acceptodds
Under review as a conference paper at ICLR 2027

Standard Evaluation Metrics for Generative Image Models Overlook Hallucinations

Abstract

This paper reveals an important problem with the standard evaluation metrics used for image generation: they cannot distinguish between two kinds of flawed samples, poor inliers and hallucinations. Poor inliers preserve the semantic properties that define the target distribution but are rendered imperfectly, with defects such as blur or warped edges. Hallucinations violate those properties, while often remaining visually plausible. The distinction is semantic rather than visual, and it significantly affects how the model should be improved. We show that the scores of the standard metrics are consistent with many combinations of the two failures, so a model that hallucinates more but produces better inliers can score the same as, or even higher than, one that hallucinates less. Having identified the problem, we analyse what determines how strongly these metrics respond to hallucinations relative to inlier quality. Experiments in controlled settings and on samples from trained generative models show that hallucinations largely go unnoticed, because the feature extractors the metrics rely on often embed them close to inliers. Even when hallucinations are embedded far from inliers, averaging over all samples hides them, while a statistic that reports only the most atypical samples recovers them and ranks models differently from the standard metrics.

Primary area:
generative models
Keywords:
evaluation, generative image models, hallucinations

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_4VPBOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper reveals a fundamental limitation in the standard evaluation metrics used for generative image models, showing that they cannot distinguish between two distinct sources of image flaws: poor inliers and hallucinations. Poor inliers are images that preserve the semantic properties of the target distribution but are rendered poorly, whereas hallucinations are images that violate those semantic properties while remaining visually plausible. The authors mathematically formulate these two sources of error by modeling the generated distribution as a mixture of degraded inliers and hallucinations, which introduces a parameter measuring the degradation level of the inliers and another parameter representing the fraction of hallucinations.

The authors show that standard feature extractors fail to represent semantic deviations and that standard metrics, which rely on averaging scores over all samples, dilute and fail to identify these errors. Consequently, a single evaluation score is consistent with many different combinations of inlier degradation and hallucination rates. This means models cannot be properly compared from a single score point, but rather require analyzing how the metric scores change across varying levels of degradation and hallucination to understand the exchanges made by the metric.

To validate their theoretical framework, the authors conduct controlled experiments by manually varying the levels of inlier degradation and hallucination, and they perform evaluations on real generative models. In both setups, standard evaluation metrics fail to distinguish between the two errors, often failing to penalize or even rewarding hallucinations, which ultimately leads to an incorrect ranking of different generative models. Finally, the authors propose a new measure based on a tail statistic that averages only the most atypical samples. This method successfully separates the two failure modes and provides a reliable way to detect hallucinations and rank models correctly.

Critical strengths

The paper is well written, organized, and clear.

  • The paper sheds light on two distinct sources of errors and their mixture, providing a clear and mathematically rigorous formulation of the combination of hallucinations and poor inliers. This provides a clean and elegant structure to the problem upon which the research community can easily build.

  • The authors provide a theoretical analysis of the limits of standard metrics, using total variation distance to bound the score gap and the score drift. This mathematical proof elegantly shows how the feature extractor itself limits the signal before any scorer is even chosen, which represents a nice theoretical contribution.

  • The experiments adequately demonstrate the described phenomenon and provide strong support for the theory presented. The authors validate their claims using both controlled simulations and samples from real generative models, ensuring that the theoretical predictions translate to practical scenarios.

  • The authors validate their theoretical findings and rankings with human evaluation, proving that the proposed tail statistic aligns with human perception and successfully corrects the ranking inversions caused by standard evaluation metrics.

  • Most importantly, the authors provide a new evaluation measure based on a tail statistic that provides a viable remedy to the discussed problem. This successfully turns a theoretical insight into a highly practical tool for the community.

Critical weaknesses
  • The proposed metric is highly sensitive to the choice of the tail fraction threshold, and the paper lacks an automated or systematic method to select this parameter.

  • The proposed tail statistic does not address the fundamental representation bottleneck of existing feature extractors. The authors mathematically prove and empirically demonstrate that if the underlying feature representations do not capture a semantic deviation, no anomaly detector or tail statistic can recover it. As a result, well rendered semantic hallucinations still go completely unnoticed, meaning the proposed solution remains highly dependent on the quality of existing feature representations.

