acceptodds
Under review as a conference paper at ICLR 2027

Normalizing Flow Maps

12 views

Abstract

Scalable approaches to learning neural dynamical systems have placed diffusion, flow matching, and few-step generators in flow maps at the forefront of generative modeling. Such state-of-the-art models have limited access to explicit model densi- ties despite evolving from classical generative modeling approaches that leveraged tractable likelihoods, e.g., Normalizing Flows and Autoregressive models. In this paper, we combine these threads of few-step generation and tractable sample likelihoods and introduce NORMALIZING FLOW MAPS (NFM), a class of exactly invertible pushforward maps that generalizes normalizing flows and flow maps. More precisely, NFMs learn the integrator of an underlying dynamical system, but crucially, they provide efficient exact likelihood evaluation by construction. To en- able scalable training of NFMs, we derive two complementary axes of consistency inherent to NFMs that yield (self) distillation objectives: (1) pointwise matching, which unlocks ℓ2-based flow-map style regression losses, and (2) distributional matching, which matches pushforward distributions to the marginals of a prescribed forward process. On conditional image generation on CIFAR-10 and ImageNet-64, NFMs achieve state-of-the-art sample quality among parameter-matched normalizing flows, while supporting likelihood-based RLHF. For equilibrium sampling in Boltzmann Generators, NFMs outperform prior state-of-the-art normalizing flows on both single-system and transferable peptide benchmarks.

Primary area:
generative models
Keywords:
normalizing flows, flow maps, flow matching, invertible generative models, generative modeling, likelihood-based training, few-step generation, boltzmann generators, post-training, fine-tuning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_mWDEOpenReview ↗
Rating
3
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces Normalizing Flow Maps which use time-conditioned invertible maps built on an iTARFlow backbone to transform noise into data along a stochastic interpolant. The goal is to combine the fast, few-step sampling of flow maps with the exact likelihood estimation of normalizing flows. The authors here introduce two training objectives NFM-R and NFM-D. NFM-R uses a regression loss inspired by MeanFlow, with a denoiser-based parameterization. NFM-D uses a conditional KL-divergence objective to match the model's output distributions to the interpolant's distributions. The paper provides results supporting both methods. On class-conditional CIFAR-10 and ImageNet-64, NFM improves FID and FLD over maximum-likelihood-trained iTARFlow. The authors also show in some capacity the reward-based reinforcement learning post-training using exact likelihoods. NFM-D improves over maximum-likelihood-trained flows on molecular sampling benchmarks involving alanine peptides, Chignolin, and transferable 4- and 8-residue peptides. However, the best-performing image model, NFM-R, has numerical invertibility issues. Noise annealing and double-precision here inference restore invertibility for this method but the invertible version performs worse.

Critical strengths
  1. The goal is clear , well motivated and nicely explained. The paper explains well why flow maps and noise-augmented normalizing flows lack clean-sample likelihoods. The paper introduces two methods with different training objectives with supporting theory.
  2. The evaluation is broad, covering image generation, RL post-training, and single-system and transferable Boltzmann generators. Baselines are carefully controlled in several places, including the iTARFlow tuning sweeps on CIFAR-10 and the SBG and TARFlow reimplementations. The Boltzmann generator results are strong on energy metrics, especially on Chignolin, and are reported with error bars.
  3. The authors openly acknowledge that NFM-R loses numerical invertibility.
Critical weaknesses
  1. The best image results come from a model that is not invertible as the top CIFAR-10 NFM-R achieves FID 3.87 and the ImageNet-64 NFM-R has FID 3.60 and NFM-R is non-invertible. The invertible NFM-R needs noise annealing and FP64 inference and drops to FID 5.17. This weakens the central claim of combining strong few-step generation with exact likelihoods.
  2. NFM-R gives the best sample quality, while NFM-D gives reliable likelihoods and is used for all Boltzmann generator results so different variants deliver different parts of the goal and there is no single best approach.
  3. The NFM-R image comparison is confounded by several factors. NFM-R is trained on ReFlow couplings generated by pretrained EDM and EDM2 models, uses auxiliary perceptual losses (Algorithm 1), and uses a modified architecture (SwiGLU, RMSNorm, RoPE, in-context conditioning). The MLE iTARFlow baseline is trained from scratch on data with none of these. The gains therefore cannot be attributed to the NFM objective alone. Table 12 shows that switching from TARFlow to iTARFlow under MLE gives most of the improvement, and NFM-D over MLE-iTARFlow is mixed (for example, TICA-W2 worsens on 4AA).
  4. The denoiser parameterization (Eq. 10) is not guaranteed to be invertible for t > 0. A weighted sum of x and an invertible map need not be invertible, and the paper notes that multi-step NFM-R densities are inaccessible. This conflicts with Definition 1, which requires every map to be a diffeomorphism.
  5. For a paper centered on exact likelihoods, no likelihood metric such as bits per dimension is reported for the image models. Also, many design choices are not ablated for NFM itself, including noise annealing, FP64, ReFlow couplings, perceptual losses, the parameterization and the optimizer.
  6. The paper has very few and limited baseline evaluation as the only parameter-matched normalizing flow baseline is the authors’ own iTARFlow run. On ImageNet-64 that run uses only two learning rates and a fixed noise level, unlike the extensive sweep on CIFAR-10.
  7. The transferable results (Tables 3, 4 and 12) have no error bars, and some margins are very small so the RL evidence isn't fully convincing. Each experiment is shown on a single prompt with very few gradient steps (5 to 40). The evaluation rewards partly overlap with the training rewards. There is no comparison with diffusion or flow post-training methods, so the claimed advantage over diffusion RL is not tested.
  8. Proposition 3 assumes L_end ≤ C_tr L_distill, a strong condition that is neither justified nor checked empirically.
  9. The evaluation could cover a more diverse range of datasets and tasks. Although the paper evaluates image generation and molecular modeling, additional experiments on diverse data types and tasks would help establish how well the method generalizes beyond the settings considered.
Questions for authors
  1. How much of NFM-R’s image performance comes from the EDM ReFlow couplings and the perceptual losses? Can you report NFM-R trained with independent couplings and no perceptual loss?
  2. Can you report bits per dimension on CIFAR-10 and ImageNet-64 for the invertible models, compared with iTARFlow?
  3. Is the denoiser map f_{s,t} invertible for t > 0? If not, which time pairs actually satisfy Definition 1 in your experiments?
  4. How often does numerical invertibility fail in FP32 versus FP64?
  5. How was the base temperature chosen in Table 3, and on which systems was it tuned?
  6. How does NFM post-training compare with a diffusion or flow RL method at matched compute, on held-out prompts?
  7. What are the wall-clock sampling times, training costs, and memory requirements compared with strong diffusion and flow-map baselines?
9 comments
AuthorsOpenReview ↗

