acceptodds
Under review as a conference paper at ICLR 2027

Making Reconstruction FID Predictive of Diffusion Generation FID

Abstract

It is well known that the reconstruction FID (rFID) of a VAE correlates poorly with the generation FID (gFID) of a latent diffusion model. We propose __interpolated FID (iFID)__, a simple variant of rFID that exhibits a strong correlation with gFID. Specifically, for each dataset element, we retrieve its nearest neighbor in latent space, interpolate between their latent representations, decode the interpolated latent, and compute the FID between the decoded samples and the original dataset. We provide theoretical intuition and examples explaining why iFID correlates well with gFID by connecting to recent results on diffusion generalization. Empirically, we construct a VAE benchmark comprising 22 pre-trained VAE models and 15 VAE metrics, along with 44 diffusion models trained for class-conditional image generation on ImageNet. Using this benchmark, we show that iFID correlates strongly with gFID, achieving Pearson and Spearman correlation coefficients of approximately . The source code is provided in supplementary materials.

Primary area:
generative models
Keywords:
diffusion models, variational autoencoder

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

This paper proposes interpolated FID (iFID), a simple VAE evaluation metric that measures the FID of images decoded from interpolated nearest-neighbor latent representations. The authors provide theoretical intuition based on diffusion generalization and demonstrate that iFID correlates strongly with diffusion generation FID (gFID) across 22 VAEs and 44 diffusion models, significantly outperforming conventional reconstruction metrics.

Critical strengths
  1. Practical significance and simplicity. The paper addresses an important problem in latent diffusion models: reconstruction quality does not reliably predict generation quality. The proposed iFID is simple to implement and provides a practical way to evaluate VAEs without training expensive diffusion models.

  2. Extensive empirical evaluation. The authors evaluate 22 diverse VAEs and compare iFID against 14 existing metrics. The consistently strong correlations across different diffusion model sizes, training settings, and additional generation benchmarks provide convincing empirical support for the proposed metric.

Critical weaknesses
  1. Limited theoretical justification. The theoretical analysis establishes that diffusion samples may lie close to local affine lines connecting neighboring latent representations. However, the proposed metric uses fixed-midpoint SLERP interpolation, which is not directly justified by the theory. Furthermore, the analysis does not formally establish a relationship between interpolation FID and generation FID.

  2. Potential confounding factors in the correlation analysis. The evaluated VAEs differ substantially in architecture, latent dimensionality, and training objectives. Consequently, the strong overall correlation may partly arise from systematic differences between VAE families rather than the ability of iFID to distinguish generation performance among similar VAEs. Additional controlled experiments are needed to establish its fine-grained predictive capability.

Questions for authors
  1. Predictive power within VAE families ( Weakness 2 ). Does iFID remain strongly correlated with gFID when evaluated within the same VAE architecture or across different training checkpoints? Such experiments would help determine whether the observed correlation is primarily driven by differences between VAE families.

  2. Theoretical consistency ( Weakness 1 ). The theoretical analysis focuses on linear interpolation between nearby latent samples, whereas SLERP substantially outperforms linear interpolation in Table 6. Can the authors explain this discrepancy and provide further justification for using SLERP with a fixed interpolation coefficient of 0.5?

  3. Generalization to text-to-image generation. While Table 7 reports promising correlations between iFID and T2I benchmarks, it remains unclear whether these correlations reflect the intrinsic predictive capability of iFID or other differences across the evaluated models. Can the authors provide a more controlled evaluation where only the VAE is varied while keeping the T2I backbone and training configuration fixed?

1 comment
AuthorsOpenReview ↗

Response to Reviewer DyWG

  • We thank the reviewer for the thoughtful and constructive feedback. We are encouraged that the reviewer recognizes the practical significance, simplicity, and extensive empirical evaluation of iFID. We particularly appreciate the questions regarding the theoretical justification, fine-grained predictive capability, and controlled text-to-image evaluation. We address these concerns below.