  • The paper lacks a dedicated limitations section. A detailed discussion outlining the boundary conditions of the proposed tail statistic, such as its sensitivity to sample size, behavior under severe class imbalance, or performance on conditional image generation tasks, is highly important for a complete scientific contribution.

Questions for authors
  1. Can the theoretical framework and the proposed tail statistic presented in this work be extended beyond evaluation to active training paradigms, such as serving as reward signals in reinforcement learning post training? A brief discussion on how the metric could guide model training would highly enrich the paper.

  2. The paper models poor inliers and hallucinations as two distinct and independent sources of error. However, is it possible that these failure modes are fundamentally coupled? For instance, when a generative model lacks the precise semantic information required for perfect fidelity, it may face a trade-off where it must either output a blurred or degraded image that remains in-distribution, or alternatively produce a sharp, visually plausible image that contains a hallucination. This suggests a potential latent variable or shared underlying cause explaining both error types, and the authors are encouraged to discuss this relationship

3 comments
AuthorsOpenReview ↗

Response to reviewer 4VPB

We thank the reviewer for the comments. We have uploaded a revised version of the paper and answer the questions below.

Weaknesses 1–3 (tail fraction, representation bottleneck, limitations). All three weaknesses concern the tail statistic, so we first clarify its role. We do not propose the tail as a metric. It demonstrates what our analysis predicts for the aggregation stage. Once the extractor and the scorer rank hallucinations above the inliers, averaging only the tail of the scores removes the dilution of the mean (Section 5.1.3). On trained generators, the tail then reverses the ranking of the standard metrics, as annotators confirm (Section 5.2.2). The limits the reviewer identifies follow from our analysis, and they are why we present the tail as a demonstration.

  • Tail fraction. The tail separates the two failures only when τ≥α\tau \geq \alpha (Appendix D). A narrower tail holds only hallucinations and stops measuring inlier quality. A wider one dilutes them again, and as τ\tau grows the exchange rate falls steadily towards that of the mean (Fig. 10). Since α\alpha is the quantity being measured, τ\tau cannot be set from it.
  • Representation bottleneck. We agree, and this is Proposition 2: no scorer and no aggregation can recover a hallucination that the extractor does not represent. Section 5.2.1 confirms this for wrong finger counts. The bound holds for every metric of the form of Eq. 1, the tail included.
  • Limitations. We added a new section, Implications for Metric Design (Section 6). It states the requirements a metric must meet to separate the two failures, and where current components fall short: no extractor we tested represents fine semantic deviations, and unsupervised scorers can miss deviations the features carry. The tail is a demonstration of the aggregation stage rather than a proposed metric, so we do not study its behaviour under sample size, class imbalance or conditional generation. These would matter for a metric that meets the requirements of Section 6.

We would be grateful if the reviewer could read the new Section 6, which we believe addresses these points directly.

Q1. Using the metric as a reward signal. Using such a metric during training lies beyond the scope of this paper, but our analysis has a direct implication. A metric that meets the requirements of the new Section 6 would separate the hallucination rate from inlier quality, and could therefore serve as a training signal that reduces hallucinations without trading them for inlier quality. The current standard metrics cannot serve this role, since a model could improve its score by improving its inliers while hallucinating more (Prop. 1).

Q2. Coupling of the two failures. Whether the two failures are coupled is a question about how generators produce them, which our paper does not study; our analysis concerns how evaluation metrics respond to them. We treat poor inliers and hallucinations as distinct kinds of output, not as products of independent causes. Eq. 2 describes the generated distribution, not the process that produces its errors. It splits the generated samples by semantics into an inlier and a hallucination component. α\alpha and δ\delta vary independently only in the sense that α\alpha is a mixture weight and δ\delta moves the inlier component, so each can be set without changing the other. This is what lets us measure how a metric responds to hallucinations at fixed inlier quality.

Studying this coupling would be a different analysis, and one that depends on ours. Suppose we ask whether a training or sampling choice improves inlier quality, lowers the hallucination rate, or trades one for the other. Answering that requires measuring the two separately, which the standard metrics cannot do (Prop. 1). Such a study would therefore need metrics that decouple the two failures, along the requirements of Section 6.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer 4VPB,

Thank you for your positive and constructive review. As the discussion period ends on November 18, we wanted to check whether our response on the choice of the tail fraction, the limitations of the analysis, and the possible coupling between the two failure modes answered your questions. We are happy to provide any further clarification.

I thank the authors for their clarifications and the changes made for the manuscript, addressing most of my concerns. I enjoy reading the paper and learning from it. I believe the paper provides a good contribution to the research community. Hence, I raised my score.

Reviewer_gc5eOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The key problem this paper proposes is that there is a problem in the existing literature on evaluation methods. Specifically, current evaluation metrics cannot distinguish between poor inliers and hallucinations. The authors point out that existing works focus on outliers and that generated outliers are a specific case of hallucinations. The authors provide a framing where they model the generated distribution as a combination of inlier components with hallucinations. The authors provide some theoretical results showcasing in ideal situations that when assuming a metric punishes hallucinations more than inliers this corresponds to a whole curve of pairs (hallucination fraction, inlier quality). They then introduce a proposition to bound the gap and the drift by TV distance in feature space. The authors run experiments with a set of hallucinations and poor inliers which are constructed. They compare to FD, MMD, Precision & Recall, and Density & Coverage. They claim the methods they test on exhibit the three failure regimes. They also have experiments with real outputs and argue they witness the same effect.

Critical strengths
  1. The problem is important and it is well motivated. It is known that existing evaluation methods for generative models suffer from several failure modes. From the angle of hallucinations this is a relatively novel view.
  2. The experiments show the failure regimes in a controlled setting which is useful. The real results are also insightful of the failure mode of existing methods.
  3. The theoretical results highlight the problems of existing methods.
Critical weaknesses
  1. The biggest argument is that this problem is not completely unknown to the literature [1]. This work talks about how FID correlates weakly with counting hallucinations but strongly with other failures. Also, the existing Precision and Recall Paper ([2]) uses a very similar formalism to that introduced in section 3.2 (eq 2). Although it is adapted to show the problem of hallucinations it is a more incremental contribution. 2.The authors have run experiments and comparisons on more dated evaluation metrics. There are numerous new methods the authors have not compared against that are also relevant. There are many knn based methods that need attention as well. If they are not relevant they should be addressed. See [3] and [4]
  2. The experiments are insightful and show promise; however, the authors only use one form of degradation which is elastic distortion. To further support the claims more forms of degradation could be added.

[1] Fu, Shuai, et al. "Counting hallucinations in diffusion models." arXiv preprint arXiv:2510.13080 (2025).

[2] Sajjadi, Mehdi SM, et al. "Assessing generative models via precision and recall." Advances in neural information processing systems 31 (2018).

[3] Kim, Pum Jun, et al. "TopP&R: Robust support estimation approach for evaluating fidelity and diversity in generative models." Advances in Neural Information Processing Systems 36 (2023): 7831-7866.

[4] Alaa, Ahmed, et al. "How faithful is your synthetic data? sample-level metrics for evaluating and auditing generative models." International conference on machine learning. PMLR, 2022.

Questions for authors

I believe the work is solid and introduces the work from a relatively novel angle. If the authors can provide some rebuttal or address my weaknesses i would be willing to adjust my score. Namely, addressing the concerns in the weaknesses and answering the questions:

  1. How do the KNN methods behave on your grid if they are applicable?
  2. How do you position yourself relative to Fu et al. (2025)?
  3. what would the experimental results look like under different types of degradation such as noise, blurring, or JPEG compression?
3 comments
AuthorsOpenReview ↗

Response to reviewer gc53

We thank the reviewer for the comments. We have uploaded a revised version of the paper and answer the questions below.

Q1. More recent metrics. Our analysis uses FD, MMD and PRDC because they are the most widely used metrics in image generation, and because the two lines of work we build on report results with them (Section 2). We ran TopP&R and α-Precision/β-Recall on our grid in all three embedding spaces, for the elastic distortion and the three degradations the reviewer proposed (see Q3), and report the results in Appendix C.4. They show the same patterns as the metrics in the main paper. We also added both in the revised Related Work (Section 2).

Q2. Positioning relative to Fu et al. (2025). The two papers take different positions and reach different conclusions.

Position. Fu et al. is a benchmark paper: it proposes datasets and a protocol for quantifying counting hallucinations in diffusion models and studies how sampling settings affect them. Its finding on evaluation is empirical: the correlation between FID and the hallucination rate is positive on one dataset and negative on another. Our paper studies the metrics themselves and explains why they fail to register hallucinations.

Outcomes. Our two central conclusions go beyond what this correlation can show:

  • A positive correlation would not make FID a hallucination measure. Fu et al. observe a positive correlation on SimObject and a negative one on RealHand (Fu et al., Table 2). Even if it had been positive on every dataset, as on SimObject, so that settings with more hallucinations always had a worse FID, a single FID value would still not tell how much a model hallucinates. The score conflates hallucinations with inlier quality: more hallucinations can be traded for better inliers at the same value (our Section 3.2, Prop. 1).
  • The correlation cannot measure how FID responds to hallucinations. In Fu et al.'s setup, changing the sampler changes two things at once: the hallucination rate (our α\alpha) and the quality of the remaining samples, for which the non-counting failure rate is a proxy (our δ\delta; our Section 3.2, Eq. 2). On RealHand, the settings with fewer non-counting failures (18% → 7%) also have more counting hallucinations (13% → 16%; Fu et al., Table 1). FID improves with these settings because these settings draw better hands (Fu et al., Table 2). FID would improve just the same if it ignored hallucinations completely, so the negative correlation says nothing about how FID responds to hallucinations. Measuring that response requires changing α\alpha while keeping δ\delta fixed. Our analysis does exactly this and shows that a metric can penalise, ignore or even reward hallucinations, and when each occurs (our Section 4, Prop. 2; Section 5.1.1, Fig. 3).

Q3. Other degradations. We repeated the full grid under Gaussian noise, Gaussian blur and JPEG compression, alongside the elastic distortion of the main text (new Appendix C.2–C.3). We also measured how far each degradation moves the embeddings of the three extractors: every one registers substantially, even when the image looks almost unchanged, and DINOv2 is the least sensitive. Across all degradations, metrics and extractors, the exchange rates fall into the same four categories as in the main text. The rare cases where a score follows hallucinations almost alone arise where the extractor barely registers the degradation, as Prop. 2 predicts. No combination isolates the hallucination rate, so the conclusions of Section 5.1 hold beyond elastic distortion.

Relation to Sajjadi et al. (2018). Our parameterization splits the generated distribution differently from Sajjadi et al., and this difference is what our paper studies. They decompose both the reference and the generated distribution by support. We decompose only the generated distribution, by semantics: a badly rendered inlier stays in PδP_\delta even when it leaves the reference support, and a well-rendered hallucination stays in QQ even when it falls inside it. The two failures also vary independently, because α\alpha is a mixture weight while δ\delta moves the inlier component itself.

In our parameters, the loss in precision of Sajjadi et al. is

1−αˉ=(1−α)Pδ(Sˉ)+αQ(Sˉ),1 - \bar{\alpha} = (1-\alpha) P_\delta(\bar{S}) + \alpha Q(\bar{S}),

where Sˉ\bar{S} is the complement of the reference support in the embedding space. This is Eq. 3 with a scorer that flags off-support samples, so our analysis applies directly. Such a split penalizes off-support degraded inliers and off-support hallucinations alike, so the same precision is reached by many pairs (α,δ)(\alpha, \delta) (Prop. 1). Hallucinations inside the reference support also count as precision, as our experiments show for Precision (Section 5.1).

Note that their QQ is the generated distribution, whereas ours is the hallucination component.

AuthorsOpenReview ↗

Follow-up on related work and additional experiments

Dear Reviewer gc5e,

Thank you for your detailed feedback. With the discussion period ending on November 18, we would like to ask whether our response resolved your concerns, in particular our positioning relative to Fu et al. (2025) and Sajjadi et al. (2018), the comparison with TopP&R and α-Precision, and the additional degradation types. If any point remains unclear, we would be glad to address it in the remaining time.

