are pairwise distinct and increase from faces to cofaces, so the perturbed map is a bifiltration without ties. MMA and the loss are computed on it, so Theorems 10-12 apply directly to the optimized loss. We are rerunning the experiments while monitoring that no ties occur. Preliminary results show an almost identical loss, and we will post the updated results when the runs are complete.

Evaluation metric. We agree that the aggregate score alone does not show which behaviour improves. The first table decomposes the score of Table 1 into the terms of Eq. 19, averaged over all 1000 test images. The second reports the score without vjv_j for each object count NN. MMA keeps a clear advantage over 1PH on the method-independent terms, and without vjv_j it has the lowest mean score at every NN, N=9,10N = 9, 10 included. We will add the full decomposition for each NN to Appendix F, updated after the reruns.

ModelUnmatchedSpuriousWrong degreeDisplacementvjv_jScore (Table 1)
MMA0.0000.0060.0090.0350.0090.059
1PH0.0030.0330.0110.0410.0110.100
Ablation0.0020.3060.1860.0250.0030.522
Score without vjv_jN=1N=2N=3N=4N=5N=6N=7N=8N=9N=10
MMA0.0380.0340.0360.0350.0450.0390.0490.0630.0770.088
1PH0.0800.1010.1100.0750.1070.0890.0800.0750.0790.092
Ablation0.5250.4560.4030.4950.4480.5270.5520.5700.5890.621

The 1PH baseline. The 1PH baseline uses the same CVAE, reconstruction and KL terms, coefficients and schedule as the MMA model (Table 2), and replaces LtopoL_{topo} with a one-parameter persistence loss in the style of Gupta et al. (2025). The filtration is given by the sublevel sets of 1−x^1 - \widehat{x} at resolution 48×4848 \times 48. The target is the pair of Betti numbers (k0,k1)(k_0, k_1) of the ground-truth mask, with no information on the location of the objects. In each degree dd, the loss keeps the kdk_d most persistent bars and pushes every other bar to the diagonal, with weight α=5\alpha = 5. The only spatial information is a pixel term that penalizes foreground where the normalized mixture gmixg_{mix} is low. We will add this description to Appendix E. We are also running the localized 1PH baseline the reviewer suggests, with this loss computed in a window around each prescribed object, and we will report its results in this thread.

Image generation and stronger generator. We are generating images from the masks of our models with ControlNet, as in Fig. 1, and we will report them in this thread and in the revised paper.

The sentence at the opening of Sec. 4 on the U-Net diffusion model of Gupta et al. (2025) is too general, and we will rewrite it. We designed the conditioning of this paper so that the topological losses and the metrics that evaluate them are easy to define and interpret, for MMA and single-parameter PH alike: it prescribes the position, the type and the number of the objects through one spatial profile per object. This conditioning is so informative that the U-Net solves the task without topological supervision. With a less informative conditioning, such as the number of objects of each type and their coarse position in the image (e.g., top right), the U-Net produces masks whose topology does not match the prescribed one, and we can provide these results on request. The CVAE fails already under the informative conditioning, so this setting measures the contribution of the topological loss where the loss and the metric are easiest to define and interpret, which in our view makes it the right setting for a first application of MMA-based losses.

Minor points. We thank the reviewer for the careful reading. The loss uses the Lorentzian profiles gjg_j of Eq. 18 and the map cc uses Gaussian bumps (App. E.1), and we will align Sec. 4.1 and Fig. 1. We will fix the garbled sentence in Sec. 4.1 and the typos. Using ∣L∣=25|L| = 25 lines in the loss and 557 in the metric is intentional, and we will state it.

Ethics flag. The paper uses only synthetic binary masks generated by the authors. We would be grateful if the reviewer could indicate the concern behind the flag.