W1 & Q2. Theoretical justification and the discrepancy between linear interpolation and SLERP

  • We agree that our theoretical analysis does not formally establish a direct relationship between iFID and gFID, nor does it specifically justify midpoint SLERP. Our theoretical results are intended to provide geometric intuition for why latent interpolation quality is relevant to diffusion generation, rather than a formal guarantee of predictive correlation. Specifically, Theorem 3 establishes that, under certain local assumptions, the ridge set lies close to the affine line connecting two nearby training samples. Since diffusion samples tend to concentrate around the ridge set, this result suggests that the quality of decoded samples along local interpolation directions can be informative about diffusion generation quality. Our proposed iFID evaluates this property through nearest-neighbor latent interpolation.
  • Regarding the choice between linear interpolation and SLERP, we emphasize that linear interpolation already achieves meaningful predictive performance, with PCC/SRCC/KRCC of 0.69/0.74/0.55 for SiT-B without CFG (Table 6). This provides empirical support for the interpolation-based intuition underlying our theoretical analysis SLERP further improves these correlations to 0.85/0.82/0.62. One possible explanation is that SLERP better preserves the geometry and latent norms associated with Gaussian-regularized VAE prior, thereby producing interpolated latents that are more representative of those encountered during diffusion generation. However, this remains an empirical observation rather than a consequence of Theorem 3.
  • We will revise the manuscript to clearly distinguish the theoretical motivation for linear interpolation from the empirical choice of SLERP.

W2 & Q1. Predictive capability within the same VAE family

  • We agree that correlations across heterogeneous VAEs may be influenced by differences in architecture, latent dimensionality, and training objectives. To directly address this concern, we conduct an additional controlled experiment involving 13 VA-VAE variants with the same VAE architecture and latent dimensionality, differing only in their foundation-model alignment configurations.

  • Specifically, we consider VAEs aligned with CLIP, DINOv2, DINOv3, and MAE using different alignment losses (cosine similarity and DMS), together with an unaligned baseline. The generation results are obtained from the publicly available benchmark [Diffusing in the Right Space: A Systematic Study of Latent Diffusability].

    Alignment modelAlignment lossgFID (SiT-B w/o CFG) ↓iFID ↓
    None–49.1443.27
    CLIPcos+dms42.8338.36
    CLIPcos37.7531.65
    CLIPdms48.6041.67
    DINOv2cos+dms29.9427.96
    DINOv2cos29.7428.47
    DINOv2dms47.6740.78
    DINOv3cos+dms39.1435.82
    DINOv3cos39.7534.16
    DINOv3dms47.9441.10
    MAEcos+dms50.9643.36
    MAEcos49.1137.22
    MAEdms48.3645.90

    Correlations between iFID and gFID:

    MetricControlled VA-VAE familyDiverse VAEs (Table 4)
    PCC0.930.85
    SRCC0.860.82
    KRCC0.740.62
  • iFID achieves even stronger correlations within this controlled VAE family than across the diverse VAEs in our original benchmark. Since the VAE architecture and latent dimensionality are held fixed, these results substantially reduce the possibility that the observed predictive correlations arise primarily from architectural differences between VAE families.

Q3. Controlled evaluation for text-to-image generation

  • We thank the reviewer for raising this important question. We would like to clarify that the text-to-image results reported in Table 7 are already obtained under a controlled diffusion training configuration, rather than from diffusion models with different backbones or training budgets. The diffusion backbone is 0.6B DDT with v-prediction. model. All models are trained with BLIP3o for 100k iterations. No REPA or REG is applied.
Reviewer_QMYXOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes interpolated FID (iFID), a proxy metric for predicting the generation quality of latent diffusion models without training the downstream diffusion model. It computes FID between real images and images decoded from interpolations between neighboring latent representations. The authors evaluate the metric across 22 pretrained autoencoders and multiple generation settings, and provide a geometric motivation based on ridge-set theory.

Critical strengths
  1. Practically relevant problem and a simple approach. The paper addresses the mismatch between reconstruction quality and downstream generation quality. By evaluating decoded latent interpolations, iFID provides an intuitive and potentially useful criterion for screening autoencoders before expensive diffusion training.
  2. Broad empirical evaluation with promising overall correlations. The benchmark covers autoencoders with diverse architectures, latent dimensions, and training objectives. Without classifier-free guidance (CFG), iFID achieves Pearson correlations of 0.85–0.88 with generation FID, substantially outperforming reconstruction FID. Additional experiments across datasets, architectures, training durations, and text-to-image settings strengthen the empirical evidence.
Critical weaknesses
  1. Strong aggregate correlations do not yet establish reliable model-selection performance. Within-family ranking failures raise concerns about practical applicability. For example, VTP-S/B/L have progressively worse iFID values of 20.70/22.92/25.85, while their SiT-XL generation FID without CFG improves from 9.55 to 7.64 to 5.79. Ranking correlations also decrease under CFG. I admit that a more detailed search for the guidance coefficient and guidance interval would significantly increase the workload of the entire process. Therefore, this is also a challenge of this paper.
  2. I suggest clearly distinguishing between the linear and SLERP implementations in the method section, and indicating that the main experiment uses SLERP, so that readers can correspond the formulas and experimental results.
Questions for authors

