acceptodds
Under review as a conference paper at ICLR 2027

AdvFD: Boosting Visual Generation via Adversarial Fr'echet Distance Loss

Abstract

Fréchet distance has recently emerged as an effective distribution-level objective for generator post-training, complementing the conventional sample-level diffusion and flow-matching losses. However, directly optimizing Fréchet objectives can cause Fréchet hacking. The target metrics keep improving, but visual quality and Fréchet alignment in other feature spaces may stagnate or deteriorate. We attribute this failure to the static pretrained feature spaces used by existing Fréchet losses. These feature spaces provide incomplete and fixed views of the differences between real and generated distributions. To address this limitation, we propose Adversarial Fréchet Distance (AdvFD), which complements the static representation targets in FD-Loss with a calibrated adversarially learned representation. AdvFD augments the original static Fréchet objective with a learnable representation that adversarially maximizes the Fréchet discrepancy between real and generated samples, while the generator minimizes the same discrepancy in the resulting adaptive feature space. To prevent the adversarial representation from trivially increasing the objective through feature amplification, we further introduce real-feature whitening, which normalizes its scale and covariance geometry and stabilizes the min–max optimization. Extensive experiments show that AdvFD consistently improves one-step generator post-training across both JiT and pMF backbones and across different model scales.

Primary area:
generative models
Keywords:
Image Generation, Fr'echet Distance

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

This paper proposes an adversarial Fréchet distance as a post-training objective for generators, aiming to address the Fréchet hacking problem. This objective can reveal the residual differences between real and generated data, and the learning is stabilized by introducing real feature whitening. Experimental results on models of multiple scales demonstrate the effectiveness of the proposed method.

Critical strengths
  1. The research question and motivation are very clear, and the paper provides a deep and insightful analysis of the problems in the current FD loss.

  2. The description of the method is fairly detailed, and no obvious errors appear in the formulas or the demonstration.

Critical weaknesses
  1. The effectiveness of the method is mainly demonstrated through indirect experimental results, which makes some of the claims potentially too absolute. The authors claim that D_adv can reveal the residual differences between the real and generated distributions, but the premise is that maximizing the Fréchet distance leads to learning differences that are meaningful for visual quality. Actually, the learned differences may also correspond to features unrelated to human visual perceptual quality; the improvements in the experimental results may also stem from the additional trainable network parameters. Currently, no direct experimental evidence shows that D_adv learns meaningful visual differences.

  2. The advantage of the proposed method over introducing more frozen encoders is unclear. The authors state that, apart from still being confined to static representations, introducing more frozen encoders also increases computational and memory costs. However, the proposed method introduces new network parameters and whitening computations. Evidence should be provided to show that the proposed method significantly outperforms introducing more frozen encoders, or to show that the overhead added by the proposed method is obviously lower than that of introducing more frozen encoders.

Questions for authors
  1. Can the authors provide more direct, metric-based evidence showing that Fréchet hacking occurs in the training of the JiT and pMF generators in the experiments, and that it is alleviated after adopting the proposed method? The current results only indicate improvements in visual quality and quality metrics, and the source of the improvement cannot be clearly identified.

  2. Can the authors show how many frozen encoders the currently added computational and memory overhead is equivalent to? It is also suggested to provide the generation results achieved after adding the corresponding number of frozen encoders.

2 comments
AuthorsOpenReview ↗

Rebuttal to KiVb

Thanks for your reviews! Here is our rebuttal.

We thank the reviewer for finding the motivation clear, the analysis of FD-loss insightful, and the method description detailed and correct. The main concern is that our evidence is indirect. Below we add direct evidence for each claim, split over two comments.


W1 & Q1 — Does the adaptive branch learn visually meaningful differences, and is there direct evidence that Fréchet hacking occurs in our main experiments and is alleviated by AdvFD? We agree that our original evidence was indirect. We add two direct analyses on JiT-B.

Evidence 1: AdvFD closes a blind direction of the static encoder. Fréchet hacking arises when the generator exploits directions that a static encoder does not penalize, so we construct such a direction explicitly. As in Figure 2(a), we optimize a universal perturbation solely to lower Inception FID, and add it to JiT-B samples generated by FD-loss. The perturbation visibly degrades the images, yet the static Inception rewards it.