Response to mWDE Part 1

We thank the reviewer for their thoughtful review and for the positive reflection of our work. We are especially grateful that the reviewer found that the main thesis of the paper to be “well motivated and nicely explained”, and that our evaluation of NFM is “broad”. In particular, we are heartened to hear that the reviewer found that our BG experiments are “strong”---especially for the largest peptide, Chignolin. We now turn to the chief unanswered questions grouped by theme.

NFM-R invertibility

We thank the reviewer for highlighting this trade-off. While our best NFM-R image results lose numerical invertibility, the invertible configurations still outperform the corresponding MLE baselines. On CIFAR-10, invertible NFM-R achieves FID 5.17, compared with iTARFlow’s 6.55 and SoftFlow’s 6.76. On ImageNet-64, invertible NFM-D achieves 4.59, improving to 4.52 with two steps, versus iTARFlow’s 4.89(Table 1). NFM-R also remains numerically invertible in our BG experiments, as demonstrated by the AL3 comparison with NFM-D (Appendix F.2, Table 13). These results demonstrate strong few-step generation alongside tractable likelihoods. We additionally investigate ImageNet-64 NFM-R’s instability in Appendix F.1 (Figure 9), showing error amplification during autoregressive inversion. We stress that the choice of invertible backbone is important for numerical invertibility, and other architectures may reveal a dramatically different invertibility-sample quality tradeoff. Finally, we reiterate that even this configuration supports likelihood-based, entropy-regularized post-training, which improves diversity at comparable reward (appendix F.1.1 Figure 12).

When to use NFM-R vs. NFM-D

We thank the reviewer for highlighting this trade-off. Both NFM-R and NFM-D support all applications considered, with different practical advantages. For images, NFM-R achieves higher sample quality (Table 1) and supports reward fine-tuning through pathwise gradients, while NFM-D supports more flexible KL-regularized finetuning (section 4.1). For Boltzmann generators, the comparison in Appendix F.2, Table 13, demonstrates both variants on AL3: their raw sampling metrics are comparable, while NFM-D performs better after resampling. We attribute this advantage to its better learned density representation, which motivates its use in our main BG experiments.

Thus, the variants offer complementary choices within the same framework, with the preferred objective depending on the application. We highlight that NFM itself is a model class, and practical choices such as an invertible backbone (iTarFlow) may yield different trade-offs. We defer a deeper exploration of the practical design space of NFMs as future work

Ablations on NFM-R

We have conducted additional ablations to further clarify the source of NFM-R’s superior sampling performance. In short, the use of synthetic EDM data and/or the architectural modifications that we make for NFM-R are not beneficial for MLE iTARFlow, indicating that their performance differences are algorithmic. We provide further details below in our response to the direct question Q1.

Clarifications on the Denoiser parametrization

We thank the reviewer for their question, as we believe the text could be clearer about this. As per Definition 1, an NFM is a model family that is restricted to invertible maps across a collection of admissible time pairs T⊂{(s,t):0≤t<s≤1}\mathcal{T} \subset \lbrace(s,t): 0 \leq t < s \leq 1\rbrace. NFM-R with a denoiser parametrization is an NFM restricted to the set T={(s,t):t=0,t<s≤1}\mathcal{T} = \lbrace(s,t): t=0, t < s \leq 1\rbrace. As an NFM, it is invertible for all (s,t)∈T(s,t) \in \mathcal{T}. For t>0t>0 the map parametrised by the denoiser is not invertible; however, this does not contradict Definition 1. NFMs using a dual time or gated parametrization (e.g. NFM-D) admit a larger set of time pairs T={(s,t):0≤t<s≤1}\mathcal{T} = \lbrace(s,t): 0\leq t < s \leq 1\rbrace. We hope this clarifies the situation adequately.

We understand how this was not sufficiently clear in the main paper, and we have updated the manuscript to highlight these distinctions.

Reporting additional Likelihood metrics

We thank the reviewer for raising this relevant question. We have conducted additional likelihood evaluation on CIFAR-10’s test set and ImageNet-64’s validation set. In short, NFM-D outperforms MLE iTARFlow on both datasets and invertible NFM-R underperforms both methods on CIFAR-10, further suggesting that it learns a worse density representation. We provide further details on these experiments in our response to Q2 below.

AuthorsOpenReview ↗

Response to mWDE Part 2

Additional Normalizing Flow baselines

We thank the reviewer for highlighting the utility of additional NF baselines. We have trained the original TARFlow on CIFAR-10 using a fork of the official repository and found that it significantly underperforms even iTARFlow (and all NFM models), achieving an FID of 10.48 and FLD of 9.47. We provide further details of the experiment setup, including learning rate sweeping, in Appendix F1. We also investigated incorporating Jet, another transformer-based NF, into our comparisons. However, we could not find any released checkpoints or reported FID, and its CIFAR-10 configuration uses over 2 billion parameters, which go well beyond our more modest parameter sizes used in the experiments. Moreover, this configuration does not produce coherent samples.