The starting point of this paper seems to be quite close to the recent work VE-loss [1]. The author could discuss the connections between these two works. Overall, I think this topic is very interesting. Taking a step in this direction is beneficial to the development of this field, although this work is still some way from being perfect. So I gave it a "weak accept".

[1] Taming Sampling Perturbations with Variance Expansion Loss for Latent Diffusion Models(CVPR 2026)

2 comments
AuthorsOpenReview ↗

Response to Reviewer QMYX

  • We thank the reviewer for the thoughtful and constructive feedback. We are encouraged that the reviewer recognizes the practical relevance of our problem, the simplicity of iFID, and the extensive empirical evaluation. We particularly appreciate the suggestions regarding fine-grained model selection, interpolation methods, and the connection to VE-Loss. We address these points below.

W1. Strong aggregate correlations do not necessarily establish reliable model-selection performance.

  • We agree that aggregate Pearson correlation alone is insufficient to establish reliable model-selection performance. To address this concern, we additionally evaluate rank correlations using Spearman (SRCC) and Kendall (KRCC) correlation coefficients, as reported in Table 4. iFID consistently outperforms all 14 competing VAE metrics in both SRCC and KRCC, demonstrating that its predictive capability extends beyond linear correlation to relative model ranking.

W1. Within-family ranking failures raise concerns about practical applicability.

  • We acknowledge the reviewer's observation that iFID does not correctly rank the VTP-S/B/L models. Although these models are named VTP family, their model architecture differ substantially, ranging from approximately 22M to 300M parameters.

  • In fact, we find that iFID provides consistent rankings within groups of VAEs with similar architectures and latent dimensionalities We highlight several examples from Tables 2 and 3, using SiT-XL without CFG:

    VAE groupVAEsiFIDgFID
    CNN, 16×32²SD3-VAE / FLUX-VAE / QW-VAE46.10 / 70.52 / 36.2226.38 / 34.06 / 23.62
    CNN, 4×32²SD-VAE / EQ-VAE / REPAE68.05 / 54.18 / 47.0225.91 / 20.81 / 12.95
    CNN, 32×16²IN-VAE / VA-VAE / PAE47.35 / 25.01 / 22.0725.65 / 8.57 / 5.65
    ViT, 128×32²MAE-TOK / DE-TOK18.26 / 21.056.27 / 11.97
  • Besides, we conduct an additional controlled experiment on 13 VA-VAE variants with the same VAE architecture and latent dimensionality (64×16²), differing only in their foundation-model alignment configurations. Specifically, we consider different alignment models (CLIP, DINOv2, DINOv3, and MAE) and alignment losses (cos similarity and dms), together with an unaligned baseline. The generation results are obtained from the publicly available benchmark [Diffusing in the Right Space: A Systematic Study of Latent Diffusability]:

    Alignment modelAlignment lossgFID (SiT-B w/o CFG)iFID
    None–49.1443.27
    CLIPcos+dms42.8338.36
    CLIPcos37.7531.65
    CLIPdms48.6041.67
    DINOv2cos+dms29.9427.96
    DINOv2cos29.7428.47
    DINOv2dms47.6740.78
    DINOv3cos+dms39.1435.82
    DINOv3cos39.7534.16
    DINOv3dms47.9441.10
    MAEcos+dms50.9643.36
    MAEcos49.1137.22
    MAEdms48.3645.90

    Correlations between iFID and gFID:

    MetricAcross 13 VA-VAE variantsAcross diverse VAEs (Table 4)
    PCC0.930.85
    SRCC0.860.82
    KRCC0.740.62
  • These results demonstrate that iFID works better within VAE families of similar model architecture and latent dimensions.

W2. Clarification of linear interpolation versus SLERP.