Feature spaceClean+ perturbationChange
Static Inception (used in training)1.000.93−7.0% (fooled)
Adaptive representation after AdvFD32.345.1+39.6%

Each row is on its own scale, so the Change column is the comparison. The adaptive representation starts from the same Inception network and never sees the perturbation during training, yet it penalizes the perturbation by 39.6%. The differences it learns therefore include visible degradation that the static space misses, which is consistent with the human preference study in Appendix C.

Evidence 2: hacking occurs in our main setting, and AdvFD removes it. Figure 2(b) uses a single Inception objective and does not include AdvFD, so we repeat the analysis with the SIM objective. We evaluate JiT-B checkpoints every 12.5k steps and track FD-r3 over ConvNeXt, DINOv2, and CLIP, none of which is used in training. The full curves will be added to the appendix of the revised paper.

FD-r3 (held-out)↓12.5k25k37.5k50k62.5k75k87.5k100k
FD-loss11.039.757.597.878.168.839.919.47
AdvFD (ours)9.578.017.246.686.776.446.426.39

Under FD-loss, FD-r3 drops to 7.59 at 37.5k steps, then rises to 9.47 (+24.7%) as training continues. The generator keeps fitting the training encoders, while alignment in unseen feature spaces gets worse. This is Fréchet hacking. Under AdvFD, FD-r3 decreases over the same interval, from 7.24 to 6.39 (−11.8%). AdvFD therefore improves results by reducing hacking, not just by training better overall.

The gain does not come from extra parameters. The adaptive branch is used only in the loss and discarded at inference, so the generator is unchanged. With the same extra Inception branch, a scalar discriminator (Inception-D) does not improve FID or FD-r6 (1.73 / 6.72 versus 1.00 / 5.53 for FD-loss), and training the Inception inside SIM-FD jointly with the generator degrades all metrics (7.76 / 12.44 / 12.06). The gain appears only when the branch is updated adversarially under the Fréchet objective.


W2 & Q2 — AdvFD versus adding more frozen encoders. We add frozen encoders to the static SIM objective: one (ResNet-101) and three (ResNet-101, Swin-T, DeiT). None of them is part of FD-r6, so FD-r3 remains fully held-out. All results below are at 75k steps, under the same protocol, so absolute values differ from Table 1, which reports 125k steps.

Setting (JiT-B, 75k steps)Frozen encodersms/stepExtra timePeak mem (GiB)Extra memFID↓FD-r6↓FD-r3↓
FD-loss baseline02599—44.2—2.438.518.84
+ ResNet-10113077+18.4%59.6+15.41.988.538.20
+ ResNet-101, Swin-T, DeiT33790+45.8%76.0+31.80.975.137.34
+ AdvFD (ours)03143+21.0%58.1+13.90.924.266.43

Overhead (Q2). The overhead of AdvFD is equivalent to about one frozen encoder: +21.0% time and +13.9 GiB, versus +18.4% and +15.4 GiB for ResNet-101. Three frozen encoders cost more than twice as much (+45.8% time, +31.8 GiB) and bring peak memory to 76.0 GiB, close to the 80 GB capacity of our GPUs.

Quality (W2). At comparable overhead, one frozen encoder improves FID to 1.98 and FD-r3 to 8.20, but leaves FD-r6 unchanged (8.51 → 8.53), while AdvFD reaches 0.92, 4.26, and 6.43. Even three frozen encoders, at more than twice the overhead, remain behind AdvFD on all three metrics (0.97 / 5.13 / 7.34). Adding encoders therefore helps to a degree, but a fixed set cannot cover every aspect of the images, so blind spots remain. AdvFD instead adapts the feature space to the current generator, and achieves better results at lower cost.

Dear Reviewer KiVb,

Thank you again for your constructive review. Following your main concern that our evidence was indirect, we have added direct analyses, which we briefly summarize here:

  • W1 & Q1 (direct evidence of Frechet hacking): (i) A universal perturbation optimized only to lower Inception FD visibly degrades the images and fools the static Inception (-7.0%), yet the adaptive representation learned by AdvFD, which never sees the perturbation, penalizes it by +39.6%. (ii) Tracking held-out FD-r3 (ConvNeXt, DINOv2, CLIP) during training, FD-loss reaches 7.59 at 37.5k steps and then rises to 9.47, a direct sign of Frechet hacking, whereas AdvFD keeps improving to 6.39 over the same interval.
  • W2 & Q2 (vs. adding more frozen encoders): The overhead of AdvFD is comparable to one extra frozen encoder, yet AdvFD is clearly better on all three metrics. Even three frozen encoders, at more than twice the overhead, remain behind AdvFD (FID 0.97 vs. 0.92, FD-r6 5.13 vs. 4.26, FD-r3 7.34 vs. 6.43).

