acceptodds
Under review as a conference paper at ICLR 2027

Normal Fréchet Drift: Measuring Post-Training Departure from Pretrained Reference Manifold

Abstract

Post-training can improve a generator’s reward while moving its outputs away from pretrained reference geometry. We distinguish this normal drift from tangential drift, which redistributes probability along the same reference manifold. We introduce Normal Fréchet Drift (NFD), the expected squared distance to a fixed reference set in a shared feature space. Its ideal value is zero for distributions confined to that set. To estimate NFD from finite pretrained samples, we construct bounded Multi-Patch reconstructions: local plane patches centered at sampled features, with finite radii that control extrapolation. We separate evaluation sampling variance from reconstruction bias and bound the bias through local coverage, curvature, tangent, and anchor errors. Controlled geometries validate the distinction between normal and tangential drift and reveal the reconstruction-resolution limit. On Stable Diffusion 3.5 Medium, FlowDPO exhibits the largest reconstruction-relative departure, with stable ordering across prompt splits, encoders, and reference refits. A FLUX.2 comparison extends the diagnostic to distillation. Blinded judgments on selected image pairs associate lower NFD with better visual quality. NFD measures geometric departure alongside distributional and perceptual evaluations.

Primary area:
generative models
Keywords:
generative model evaluation, manifold drift, post-training, preference optimization, diffusion models, text-to-image generation, reward hacking

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_niWROpenReview ↗
Rating
3(+2)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2(+1)
Read review
Paper summary

Final justification after rebuttal


I appreciate the authors' efforts and detailed responses during the rebuttal.

I now believe that NFD offers a meaningful contribution as a diagnostic metric for measuring departures from pretrained reference geometry, particularly given its demonstrated value beyond existing metrics.

I also appreciate the authors' explicit acknowledgement that the causal relationship between normal drift and quality degradation remains to be established. I hope the authors will clearly communicate this limitation in the final manuscript to avoid potential misinterpretation by readers.

I trust that the authors will incorporate these new findings and their limitations into the revised manuscript (I hope to see it before the end of the discussion period).

Thank you again for your efforts.


This paper introduces NFD, which measures the normal drift of post-trained diffusion models (e.g., after DPO) to distinguish off-manifold drift (i.e., normal drift) from on-manifold drift (i.e., tangential drift), beyond what FID can distinguish.

Unlike FID, which gives the same value for equal-magnitude tangential and normal shifts in the paper's line-translation example, NFD measures normal departure while remaining zero for ideal tangential drift.

Since the exact reference manifold is unknown, the authors approximate it from finite pretrained samples using anchor-centered bounded local linear patches (conceptually similar to piecewise-linear approximations). Due to this approximation, they also report a nonzero NFD for the Base model.

What they find is that (i) the NFD-based model ranking is consistent across different prompt splits and etc, and (ii) NFD has some correlation with preference and aesthetic scores (e.g., Spearman ρ = 0.195 for ImageReward).

They also provide a set of qualitative results suggesting that overall visual quality decreases as NFD increases, with OCR and PickScore (PS) reported alongside the images.

Critical strengths

Critical Strength / Problem Formulation

  • The problem formulation itself is interesting, since I also feel that FID cannot properly represent overall generation quality (e.g., for large models such as FLUX, whose training data are not fully disclosed; calculating FID against ImageNet or other publicly available datasets may not be very meaningful when the reference distributions are mismatched).
  • Of course, many alternative measures have been proposed to address this issue (e.g., Jayasumana et al., 2024). Still, I think this kind of manifold-wise approach itself has great potential worth exploring.
Critical weaknesses

Critical Weakness / Soundness

TLDR: Overall, the experiments cannot fully support the paper's main claim.

Before discussing the experiments, let me reiterate how I understood the main problem formulation of this manuscript.

  1. FID cannot properly distinguish tangential drift from normal drift in certain cases.
  2. Tangential drift redistributes probability mass along the pretrained M\mathcal M, which could be a desirable effect.
  3. However, normal drift moves samples away from the pretrained M\mathcal M, which may lead to undesirable artifacts (e.g., reward hacking).
  4. Therefore, we need NFD.

At this point, I generally agree with this assumption. However, since the paper's main contribution relies on distinguishing these two types of drift, I think at least a controlled experiment showing the differences between tangential and normal drift under similar levels of FID discrepancy is necessary, especially for actual diffusion-generated images.