Clarification on experimental setup for Transferable and RL experiments

We thank the reviewer for highlighting this point. Following prior work, transferable results are reported as point estimates averaged over 30 held-out peptides per group. We have clarified this convention in Section 4.2 (Transferable generation), explicitly noting that these averages do not quantify variability across training seeds.

The RL experiments demonstrate the use of endpoint densities for post-training, something that diffusion and flow matching models are unable to do. Specifically, diffusion and flow matching approaches approach post-training in a fundamentally different manner and construct a soft value function from intermediate states to avoid computing endpoint likelihoods. If they could offer endpoint likelihoods, they would match the mature post-training stack of existing Autoregressive models. As a result, NFMs offer a very different mechanism—an alternative approach to post-training that prevents straightforward comparison.

With regard to rewards, we clarify that the only evaluation reward for NFM-D in Figure 3 is Skywork, which is an entirely separate VLM whose reward we do not train on. The ImageReward, Pickscore, and Aesthetic reward curves are provided to demonstrate consistent improvement over gradient updates. We appreciate that this was not sufficiently clear in the original draft, and we have now added extra clarification in Figure 3 (Section 4.1) the updated draft. Finally, we note that our demonstration here aims to prove out capabilities of a new model class, which is the primary goal of the evaluation, which includes both quantitative metrics and qualitative comparisons in Fig. 2 that very clearly illustrate the dramatic difference between pre-and-post RL training.

Theory clarifications

We thank the reviewer for this observation. We have added an empirical assessment of the endpoint control condition using 4,096 noise–image couplings at an ImageNet-64 NFM-R checkpoint (see appendix A.1, Empirical assessment of endpoint control). Under uniform time sampling, the estimated ratio Lend/Ldistill=3.56L_{end}/L_{distill} = 3.56, with a bootstrap 95% interval of [1.22,9.60][1.22, 9.60]. This provides evidence supporting the condition, which remains an explicit assumption of Proposition 3.

Diversity of evaluation

We thank the reviewer for this suggestion. We wish to very politely push back against this assertion. Our evaluation already spans image generation, reward-based image fine-tuning, and molecular sampling, covering a broader range of applications than is typically considered in either Flow Map or Boltzmann generator studies. To the best of our knowledge, no single paper that contains BG also demonstrates image generation and post-training capabilities. However, we do agree that exploring additional data types and tasks would further establish NFM’s generality and enhance its downstream use. We leave these as exciting directions for future work.

AuthorsOpenReview ↗

Response to mWDE Part 3

Questions

Q1.

How much of NFM-R’s image performance comes from the EDM ReFlow couplings and the perceptual losses? Can you report NFM-R trained with independent couplings and no perceptual loss?

The ability to incorporate couplings and auxiliary sample-based losses (e.g. LPIPS) , both standard in the few-step generative modelling literature, is an important advantage of NFM-R over TARFlow models trained with MLE. TARFlow-based NFs (like MAFs) have a fast parallel transformation and a slower autoregressive transformation which are inverses. MLE-based training uses the fast parallel transformation to map data to noise and samples with the slow autoregressive transformation, making sample-level losses infeasible as computing gradients would require backpropagating through the autoregressive rollouts in each block. Furthermore, the use of Tweedie’s trick to denoise samples further increases memory usage and computational cost (Appendix F.1, Figure 8). In contrast, NFM-R’s sample-level objective enables the fast transformation to be used for both training and sample generation, unlocking the use of auxiliary sample-based losses. Furthermore, since NFM-R optimizes an average velocity, its training can benefit from the use of couplings.

To further isolate the source of NFM-R’s advantage over iTARFlow trained with MLE we re-train three MLE iTARFlow baselines on CIFAR-10:

with the same EDM ReFlow data, with the same architectural modifications with both EDM data and architectural modifications

The exact experimental setup is detailed in Appendix F1 and the results in Table 8 indicate that they are not beneficial for MLE iTARFlow training, giving further evidence that it is NFM-R’s algorithmic differences which primarily account for its improved performance.

ModificationFID
Baseline6.55
w/ EDM data6.68
w/ NFM-R arch8.21
w/ both8.23

Q2.

Can you report bits per dimension on CIFAR-10 and ImageNet-64 for the invertible models, compared with iTARFlow?

We thank the reviewer for raising this important question. In this work we seek to develop models which can produce both high quality samples and evaluate clean data likelihoods. For this reason we do not train with dequantization noise or train separate likelihood-only models. Hence, we cannot report BPD with dequantization noise, as has been done in prior works (e.g. TARFlow). We now report clean data NLL in nats per dimension below and in Appendix F1 (‘Image likelihood evaluation’):

CIFAR-10:

MethodNLL (nats per dim)
MLE iTARFlow-1.53
NFMD-2.67
NFMR (FP64)29.73

ImageNet-64:

MethodNLL (nats per dim)
MLE iTARFlow-1.60
NFMD-2.60

Q3.

Is the denoiser map f_{s,t} invertible for t > 0? If not, which time pairs actually satisfy Definition 1 in your experiments?

Please see the response above.

Q4.

How often does numerical invertibility fail in FP32 versus FP64?

For configurations reported as numerically invertible, we observed no inversion failures at the precision used for evaluation. The invertible NFM-R configuration uses FP64, while NFM-D remains numerically invertible in FP32.

Q5.

How was the base temperature chosen in Table 3, and on which systems was it tuned?

We choose the temperature by sweeping T∈[0.9,1.1]T\in[0.9,1.1] independently on 4AA and 8AA validation sets. We now state this in Section 4.2 (Temperature tuning), and we report a more detailed description in Appendix F.3 (Temperature tuning settings). We additionally report in Figure 20 E−W2\mathcal{E}-{W}_2 and T−W2\mathbb{T}-{W}_2 as a function of temperature.

Q6.

How does NFM post-training compare with a diffusion or flow RL method at matched compute, on held-out prompts?

We thank the reviewer for this important question. Our image models are class-conditional and therefore cannot be evaluated on held-out prompts. When conducting our RL experiments, we align the outputs of our model with a text prompt purely through reward signal; the model does not directly condition on the text. In future we hope to develop text-to-image NFM models that would enable robust evaluations on held-out prompts. Finally, we note that our demonstration here aims to prove out capabilities of a new model class, which is the primary goal of the evaluation, which includes both quantitative metrics and qualitative comparisons in Fig. 2 that very clearly illustrate the dramatic difference between pre-and-post RL training.

Reviewer_ed8LOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper introduces Normalizing Flow Maps, a framework aimed at combining few-step generation with tractable, normalized model densities. It develops two training approaches: NFM-R uses pointwise flow-map regression, while NFM-D matches distributions across noise levels. The authors also derive endpoint error bounds under regularity assumptions.

Experiments cover image generation, reward-based post-training, and moleculal sampling on single-system and transferable peptide benchmarks. NFM-R achieves the strongest image quality but lose numerical invertibility. NFM-D retains tractable densities, enabling likelihood-based RL and importance-sampling corrections for molecular generation.

Critical strengths

Combining generation with tractable sample densities is a useful contribution, with demonstrated applications to reinforcement learning and molecular sampling corrections.

Critical weaknesses

W1) The fidelity–invertibility tradeoff is insufficiently reflected in the abstract. The strongest NFM-R image results lose numerical invertibility, so the best sample-quality results do not simultaneously retain the numerical invertibility needed for the stated REINFORCE and KL-regularized updates.

W2) Several motivating applications do not require normalized densities. Proportional densities suffice for self-normalized importance sampling and the stated leave-one-out REINFORCE update, assuming correct model sampling within each conditioning group. The paper should clarify when normalized likelihoods justify the constraints imposed by its invertible architecture.

W3) Relevant energy-based alternatives are missing from the discussion. Energy Matching [1] learns an unnormalized density and reports lower CIFAR-10 FID. Equilibrium Matching [2], including its explicit-energy variant, also warrants discussion. Their tradeoffs in fidelity, sampling cost, and downstream applicability should be addressed, while distinguishing energy values from the density of actual sampler outputs.

W4) Computational advantages are not sufficiently measured. Few sampling steps do not establish low wall-clock cost, particularly for the invertible NFM-R configuration requiring FP64. Matched timing comparisons for generation, likelihood evaluation, and post-training are missing.

References:

[1] Balcerak et al. Energy Matching: Unifying Flow Matching and Energy-Based Models for Generative Modeling. NeurIPS, 2025.

[2] Wang et al. Equilibrium Matching: Generative Modeling with Implicit Energy-Based Models. NeurIPS, 2026.

Questions for authors

Q1) Which demonstrated applications require normalized densities rather than (less constrained) proportional densities?

Q2) How does the method compare with Energy Matching [1] and/or Equilibrium Matching [2] (or alternatives) in fidelity, sampling cost, and suitability for downstream applications? Basically what are the differences vs energy-based models which is a large established field.

Q3) Can you report wall-clock costs for generation, likelihood evaluation, and RL post-training, including FP64 and fine-tuning costs?

7 comments
AuthorsOpenReview ↗

Response to ed8L Part 1

We are greatly appreciative of the reviewer's time, effort, and dedication to the review process and their positive endorsement of our work. In particular, we thank the reviewer for acknowledging that “combining generation with tractable sample densities is a useful contribution”, which is demonstrated through our experiments on KL-constrained post-training and conformational sampling with Boltzmann Generators. We now turn to answering the main outstanding questions from the review.

Fidelity-Invertibility Tradeoff of NFM-R

We thank the reviewer for raising this point about the fidelity–invertibility trade-off. The abstract describes capabilities demonstrated by the NFM family, rather than attributing every capability to the highest-fidelity configuration. NFM-D retains numerical invertibility, outperforms the corresponding MLE baseline (Table 1), and supports KL-regularized REINFORCE on CIFAR-10 (Figure 10) and ImageNet-64 (Figure 3). Noise-annealed NFM-R also supports these updates on CIFAR-10 (Appendix E.5, Figure 10). NFM-R also remains numerically invertible in our BG experiments (Appendix F.2, Table 13) as well as CIFAR. Meanwhile, in ImageNet-64 NFM-R supports likelihood-based post-training through entropy regularization (Appendix E.5), which improves diversity at comparable reward (Figure 12). Table 1 explicitly distinguishes invertible configurations. Our blockwise analysis in Appendix F.1 (Figure 9) suggests that ImageNet-64 NFM-R’s numerical instability arises from the amplification of reconstruction errors during autoregressive inversion of the iTARFlow backbone. We highlight that such practical choices matter and that other invertible backbones may incur a different numerical invertibility tradeoff. We leave this exploration as future work.

The need for normalized densities

We thank the reviewer for this valid observation. In principle, unnormalized densities are sufficient to perform SNIS and the leave-one-out REINFORCE update. In practice, the chief requirement is both fast sampling and fast (un)normalized densities. Unfortunately, drawing samples from such a model class that parametrizes unnormalized densities in a fast manner is quite challenging. The gold standard for drawing samples from an unnormalized density is Markov Chain Monte Carlo (MCMC), which is known to have long mixing times. This, unfortunately, rules out practical applications such as conformational sampling with Boltzmann Generators. The practical advantage of NFM is access to this sampling distribution’s tractable density, without requiring equilibration or a separate density estimator. We understand how this may not have been sufficiently clear in the original manuscript, and in the rebuttal we have expanded the discussion in Appendix D (Extended Related Work) and address this distinction further in Q1 below.

Energy-based baselines

We thank the reviewer for pointing us to these relevant and interesting alternatives. We have expanded Appendix D (Extended Related Work) to discuss Energy Matching [1] and Equilibrium Matching [2], distinguishing their sampling procedures and learned energies from the density of their generated outputs. We summarize the implications for our applications in Q2 below and note that first EBMs have not been scaled to larger BG systems like Chignolin due to long mixing times and a bias that is difficult to characterize for any finite kk steps of Langevin. Thus, the actual distribution used for importance weighting is not the same as the one used in exact likelihood methods—i.e., the estimator is biased and not consistent. Second, drawing samples from an EBM is slow, which renders applications in BGs and RL based post-training more computationally demanding than NF and NFM-based alternatives.