Apologies for the late reply. I wanted to thank the authors for taking the time to revise their work and also taking my questions and comments into account.

I feel the revised version and the answers they provided answers all of my questions and I am happy to change my rating from weak reject to weak acceptance. I hope to see this work get accepted as I believe it solves an important problem.

Reviewer_81EUOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper area of research is evaluation of image generative models. In particular, standard generative metrics can be broadly separated as measures of two quantities: fidelity (how realistic a sample looks) and coverage (how well the generated samples cover the real distribution space). The paper argues that fidelity itself can be split into two further failure modes:

  • poor inliers: samples that satisfy the semantic properties of the target distribution but are poorly rendered.
  • hallucinations: samples that do not satisfy one or more (even all) the semantic properties of the target distributions.

From these premises, the authors proceed to decompose the fidelity metric into two components (one per failure mode) and with further mathematical reasoning, they show that Fidelity alone (1 value) cannot properly distinguish between poor inliers and hallucinations, as a specific fidelity value can be obtained from different components of the two failures. Furthermore, based on the value of ρ\rho, the exchange rate of the metric between the two failures, it may happen that hallucinations improve generative scores. Finally, being severe hallucinations rare, the mean score over features dilutes them in typical metrics.

Experimentally, the paper validates its claim in a controlled generated scenario, and further on real generative models. The final recommendation is to report tail statistics to evidence hallucinations, while the main identified gaps in current metrics remain unaddressed.

Critical strengths

Overall, the paper is very well written and it is truly interesting to follow it, finding out where the derivations lead and their conclusions.
The mathematical derivations provide a structural contribution and are relevant to better understand fidelity scores in generative metrics. The identified gaps are sound and well describe the current limitations of the existing scores, being also empirically supported by the experimental analysis.

More specifically, once set up that fidelity can be affected by two different failure modes, it becomes clear that one score cannot separate them faithfully. The true additional value of the paper is that the mathematical derivations lead to interesting conclusions. In particular, I find relevant the fact that depending on the "exchange rate" of the metric, it may happen that hallucinations are rewarded.

Critical weaknesses

The main weakness of the paper is that it stops right before the most critical and relevant question: how can we design better fidelity metrics, that are truly able to discriminate between the failure modes and improve evaluation of generative metrics?

In the conclusion, the authors limit to suggest the report of a tail statistics, which solves the problem of diluted scores over sparse hallucinations. However, clear indications on how to design better metrics in the future, or a better metric itself are not discussed in the paper. This directly impacts the relevance of the paper itself, which becomes a "diagnostic analysis" with no clear proposed solutions.

Questions for authors

Clarification requests
What do the authors consider to be the paper contributions, ordered by importance/relevance?
Have the authors a clear and sound answer to the main doubt identified in weaknesses?

A convincing answer to the second question is what can lead to a score increase from my side.

3 comments
AuthorsOpenReview ↗

Response to reviewer 81EU

We thank the reviewer for the comments. We have uploaded a revised version of the paper and answer the questions below.

Q1. Contributions, in order of importance.

1. Standard metrics conflate two different failures. We show that the standard fidelity metrics react both to poor inliers, samples that keep the semantics of the target but are rendered badly, and to hallucinations, samples that violate those semantics. The distinction is semantic, not visual. We show that the metrics respond to both, with inlier degradation modelled in four ways that mimic how real generative models lose fidelity (elastic distortion, plus three added in the revision: additive Gaussian noise, Gaussian blur and JPEG compression). A single score is therefore underdetermined: many combinations of hallucination rate and inlier quality give the same value (Prop. 1). This matters because two lines of work, outlier reduction and hallucination mitigation (Section 2), report progress with exactly these metrics.

2. Standard metrics as aggregated anomaly scores. We break every standard metric down into three choices: a feature extractor, an anomaly scorer fitted on the reference, and a statistic that aggregates the scores (Eq. 1). We then locate a failure at each stage: the extractor may not represent hallucinations, the scorer may not read them, and the mean dilutes them (Prop. 2, Sections 5.1.2–5.1.3, Section 5.2.1). To our knowledge, this framing of evaluation metrics as aggregations over an anomaly scoring function is new, and it gives a concrete starting point for new evaluation methods.

3. Hallucinations can lead to unwanted high scores. We show, theoretically and empirically, that hallucinations are not only overlooked. Depending on the sign of the exchange rate, adding hallucinations can improve a metric (Eqs. 4–5, Section 5.1.1).

4. A condition under which the two failures separate. Although the paper is diagnostic, it identifies when a metric can isolate the hallucination rate. If the extractor distinguishes hallucinations and the scorer ranks them above the degraded inliers, then a statistic other than the mean, such as the mean of the score tail, separates the hallucination rate from inlier quality (Section 5.1.3). On trained generators this holds for StyleGAN2's degenerate samples, and the tail reverses the ranking the standard metrics give, as annotators confirm (Section 5.2.2).

Q2. Guidance for designing better metrics.

We added a new section, Implications for Metric Design (Section 6), that turns our findings into requirements for a metric that discriminates between the failure modes. Since a single value cannot identify both failures (Prop. 1), evaluation needs three axes rather than two: the hallucination rate, inlier quality and diversity. A metric that separates them should:

(1) measure the hallucination rate in a representation that responds to semantic deviations much more than to rendering errors (a large G(0)G(0) and a small DD, Prop. 2);
(2) measure inlier quality in a separate representation that retains low-level information;
(3) detect hallucinations without examples of them and aggregate the scores without diluting them, which a tail statistic achieves once the scorer ranks hallucinations above the inliers (Section 5.1.3);
(4) remain reproducible.

The first requirement draws on a new analysis in Appendix C, where we measure how far four rendering errors move the embeddings of each extractor. DINOv2 is the least sensitive of the three, which this requirement asks for. It still fails to represent fine semantic deviations such as a wrong number of fingers or objects (Section 5.2.1), so no extractor we tested meets the requirement in full.

We would be grateful if the reviewer could read Section 6, which we believe directly addresses the main concern of the review.

AuthorsOpenReview ↗

Follow-up on metric design question

Dear Reviewer 81EU,

Thank you again for your thoughtful review. As the discussion period closes on November 18, we wanted to check whether our response addressed your main question, on what our analysis implies for designing better fidelity metrics. We would be happy to clarify anything further or discuss any remaining concerns before the deadline.

Dear Authors,

thank you for the rebuttal and for having included a new section in the paper. It directly addresses my main concern, and I therefore have no further major weaknesses to raise. In light of this, I am considering a score increase. I will however wait for the Reviewers-AC discussion period before giving my final assessment, in the hope to reach a common consensus with the remaining Reviewers. Thank you again for your work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We thank all reviewers for their careful reading and constructive feedback. We have uploaded a revised version of the paper, with all changes highlighted in blue. Below we summarize the revisions for the convenience of the reviewers and the AC.

  • New Section 6, Implications for Metric Design. It turns our analysis into concrete requirements for a metric that separates hallucinations from poor inliers, and states where current extractors and scorers fall short. (Reviewers 81EU and 4VPB)
  • Three additional degradations. We repeated the full grid under Gaussian noise, Gaussian blur and JPEG compression, alongside elastic distortion, and measured how far each degradation moves the embeddings of each extractor (Appendix C.2–C.3). The conclusions of §5.1 hold across all of them. (Reviewer gc5e)
  • Recent metrics. We evaluated TopP&R and α-Precision/β-Recall on the full grid, in all three embedding spaces and under all four degradations (elastic distortion plus the three newly added), and added both to the related work (Appendix C.4). They show the same patterns as the metrics in the main text. (Reviewer gc5e)
  • Positioning. We clarified our relation to Fu et al. (2025), whose correlation analysis varies the hallucination rate and inlier quality together, whereas we isolate the metric's response to hallucinations at fixed inlier quality. We also clarified our relation to Sajjadi et al. (2018). (Reviewer gc5e)

We hope these revisions address the reviewers' concerns and would be happy to clarify anything further before the discussion period closes.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.