Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models

We systematically study a wide variety of generative models spanning semantically-diverse image datasets to understand and improve the feature extractors and metrics used to evaluate them. Using best practices in psychophysics, we measure human perception of image realism for generated samples by conducting the largest experiment evaluating generative models to date, and find that no existing metric strongly correlates with human evaluations. Comparing to 17 modern metrics for evaluating the overall performance, fidelity, diversity, rarity, and memorization of generative models, we find that the state-of-the-art perceptual realism of diffusion models as judged by humans is not reflected in commonly reported metrics such as FID. This discrepancy is not explained by diversity in generated samples, though one cause is over-reliance on Inception-V3. We address these flaws through a study of alternative self-supervised feature extractors, find that the semantic information encoded by individual networks strongly depends on their training procedure, and show that DINOv2-ViT-L/14 allows for much richer evaluation of generative models. Next, we investigate data memorization, and find that generative models do memorize training examples on simple, smaller datasets like CIFAR10, but not necessarily on more complex datasets like ImageNet. However, our experiments show that current metrics do not properly detect memorization: none in the literature is able to separate memorization from other phenomena such as underfitting or mode shrinkage. To facilitate further development of generative models and their evaluation we release all generated image datasets, human evaluation data, and a modular library to compute 17 common metrics for 9 different encoders at https://github.com/layer6ai-labs/dgm-eval.

Paper

References (100)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer SaoU6/10 · confidence 4/52023-06-20

Summary

The authors performed extensive experimental study on various image-based generative models. Based on the study, it showed that no existing metric strongly correlated with human evaluations. The authors also included alternative self-supervised features extractors for evaluation. Additionally, data memorization was investigated. The experiments revealed limitations of existing evaluation metrics for generative models.

Strengths

The authors implemented extensive experiments to evaluate various image-based generative models from different perspectives, eg., encoders, human evaluations, diversity, and memorization. The evaluation was performed using several state-of-the-art evaluation metrics including humans. The comparison results are expected to have a high impact on evaluating existing generative models.

Weaknesses

The experiments showed that no existing metric strongly correlated with human evaluations. There is some concern about how the human error rate is calculated. The reviewer wonders whether the conclusion would change by improving the calculation of the human error rate. See detailed comments under Questions.

Questions

. Human error rate: this metric is calculated as the fraction of images which were incorrectly classified. The incorrectly classified cases include both real-->fake and fake-->real. From the reviewer's perspective, these two cases should be separately treated, and the fake-->real case should be more important for evaluation. For example, if there are a lot of real-->fake cases from one human participant, this probably indicates that this human participant has some judgment issues and his/her results cannot be trusted. One thing we probably can try is to use the real-->fake case to evaluate whether this human participant can be trusted or not, and then use the fake-->real case as the human error rate. Of course, there are definitely other ways to improve that. . Memorization: how to evaluate pixel-wise memorization and reconstructive memorization mathematically? Based on these two metrics, how to set the threshold values to determine memorized samples?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors have discussed the limitations of their work, and probably will address them in the future work.

Reviewer uTCz7/10 · confidence 2/52023-06-29

Summary

This paper try to get rid of the limitations of current evaluation metrics for generative models and focuses on the perceptual fidelity of diffusion models. The authors conduct an extensive study using a wide range of image-based generative models across diverse datasets. They employ psychophysics to measure human perception of image realism and compare it with existing evaluation metrics. The paper reveals that commonly used metrics, such as FID, do not align with human evaluations of perceptual realism in diffusion models. The authors attribute this discrepancy to over-reliance on the Inception-V3 network. They propose using alternative self-supervised feature extractors to improve the evaluation of generative models. Additionally, the paper explores the issue of data memorization in generative models and highlights the limitations of current metrics in detecting memorization accurately. The authors release the generated image datasets, human evaluation data, and a library to compute 15 common metrics for 8 different encoders, facilitating further research in this area.

Strengths

- The author comprehensively points out the issues with the existing evaluation metrics for generative models. - The article presents a considerable number of insights and involves a substantial amount of work.

