acceptodds
Under review as a conference paper at ICLR 2027

COAL: Jacobian-Free Likelihood Evaluation for Rectified Flows

Abstract

Flow-based generative models achieve strong sample quality, but per-sample likelihood evaluation under the instantaneous change-of-variables formula remains bottlenecked by an divergence integral. Hutchinson's stochastic trace estimator partially mitigates this, at the cost of backpropagations through the velocity network per ODE step and a noisy integrand. We show that the divergence of a Rectified Flow velocity admits a deterministic closed form: by Tweedie's formula, it equals the trace of the conditional covariance of the velocity residual, scaled by a known function of. We amortize this trace with a lightweight per-coordinate variance head trained jointly with the velocity, yielding Coal (covariance-amortized likelihood): a deterministic per-sample likelihood estimator that runs in a single forward pass per ODE step, with no input-space backpropagation and no stochastic probes. On class-conditional ImageNet with a DiT-XL latent rectified flow, Coal matches the Hutchinson reference at less evaluation compute.

Primary area:
generative models
Keywords:
rectified flow, likelihood evaluation, generative modeling

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_f36YOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes COAL (Covariance-Amortized Likelihood), a Jacobian-free likelihood estimation method for Rectified Flow models. The key idea is to exploit a relationship between the divergence of the optimal Rectified Flow velocity field and the trace of the conditional covariance of its velocity target. Based on this identity, the authors introduce an auxiliary variance prediction head trained with a Gaussian negative log-likelihood objective. This enables deterministic likelihood estimation without computing Jacobian-vector products or using Hutchinson trace estimators during inference. Experiments on a 2D checkerboard distribution and ImageNet-256 demonstrate that COAL achieves accurate likelihood estimates with substantially reduced computational overhead. In particular, the method achieves comparable BPD-RMSE to Hutchinson estimation with approximately 3.5× less evaluation compute.

Critical strengths
  1. The paper introduces an approach to likelihood estimation by leveraging the mathematical structure of Rectified Flows. Although the underlying conditional covariance identities and Gaussian variance estimation techniques are established, combining them into a Jacobian-free likelihood estimator is a meaningful methodological contribution.

  2. Proposition 3.1 provides a clear mathematical foundation for the proposed estimator under the stated assumptions. The resulting algorithm is relatively simple, requiring only an additional variance prediction head and a modified training objective, without introducing complicated inference procedures.

  3. The experiments demonstrate close agreement with the Hutchinson reference while significantly reducing computational cost. The reported improvements in GPU throughput and memory efficiency are particularly valuable for likelihood evaluation in high-dimensional generative models.

  4. The paper investigates several important design choices, including per-coordinate variance prediction, variance branch depth, and boundary parametrization. The additional bias-variance analysis provides useful evidence for understanding the estimator's empirical behavior.

Critical weaknesses
  1. Unlike Hutchinson estimation, which can be directly applied to pretrained velocity networks, COAL requires training an auxiliary variance prediction head.

  2. The method relies on the specific structure of Rectified Flow training, which may limit its applicability to other continuous normalizing flow models.

Questions for authors
  1. Could the authors provide a comparison of training overhead and inference-time savings, particularly for scenarios where likelihood evaluation is performed infrequently?

  2. Since the theoretical identity holds for the optimal velocity field, it would be interesting to investigate whether the likelihood estimation error changes significantly with the quality or convergence of the velocity model.

Reviewer_5eCFOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a method to compute the per-sample log-likelihood of a Rectified Flow model without computing the exact divergence and without the Hutchinson estimator. Using Tweedie's formula, the authors show that the divergence of the optimal velocity depends only on the trace of the conditional covariance of X_1 - X_0 given x_t. They train an extra variance head jointly with the velocity under a per-coordinate Gaussian NLL, and the sum of the predicted variances gives this trace. The likelihood can then be computed with one forward pass per ODE step. They claimed that COAL reaches the same per-sample BPD error as Hutchinson with about 3.5x less compute.

