acceptodds
Under review as a conference paper at ICLR 2027

Your FID Cannot See the Tail: Auditing and Repairing the Rare-Event Calibration of Large Generative Models

17 views

Abstract

Generative models are increasingly used as data: synthetic corpora, simulators, and world models supply samples that downstream systems treat as draws from the real world. For these uses, the question is not only whether typical samples look good, but whether events of a chosen severity occur at the right rate. We study this question as tail calibration. Given a black-box sampler , a reference corpus , and a fixed severity functional , we audit the exceedance-rate ratio at named tail depths. We first show why this rate must be measured directly: FID-style finite-moment metrics can agree while tail mass changes, and any sample-based test pays a rare-event cost that scales with the inverse tail probability. The audit spends its sample budget on the tail indicator, reports intervals that combine exceedance-count error with threshold-estimation error, and measures its own error on real-versus-real splits before judging models. We also give a black-box wrapper that remaps generated samples to a calibrated severity law without retraining or weight access. The wrapper fixes exactly the severity law that the audit certifies, up to the audit's own measured error, and along shape coordinates that its transport preserves, no severity-keyed post-processor can produce conditional tail shapes beyond mixtures of the sampler's own. Across public image-generation samplers, two reference corpora, and multiple severity functionals, the certified ratios range from nearly sixfold under-production to more than fivefold over-production. Bulk-quality scores alone do not explain these errors, and the verdict can change when the severity functional changes. Even when exceedance rates are calibrated, conditional tail shapes remain compressed. These results suggest that samplers used as data should be reported with per-(configuration, functional) tail-calibration certificates, together with the audit's own measured error.

Primary area:
generative models
Keywords:
tail calibration, rare events, generative model evaluation, black-box recalibration, finite-sample certificates, synthetic data

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_D3SHOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies FID evaluation from a new perspective: whether the extreme samples generated by the model have the correct occurrence frequencies. The paper argues that FID cannot guarantee that the tail occurrence rates are consistent with real data, and proposes a tail calibration audit method to recalibrate large generative models. Experimental results demonstrate the effectiveness of the proposed method.

Critical strengths
  1. The research motivation of the paper is clear and has strong practical significance.

  2. The theoretical justification of the method in the paper is thorough, the method design exhibits good logical coherence, and the delineation of the method’s scope of applicability is also accurate.

  3. The experimental design is well-targeted and demonstrates the effectiveness of the proposed method.

Critical weaknesses

The motivation of this paper lies in the fact that when generative models provide samples for downstream systems, the accuracy of rare events is of great importance. One key design of the method design is the definition of severity functionals. However, no close connection has been established between these two; the severity functionals mainly consider statistical anomalies rather than the direct representation of extreme events in downstream systems. Considering that real-world extreme events are difficult to represent directly and accurately, it is suggested that the authors discuss the motivation behind the design of the severity functionals and clarify that this is not a scheme that can be fully and directly oriented toward downstream systems.

Questions for authors

There are no further questions.

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

This paper studies tail calibration in generative models, arguing that conventional metrics such as FID may overlook discrepancies in tail-event frequencies. The authors propose a statistical audit based on exceedance-rate ratios, along with a black-box recalibration wrapper that adjusts generated samples through quantile mapping and transport. Experiments across multiple image generators and severity functionals reveal substantial tail-frequency discrepancies, while recalibration reduces these discrepancies with minimal changes in FID. The authors also identify limitations in correcting conditional tail diversity.

Critical strengths

The paper provides a well-structured theoretical analysis of the limitations of conventional evaluation metrics and the sample complexity of rare-event detection. The proposed audit carefully accounts for sampling and threshold-estimation uncertainty, with additional self-validation experiments. The analysis of conditional tail diversity also helps clarify what the proposed recalibration procedure can and cannot guarantee.

Critical weaknesses

The practical relevance of the proposed evaluation framework is insufficiently justified. The paper motivates rare-event calibration through prior studies on safety-critical and application-specific rare events. However, the proposed severity functionals measure statistical atypicality in pixel or representation spaces, without validating whether these extremes correspond to the practically important rare events discussed in the motivating literature. This creates a fundamental gap between the motivation and the empirical evaluation. Furthermore, while the paper identifies statistically significant tail-frequency discrepancies, it does not demonstrate that these discrepancies affect downstream performance or model selection. Consequently, the practical necessity of introducing a separate evaluation criterion remains unclear.