For example, I would expect a comparison along the following lines:

MetricBaseTangential-dominatedNormal-dominated
FIDReferenceMatchedMatched
NFDReferenceReportReport
ImageRewardReferenceReportReport
CLIPScoreReferenceReportReport
Diversity / RecallReferenceReportReport
Human visual qualityReferenceReportReport

Here, the two types of drift should be controlled or identified independently of the proposed NFD.

  • Furthermore, looking at Table 10, I find it somewhat unclear what additional benefits NFD provides over existing metrics. The reported Spearman rank correlations between NFD and MMD$^2$, and etc are all 1.0, indicating that these metrics produce identical model rankings in this experiment. Given that the paper motivates NFD by highlighting the limitations of existing metrics, I believe it would be important to demonstrate what additional information or insights NFD can provide.

  • Additionally, I am not fully convinced that Tables 11 and 12 sufficiently support the manuscript's main claim. Table 11 reports correlations between NFD and preference/AES scores, while Table 12 presents human evaluations on selected FlowDPO-related pairs. Although these results suggest an association between NFD and image quality, I think further evidence is needed to establish why distinguishing normal drift from tangential drift is particularly important.


Critical Weakness / Insufficient Controlled Ablation

TLDR: At least the effects of feature normalization and norm-related artifacts should be examined.

Looking at Figure 6 and other qualitative results, some high-NFD images seem to exhibit oversaturation or distortion similar to the artifacts discussed in previous diffusion guidance studies, e.g., Sadat et al., 2025 and Jin et al., 2025.

In particular, Jin et al. (2025) discussed how latent norm amplification can lead to color distortions under strong guidance. Although the setting is different, I wonder whether similar norm-related effects might contribute to some of the observed qualitative results.

However, I could not find an ablation comparing NFD with and without explicit L2 normalization of the projected features, or an analysis of the relationship between the generated latent norm and NFD.

Since NFD directly measures Euclidean distances in feature space, I think it would be helpful to clarify the normalization procedure and, if possible, provide additional ablations to better understand the potential influence of magnitude-related effects on the reported results.

Questions for authors

Q1. I think this kind of measurement could also be applicable to inference-time guidance methods, since the evaluation only relies on generated samples and a fixed pretrained reference. Thus, it could be meaningful to evaluate other manifold-related guidance methods, such as Sun et al., 2023 and He et al., 2024, and could provide further insights into the role of manifolds in diffusion models.

Could the authors clarify whether NFD can be directly applied to these guidance methods, and whether there are any results demonstrating its advantages over existing metrics in such settings?


Q2. My main concern, as outlined in the Weaknesses section, is that the problem formulation itself is interesting, but the resulting experiments do not sufficiently demonstrate the practical importance of distinguishing normal drift from tangential drift. Consequently, I remain uncertain about the additional value of NFD beyond existing metrics.


Could the authors clarify whether I have misunderstood any aspect of the results or their implications? I would appreciate a clearer explanation of what NFD reveals that existing metrics cannot capture, particularly in actual diffusion-generated images.

I am open to the possibility that I have underestimated the value of this paper. A convincing clarification of its practical significance and the main experimental findings in the rebuttal could positively affect my assessment, and I would be willing to reconsider my score accordingly.

Reviewer_e48ROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

While post-training can improve a generator's reward, it can push the samples out of the pretrained sample manifold. The authors propose NFD (Normal Frechet Drift), which is the expected squared distance of the post-trained samples from the pretrained manifold, to measure this discrepancy. To compute NFD with finite samples, the pretrained manifold is reconstructed as a union of bounded local patches for each condition cc. NFD computed on the reconstruction is proven to be within an error bound with the true NFD value.

The experiments show that the proposed method to compute NFD is valid under different simple geometries. NFD provides a comparison of off-manifold drift among post-trained checkpoints, and the checkpoints demonstrate a bigger drift compared to held out pre-trained samples. This finding stays robust with the choice of feature encoder or cc. Furthermore, NFD agrees with existing metrics and correlates with perceptual quality measured by human raters. Therefore NFD establishes a geometric method of quantifying off-manifold drift of the post-trained model.

Critical strengths
  • S1. The paper establishes a new method of measuring off-manifold drift of post-trained checkpoints.

Post-training or distillation of models can often lead to loss of quality or hallucination, which can be modeled as a off-manifold drift. The paper suggests a way to quantify this drift, allowing model developers to measure how much error that post-training have introduced.

  • S2. The paper provides a proof for error bounds of the proposed estimator of NFD.

Reconstructing a manifold from finite samples can lead to various kinds of errors, which can root from curvature and other reasons. The authors present an error bound of the NFD measured with a reconstructed manifold, which allows model developers to better rely on the computed NFD value.

Critical weaknesses
  • W1. The presentation of the paper is rather hard to follow at first read.

For example, does "ideal" mean the continuous true manifold whereas "reference" mean "sampled from the pretrained model" (line 110)? Why does ψ\psi have to be Π∘ϕ\Pi \circ \phi (line 111)? Why is a post-trained checkpoint first written θ\theta (line 111) and then ll (line 235)? Why should an error bound be defined as a "root expected sum of its squared upper- and lower-error bounds" (line 256)? While the appendices cover the definition and details, it is hard to follow the intuition behind the developed mathematical framework. I recommend the authors to clearly define important terminology before introducing new ideas, and explain why each new idea is required.

  • W2. Statistical power of Spearman's rho is low.

To check the robustness of NFD, the authors compare six model checkpoints and test if NFD variants rank them differently (Section 4.3). However, the statistical power of Spearman's rho is low with only six samples. In particular, there is a 1/7201/720 chance that a random ranking gives ρ=1\rho=1.

  • W3. Previous work on reconstructing manifolds.

Reconstructing a manifold with multiple local bounded patches is a recurring concept in the field of machine learning [1, 2]. The authors should cite previous work and show how the suggested MP reconstruction coincides or differs from them.

[1] Allard, William K., Guangliang Chen, and Mauro Maggioni. "Multi-scale geometric methods for data sets II: Geometric multi-resolution analysis." Applied and computational harmonic analysis 32.3 (2012): 435-462.

[2] Yao, Zhigang, and Jiaji Su. "Manifold Fitting: A Review of Methods and Applications." arXiv preprint arXiv:2606.22356 (2026).

Questions for authors
  • Q1. Quality of reconstruction is not sufficiently explained.

In real use cases, the features live in a high-dimensional ambient space; for example ambient dimension D=2048D=2048 for InceptionV3 features used in FID. Therefore, showing the validity of NFD estimation at low dimensional example manifolds (Section 4.1) does not show the validity at real feature spaces. Are there enough samples to define the reconstructed manifold for each condition cc, and what is the cost to obtain such samples? How much variance is lost when PCA is done to extract d=16d=16 dimensions (line 353)? What is the intrinsic dimension of the manifolds?

  • Q2. The advantage of NFD over existing metrics is not clear.

While the authors show that NFD agrees with existing metrics (Section 4.4), it is not clear when NFD captures something that existing metrics cannot. If NFD always agrees with existing metrics, there is no advantage of using a complicated reconstruction and estimation algorithm. Is there a real example in which NFD distinguishes normal drift and tangential drift while FID cannot? This should help convince model developers to use NFD in their work.

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

This paper introduces Normal Frechet Drift NFD. The goal of this metric is to measure drift from the retrained manifold when doing fine-tuning. It’s defined as the distance to the nearest pretrained point.

Critical strengths
  • Motivation seems pretty important. The goal of post training is to improve the quality of the output why not going away too much from the pertaining
  • The estimator construction makes sense to me
Critical weaknesses
  • I feel that this paper lacks empirical evidence to asses they claimes. For example, the user study says it correlate highly with NFD. But there’s a huge bias there, since it’s always the same methods against others (FLowDPO vs better methods). Always choosing the other methods should give you good correlation as well. I expect to
    • 1: Have different methods pairs / different models / diversity of prompts
    • 2: Compare the correlation of other metrics to this user study.
  • My feeling is that FID between the pretrained images and post-trained ones should yield similar results. Comparaison is done from a theoretical point of view but not a practical one. Same thing could be done with MMD, MIND or any distributional distance between the 2 sets and I think it should be studied
  • The transport characterization in Proposition 1 closely matches Cañas and Rosasco (2012, Lemmas 3.1–3.2), which should be cited and discussed.
Questions for authors
  • Can you fix the empirical evidence for the claims (user study)
  • Show empirically why NFD is better than the existing metrics if compted with the pretrained set and the post trained set.
  • My feeling is that NFD is a measure of reward hacking. If the authors manage to use it as a training regularizer on some overfitting rewards (Aesthetics Scores tend to easily overfit) this would give a new dimension to the paper

Comments on the paper

1 comment
Area_Chair_Ydu2OpenReview ↗

Author-Reviewer Discussion

Dear Reviewers,

Thank you for your time and thoughtful reviews. With the author–reviewer discussion period ending on November 18 (AoE), please review the authors' rebuttal and share your updated assessment. Specifically, let us know whether your main concerns have been resolved, and if not, which points remain open.

Your input before the deadline is greatly appreciated!

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.