- 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
- 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.
- 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.
- The authors openly acknowledge that NFM-R loses numerical invertibility.
Critical weaknesses
- 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.
- 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.
- 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).
- 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.
- 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.
- 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.
- 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.
- Proposition 3 assumes L_end ≤ C_tr L_distill, a strong condition that is neither justified nor checked empirically.
- 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
- 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?
- Can you report bits per dimension on CIFAR-10 and ImageNet-64 for the invertible models, compared with iTARFlow?
- Is the denoiser map f_{s,t} invertible for t > 0? If not, which time pairs actually satisfy Definition 1 in your experiments?
- How often does numerical invertibility fail in FP32 versus FP64?
- How was the base temperature chosen in Table 3, and on which systems was it tuned?
- How does NFM post-training compare with a diffusion or flow RL method at matched compute, on held-out prompts?
- What are the wall-clock sampling times, training costs, and memory requirements compared with strong diffusion and flow-map baselines?
9 comments
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 , with a bootstrap 95% interval of . 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.
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.
| Modification | FID |
|---|---|
| Baseline | 6.55 |
| w/ EDM data | 6.68 |
| w/ NFM-R arch | 8.21 |
| w/ both | 8.23 |
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:
| Method | NLL (nats per dim) |
|---|---|
| MLE iTARFlow | -1.53 |
| NFMD | -2.67 |
| NFMR (FP64) | 29.73 |
ImageNet-64:
| Method | NLL (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 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 and 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.
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}. NFM-R with a denoiser parametrization is an NFM restricted to the set T={(s,t):t=0,t<s≤1}. As an NFM, it is invertible for all (s,t)∈T. For t>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}. 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.