Questions for authors
  • Is there evidence that the rare events identified by R1, R2, and R3 correspond to the practically important rare events discussed in the motivating literature, rather than merely statistical outliers?

  • Is there evidence that the observed tail-calibration discrepancies affect downstream performance or lead to different model-selection decisions compared with existing evaluation metrics?

  • What evidence supports the practical advantage of the proposed image-level transport over simpler alternatives such as rejection sampling or importance reweighting?

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

The paper looks at the ability of moment based metric such as FID to properly measure tail events. They do some basic theoretical analysis, propose an evaluation tool and a way to repair the calibration. They then test it on real images but with very based metrics for which they look at tail events.

Critical strengths
  • The question of proper evaluation of generative models, and specifically looking at tail events is an important question.
  • The statistical analysis they preform appears to be sound to the best of my knowledge.
Critical weaknesses
  • The recalibration wrapper is very limited as, unless we have basic (and not very interesting) tail events like the ones presented here it requires to do a very high level editing of the image which is far from trivial as the paper makes it to be
  • The theory doesn't really say anything interesting. Everything is a direct consequence of assumption 1 which is too strong to make the following work of any serious value. If you start that there exists two distributions with exactly the same mean statistics but extreme opposite support on the tail distribution then of course anything you compute from the mean values cannot give you any information about the tail.
  • If you ignore the wrapper and theory (which I explained why they are very weak), you are left with a basic statistical analysis that doesn't add any novelty and is not a strogn enough contribution to carry this paper.
Questions for authors
  • How do you make sure the transport produces natural images and not some weird artifact that matches the tail statistics?
  • Why not to resample to reweighing the generated images?

Comments on the paper

2 comments

I was drawn to this manuscript with strong interest due to its overlap with my own current ongoing research in generative model assessment. I have not examined the technical aspects of this manuscript carefully and so my comments do not regard the technical strengths or weaknesses of this paper. The referees and the author have had very insightful discussions about the pros and cons of the proposed strategy, which I learned a great deal from and won’t comment upon.

I do believe the topic under consideration is an important and timely one—assessing generative models and synthetic data using aggregate scores such as the FID is helpful to some extent, but it lacks the capability of providing more detailed insights to help understand the adequacy of the synthetic data and generative model for a given downstream application.

This particular manuscript considers one such scenario, when the downstream application requires accurate generation of “tail” or “extreme” events. Global metrics such as the FID generally characterize the overall distance between the real and synthetic data. As such, models that are excellent as measured by FID, could misrepresent application-relevant characteristics such as the tail behaviors. This paper proposes one possible method to remedy this issue, which has its own strengths and weaknesses.

More broadly, I believe application-driven assessment of generative models that goes beyond using standard scores such as the FID is a very important area that deserves more attention and effort from our community.

Thank you for taking the time to read the discussion and for your comment.

We agree with your broader point. A global score such as FID is a useful summary of overall distance, but it does not tell a user whether synthetic data are adequate for a particular downstream use. Our paper studies one such use, where the application needs rare events at the right frequency, and we see it as one instance of the larger question you describe.

This is why the audit takes the severity functional RR as an input from the application and reports each result together with its RR. In our experiments the verdict for the same sampler changes with the event that is specified, so an assessment has to be tied to the intended use. We also agree that our proposal is one possible approach with its own strengths and weaknesses; the discussion with the reviewers has helped us state its scope more precisely.

Your comment also drew our attention to your work on distributional evaluation via the relative density ratio (arXiv:2510.25507). Two aspects stood out to us. First, the ratio is evaluated at the sample level, so it shows where in the data space a generator over- or under-represents the real distribution, for example the under-represented digits in the MNIST study. A single score cannot give this. Second, relating the ratio to covariates, as in the attribute regression on CelebA, expresses the difference in terms that a user of the data cares about. Your conclusion that generators producing realistic samples can still be biased at the distributional level matches what we observe for tail events among samplers with similar FID.

We see the two views as complementary. A sample-level ratio describes where real and generated data differ without naming an event in advance. An audit answers a narrower question: for an event named in advance, how often it occurs relative to real data, with an interval. A natural combination would be to use the first to suggest application-relevant events, then fix their definitions and audit them on independent data.

Thank you again for the encouragement to pursue application-driven assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.