Weaknesses

- Figure 1 lacks intuitiveness, and it is recommended to further improve it so that readers can grasp its meaning within a short period of time. - The conclusion of the entire article is extensive, making it challenging to grasp the main points. It is suggested to enhance the writing by highlighting key conclusions, particularly in the introduction section. - If possible, I suggest conducting a bias analysis on the 1000 paid participants, such as examining whether these participants are all students or individuals involved in the field of artificial intelligence, or if the age distribution is concentrated within a specific range, like 20-22 years old, and so on. This preliminary examination of bias can enhance the reliability of the experimental results. Of course, privacy concerns should be carefully addressed during the analysis. Overall, I greatly appreciate this work and recommend acceptance of the article. Furthermore, I strongly suggest the authors share the code for analyzing the figures and charts in the article to facilitate further analysis by other researchers.

Questions

see weakness.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Despite the considerable experimental analysis and insights presented by the authors in this paper, I suggest that they provide a detailed discussion of the limitations of the content being discussed. It would be preferable to dedicate a separate section at the end of the article specifically for this purpose.

Reviewer 6HZb7/10 · confidence 5/52023-07-04

Summary

In this paper, the authors conduct a thorough investigation into the limitations of the Frechet Inception Distance (FID) metric for evaluating generative models. They address this issue by performing human evaluation and proposing a superior alternative for automatic generative model evaluation. Through dedicatedly designed experiments, the authors empirically demonstrate that FID, which relies on a pre-trained InceptionV3 model, exhibits a weak correlation with human evaluation. However, by replacing the InceptionV3 with a self-supervised model like DINO, the automatic evaluation becomes more closely aligned with human evaluation. The insights presented by the authors offer valuable guidance on the appropriate approach for evaluating image generative models.

Strengths

[1] This paper addresses a highly significant issue within the generative modeling community. The authors' findings, demonstrating the lack of correlation between FID and human fidelity judgment, are particularly intriguing to me. [2] The proposed alternative, FD_{DINO}, appears to be a sensible solution and holds promise for future evaluations of generative models. [3] I think the provided benchmark tables are helpful for researchers trying to evaluate their generative models. [4] In my opinion, the paper is well-written and makes a noteworthy contribution to the community.

Weaknesses

[1] One potential concern is that all interpretations in the paper are based on the assumption that human evaluation is entirely correct, which may introduce some inherent risks. [2] Providing further explanations of the evaluation metrics would enhance the paper's accessibility for individuals who are not experts in this particular field. [3] Adding more detailed explanations about the differences between the authors' paper and closely related work [R1] would be beneficial for better understanding the novelty and contributions of the presented work. [R1] M. Yang, C. Yang, Y. Zhang, Q. Bai, Y. Shen, and B. Dai. Revisiting the evaluation of image synthesis with GANs. arXiv preprint arXiv:2304.01999, 2023.

Questions

please refer to the weaknesses.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

I believe this paper delivers a crucial message to the community by emphasizing the inadequacy of FID as a metric for evaluating image generative models. Additionally, the authors propose a substantial improvement by changing from InceptionV3 to DINO as the backbone, which significantly mitigates this issue. While it may not completely resolve the problem, it remains a commendable contribution. Considering these contributions, I would give a score of 7 to this paper.

Reviewer uifn7/10 · confidence 4/52023-07-06

Summary

This paper initially demonstrates which embedding space is similar to human evaluation criteria by utilizing various datasets and image generation models. It reveals that the embedding space of DINOv2 aligns most closely with the tendencies identified through a large-scale human survey. Moreover, it highlights the fact that most of the commonly used metrics do not align with human preferences. Additionally, the paper proposes approaches to address the issue of memorization, assessing whether the models are simply reproducing existing images.

Strengths

The limitations of the traditional FID metric were widely recognized, and several papers were already trending towards using CLIP FD as the primary metric. While FID captures typical trends well, its clear limitations prompted this paper to explicitly demonstrate these constraints and advocate for the use of specific models' embedding spaces. The paper made significant contributions by conducting large-scale surveys to create metrics for human evaluation and by publicly sharing the generated images used in subsequent analyses, which is particularly valuable considering the time-consuming nature of Diffusion models. Furthermore, the paper addressed the issue of memorization and provided a quantitative representation of it.