AuthorsOpenReview ↗

Updated results

Updated results. The reruns with the perturbed bifiltration described in our first response are complete, and we have uploaded the revised paper. In all reported MMA runs we compute MMA and the loss on the perturbed bifiltration, and we monitored the bifiltrations during training: no ties occurred. Table 1 of the revised paper (first table below) compares MMA with four baselines: 1PH, windowed 1PH, Betti matching and the ablation. MMA has the lowest mean score at every object count NN, N=9,10N = 9, 10 included. Paired two-sided Wilcoxon signed-rank tests on the same test images show that MMA improves on 1PH and windowed 1PH for N=2,…,9N = 2, \dots, 9 (p≤3.2⋅10−3p \leq 3.2 \cdot 10^{-3}); at N=1N = 1 and N=10N = 10 the differences are not significant.

ModelN=1N=2N=3N=4N=5N=6N=7N=8N=9N=10
MMA0.06 ± 0.080.05 ± 0.060.05 ± 0.080.06 ± 0.070.05 ± 0.060.05 ± 0.050.06 ± 0.050.07 ± 0.070.07 ± 0.050.09 ± 0.08
1PH0.09 ± 0.200.11 ± 0.200.12 ± 0.170.09 ± 0.100.12 ± 0.150.10 ± 0.100.09 ± 0.080.09 ± 0.070.09 ± 0.070.10 ± 0.08
Windowed 1PH0.11 ± 0.300.09 ± 0.140.12 ± 0.190.10 ± 0.120.10 ± 0.100.11 ± 0.120.08 ± 0.060.08 ± 0.080.09 ± 0.070.10 ± 0.07
Betti matching0.45 ± 0.740.55 ± 0.550.49 ± 0.440.65 ± 0.410.62 ± 0.360.66 ± 0.320.69 ± 0.360.64 ± 0.300.66 ± 0.270.56 ± 0.28
Ablation0.53 ± 0.860.46 ± 0.510.41 ± 0.460.50 ± 0.390.45 ± 0.290.53 ± 0.320.56 ± 0.340.57 ± 0.270.59 ± 0.290.62 ± 0.29

Evaluation metric. The second table, reported in Appendix F, decomposes the score into its terms, averaged over the 1,000 test images, and the third gives the score without vjv_j for each NN. They replace the tables of our first response. MMA keeps a clear advantage over the two 1PH baselines on the method-independent terms, mainly through spurious features, and without vjv_j it has the lowest mean score at every NN.

ModelUnmatchedSpuriousWrong degreeDisplacementvjv_jScore without vjv_jScore
MMA0.0000.0060.0120.0350.0080.0530.061
1PH0.0030.0330.0110.0410.0110.0890.100
Windowed 1PH0.0030.0300.0140.0420.0090.0890.098
Betti matching0.0090.3410.2250.0200.0030.5950.598
Ablation0.0020.3060.1860.0250.0030.5190.522
Score without vjv_jN=1N=2N=3N=4N=5N=6N=7N=8N=9N=10
MMA0.0520.0430.0500.0530.0480.0460.0520.0560.0550.078
1PH0.0800.1010.1100.0750.1070.0890.0800.0750.0790.092
Windowed 1PH0.1000.0800.1120.0860.0910.0980.0750.0760.0840.086
Betti matching0.4460.5410.4820.6450.6190.6610.6920.6410.6620.560
Ablation0.5250.4560.4030.4950.4480.5270.5520.5700.5890.621

Localized 1PH baseline. Windowed 1PH computes the 1PH loss in a square window of half-side round(Rj)+5\mathrm{round}(R_j) + 5 around each target jj, with the Betti numbers of target jj as target, and an outside term sets every window to background and computes the same loss with a blank target. Its scores are close to those of 1PH, and MMA improves on it at the same object counts. Appendix E now describes both 1PH baselines.

Image generation. Figure 3 of the revised paper shows images generated by ControlNet from the masks of MMA, 1PH and the ablation, for two scenes: crater lakes and mountains, and donuts and krapfen.