  • Thank you for pointing out this ambiguity, we will update our description in Eq. 4 and Section 3.1.

Q1. Connection to VE-Loss (CVPR 2026).

  • We find VE Loss [CVPR 26 Taming Sampling Perturbations with Variance Expansion Loss for Latent Diffusion Models] indeed very interesting and relevant. Most previous works focus on latent alignment, while VE Loss focus on KL divergence penalization, a more fundmental part of VAE. We will include VE Loss into discussion for sure, as it also deals with latent robustness. We are working on integrating the VAE with VE Loss into our benchmark.

Thank you for your detailed reply, and my concerns have been addressed.

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

This paper proposes iFID, a metric to predict downstream generation performance given a pre-trained VAE model. iFID measures the FID scores between generated images and a reference image set. The image set is constructed by finding the nearest neighbor of each generated samples in the latent space of the VAE model.

Critical strengths
  1. The motivation is clear and the idea is simple and easy to implement
  2. The benchmark is comprehensive, covering 22 VAEs, 15 metrics, and 44 diffusion models to thoroughly test the method.
Critical weaknesses

Weakness

  1. Computing iFID is expensive. By default, it requires 200K images for reference and takes ~46 minutes to evaluate on even 8xA800. Although the authors conduct a simple ablation using 50K reference images, it remains unclear how the reference set size impacts the predictive accuracy of iFID across different VAE models. oes the correlation/precision of iFID monotonically increase as the number of reference images grows? Furthermore, iFID is highly sensitive to the interpolation strength parameter α\alpha, implying that exhaustive hyperparameter tuning may be necessary before applying iFID reliably

  2. The absolute magnitude of iFID does not consistently reflect the downstream generation performance (gFID) of diffusion models. For instance:

    1. In Table 3, VA-VAE and VA-VAE-64 achieve similar iFID scores (25.01 vs. 23.04). However, as shown in Table 2, the downstream SiT models trained on VA-VAE achieve better or significantly different gFID scores compared to VA-VAE-64 (e.g., 8.57 vs. 15.09 for SiT-XL without CFG)
    2. Similar inconsistencies between absolute iFID rankings and downstream gFID can be observed when comparing SOFT-VQ vs. VA-VAE-64, as well as RAE vs. SVG-T21. This suggests that while iFID exhibits high overall correlation across a large set of VAEs, local relative rankings based on absolute iFID values may not always hold true.
  3. Consider the above weaknesses, it is crucial to further clarify the boundary conditions and limitations of iFID. Specifically, does iFID perform significantly better on specific family of VAEs (e.g., continuous vs. discrete, 1D vs. 2D latents, generative v.s. semantic) while becoming less effective on others? What underlying properties of the latent manifold cause these discrepancies?

Questions for authors

Please see weakness.

6 comments
AuthorsOpenReview ↗

Response to Reviewer bCie

  • We thank the reviewer for the detailed and constructive feedback. We are encouraged that the reviewer recognizes the clear motivation, simplicity, and comprehensive evaluation of iFID. We appreciate the concerns regarding computational efficiency, hyperparameter sensitivity, and fine-grained predictive reliability. We address each point below.

W1. Computational cost of iFID

  • Thank you for raising this concern. Upon re-examining the runtime measurements in Table 6, we found that the reported values inadvertently included the computation of other VAE evaluation metrics in addition to iFID. This explains why the reported runtime remained nearly constant as the reference dataset size increased.
  • We have corrected the evaluation script to measure the computational cost of iFID alone. The corrected results show that iFID with 200K reference images requires only 29.01 minutes on a single A100 GPU, rather than the previously reported 46 minutes on eight A800 GPUs.
    Reference imagesOriginal (8×A800) Time / MemoryCorrected (1×A100) Time / Memory
    50K46.18 min / 120.08 GB16.43 min / 17.60 GB
    200K46.05 min / 132.24 GB29.01 min / 36.16 GB
    1M46.62 min / 224.98 GB94.38 min / 74.82 GB
  • iFID requires no diffusion model training, making it substantially less expensive than evaluating VAE quality through downstream diffusion training.

W1. Sensitivity to reference dataset size

  • As shown in Table 6, increasing the reference dataset size from 50K to 200K improves the Pearson correlation between iFID and gFID from 0.79 to 0.85 for SiT-B and from 0.82 to 0.88 for SiT-XL. However, increasing the reference size further to 1M yields no additional improvement in Pearson correlation. This result indicates that a moderate reference dataset is sufficient to achieve strong predictive performance.