Weaknesses

The setting of using DINOv2's gradCAM is inevitably beneficial. It would be advantageous to include comparisons with metrics such as LPIPS as well.

Questions

Is there an intention to make all the features publicly available? Additionally, what distinguishes this paper from the next paper? https://dreamsim-nights.github.io/

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The novelty may not be significant; however, I believe the experiments conducted in this paper hold sufficient value.

Reviewer xpCe4/10 · confidence 4/52023-07-07

Summary

This paper constructs an image dataset sampled from various generative models and scored by human participants in terms of their fidelity, and argues that existing metrics do not correlate well with this notion of fidelity. Then, it investigates how different choices of embedding space (i.e. different encoders) affect the metrics in terms of their ability to measure diversity and fidelity, and concludes that DINOv2 can be a superior choice of encoder. Finally, the paper studies whether existing memorization metrics can effectively measure memorization in high resolution datasets, and argues that none are reliable.

Strengths

The paper asks interesting and important questions regarding the reliability of existing metrics of generative performance, and provides many valuable experiments with unique insights on the limitation of existing metrics for measuring fidelity and diversity of generative models. The proposed dataset of human evaluation can be a useful addition to the existing datasets for measuring and improving generative performance.

Weaknesses

My main criticism of this paper is that it is not one coherent paper, rather parts of three different papers: one paper on constructing a dataset of human evaluation of generative models, a second paper on how the choice of encoders affects existing metrics, and a third paper on measuring memorization in high resolution datasets. As a result of this, many important results are shifted to Appendix (Appendices B and D in particular), and several observations are poorly explored and studied. I think each of these three directions deserves its own focused paper, carefully considering the caveats, and avoiding broad unjustified claims. I’ll elaborate on my specific concerns below. ***Regarding the dataset:*** 1. I think the human experiment is flawed in the sense that “choose the fake sample” could result in choosing the unlikely samples as fake (confusing fakeness with unlikeliness). For example, a contorted low quality image of the common crow can be rated as “real” whereas a high quality image of an exotic rare parrot can be rated as “fake”. Therefore, a diverse model can receive a lower human error rate compared to a model of lower quality but less diversity. The experiments must propose a way to control for the effect of rarity, otherwise the dataset can be misleading. 2. Another issue with the human experiment is that it is unclear how the participants were selected. For example, if participants are mostly of one ethnicity, it is possible that they penalize diversity in place of fakeness. I understand that this issue is not easy to solve, but at the very least, I expect to see some effort in making sure the human participants are of sufficiently diverse backgrounds and origins. 3. For the results in Table 1, the difference being statistically significant alone is not enough, the amount of difference is itself of value. For example, a 1 percent difference is not as important, albeit statistically significant, as a 10 percent difference in human error rates. Reporting the actual numbers in a table similar to Table 1 will clarify this matter. 4. I am not sure how the main claim of “unfair treatment of diffusion models” in the title is justified. The brief explanation in L198-199 is too broad: “Coupling the results in Table 1 with the FID rankings in Figure 2, we conclude that current diffusion models produce the most realistic images, but are unfairly downranked by FID”. Could you be more specific about the connection from evidence to this claim? ***Regarding the metrics and encoders*** 5. L225 claims that “We find that Inception does not perceive a holistic view of images even on its ImageNet training set”, however I don’t think the qualitative results in Fig 3 can support such a strong statement. A quantitative experiment is required to be able to claim this finding (I don’t see how the quantitative results in Appendix D.0.2 can support this claim, if so, please elaborate.) 6. L299 claims that: “We see this as strong evidence of a fundamental limitation of the use of the Inception network when computing FID”. I don’t see how this shows a fundamental limitation. To me, it only shows that for some generative models, FID seems to not correlate well with per class Vendi score. If the intention is to show that FID does not correlate with Vendi in general, why not explicitly report correlation? But even then, can you elaborate what “fundamental” means in this context? I do understand that different embeddings focus on different features, that is given by design, but that one embedding is fundamentally weaker than another needs a more formal argument. 7. The conclusion that DINOv2 is better than Inception is not well justified to me. The main evidence for this strong claim seems to be Figure 6, which is at best a motivating observation (reporting the correlation coefficients could help make Figure 6 more substantial). A systematic ablation study for the encoder on the effect of a) the number of training classes, b) various loss functions, and c) various architectures (only changing one factor at a time and fixing the other factors), on the performance of metrics can be more conclusive regarding which encoder is more reliable. 8. For a paper about metrics, the definition of considered metrics should be restated to facilitate the clarity and readability of the paper. This can be done for the main metrics in the body of the paper (FD, P/R, Vendi) and for others in the appendix. The current definitions in Appendix B lack the exact mathematical definitions except for FD. ***Regarding memorization*** 9. The paper does not discuss why each of the considered memorization metrics fail, that is, what assumptions in the definition of each metric deviates from practice and potentially causes the observations. The lack of such discussions and followup experiments to pinpoint the cause, makes the empirical results inconclusive in general.