Other points. We rewrote the opening of Section 4 on the U-Net diffusion model, aligned Section 4.1 and Figure 1 on gjg_j (Lorentzian profiles in the loss, Gaussian bumps in the conditioning map), fixed the garbled sentence and the typos, and stated in Appendix F that the metric uses its lines independently of the ∣L∣=25|L| = 25 lines of the training loss.

Reviewer_JhN8OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
1
Read review
Paper summary

This paper develops Multi-parameter Module Approximation (MMA) as a differentiable tool for topological learning. Theoretically, it establishes local regularity properties of MMA under genericity assumptions and derives conditions for MMA-based losses to be locally Lipschitz and definable. Practically, it introduces an MMA-based loss for conditional VAE mask generation that associates prescribed topology with prescribed spatial locations, and evaluates it against single-parameter persistent homology (1PH) and a non-topological baseline.

Critical strengths
  1. Nontrivial theoretical analysis of MMA.

The block-based representation and local regularity analysis provide a meaningful theoretical contribution toward using multiparameter persistent-homology descriptors in gradient-based learning.

  1. A natural use of multiparameter persistence for localized topology.

Using a second filtration to encode where a topological feature should occur addresses a genuine limitation of global topological losses and leads to an intuitive learning objective.

Critical weaknesses
  1. It is unclear whether Assumption 6, and consequently the main regularity theory, applies to the cubical filtrations used in the experiments.

Assumption 6 requires

hl(σ1)≠hl(σ2)h^l(\sigma_1) \neq h^l(\sigma_2)

for every sampled line ll and every pair of distinct simplices σ1,σ2\sigma_1,\sigma_2. The paper motivates this as a mild genericity assumption because the exceptional set has Lebesgue measure zero in the ambient space of arbitrary filtrations.

However, the experimental filtration is highly structured rather than arbitrary. Appendix E.4 defines the bifiltration on pixels and extends it to edges and squares by the coordinatewise maximum over their vertices. This construction can create exact face/coface ties on sets of positive measure.

For example, consider an edge e=[v1,v2]e=[v_1,v_2] with

f(v1)=(a1,a2),f(v2)=(b1,b2).f(v_1)=(a_1,a_2), \qquad f(v_2)=(b_1,b_2).

Whenever

a1>b1,a2>b2,a_1>b_1, \qquad a_2>b_2,

the max-extension gives f(e)=(max⁡{a1,b1},max⁡{a2,b2})=f(v1)f(e)=(\max\{a_1,b_1\},\max\{a_2,b_2\})=f(v_1). Consequently,

fl(e)=fl(v1)f^l(e)=f^l(v_1)

for every diagonal line ll, directly violating Assumption 6.

Importantly, this is not an accidental equality requiring a measure-zero event. The inequalities above define an open set and can occur with positive probability. Thus, although nongeneric filtrations may form a measure-zero subset of the ambient filtration space, the structured family induced by the cubical max-extension can map positive-measure sets into this exceptional set.

This distinction matters because the paper later states that the theoretical conclusions apply to the experimental loss at parameters for which all constituent bifiltrations belong to Ω\Omega, but it does not establish that this condition is generic, or even sufficiently common, for the actual cubical construction. If structural face/coface ties are present, the current argument does not appear to justify applying Theorems 10--12 to the practical loss.

Since connecting MMA regularity to gradient-based learning is a central theoretical contribution of the paper, I consider this issue consequential to my assessment.

  1. The principal empirical comparison uses an evaluation metric that is partially aligned with the proposed MMA representation.

The main benchmark score combines method-independent quantities such as missing/spurious objects, incorrect homological degree, and spatial displacement with an MMA-derived "quality deficit" vjv_j. The latter is computed by applying MMA again to the generated output and examining the dominant MMA summand.

Since the proposed model is itself explicitly trained using an MMA-based loss, whereas the main competing method uses single-parameter persistent homology, the evaluation is not fully method-neutral. This creates the possibility that part of the reported MMA-vs-1PH advantage reflects alignment between the training representation and the evaluation representation.

I do not claim that the reported improvement is therefore invalid. Missing/spurious objects and incorrect topology receive substantial independent penalties, so the MMA-derived term need not dominate the score. However, because only the aggregate score is reported, it is impossible to determine how much of the performance gap comes from genuinely better object recovery, topology, and localization, versus improvement specifically in the MMA-derived component.

This issue should be relatively easy to resolve: reporting the method-independent components of the existing score separately, or additionally reporting the score with the MMA-derived quality term removed, would allow the central MMA-vs-1PH comparison to be evaluated independently of the proposed representation.

I view other limitations of the empirical study, such as the use of a single synthetic benchmark, the absence of repeated training runs across multiple seeds, and the lack of runtime or broader downstream image-generation evaluation, as minor. They do not materially affect my recommendation, and I do not consider a broad additional experimental campaign necessary for acceptance.

Questions for authors
  1. Clarification regarding Assumption 6 and the experimental filtration.

How can Assumption 6 hold for the coordinatewise-max cubical filtration used in Appendix E.4? In particular, if one endpoint of an edge is larger than the other endpoint in both filtration coordinates, the edge inherits exactly that vertex's bifiltration value, yielding a face/coface tie on every sampled line. This occurs under strict inequalities and therefore can have positive probability rather than being a measure-zero degeneracy.

Do the authors apply a symbolic perturbation or another modification of filtration values to eliminate such ties? If so, please clarify whether the MMA descriptor and loss actually optimized during training are computed from this perturbed filtration, and how the theoretical guarantees transfer to the implemented objective. If no such modification is used, what argument establishes that the experimental bifiltrations satisfy the genericity condition required by Theorems 10--12?

The answer to this question would materially affect my assessment of the theoretical contribution.

  1. Clarification regarding the MMA-derived evaluation term.

How much of the reported MMA-vs-1PH performance difference remains when the MMA-derived quality deficit vjv_j is excluded from the evaluation?

A decomposition of the already reported score into its existing components -- e.g., missing/spurious objects, topology mismatch, spatial displacement, and MMA quality deficit -- or the same aggregate score with vjv_j removed would be sufficient. I am not requesting a broader new experimental study. If MMA retains a clear advantage under the method-independent components, this concern would largely be resolved.

3 comments
AuthorsOpenReview ↗

Response to Reviewer JhN8

Assumption 6 and the cubical bifiltration. We thank the reviewer for this question, and the reviewer is right. Before submission, we planned to perturb the bifiltration so that no two simplices take the same value on a line of LL, but an unresolved bug prevented this, so the submitted code admits ties, including the face/coface ties the reviewer describes. We have since fixed the bug: each simplex σ\sigma receives a fixed offset cσ=10−6r(σ)/Nc_\sigma = 10^{-6} r(\sigma)/N on both coordinates, where r(σ)r(\sigma) is its rank in an order of the NN simplices that lists faces before cofaces. The offsets are pairwise distinct and increase from faces to cofaces, so the perturbed map is a bifiltration without ties. MMA and the loss are computed on it, so Theorems 10-12 apply directly to the optimized loss. We are rerunning the experiments while monitoring that no ties occur. Preliminary results show an almost identical loss, and we will post the updated results when the runs are complete.

Evaluation metric. We agree. The first table decomposes the score of Table 1 into the terms of Eq. 19, averaged over all 1000 test images. The second reports the score without vjv_j for each object count NN. MMA keeps a clear advantage over 1PH on the method-independent terms, and without vjv_j it has the lowest mean score at every NN, N=9,10N = 9, 10 included. We will add the full decomposition for each NN to Appendix F, updated after the reruns.