We would greatly appreciate it if you could let us know whether these responses address your concerns, or if anything remains unclear. We are happy to provide further clarification during the discussion period.

Best regards, Authors

Reviewer_7qF1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies Frechet hacking where generator's trained to minimize Frechet distance in a fixed pretrained feature spaces can degrade with visual artifacts. The paper proposes AdvFD to augment the original FD-loss with an adversarial objective to minimize Frechet hacking. Specifically, the feature space is trained to maximize the difference under Frechet distance. To avoid the distance becoming arbitrarily large simply by scaling the magnitude of the feature vector, the paper proposes feature whitening to normalize the features.

Critical strengths
  • The paper clearly identifies FD-loss using fixed pretrain feature space can result in hacking in Figure 2.
  • The proposal to include adversarial objective to train the feature space is sound and intuitive.
  • The proposed AdvFD improves every reported metrics over six tested generator configurations.
  • The paper has also conducted a human preference study.
Critical weaknesses
  1. From a theoretical perspective, AdvFD does not completely prevent hacking either. Consider if two different distributions yield the same first-order and second-order statistics in the feature space, it will result in a zero frechet distance, stalling the optimization for both the generator and the feature encoder. In this case, the generator can still hack the FD loss. This is a fundamental limitation of the frechet distance itself which only considers the first and second-th moments.

  2. The final generator objective uses both the frozen feature space and the learned feature space. as Dstatic(p,q)+λ∗Dadv(p,q)D_{static}(p, q) + \lambda * D_{adv}(p, q). The paper does not provide an ablation study of only using the learned feature space. So a natural question is, why do we still need the pretrained frozen feature space? Is it necessary?

  3. The paper using both frozen and learned feature space. Technically we could simply do frozen FD + regular adversarial discriminator? This may bypass the limitation of frechet distance as discussed in weakness 1? This also leads to a question: how does adversarial frechet distance compare to just regular adversarial training? What's the advantage of AdvFD? The paper does not offer exploration for this question.

  4. All the experiments are done on ImageNet 256. (I am fine with this setting, but the original FD-loss paper have also shown 512px results besides 256px).

  5. All the experiments are done using pixel-space 1NFE models. (I am fine with this setting as it follows the original FD-loss. But this shows how impractical this line of method is. This is not a criticism to this particular paper, as it's contribution is just to improve the FD-loss.)

Questions for authors

Can you please address weakness 2 and 3.

2 comments
AuthorsOpenReview ↗

Rebuttal to 7qF1

Thanks for your reviews! Here is our rebuttal.

We thank the reviewer for recognizing that the paper clearly identifies Fréchet hacking under fixed feature spaces, that the adversarial objective is sound and intuitive, and that AdvFD improves every reported metric across all six generator configurations, supported by a human preference study. The two questions are answered in W2 and W3 below.

W1 — Fréchet distance only matches the first two moments. We thank the reviewer for this observation and agree that it is a limitation of the Fréchet distance. AdvFD mitigates Fréchet hacking but does not eliminate it, and we do not claim a guarantee. The scenario described, however, is hard to reach in AdvFD. With a fixed feature space, the generator only needs to match the first two moments in that one space. In AdvFD the adaptive representation keeps changing and keeps looking for a space in which the two distributions still differ. For the generator to stall the optimization, the two distributions would need to agree in the first two moments across all the spaces the representation visits. If two distributions look the same in many different feature spaces, they are already close to each other. We will state this limitation explicitly in Section 4 and Appendix I, and make clear that AdvFD mitigates rather than eliminates Fréchet hacking.

W2 — Is the frozen static objective necessary? Yes. We add an ablation that removes the static SIM FD objective and keeps only the adaptive branch.

Setting (JiT-B)Static SIM FDAdaptive branchFID↓FD-r6↓FD-r3↓
FD-loss✓—1.005.538.45
Adaptive only—✓30.7047.4747.23
AdvFD (ours)✓✓0.793.926.03

Without the static objective, training collapses: FID rises to 30.70, far worse than both FD-loss and AdvFD. The two terms therefore play complementary roles. The static objective keeps the generator aligned in stable feature spaces, while the adaptive branch exposes the discrepancies these spaces miss. The adaptive branch alone is a pure min–max game with nothing to hold the generator in place, so it cannot drive post-training on its own. Table 5 shows the same trend: when the adaptive weight increases to 0.20 and the adaptive term starts to dominate, all metrics are worse than with 0.10. AdvFD is designed to complement the static objective, not to replace it.

W3 — Comparison with frozen FD plus a regular adversarial discriminator. Table 2(a) already contains this setting: FD-loss + PatchGAN is exactly frozen FD plus a standard discriminator. Since PatchGAN does not share AdvFD's backbone, we further add Inception-D, which uses the same pretrained Inception backbone as AdvFD but replaces its FD critic with a scalar GAN head trained with a hinge loss, following the PatchGAN setup. Only the objective differs.

Setting (JiT-B)Adv. backboneFID↓FD-r6↓FD-r3↓
FD-loss—1.005.538.45
+ PatchGANCNN0.905.578.90
+ Inception-DInception1.736.726.31
+ AdvFD (ours)Inception0.793.926.03

With the same backbone, Inception-D worsens FID and FD-r6, while AdvFD improves all three metrics. As analysed in Section 4.2, a regular discriminator assigns a scalar score to each sample and leaves the comparison geometry unchanged. AdvFD instead learns the feature space in which the distributions are compared. A scalar discriminator can in principle look beyond two moments, but in practice it trades the distributional alignment of the static objective for a realism signal. AdvFD refines that alignment instead. We will add the configurations of PatchGAN and Inception-D to Appendix H.1.

W4 — Higher resolution. We add results on ImageNet 512×512 with pMF-B, following the 512px protocol of FD-loss.

Setting (pMF-B, 512×512)FID↓FD-r6↓FD-r3↓
FD-loss0.873.824.80
AdvFD (ours)0.833.314.43

AdvFD improves all three metrics at 512px, reducing FID by 4.6%, FD-r6 by 13.4%, and the held-out FD-r3 by 7.7%. The gains are consistent with those at 256px, showing that AdvFD transfers to higher resolution. All metrics are evaluated with the same 512px reference statistics, using the released FD-loss checkpoint for the baseline. We will add these results to the paper.

W5 — Pixel-space one-step models. We thank the reviewer for this note. AdvFD changes neither the generator architecture nor the sampling procedure, so it applies to any generator post-trained with FD-loss. We see broader backbones as natural future work.

Dear Reviewer 7qF1,

Thank you again for your thoughtful review. We have posted our responses to W1-W5, and briefly summarize the main points here:

  • W1 (moment matching): We agree this is a limitation of the Frechet distance. AdvFD mitigates rather than eliminates Frechet hacking, and we will state this explicitly in Section 4 and Appendix I.
  • W2 (necessity of the static objective): Removing the static SIM-FD objective makes training collapse (FID 30.70), confirming that the static and adaptive terms play complementary roles.
  • W3 (frozen FD + regular discriminator): Beyond FD-loss + PatchGAN in Table 2(a), we added Inception-D, which shares AdvFD's Inception backbone and differs only in the objective. It worsens FID (1.00 -> 1.73), whereas AdvFD improves all three metrics.
  • W4 (higher resolution): On ImageNet 512x512 with pMF-B, AdvFD improves FID (0.87 -> 0.83), FD-r6 (3.82 -> 3.31) and FD-r3 (4.80 -> 4.43) over FD-loss.

We would greatly appreciate it if you could let us know whether these responses address your concerns, or if anything remains unclear. We are happy to provide further clarification during the discussion period.

Best regards, Authors

Reviewer_ME7YOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses the discrepancy between the Fréchet distance (FD) metric and perceived visual quality. In many cases, the FD—calculated within a fixed feature space—may consistently decrease during training, yet the generated images still exhibit artifacts; switching to a different feature space does not necessarily resolve this issue. The authors propose "AdvFD," which augments the standard static FD objective (based on MAE, SigLIP, and Inception features) with a learnable feature branch. This branch identifies the current generator's shortcomings by maximizing the FD between the real and generated distributions, while the generator, in turn, seeks to minimize this distance. To prevent the adversarial branch from artificially inflating the loss simply by scaling up the feature values, the authors incorporate a whitening process based on the statistics of the real-world features. Experiments were conducted on the ImageNet 256×256 dataset across multiple model scales and architectures. The effectiveness of the approach is validated through evaluation using independent feature-based metrics (not used during training), human preference assessments, and ablation studies.