Questions

See the weaknesses section.

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

3 good

Contribution

3 good

Limitations

See the weaknesses section, particularly on the collection of the dataset.

Reviewer xpCe2023-08-11

Thank you for your response

Regarding “dataset”, “metric” and “memorization” sections being part of a coherent paper, I respectfully disagree with the authors: the “metric investigation” section, uses the dataset to study several metrics, but rather than focusing on why existing metrics do not correlate with Human Error and if that reveals a limitation in the Human Error metric itself or the existing metrics (since these metrics are backed by their own user studies and theories), it quickly takes Human Error as ground truth and moves on to improving metrics (hence a separate paper in my opinion); b) the “memorization” section does not use the proposed dataset or the Human Error metric, it does not even mention them. 1.1. You claim “each participant sees a total of 250 images from ImageNet (which has 1000 distinct classes), making it essentially impossible to learn any diversity within any class, and thus forcing a focus on fidelity”, the issue is not that your participants will learn to focus on diversity or not, it is that they might already associate rarity with fakeness, so you need to provide evidence for the claim that “... forcing a focus on fidelity”: one way to do so is to report Human Error rate on rare real samples versus common real samples (e.g. using Rarity Score, Han 2022), and see if the error rate is the same on both sets. Lack of such experiments, and any discussion of this potential issue in the paper, is why I do not think you can claim that “our methodology accounts for and negates such an effect” at present. 1.2. The parrot/crow is of course just an extreme example to clarify my point, within each ImageNet class you can find unlikely as well as likely samples. 1.3. You claim “if participants were confusing fakeness with unlikeliness, this effect would be consistent across models and therefore would not alter rankings”, I don’t see why the effect would not alter rankings: if participants confuse fakeness with rarity, they will make less mistakes on models that generate more rare samples (because they will just flag them as fake based on the samples being unlikely), so your score would unfairly rank models that are more diverse worse. 1.4. I don’t think being similar to HYPE or aided by psychophysicist answers any of my very specific concerns. What is lacking here is quite straightforward, as I explained in 1.1, you should provide a control experiment, otherwise the dataset – which I want to emphasize that I think is valuable and interesting – will incentivise a series of misleading and incorrect followup works. 2.. I don’t understand what this means: “The Normalized Error Rate accounts for the varying difficulty of tasks over different combinations of (dataset, model).”, please elaborate. 3.1. You claim “Figure 2 clearly shows that diffusion models score the highest human error rates, and that GANs often score lower human error rates yet achieve a better FID ranking”, but in Figure 2 CIFAR10 the diffusion model is ranking best in terms of FID too, so FID is not unfair. Same in ImageNet. My point is by just looking at Figure 2, there is no concrete evidence to back “unfairness”. You need to be more specific about what is the mathematical definition of “unfair” in your work, and how you measure it. For example, you could define unfair as low correlation between FID and Human Error Rate, and then report the correlation coefficients and claim that the correlation is higher for GANs, but lower in Diffusions. 3.2. You claim “Table 1 summarizes the statistical significance”, yet you report no significance test results, it is unclear what the ordering in this table are based on. Table 11 also does not clearly show “unfairness” towards diffusion models. 6.. I understand your motivation, but what I still do not understand is what “fundamental” means in this context. In any way, I do not consider this a main concern, I acknowledge that whether something is a “fundamental” flaw, or simply a lack of correlation between some metrics, is subjective. I appreciate the additional explanations by the authors. 7.. If computational restrictions do not allow you to sufficiently support the claim that DINOv2 is better than Inception, please avoid making that claim in your abstract. If Figure 6 is able to sufficiently support this claim, please elaborate. 8.. I agree that there might be interesting connections, but lack of a study on those connections makes the memorization results inconclusive. Han, Jiyeon, et al. "Rarity score: A new metric to evaluate the uncommonness of synthesized images." ICLR 2022.