ModelUnmatchedSpuriousWrong degreeDisplacementvjv_jScore (Table 1)
MMA0.0000.0060.0090.0350.0090.059
1PH0.0030.0330.0110.0410.0110.100
Ablation0.0020.3060.1860.0250.0030.522
Score without vjv_jN=1N=2N=3N=4N=5N=6N=7N=8N=9N=10
MMA0.0380.0340.0360.0350.0450.0390.0490.0630.0770.088
1PH0.0800.1010.1100.0750.1070.0890.0800.0750.0790.092
Ablation0.5250.4560.4030.4950.4480.5270.5520.5700.5890.621

Can you please respond to the authors even in the last minute?

Thank you

AuthorsOpenReview ↗

Updated results

Updated results. The reruns with the perturbed bifiltration described in our first response are complete, and we have uploaded the revised paper. In all reported MMA runs we compute MMA and the loss on the perturbed bifiltration, and we monitored the bifiltrations during training: no ties occurred. The tables below replace those of our first response. Table 1 of the revised paper also includes two new baselines, windowed 1PH and Betti matching.

Evaluation metric. Appendix F of the revised paper reports the decomposition averaged over the 1,000 test images (first table), and the tables that follow give it for each object count NN. MMA keeps a clear advantage over the two 1PH baselines on the method-independent terms, mainly through spurious features, and without vjv_j it has the lowest mean score at every NN.

ModelUnmatchedSpuriousWrong degreeDisplacementvjv_jScore without vjv_jScore
MMA0.0000.0060.0120.0350.0080.0530.061
1PH0.0030.0330.0110.0410.0110.0890.100
Windowed 1PH0.0030.0300.0140.0420.0090.0890.098
Betti matching0.0090.3410.2250.0200.0030.5950.598
Ablation0.0020.3060.1860.0250.0030.5190.522

MMA

TermN=1N=2N=3N=4N=5N=6N=7N=8N=9N=10
Unmatched0.0000.0000.0000.0000.0000.0000.0000.0000.0010.001
Spurious0.0000.0050.0130.0050.0060.0030.0040.0060.0030.015
Wrong degree0.0150.0050.0020.0150.0070.0080.0110.0130.0140.026
Displacement0.0370.0330.0350.0330.0350.0350.0360.0370.0370.037
vjv_j0.0050.0040.0030.0050.0060.0090.0080.0110.0150.014
Score without vjv_j0.0520.0430.0500.0530.0480.0460.0520.0560.0550.078
Score0.0560.0460.0530.0580.0540.0540.0600.0670.0700.093

1PH

TermN=1N=2N=3N=4N=5N=6N=7N=8N=9N=10
Unmatched0.0000.0000.0000.0000.0080.0030.0060.0010.0040.008
Spurious0.0300.0550.0530.0280.0460.0350.0210.0210.0200.023
Wrong degree0.0100.0050.0180.0080.0120.0100.0120.0100.0110.018
Displacement0.0400.0410.0390.0400.0410.0410.0410.0430.0430.043
vjv_j0.0090.0100.0070.0120.0110.0110.0120.0120.0130.012
Score without vjv_j0.0800.1010.1100.0750.1070.0890.0800.0750.0790.092
Score0.0900.1110.1180.0870.1170.1000.0930.0870.0920.104

Windowed 1PH

TermN=1N=2N=3N=4N=5N=6N=7N=8N=9N=10
Unmatched0.0000.0000.0030.0000.0000.0020.0040.0040.0010.011
Spurious0.0500.0250.0570.0320.0320.0370.0140.0160.0210.015
Wrong degree0.0100.0150.0120.0120.0170.0180.0130.0120.0180.015
Displacement0.0400.0400.0400.0410.0420.0420.0430.0440.0440.045
vjv_j0.0090.0090.0060.0110.0080.0090.0100.0090.0110.010
Score without vjv_j0.1000.0800.1120.0860.0910.0980.0750.0760.0840.086
Score0.1090.0880.1180.0970.0990.1070.0850.0850.0950.096

Betti matching

TermN=1N=2N=3N=4N=5N=6N=7N=8N=9N=10
Unmatched0.0000.0000.0000.0020.0020.0030.0090.0120.0210.038
Spurious0.2500.3150.2730.3800.3660.3980.4240.3620.3690.271
Wrong degree0.1700.2050.1870.2440.2300.2400.2400.2480.2540.232
Displacement0.0260.0210.0220.0190.0210.0190.0190.0180.0180.019
vjv_j0.0010.0090.0070.0040.0020.0010.0010.0010.0000.000
Score without vjv_j0.4460.5410.4820.6450.6190.6610.6920.6410.6620.560
Score0.4480.5500.4890.6490.6210.6620.6930.6430.6620.560

Ablation

TermN=1N=2N=3N=4N=5N=6N=7N=8N=9N=10
Unmatched0.0000.0000.0000.0000.0000.0000.0010.0020.0020.009
Spurious0.3100.2750.2370.2900.2520.3100.3330.3350.3420.373
Wrong degree0.1900.1550.1400.1810.1700.1920.1920.2080.2210.215
Displacement0.0250.0260.0260.0240.0260.0250.0250.0250.0240.024
vjv_j0.0020.0040.0030.0050.0040.0030.0030.0040.0030.002
Score without vjv_j0.5250.4560.4030.4950.4480.5270.5520.5700.5890.621
Score0.5270.4600.4060.5000.4520.5300.5550.5740.5920.624

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

We thank the reviewers and the area chair. We have uploaded a revised version of the paper. Below are the changes and the new experiments, each with the reviewers who asked for it.

  • Ties in the cubical bifiltration (JhN8, WPbu). Each simplex now receives a fixed offset that removes all ties, and we compute MMA and the loss on this perturbed bifiltration, so the regularity results of Section 3 apply to the trained loss. We reran all MMA experiments and monitored the bifiltrations during training: no ties occurred (Section 4.1, Appendix E.4).
  • MMA term vjv_j in the metric (JhN8, WPbu). Appendix F decomposes the score into its terms and reports it without vjv_j. Without vjv_j, MMA keeps the lowest mean score at every object count, and its advantage over the 1PH baselines comes mainly from spurious features.
  • 1PH baseline (WPbu, 1ghd). Appendix E.4 now describes the 1PH objective exactly.
  • New baselines (WPbu, 1ghd). Table 1 adds windowed 1PH, the localized baseline suggested by WPbu, and Betti matching, suggested by 1ghd. We also trained the signed measure baseline of Scoccola et al. requested by 1ghd, and report it in our reply to 1ghd.
  • Significance tests (1ghd). Paired Wilcoxon signed-rank tests show that MMA improves on 1PH and windowed 1PH for N=2,…,9N = 2, \dots, 9, and on Betti matching and the ablation at every NN (Appendix F).
  • Runtime (1ghd, JhN8). Appendix E.4 reports the time per optimization step of the CVAE, 1PH and MMA.
  • Image generation (WPbu, 1ghd, JhN8). Figure 3 shows images generated by ControlNet from the masks of MMA, 1PH and the ablation.
  • Scope and framing (WPbu, 1ghd). We restricted the framing to generative models without a ground-truth mask, cited induced matchings, and rewrote the opening of Section 4 on the U-Net diffusion model.
  • Writing (1ghd, WPbu). We rewrote the introduction, made the vague statements precise, corrected the citation commands, made gjg_j consistent, and fixed the typos.

Summary. The revised paper computes MMA and the loss on a bifiltration without ties, so the theory applies to the trained loss. It adds two baselines to Table 1, windowed 1PH and Betti matching, and we report a third one, the signed measure, in our reply to 1ghd. It also adds the score decomposition, significance tests, the runtime and ControlNet images. Reviewer WPbu flagged the paper for ethics review without giving a reason, and our request for clarification is still unanswered; the paper uses only synthetic binary masks generated by the authors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.