Critical strengths

Ablation studies provide direct support for the core design, with improvements achieved without increasing the number of inference steps. As shown in Table 2, removing the whitening operation from the adversarial branch leads to a significant degradation in training results. Table 6 presents comparisons—such as freezing additional branches—that account for total computational cost, further validating the design choice of allowing the comparison features to evolve alongside the generator, rather than relying on a static supervision term. Moreover, the method preserves the original generator architecture and single-step sampling, making it highly practical for post-training existing generative models.

Evidence for performance gains extends beyond the training objectives themselves, as illustrated in Table 1. Improvements are observed across all six model variants for both JiT and pMF. This is further corroborated by human preference evaluations in Appendix C, where AdvFD achieves favorable preference rates in four comparative settings.

Critical weaknesses

Existing comparisons have not sufficiently disentangled the respective contributions of "using adaptive features" and "employing the proposed adversarial FD objective." While the "frozen branch" ablation in Table 6 demonstrates the utility of updating features, it fails to clarify whether this specific adversarial objective outperforms standard adversarial training under equivalent conditions. Similarly, although Table 2 compares PatchGAN and DMD, it lacks sufficient detail regarding network architectures, loss weighting, and hyperparameter tuning; furthermore, it does not standardize the backbone—specifically, it fails to ensure the use of the same pre-trained Inception feature extractor employed by AdvFD. Consequently, it is difficult to discern how much of the observed performance gain stems from the FD and whitening design, and how much is attributable to more favorable feature representations or training configurations. Moreover, there remains a gap between the theoretical properties of whitening and its actual implementation during updates. While Section 4.1 and Appendix E posit scale invariance—relying on the assumption that ground-truth feature statistics evolve in tandem with the current representations—Appendix H clarifies that gradients are stopped for the ground-truth features, and that the whitening process utilizes "detached" EMA (exponential moving average) statistics. Although the authors address the lag associated with EMA in Appendix E, the use of a stop-gradient operation fundamentally alters the discriminator update step (D-step): even if the EMA has converged to the current representation, the whitening matrix is ​​treated as a constant during backpropagation, meaning that a simple scaling of features can still generate gradients that increase the loss. Thus, the fact that the forward-pass distance remains approximately scale-invariant does not automatically imply that the actual update process eliminates this scale dependency.

Questions for authors

What are the actual training durations, the number of additional steps, and the proportions of computational overhead presented in Table 6? Furthermore, when the training duration for the FD-loss is extended, is the cosine learning rate schedule adjusted accordingly? Since the baseline performance actually deteriorates following extended training, these details are crucial for interpreting the comparison.

3 comments
AuthorsOpenReview ↗

Rebuttal to ME7Y for W1 and W2

We thank the reviewer for recognizing that AdvFD improves all six generator configurations, that the human preference study corroborates the metric gains, and that it preserves the original architecture and 1-NFE sampling. We also thank the reviewer for the precise observation on detached whitening, which sharpened our claim.

W1 — Disentangling the objective from the feature extractor. On JiT-B with the same static SIM-FD objective (SigLIP, Inception, MAE), we add two variants. Inception-D uses the same pretrained Inception backbone as AdvFD, but replaces its FD critic with a scalar GAN head trained with a hinge loss, following the PatchGAN setup. Trainable Inception unfreezes the Inception inside SIM-FD and trains it jointly with the generator to minimize the same FD loss, so the feature extractor adapts without any adversarial objective. Both variants are evaluated with a separate frozen Inception.

SettingAdv. backboneFID↓FD-r6↓FD-r3↓
FD-loss—1.005.538.45
+ PatchGANCNN0.905.578.90
+ Inception-DInception1.736.726.31
+ Trainable InceptionInception7.7612.4412.06
+ AdvFD (ours)Inception0.793.926.03

With the same Inception backbone, Inception-D worsens FID (1.00 → 1.73) and FD-r6 (5.53 → 6.72) and improves only FD-r3, whereas AdvFD improves all three metrics (0.79 / 3.92 / 6.03). As analysed in Section 4.2, Inception-D learns a scalar potential under a fixed comparison geometry, while AdvFD learns the geometry itself. Trainability alone does not explain the gain: when the Inception in SIM-FD is trained jointly with the generator to minimize the same FD, all metrics degrade sharply (7.76 / 12.44 / 12.06). The improvement therefore comes from the adversarial objective, not from making the feature extractor trainable. We will add the configurations of PatchGAN (architecture, loss, weight sweep) and of both new variants to Appendix H.1.


W2 — Scale invariance under detached EMA whitening. We agree that Section 4.1 was imprecise. Within a single D-step, the whitening statistics are treated as constants, so scaling up the features does increase the objective. This gain is temporary. The EMA statistics are updated at every step and catch up with the new scale within about 100 updates (β = 0.99), after which the objective returns to its original value. The representation can therefore keep increasing the objective only by finding real differences between real and generated samples. Stopping the gradient through these statistics is intentional: otherwise the D-step could raise the objective by changing how features are normalized instead of finding new discrepancies. This follows standard practice (Spectral Normalization, Miyato et al., 2018; DINO, Caron et al., 2021). Together with gradient clipping, it keeps feature norms bounded, as shown in Figure 5(b). We will rewrite the scale-invariance statement in Section 4.1 and add the derivation to Appendix E.

AuthorsOpenReview ↗

Rebuttal to ME7Y for Q1

Q1 — Training cost and learning-rate schedule in Table 6. All settings use JiT-B on 8×NVIDIA A800 (80 GB) GPUs with a global batch size of 1024. FD-loss is trained for T = 125k steps, which is the end of its cosine schedule. We time 200 training iterations after 50 warm-up iterations, including the generator backward pass, the FD covariance and eigendecomposition, and the D-step. AdvFD takes 3177 ms per step versus 2634 ms for FD-loss (1.21×), and peak memory per GPU rises from 44.2 to 58.1 GiB. About two thirds of this overhead comes from the extra Inception branch itself: keeping it frozen already costs 1.14× and 57.8 GiB. As listed in Table 7, the D-step runs once every two generator steps and takes 362 ms per update, which is 5.7% of the amortized step time. To match the wall-clock time of AdvFD, we train FD-loss for N ≈ 26k additional steps (T + N ≈ 151k) and re-define the cosine schedule over the full T+N horizon, so the learning rate fully decays at the end of the longer run.

SettingStepsms/stepWall-clockPeak mem (GiB)FID↓FD-r6↓FD-r3↓
FD-lossT = 125k26341.00×44.21.005.538.45
FD-lossT+N = 151k26341.21×44.21.717.708.63
AdvFD (ours)T = 125k31771.21×58.10.793.926.03

With the same wall-clock budget and a fully annealed schedule, extending FD-loss from 125k to 151k steps still degrades all three metrics. The degradation is therefore not caused by an unfinished learning-rate decay, and is consistent with Fréchet hacking under prolonged optimization. At the same wall-clock cost, AdvFD improves all three metrics, so its gain does not come from extra compute. We will add the timing details to Appendix H.1 and the schedule to the caption of Table 6 (Appendix G), and annotate Algorithm 1 with the D-step frequency.

Dear Reviewer ME7Y,

Thank you again for your careful review. We have posted our responses to W1, W2 and Q1, and briefly summarize them here:

  • W1 (objective vs. feature extractor): With the same Inception backbone, replacing the AdvFD critic with a scalar GAN head (Inception-D) worsens FID (1.00 -> 1.73), and making the Inception trainable under the same FD loss degrades all metrics sharply, whereas AdvFD improves all three. This indicates the gain comes from the adversarial objective, not from the backbone or its trainability.
  • W2 (scale invariance): We clarified that, with detached EMA whitening, feature scaling yields only a transient gain that vanishes once the EMA statistics catch up, and we will rewrite the statement in Section 4.1 with a derivation in Appendix E.
  • Q1 (training cost): AdvFD costs 1.21x wall-clock per step. Under a matched wall-clock budget with a fully annealed cosine schedule, FD-loss still degrades on all three metrics, so AdvFD's gain does not come from extra compute.

We would greatly appreciate it if you could let us know whether these responses address your concerns, or if anything remains unclear. We are happy to provide further clarification during the discussion period.

Best regards, Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.