W1. Sensitivity to interpolation strength α

  • We agree that the interpolation strength affects the predictive performance of iFID. However, our method does not require exhaustive hyperparameter tuning.
  • We adopt α = 0.5 as a fixed default throughout our main experiments, without tuning it separately for different VAEs or downstream diffusion models. This choice corresponds to symmetric interpolation between a latent representation and its nearest neighbor, i.e., when α=0.0\alpha=0.0, the interpolation is latent and when α=1.0\alpha=1.0, the interpolation is nearest neighbor. The experiments with α\alpha= 0.1 and 0.3 in Table 6 are ablations intended to characterize the effect of interpolation strength, rather than results of a hyperparameter optimization procedure.

W2. Inconsistencies between iFID and gFID rankings

  • We agree that iFID is not a perfect predictor of the absolute magnitude or relative ranking of gFID for every pair of VAEs.
  • Our results suggest that iFID can provide more reliable absolute magnitude and relative rankings when comparing VAEs with similar architectures and latent dimensionalities. We highlight several examples from Tables 2 and 3, using SiT-XL w/o CFG:
    VAE groupVAEsiFIDgFID
    CNN, 16×32²SD3-VAE / FLUX-VAE / QW-VAE46.10 / 70.52 / 36.2226.38 / 34.06 / 23.62
    CNN, 4×32²SD-VAE / EQ-VAE / REPAE68.05 / 54.18 / 47.0225.91 / 20.81 / 12.95
    CNN, 32×16²IN-VAE / VA-VAE / PAE47.35 / 25.01 / 22.0725.65 / 8.57 / 5.65
    ViT, 128×32²MAE-TOK / DE-TOK18.26 / 21.056.27 / 11.97
  • Within each of these groups, iFID correctly predicts the relative ordering of downstream gFID with reasonable absolute magnitude. These examples suggest that iFID can be particularly informative when comparing VAEs with similar architectural configurations.

W3. Boundary conditions and limitations of iFID

  • Our results indicate that iFID is particularly effective when comparing VAEs with similar architectures and latent dimensionalities, as illustrated by the examples above. In contrast, discrepancies can arise when comparing VAEs with substantially different latent representations or architectures.

Thank you for your detailed responses to my questions. I have two remaining concerns

  1. In your response, you refer to “VAEs with similar architectures and latent dimensionalities.” Could you clarify in a more principled way on how to group different VAEs and explain how these groupings should guide the practical application of iFID? It's ok that every method has its strengths and limitations. However, a more in-depth analysis is needed to characterize the conditions under which iFID provides meaningful and reliable predictions for downstream generation performance, as well as the limits of its applicability.

  2. The experimental results would benefit from clearer organization. Rather than presenting all of them in a single large table, please consider grouping them into smaller, focused tables or figures, each addressing a specific research question and highlighting the key comparisons and takeaways.

AuthorsOpenReview ↗

Response to Reviewer bCie II part 1/2

Q1. Principled VAE grouping and practical applicability of iFID

  • We thank the reviewer for this valuable suggestion. To systematically characterize the conditions under which iFID provides reliable predictions, we exhaustively group all VAEs in Table 1 according to two objective criteria: (1) model architecture (CNN or ViT) and (2) latent dimensionality and shape. We exclude singleton groups, for which within-group ranking cannot be evaluated. This procedure yields five groups, covering all VAEs that share both properties with at least one other model. The results are summarized below (gFID is evaluated using SiT-XL without CFG):

    VAE groupVAEsiFID ↓gFID ↓
    CNN, 16×32²SD3-VAE / FLUX-VAE / QW-VAE46.10 / 70.52 / 36.2226.38 / 34.06 / 23.62
    CNN, 4×32²SD-VAE / EQ-VAE / REPAE68.05 / 54.18 / 47.0225.91 / 20.81 / 12.95
    CNN, 32×16²IN-VAE / VA-VAE / PAE47.35 / 25.01 / 22.0725.65 / 8.57 / 5.65
    ViT, 128×32²MAE-TOK / DE-TOK18.26 / 21.056.27 / 11.97
    ViT, 64×16²VTP-S / VTP-B / VTP-L20.70 / 22.92 / 25.859.55 / 7.64 / 5.79
  • iFID correctly predicts the relative gFID rankings in four of the five groups, suggesting that its predictive reliability is generally stronger when comparing VAEs with similar architectures and latent configurations.