Authorsrebuttal2023-08-14

Second Reply (1/2)

Thank you for replying quickly and engaging in discussion, we appreciate the added clarification and hope to continue discussion on any lingering concerns. 0. You are correct that we take human assessment of fidelity as ground truth, we will make this more explicit in the final version of our paper. We believe this is an extremely reasonable assumption though, as humans are the end users of these models. The goal of our diversity experiments was to verify if this assumption holds: to see if diversity (which we do not believe humans are particularly good at detecting, nor were our experiments designed to detect) could potentially explain discrepancies between FID and human error rate (HER; see point 6 of our rebuttal). The memorization part of the paper indeed does not use our FD_DINOv2 score or HER because neither FD metrics nor HER are expected to detect memorization (note we include the value of all memorization metrics for DINOv2 in Table 11). Our main goal with these experiments is not to further validate the FD_DINOv2 score itself, but rather to ensure models at the top of the FD_DINOv2 leaderboard did not “cheat” their way in by memorizing training data. We see this as a relevant step to establish community trust in the leaderboard. We emphasize once again that we will update our manuscript to better convey this narrative. 1. Thank you for clarification, and for suggesting the Rarity Score (RS) of Han et al. 2022. We agree this analysis, which we have now performed, provides more evidence for whether participants confuse fakeness with unlikeliness. Our rebuttal pdf cannot be updated and no further attachments can be added, so here we describe the results in detail and outline what we will add to the final paper. *Experiment*: We focus on your suggestion of “Error rate on rare real samples versus common real samples”. For each of the 2000 real images we used from each dataset (evaluated by an average of 13 humans), we determined the fraction of humans that labeled it as fake, as well as its RS. We performed the calculation using both Inception and DINOv2 to quantify the dependency of RS on the embedding space. *Results*: We find no correlation between HER and RS on ImageNet and FFHQ. On CIFAR-10 and LSUN-Bedroom we find a small (e.g. see Table 1 in https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3576830/) but statistically significant correlation, which we identify as driven by dataset issues: the non-zero correlation is caused by a very small percentage of “real” images which are clearly taken from 3D-generated scenes (instead of bedroom photographs in LSUN-Bedroom), or from 2D-generated scenes or low quality (extremely blurry) images (in CIFAR-10). These results show that 1.) humans are more likely to label generated scenes as fake/generated (LSUN-Bedroom, CIFAR-10), and 2.) humans are more likely to label low-quality images as fake/generated (CIFAR-10). Such images have a higher than average RS, and hence *the small correlation between human evaluation and RS on CIFAR10 and LSUN-Bedroom is due to humans properly identifying these dataset issues*. We find that removing just 6% of the “fakest” (as measured by humans) real images on LSUN-Bedroom removes the correlation of RS and HER - quantitative proof that the small correlation is driven by dataset issues, and not due to humans associating diversity with fakeness. Rare defects in the training set are not enough to affect our results: being so rare (~6%) means they barely affect the average error rate; and *this training-set effect is the same for every generative model evaluation and thus does not change their rankings*. We have prepared a few additional scatter-plots and image visualizations for the final version of our paper (the lack of correlation is visually evident), and summarize the correlations and their significance on the table below. We will also include the rarity score in our public codebase. Table 12: Pearson correlation of the fraction of humans that labeled a real image as fake, and the Rarity Score (RS; Han 2022) of that image. The RS can only be determined for images that fall “on manifold” | | | % on manifold | r | p-value | r (94%) | p-value (94%) | |--------------|-----------|-----------------|--------|---------|----------|----------------| | CIFAR10 | Inception | 76 | 0.28 | 0.00 | 0.216 | 0.000 | | | DINOv2 | 88 | 0.062 | 0.01 | 0.00 | 0.86 | | ImageNet | Inception | 82 | -0.03 | 0.16 | -0.00 | 0.92 | | | DINOv2 | 93 | 0.01 | 0.73 | -0.01 | 0.62 | | LSUN-Bedroom | Inception | 70 | 0.11 | 0.00 | 0.05 | 0.11 | | | DINOv2 | 90 | 0.10 | 0.00 | 0.02 | 0.39 | | FFHQ | Inception | 79 | -0.03 | 0.29 | -0.02 | 0.36 | | | DINOv2 | 90 | -0.02 | 0.43 | -0.04 | 0.10 |