AuthorsOpenReview ↗

Response to ed8L Part 2

Computational advantages

Computational advantages are not sufficiently measured. Few sampling steps do not establish low wall-clock cost, particularly for the invertible NFM-R configuration requiring FP64. Matched timing comparisons for generation, likelihood evaluation, and post-training are missing.

We thank the reviewer for inquiring about this relevant practical issue. We have added additional computational evaluations in Figure 8 (Appendix F.1) comparing the efficiency of the ImageNet NFM models and iTARFlow on a single NVIDIA RTX Pro 5000 Blackwell (48 GB). We do all our training evaluations in mixed bf16 precision and our inference-time evaluations in fp32. In terms of training latency, we note that this is confounded by our architectural choice of using a SOTA invertible model in iTarFlow which is slow in one direction. A different architecture e.g., affine coupling flows would not have the same restrictions, but likely worse sample quality. Consequently, we find that NFM-D (one-step) and Flow Matching are 15% to 20% slower than the MLE train step at their maximum batch size of 64, which we attribute to their use of the Muon optimizer. In contrast, the pixel mean flow training step (improved mean flow with guidance and perceptual loss) halves the latency but only fits a significantly smaller batch size of 8. However, our NFM-R training recipe only uses 125 epochs of FM followed by 25 epochs of pMF (as opposed to 300 epochs of MLE or NFMD), making the overall training time lower. Likelihood evaluation on samples not produced by the model requires that the NF’s data-to-noise transformation is used, for MLE and NFM-D this is the fast parallel TARFlow transformation and for NFM-R, this is the slow, autoregressive transformation. Hence, MLE iTARFlow and NFM-D evaluate a single sample’s likelihood in 20ms, whereas NFMR takes 2.5 seconds. In contrast, sampling requires mapping noise-to-data, making NFM-R superior in both sample latency and throughput. Whilst MLE iTARFlow and NFM-D take 4 seconds to generate a single sample with guidance, NFM-R requires only 22ms. Furthermore, the need for Tweedie denoising means that MLE iTARFlow can only fit 8 samples into memory leading to a large difference in sampling throughput; NFM-R takes 15 min to sample 50 000 samples, NFM-D 320 min, and MLE iTARFlow 690 min. Finally, we note that generating a sample also provides its likelihood, hence in settings like Self-Normalized Importance Sampling and MaxEnt RL NFM-R would be superior in terms of computational efficiency.

Therefore, from both a performance and efficiency point of view, when NFMs are implemented with TARFlow architectures, NFM-R is better suited to sampling and NFM-D to density evaluation on unseen samples.

Lastly, we also provide new sampling/likelihood evaluation timing for Boltzmann generators in Appendix F.3 (Inference efficiency), Table 16 , which highlights the significant speed advantage of NFMs over existing methods. For simplicity, we report sampling time for 10,000 4AA samples here:

MethodGeneration time (s)
ECNF++>100,000
Prose29.0
Robin130.6
NFMD22.0

Questions

Q1

Which demonstrated applications require normalized densities rather than (less constrained) proportional densities?

As discussed above, proportional densities suffice for SNIS and our leave-one-out REINFORCE update. However, these densities must correspond to the distribution actually sampled. For EBMs sampled through finite Langevin runs, incomplete mixing can introduce a mismatch between the generated samples and the learned density, introducing bias—that cannot be removed with more Monte Carlo samples—when that density is used for SNIS or policy-gradient updates. This step, as a result, leads to slow sample generation while practically incurring an unquantifiable bias that prevents scalable importance sampling in BGs or post-training. We highlight that there are no EBM models that have been successfully trained for Chignolin or transferable peptides [2]. In addition, only recent concurrent work has demonstrated the capability of EBMs for larger-scale image generation, but these demonstrations have not yet proven the use of their unnormalized densities [1,2]. In contrast, NFMs directly generate samples with their corresponding tractable densities, avoiding this equilibration requirement.

AuthorsOpenReview ↗

Response to ed8L Part 3

Q2

How does the method compare with Energy Matching and/or Equilibrium Matching (or alternatives) in fidelity, sampling cost, and suitability for downstream applications? Energy Matching and Equilibrium Matching are interesting alternatives with strong image-generation results, but their downstream use requires distinguishing the learned energy from the sampling distribution:

  • Energy Matching: its unnormalized density can support SNIS in principle. However, finite Langevin runs may not equilibrate, producing incorrect basin populations relative to the learned density and consequently biased reweighting. This bias is not the same bias as SNIS which can be alleviated through more MC samples.
  • Equilibrium Matching: samples are generated through gradient descent, which does not generally sample the Boltzmann distribution associated with its energy. Its explicit energy therefore cannot directly provide the proposal density required for SNIS.
  • Molecular applications: novel EBM approaches provide preliminary results on dipeptides, but, to our knowledge, their effectiveness on the larger peptides and transferable equilibrium-sampling tasks considered here remains to be established. Moreover, any finite-sample Langevin may have a bias that prevents its systematic use in higher-dimensional importance sampling. NFM demonstrates both settings.

Q3

Computational advantages are not sufficiently measured. Few sampling steps do not establish low wall-clock cost, particularly for the invertible NFM-R configuration requiring FP64. Matched timing comparisons for generation, likelihood evaluation, and post-training are missing.

Please, see the response above.

Closing remarks

We thank the reviewer for their time and effort in reviewing this paper. We believe we addressed all the reviewer's points in our rebuttal and added new training, sampling and likelihood evaluation computational evaluations. We would be thrilled if the reviewer would potentially consider upgrading their score, and we remain eager to help clarify any additional questions if they arise. Please let us know.