  • The exception is the VTP-S/B/L family. Although these models share the same broad ViT architecture and latent shape, their model capacities differ substantially, ranging from approximately 22M to 300M parameters. This observation suggests that matching architecture type and latent shape alone may not be sufficient, and differences in model capacity or representation characteristics may also affect downstream generation performance.

  • To further investigate this issue under more tightly controlled conditions, we conduct an additional experiment on 13 VA-VAE variants sharing the same architecture and latent dimensionality (64×16²). These variants differ only in their foundation-model alignment configurations, including alignment models (CLIP, DINOv2, DINOv3, and MAE) and alignment losses (cosine similarity and DMS). The generation results are obtained from the publicly available benchmark [Diffusing in the Right Space: A Systematic Study of Latent Diffusability].

    Alignment modelAlignment lossgFID (SiT-B w/o CFG) ↓iFID ↓
    None–49.1443.27
    CLIPcos+dms42.8338.36
    CLIPcos37.7531.65
    CLIPdms48.6041.67
    DINOv2cos+dms29.9427.96
    DINOv2cos29.7428.47
    DINOv2dms47.6740.78
    DINOv3cos+dms39.1435.82
    DINOv3cos39.7534.16
    DINOv3dms47.9441.10
    MAEcos+dms50.9643.36
    MAEcos49.1137.22
    MAEdms48.3645.90

    Correlations between iFID and gFID:

    MetricControlled 13 VA-VAE variantsDiverse VAEs (Table 4)
    PCC ↑0.930.85
    SRCC ↑0.860.82
    KRCC ↑0.740.62
  • Notably, iFID achieves stronger correlations within this controlled VAE family than across the diverse VAEs in our original benchmark. This provides additional evidence that its predictive capability is not solely driven by differences between VAE architectures or latent dimensionalities.

  • Taken together, these results suggest a practical guideline: iFID is particularly informative when comparing VAEs with similar architectures, latent configurations, and model capacities. For VAEs with substantially different configurations, iFID remains a useful screening metric, as demonstrated by its strong aggregate correlations, but individual rankings should be interpreted more cautiously.

  • We emphasize that these are empirically supported guidelines rather than strict theoretical guarantees. We will clarify these applicability conditions and limitations in the revised manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response to Reviewers

We sincerely thank all reviewers for their constructive feedback and recognition of the motivation, simplicity, and extensive evaluation of iFID. We have carefully addressed the concerns raised in the reviews and updated the PDF. We summarize the key updates and clarifications below:

  • Corrected computational cost. We identified that the original runtime measurements included other VAE evaluation metrics. After isolating iFID, evaluation with 200K reference images requires 29.01 minutes and 36.16 GB of memory on a single A100 GPU. Furthermore, predictive performance saturates at 200K reference images, providing a favorable balance between accuracy and efficiency.
  • New controlled within-family evaluation. We conduct additional experiments on 13 VA-VAE variants sharing the same architecture and latent dimensionality but differing in their alignment objectives. iFID achieves PCC/SRCC/KRCC of 0.93/0.86/0.74, exceeding the correlations observed across the original diverse VAE benchmark (0.85/0.82/0.62). These results provide stronger evidence of iFID's fine-grained predictive capability.
  • Clarified theoretical motivation and interpolation. We distinguish the theoretical motivation based on linear interpolation from the empirical use of SLERP with a fixed interpolation coefficient of α=0.5\alpha = 0.5. We also clarify that our theoretical analysis provides geometric intuition rather than a formal guarantee relating iFID to gFID.
  • Clarified text-to-image evaluation. We confirm that the text-to-image experiments use a shared diffusion backbone (0.6B DDT), dataset (BLIP3o), training budget (100K iterations), and prediction objective (v-prediction), without REPA or REG.
  • Clarified limitation and related works We additionally clarify the limitations of iFID in ranking VAEs with substantially different architectures and discuss its connection to VE-Loss. Detailed results and explanations are provided in our individual responses.

We thank the reviewers again for their valuable suggestions, which have helped us strengthen the empirical evidence and clarify the scope and practical applicability of our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.