Authorsrebuttal2023-08-14

Second Reply (2/2)

2\. Tasks (defined as dataset/model pairs) have varying difficulties - if a generative model is very poor, humans have low error rates as they easily distinguish fake samples from real ones. Thus we cannot average error rates across tasks, as these numbers are not directly comparable. To avoid this issue, if $x_i$ is the error rate of participant $i$ at task $t(i)$ (the task performed by participant $i$), the normalized score is given by $(x_i - \mu_{t(i)}) / \sigma_{t(i)}$, where $\mu_{t(i)}$ and $\sigma_{t(i)}$ are the mean and standard deviation, respectively, of the error rate at task $t(i)$ across participants. Normalized scores are comparable across participants and tasks. 3\.1 Note that we discuss CIFAR-10 being an exception in Fig 2 on L249-252. Nonetheless, our argument is *not* that Fig 2 shows unfair treatment of diffusion models by FID, but the figure does show that FID and human error rate (HER) are uncorrelated (except on CIFAR-10). This is what we meant in point (3 & 4) of our rebuttal: we agree that at this point of the paper, there is not yet enough evidence to call the lack of correlation between FID and HER unfair. However, we believe that in the rest of the paper, we present enough evidence to call this unfair: (a) the Inception network focuses on the wrong aspects of images (again, we do assume here that humans provide ground truth, and understand “wrong” as “unlike humans”), which is shown in Sec 4.1; (b) metrics which *do* focus on the correct parts of an image, like DINOv2, have a much stronger correlation with HER, which is shown in Sec 4.2; and (c) diversity is *not* the reason FID is uncorrelated with HER, which is shown in Sec 4.2.1. Together, all this evidence does allow us to say the treatment given by FID to diffusion models is indeed unfair, even if we have no formal mathematical definition of what “unfair” means. Again, we will more clearly convey this narrative in the final version of the paper. 3\.2 We apologize for poor phrasing in our rebuttal. Table 11 (sorted in decreasing order of FD_DINOv2) summarizes the mean and standard error HER (which is what you asked us for in your original review), not statistical significance of the tests from Table 1 (which as mentioned in L191-193, are all highly significant). Again, we do not see Table 11 as showing diffusion models are treated unfairly by FID: we believe our paper as a whole establishes this conclusion, not just the lack of correlation between FID and HER or the numbers in this table. 6\. We will happily change the phrasing from “fundamental” to something less ambiguous. 7\-1 We believe the claims in our paper allow us to say that DINOv2 is better than Inception, *but not just because of Fig 6*: (a) Figs 2 and 4 show that FID does not correlate with HER, whereas FD_DINOv2 does; (b) Fig 3 shows that DINOv2 focuses on more human-relevant aspects of images than Inception; (c) Fig 5 shows that diversity does *not* explain differences between HER and FID: for example, let’s focus on LDM (a diffusion model) and StyleGAN-XL (a GAN). LDM has both better HER than StyleGAN-XL (Fig 2), and has diversity that more closely matches the data’s (Fig 5), yet has a worse FID (Fig 2). FD_DINOv2 does not exhibit this “unfair” ranking of LDM and StyleGAN-XL. (d) Fig 6 shows that diversity *does* explain some differences between HER and FD_DINOv2 score (e.g. DiT-guided has better HER than DiT, yet has worse FD_DINOv2 score as its diversity is a much worse match to the data’s than DiT’s). Together, all this evidence does let us say DINOv2 is better than Inception. Again, we will make sure this line of thinking is clearer in the final version of our paper. 7\-2 Our claims about computational restrictions in the rebuttal are simply about the impossibility of a full ablation study of our own foundational models to understand exactly *why* they outperform Inception, and have no bearing on whether the evidence in our paper is enough to “sufficiently support the claim that DINOv2 is better than Inception”. We specifically designed the selection of encoders in our study to shed light on what components of training/architectures benefit generative evaluation. As discussed in L214-240, an improved supervised model (ConvNeXt) indicates whether more modern supervised models share the same issues as Inception, we chose both self-supervised CNNs and ViTs trained with different objectives, models trained on ImageNet and those trained on internet-scale datasets, and performed a study of the effect of architecture in App B.4.1 (6 varieties of CLIP and 4 of DINOv2). 8\. We agree our experiments on memorization are inconclusive *in terms of establishing the reason or a solution for the failures of memorization metrics*, but not in terms of establishing that models at the top of the DINOv2 leaderboard did not “cheat their way to the top” through memorization, which was the main objective of these experiments. We will further clarify this in the paper.