Critical strengths
  1. The method is simple. It only adds a small variance branch to the DiT backbone so it is easy to add to existing DiT architectures.
  2. At inference, the estimator is deterministic and needs only one forward pass per ODE step. It uses much less memory than Hutchinson.
  3. The evaluation is large-scale, a DiT-XL model on ImageNet-256, using all 50k validation images.
Critical weaknesses
  1. The novelty is smaller than the paper claims. The recent SCALLOP paper [1] already trains a divergence head with a target that needs no Jacobian and no Hutchinson probes, in the same setting. With Y = X_1 - X_0, its per-coordinate target uses X_{0,i}(v_i - Y_i), while COAL uses t(Y_i - v_i)^2. At the optimal velocity these two differ only by a zero-mean term, so they have the same conditional expectation. I understand that [1] may count as concurrent work, so I do not weigh this heavily. But it should be cited and discussed, and claims such as "the first deterministic, Jacobian-free per-sample likelihood estimator" should be softened.
  2. COAL estimates the divergence of the optimal velocity v*, not of the learned velocity v_theta that is integrated in the ODE. If v_theta differs from v*, the integrand is not div v_theta. I therefore expect an error floor that does not go away with more integration steps. The paper stops at N = 64. Could you show the error at N = 256, 512 and 1024?
  3. There is no FID for the DiT-XL model, so we cannot see the effect of COAL on sample quality. This matters because the joint Gaussian NLL reweights the velocity loss.
Questions for authors
  1. How does COAL differ from [1] in practice? For example, does the squared-residual target t(Y_i - v_i)^2 give lower variance or better accuracy than the cross-term target X_{0,i}(v_i - Y_i) of [1]?
  2. Could you report FID for the same DiT-XL model to show the sample quality? Or some image showcases?

[1] R. OuYang et al. Few-Step Boltzmann Generators via Scalable Likelihood Flow Maps. arXiv:2606.29110, 2026 (NeurIPS 2026).

2 comments

Thank you for the rebuttal! I will raise my score to weak accept (3).

Dear Reviewer 5eCF,

Thank you very much for your constructive discussions and the updated assessment. Your comments helped us better position COAL to related works. We will add the SCALLOP discussion and the new results to the revised paper.

We truly appreciate your careful review and wish you all the best in your research.

Best regards, Authors

Reviewer_JgXCOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes COAL, a method for estimating log-likelihoods in Rectified Flow models without computing Jacobian traces. It relates the divergence of the optimal velocity field to the conditional covariance of the velocity target, then learns this covariance through a per-coordinate variance head trained jointly with the velocity network. The method also uses boundary parametrization to stabilize evaluation near the data endpoint. Experiments on a 2D checkerboard and class-conditional ImageNet-256 show that COAL approximates a high-compute Hutchinson reference, with roughly 3.5× lower evaluation cost at matched likelihood estimation error on ImageNet.

Critical strengths
  1. The paper turns a conditional covariance identity into a practical likelihood estimator for Rectified Flows. Learning the covariance trace through a variance head avoids input-space differentiation during evaluation, offering a useful alternative to Hutchinson estimation.

  2. The ImageNet experiments demonstrate a meaningful efficiency gain at realistic scale: approximately 3.5× lower evaluation compute at matched error against a high-compute Hutchinson reference. Comparing both estimators on the same velocity model makes this a reasonably controlled test of the proposed estimator.

Critical weaknesses
  1. The practical motivation for per-image log-likelihood estimation could be clearer. The paper demonstrates faster evaluation, but does not show a downstream application that benefits from it. Why is this quantity useful in practice, and what does the proposed speedup enable? My background is primarily in practical diffusion models for generation and editing rather than theory, so I may be missing an established motivation. I apologize if this question overlooks something obvious; the authors may disregard it if so.

  2. Unlike Hutchinson estimation, COAL requires an additional variance branch and training. It is unclear how this affects generation quality: the shared backbone must support both tasks, and the Gaussian NLL also changes the velocity training objective. Please compare standard flow matching and COAL at matched total model size and training budget, reporting FID and likelihood estimation error. This would establish whether the evaluation speedup comes at the expense of generation quality or additional training cost.

  3. The method relies on the learned variance producing an accurate divergence estimate, but the reported likelihood error does not reveal where this approximation succeeds or fails. Please fix a velocity checkpoint and compare COAL against a high-precision Jacobian-trace reference along the same trajectories, plotting divergence bias and RMSE against time. This would test whether the estimator is accurate across noise levels, particularly near the data endpoint, and whether small final likelihood errors conceal cancellation between errors at different times.

Questions for authors
  1. Could you clarify the practical motivation for per-image log-likelihood estimation and which applications would benefit from COAL’s evaluation speedup?

  2. Could you compare standard flow matching and COAL at matched total model size and training budget, reporting FID and likelihood estimation error to assess the trade-off between generation quality and evaluation efficiency?

  3. For a fixed velocity checkpoint, could you plot divergence bias and RMSE against time along identical trajectories, using a high-precision Jacobian-trace reference? This would reveal whether errors concentrate near the data endpoint or cancel during integration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.