References

[1] Balcerak et al. Energy Matching: Unifying Flow Matching and Energy-Based Models for Generative Modeling. NeurIPS, 2025.

[2] Wang et al. Equilibrium Matching: Generative Modeling with Implicit Energy-Based Models. NeurIPS, 2026.

[3] Aggarwal et al. BoltzNCE: Learning Likelihoods for Boltzmann Generation with Stochastic Interpolants and Noise Contrastive Estimation. NeurIPS, 2025.

Reviewer_6TMGOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper introduces Normalizing Flow Maps (NFMs), a family of time-conditioned invertible transformations intended to combine few-step generation with tractable exact model densities. It develops two training approaches: NFM-R adapts MeanFlow-style regression to an invertible backbone, while NFM-D uses likelihood-based training to match the marginal distributions of a stochastic interpolant. The paper also provides endpoint-distribution bounds under regularity assumptions. Experiments cover class-conditional generation on CIFAR-10 and ImageNet-64, reward-based image post-training, and molecular equilibrium sampling. The results show improvements over several normalizing-flow baselines, although the strongest NFM-R image-generation results lose numerical invertibility.

Critical strengths
  1. The combination of direct generation and tractable densities addresses a meaningful practical need. NFM-D produces samples without a separate Tweedie denoising transformation, preserving the change-of-variables density for the output distribution. The paper demonstrates concrete uses of this capability through KL-regularized REINFORCE and importance sampling. These applications make likelihood access consequential rather than merely an architectural property.

  2. On Chignolin, NFM-D with SMC achieves an energy Wasserstein distance of (3.42 - 0.27), compared with (10.87 - 10.18) for the authors’ matched SBG-iTARFlow implementation, while also improving the torsional metric. The transferable evaluation covers 30 unseen four-residue and 30 unseen eight-residue peptides. Furthermore, Table 12 separates some gains from architecture and optimizer changes from those attributable to the NFM-D objective.

Critical weaknesses
  1. The central combination of sample quality, reliable likelihood access, and computational efficiency is only partially demonstrated. The best NFM-R image results—FID 3.87 on CIFAR-10 and 3.60 on ImageNet-64—are explicitly classified as numerically non-invertible. Appendix D.5 acknowledges that this prevents score-function updates and KL-regularized post-training, requiring a switch to differentiable rewards and entropy regularization. Recovering numerical invertibility on CIFAR-10 requires noise annealing and FP64 inference, with FID worsening to 5.17; an analogous invertible NFM-R result is not reported for ImageNet-64. NFM-D retains the intended capability, but its ImageNet-64 FID improvement over iTARFlow is relatively modest, from 4.89 to 4.59. Quantitative inversion-error measurements and sampling/likelihood-evaluation timings are missing. These omissions matter because the paper’s main practical claim concerns obtaining these properties simultaneously.

  2. The methodological contribution and the sources of the image-generation gains are not sufficiently isolated. The single-time NFM-D objective in Eq. (11) is conditional maximum likelihood on corrupted inputs, closely related to SoftFlow, which already trains noise-conditioned normalizing flows and samples at a small or zero noise level. The paper acknowledges this relationship, but the precise advance beyond the corruption schedule and parameterization remains insufficiently established; the direct SoftFlow comparison is limited to CIFAR-10. For NFM-R, Appendix E.1 reveals the use of EDM/EDM2-generated ReFlow couplings, while Appendix D.2 describes substantial architectural changes and the training recipe also differs from the MLE baseline. The optimizer sweeps are useful, but they do not isolate these factors. Without stronger controls against the inherited MeanFlow training approach and an equivalently configured likelihood baseline, it is difficult to determine how much improvement comes specifically from the proposed NFM formulation.

Questions for authors
  1. Which configurations in Table 1 support reliable inverse evaluation and likelihood-based KL post-training?

  2. How much improvement remains when the training data/couplings, architecture, optimizer, and guidance settings are controlled? In particular, what happens without EDM/EDM2 ReFlow couplings, and how does NFM-R compare with an otherwise matched MeanFlow-style model?

  3. What specifically distinguishes single-time NFM-D from SoftFlow beyond the corruption path and noise-level sampling distribution? Which reported results use the two-time extension, and does that extension improve over single-time conditional likelihood training under matched settings?

5 comments
AuthorsOpenReview ↗

Response to 6TMG Part 1

We thank the reviewer for their time, effort, and positive appraisal of our work. We are delighted that the reviewer agrees with the main thrust of our paper in that “direct generation and tractable densities addresses a meaningful practical need”, and that our experiments with the NFM-D model demonstrate “concrete uses of this capability”. We are also grateful that the reviewer appreciated our Boltzmann Generator experiments and remarked on the tangible improvements of NFM-D over previous SOTA flow-based baselines on Chignolin and Transferable settings. We next address the main clarification points grouped by theme.

NFM-R Analysis

We acknowledge the reviewer’s comment that NFM-R’s lack of numerical invertibility in ImageNet-64 may initially partially obscure NFM’s claims of providing tractable densities. We first reiterate that for Boltzmann Generators (Appendix F.2, Table 13) and CIFAR-10, NFM-R is invertible. We also underscore that a loss of numerical invertibility in invertible neural networks is not a new phenomenon and has been highlighted by several prominent works [1,2], that link this property to the Lipschitz constant of a map. Controlling this Lipschitz constant is often necessary in practice through regularization schemes, noise injection (as in TarFlow), or higher numerical precision.

In this rebuttal, we investigate ImageNet-64 NFM-R through blockwise inversion-error measurements (Appendix F.1, Figure 9), which suggest that reconstruction errors amplify during autoregressive inversion of the iTARFlow backbone. As a result, we argue that NFM-R as an object is invertible, but practical choices such as which invertible architecture used can impact numerical invertibility in certain applications e.g., CIFAR vs. ImageNet 64 vs. BGs.

Finally, we also conduct a new ablation for Max-Ent post-training on ImageNet-64 with and without entropy regularization (Appendix F.1.1, Figure 12), finding that entropy allows more diverse samples at comparable reward. This highlights the usefulness of NFM-R’s sample likelihoods, which remain accessible through forward log-determinants without numerical inversion (Appendix E.5).

NFM-D

Regarding NFM-D, we have also now included a new two-time formulation of NFM-D (see Section 3, and Q3 below), which improves upon our original single-time loss, obtaining improvements in FID for CIFAR-10. We have also included an original TarFlow baseline for CIFAR-10, which obtains an FID of 10—considerably worse than any of our NFM models. We hope this now fully addresses the reviewer’s initial question.

Timing

We value the reviewer’s question regarding timing. We now provide new sampling/likelihood evaluation timing for Boltzmann generators in Appendix F.3 (Inference efficiency), Table 16 , which highlights the significant speed advantage of NFMs over existing methods. For simplicity, we report sampling time for 10,000 4AA samples here:

MethodGeneration time (s)
ECNF++>100,000
Prose29.0
Robin130.6
NFMD22.0

We also provide new sampling/likelihood evaluation timing for our ImageNet-64 experiments in Appendix F.1 (‘Training, sampling and likelihood evaluation efficiency’). For brevity, we report here the time each method takes to generate 50 000 samples at the maximum batch size they can fit into a single GPU.

MethodGeneration time (min)
MLE iTARFlow689.69
NFM-D (one step)319.02
NFM-R (one step)15.15

We clearly observe that NFM is the most efficient model in both BGs and Images for comparable model classes and parameter counts.

Ablations of Method on Image Generation

We value the reviewer’s feedback that it is important to ablate various design choices to understand performance gains in our image experiments better. We first highlight that we introduce a new two-time formulation of our NFM-D that trains the NFM between two intermediate marginals s→ts \to t, rather than leveraging only t=1t=1. This formulation, introduced in an updated Section 3.3, is dramatically different from SoftFlow and MLE, is only possible with an NFM, and crucially also leads to slightly improved performance on CIFAR.

To further ablate various configurations, we focus on empowering the iTarFlow baseline with (1) samples from EDM generations (couplings), (2) the same architectural modifications of NFM-R, and (3) both settings. We detail these results in our Q2 response below, but in short, we find that these knobs do not meaningfully improve our reported iTarFlow. We conclude that NFM, amongst invertible models, uniquely benefits from our chosen modifications.

AuthorsOpenReview ↗

Response to 6TMG Part 2

Questions

Q1.

Which configurations in Table 1 support reliable inverse evaluation and likelihood-based KL post-training?

In Table 1(a) all the models under the heading ‘Invertible NFs’ are numerically invertible, and it is possible to achieve KL-regularized RL. This is shown in Appendix F.1.1 (Figure 10). Regarding ImageNet-64 (Table 1(b)), NFM-D is numerically invertible, and it is possible to post-train it with KL-regularized RL. NFM-R, instead, while being analytically invertible, is not numerically. We can post-train NFM-R with entropy-regularized RL, which is still likelihood-based (as explained in Appendix E.5). We show in Appendix F.1.1 (ImageNet-64: Effect of entropy regularization) that entropy-regularized RL obtains a better reward-diversity trade-off, if compared to the non-regularized case (Figure 12).

Q2.

How much improvement remains when the training data/couplings, architecture, optimizer, and guidance settings are controlled? In particular, what happens without EDM/EDM2 ReFlow couplings

This is a great question! We report an optimizer ablation for our MLE baseline in Table 7, where we see that unlike NFM, MLE iTARFlow does better with AdamW. Guidance settings were swept for all methods (Figures 6 and 7 in the Appendix), and we note that NFM-D and MLE TARFlow methods use the same form of guidance. For NFM-R, guidance operates on the average velocity, and thus settings are not directly comparable. Furthermore, we note that the only architectural difference between NFM-D and iTARFlow is the use of timestep embeddings, and that they share the same datasets. To further isolate the sources of NFM-R’s advantage over iTARFlow, we re-train three MLE iTARFlow baselines on CIFAR-10:

  1. with the data component of the same EDM ReFlow couplings,
  2. with the same architectural modifications
  3. with both EDM data and architectural modifications

The exact experimental setup is detailed in Appendix F1, and the results in Table 8 (reproduced below) indicate that they are not beneficial for MLE iTARFlow training, giving further evidence that it is NFM-R’s algorithmic differences which primarily account for its improved performance.

ModificationFID
Baseline6.55
w/ EDM data6.68
w/ NFM-R arch8.21
w/ both8.23

How does NFM-R compare with an otherwise matched MeanFlow-style model?

The closest MeanFlow-style model is Pixel Mean Flow (pMF), whose training recipe is adapted by NFM-R. pMF does not report CIFAR-10 results, and its reported ImageNet-64 FID of 3.80 is worse than NFM-R’s 3.60. However, these results must be contextualized, as there are substantial architectural differences (DiT backbone, full self-attention vs. causal) and training recipes.

Q3.

What specifically distinguishes single-time NFM-D from SoftFlow beyond the corruption path and noise-level sampling distribution? Which reported results use the two-time extension, and does that extension improve over single-time conditional likelihood training under matched settings?

NFM is a general flow map framework for normalizing flows that encompasses both NFM-R and NFM-D. NFM-D learns an invertible map between any two time steps of an interpolant, as we now write in Section 3.3 (Two-time extension), allowing not only data likelihood evaluation, but also likelihood-based inference time steering. Limiting NFM-D to a single-time map, we recover a SoftFlow-like loss, but we show in Table 1(a) that NFM-D still achieves lower FID than SoftFlow. Additionally, we report in Section 3.3 and Appendix A.2 a novel theoretical justification for NFM-D. Finally, the original SoftFlow manuscript reports normalizing flows acting on a 3-dimensional space, while we show NFM-D is scaled to large peptides and image datasets.

We now report in Table 1(a,b) the results from 2-step inference NFM-D, which generally yield a lower FID but higher FLD compared to a single-step version. For simplicity, we report a cutout of the table here.

CIFAR-10:

MethodFID
NFM-D, 1 step6.26
NFM-D, 2 steps6.05

ImageNet-64:

MethodFID
NFM-D, 1 step4.59
NFM-D, 2 steps4.52

Closing remarks

We thank the reviewer for their time and effort in reviewing this paper. We believe we addressed all the reviewer's points in our rebuttal and added new methodological ablations on images, additional Normalizing Flow baselines, and a new invertibility analysis of NFM-R. We would be thrilled if the reviewer would potentially consider upgrading their score as they originally suggested, and we remain eager to help clarify any additional questions if they arise. Please let us know.

References

[1] Behrmann et al. "Understanding and mitigating exploding inverses in invertible neural networks." AISTATS 2021.

[2] Cornish et. al “Relaxing Bijectivity Constraints with Continuously Indexed Normalising Flows.” ICML 2020.

Dear Reviewer, thank you again for your careful review and constructive suggestions. Your feedback motivated additional controlled ablations, numerical inversion analysis, and computational benchmarks, alongside a better explanation of the two-time NFM-D formulation. We hope these additions help address the points you raised. Should you have an opportunity to revisit our responses before the discussion closes, we would be grateful for any further thoughts and would be happy to clarify any remaining questions. Thank you again for your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response to All Reviewers

We thank all reviewers for their time, care, and constructive feedback, and the Area Chair for overseeing the review process. We are encouraged that the reviewers recognize the value of combining direct generation with tractable sample densities: 6TMG identifies a “meaningful practical need,” ed8L calls this combination a “useful contribution,” and mWDE finds the motivation “well motivated and nicely explained.” We also appreciate their recognition of the broad evaluation and strong Boltzmann generator results. We have carefully addressed the concerns raised through additional experiments, analysis, and clarifications, summarized below

Numerical invertibility and likelihood-based post-training (6TMG, ed8L, mWDE)

We appreciate the reviewers’ concerns about demonstrating sample quality and likelihood access together, and now provide additional inversion analysis and likelihood-based RL experiments.

We clarify which configurations support numerical inversion and each post-training objective. Invertible NFM-R achieves CIFAR-10 FID 5.17, versus 6.55 for iTARFlow and 6.76 for SoftFlow; NFM-D retains invertibility on ImageNet-64 and improves over iTARFlow (Table 1). NFM-R also remains numerically invertible in our Boltzmann Generator experiments (Appendix F.2, Table 13).

To further analyse numerical invertibility of NFM-R on ImageNet-64 we add blockwise inversion error measurements suggesting that ImageNet-64 NFM-R’s instability arises from error amplification during autoregressive inversion of the iTARFlow backbone (Figure 9). This highlights that the choice of invertible backbone plays a key role rather than NFM-R as an mathematical object being non-invertible, which substantiates our CIFAR experiments.

Lastly, we also add an RL post-training ablation showing that entropy regularization leveraged by NFM-R preserves greater diversity at comparable reward (Figure 12), demonstrating the usefulness of NFM-R’s sample likelihoods even when numerical inverse evaluation is unstable.

Methodological distinction and controlled comparison (6MTG and mWDE)

We value the reviewers’ feedback on distinguishing NFM’s contribution from existing methods, and have added controlled ablations, further baselines, and a two-time NFM-D formulation.

We introduce a two-time NFM-D objective, with its derivation in section 3.3 and Appendix A.2. Two-step generation preserves invertibility and improves FID from 6.26 to 6.05 on CIFAR-10 and 4.59 to 4.52 on ImageNet-64 (Table 1). To investigate the sources of NFM-R’s gains, we additionally train MLE iTARFlow with EDM-generated training data, NFM-R’s backbone modifications, and both together (Appendix F.1, Table 8). NFM-R outperforms every configuration examined, showing that these data and architecture changes alone do not recover its sample quality under MLE training. We also add the original TARFlow baseline, which NFM-R and NFM-D both outperform on CIFAR-10 (Table 1). Finally, on CIFAR-10’s test set NFM-D achieves a nats/dim of -2.67 vs MLE iTARFlow’s -1.53 and on ImageNet-64’s validation set achieves -2.60 vs -1.60.

Computational efficiency (6TMG, ed8L and mWDE)

We agree that computational advantages should be measured directly, and now benchmark training, sampling, and likelihood evaluation to substantiate the practical benefits.

We add image model benchmarks covering training step latency, sampling throughput, memory limited batch sizes, and likelihood evaluation latency (Appendix F.1, Figure 8). In particular, we highlight that that on ImageNet-64 NFM-D more than doubles the sampling throughput of MLE iTARFlow and is able to fit a batch 64×64\times larger, whilst NFM-R is 40×40\times faster with a maximum batch size that is 32×32\times larger.

For molecular sampling, generating 10,000 tetrapeptide conformations and evaluating their densities takes 22.0 seconds for NFM-D, compared with 29.0 for Prose and 130.6 for Robin (Appendix F.3, Table 16). We explicitly discuss the different sampling and density-evaluation costs of the two variants.

Theoretical assumptions and experimental scope (ed8L and mWDE)

We thank the reviewers for prompting us to empirically assess the endpoint-control assumption, clarify molecular evaluation protocols, and discuss the downstream applicability of energy-based alternatives.

We empirically assess Proposition 3’s endpoint-control condition using 4,096 noise–image couplings, obtaining an endpoint-to-distillation loss ratio of 3.56 (Appendix A.1). We retain the condition as an explicit assumption. We also clarify transferable result reporting and validation based temperature selection (section 4.2, Appendix F.3), and expand the discussion of energy-based alternatives and their sampling density requirements (Appendix D).

We believe these additions substantially strengthen the evidence for NFM’s contribution while making its numerical properties, computational trade-offs, and theoretical scope clearer.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.