Reviewer xpCe2023-08-20

Thank you for the additional results, I will slightly increase my score in recognition of the additional results. However, my concerns regarding stating claims without clear definitions and measures remain (points 3 and 4 in my response). I also disagree with the authors' justification of treating human rating as ground truth, namely that: "We believe this is an extremely reasonable assumption though, as humans are the end users of these models." Humans are not always the end users, for example: the use of generative models as augmentations to improve performance of classifiers, or as debiasing methods to improve the fairness of classifiers, in both cases the end user is the downstream classifier which does not necessarily benefit from generative models that humans rate as being better.

Authorsrebuttal2023-08-21

Third Reply

Thank you very much for increasing your score and for continuing to discuss! Our view is that any attempt to compare two probability distributions (in this case the model’s and the ground truth) using a single scalar score must make inherent trade-offs as to how different attributes of these distributions are “weighted” when computing the score. We also believe that a “weighting” being sensible is task-dependent, and thus we do not think a single metric can be designed to be useful at every task. In this sense, we agree with you that human error rate and FD_DINOv2 need not be useful metrics for every task (and the same is true of any metric). We also agree that data augmentation for classification accuracy is an example task where FD-based metrics do not make sense: the distribution of augmented images need not be similar to the target distribution (e.g. AugMix [A]), yet can still help accuracy. Similarly, if attempting to debias a model which correctly learned a distribution from a biased dataset, FD-based metrics need not be appropriate. The same is true of dataset distillation [B], where the distribution of distilled images can be very different from the true distribution of images, yet still be useful. Yet, one of the main uses of image-based generative models is to generate realistic images, which is what we meant by saying humans are the end users. FID (which is not used as a metric for data augmentation, bias, or dataset distillation) is often thought about as a way to compare distributions in such a way that realism is “highly weighted”. Our claims on FD_DINOv2 are not that it should be used as the one and only metric to evaluate generative models, but that it is much better than FID as a metric of generative quality that “highly weights” realism. We will happily make this point clearer in the final version of the paper. [A] *AugMix: A Simple Data Processing Method to Improve Robustness and Uncertainty*, Hendrycks et al., ICLR 2020 [B] *Dataset Distillation by Matching Training Trajectories*, Cazenavette et al., CVPR 2022

Reviewer SaoU2023-08-11

comments on responses

Thanks. The reviewer has read all responses, which has generally addressed the concerns. By reading other reviewers' comments, the reviewer decided to keep the current score.

Reviewer uTCz2023-08-17

re: Rebuttal by Authors

Thank you for your reply and I have also read the comments of other reviewers. I tend to keep my score.

Reviewer 6HZb2023-08-18

re: Rebuttal by Authors

The authors addressed all my concerns in the rebuttal. I am inclined to maintain my original score of 7.

Reviewer uifn2023-08-19

Thank you. I decide to keep the current score.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC