We thank Reviewer QbtM for devoting time to this review and providing valuable comments.

Q1:"Why is there no computational cost analysis?"

Thank you for the reminder. This part will be included in the supplementary material.

• Training stage: The training process is composed of two distinct stages, conducted on a single NVIDIA RTX 3090:

  1. Stage 1: The model was trained for 100 epochs using a single RGB image with a batch size of 28 and image size of 256x256, requiring approximately 24 hours.

  2. Stage 2: The weights of the other parts were frozen, and only the BEGA module was trained for 40 epochs using two RGB images and the corresponding event streams, with a batch size of 32. This stage took about 48 hours to complete.

• Inference Cost and Comparison: The table below details the resource consumption compared to other methods. Our approach achieves a highly competitive runtime and model size, demonstrating an excellent balance between performance and efficiency.

$\dagger$ represents trainable parameters. PSNR and LPIPS scores come from the real dataset HS-ERGB.

Q2:"How does the model perform in terms of video-level consistency?"

Thank you for the question.

• Due to conference policy, we cannot share external links at this stage, but they will be included in the supplementary material. For a quantitative measure, we rely on the FloLPIPS metric, which is specifically designed to measure perceptual coherence across frames. As shown in Tab 1 & 2 of our paper, our method achieves state-of-the-art FloLPIPS scores on all datasets, quantitatively confirming its excellent temporal consistency.

Q3:"How does it compare with recent large-motion-capable VFI methods [1,2]?"

Thank you for bringing these highly relevant, recent works to our attention. We have conducted experiments on the BS-ERGB dataset, with results shown in the table below.

The results demonstrate our method's superior performance on this challenging real-world dataset with degradation. It is noteworthy that [1] also achieves competitive LPIPS and DISTS scores, reinforcing the importance of evaluating perceptual quality—a direction we also champion in our work. We believe this new comparison provides valuable context, and we will incorporate these important results and this analysis into the main paper.

Q4:"The related work section lacks coverage of prior efforts on unsupervised event-based VFI [3] for degraded inputs, and those further exploiting the hidden temporal information in degraded inputs [4–6]."

We appreciate the reminder and agree that these citations are important for a complete literature review. We will revise the related work section in the final manuscript to include and properly discuss these papers. Thank you for helping us strengthen our paper.

Thanks to the authors for their efforts.

Given the promise of discussion and comparisons in the main paper, as well as the computational costs and currently unavailable qualitative comparisons, this reviewer will withhold the rating.

We thank Reviewer GPzr for devoting time to this review and providing valuable comments.

Q1:"The proposed framework lacks a natural connection between motion-induced image degradation (e.g., blur) and event generation, and the two core modules are relatively loosely coupled."

Thank you for this question.

• We believe the two core modules are, in fact, tightly and logically coupled to solve the problem. Our approach decomposes the complex task of "interpolating between degraded frames" into two distinct, manageable sub-problems: (1) Extracting robust semantic information from the (potentially degraded) keyframes. (2) Accurately aligning this semantic information through reliable spatial-temporal information to generate the intermediate frame.

• The key point is that only with events assistance, the model can obtain the precise inter-frame motion information required to perform the alignment in the second sub-problem. As we argue in the paper, event data's sensitivity to moving objects aligns perfectly with the changing-sensitive nature of semantic features from models like DINO. This enables a robust, temporally coherent alignment in the feature space, which is far more reliable than aligning degraded pixels. Therefore, the DINO features and the event-guided alignment are inseparable components for generating the final, high-quality interpolated frame.

Q2:"The DINO-based feature extraction module (Sec 3.1) is primarily used to learn degradation-robust features from images and is unrelated to events. Similar strategies have been adopted in prior work. The authors should clarify the unique contribution of this component in their pipeline."

• The unique contribution is not VFM (e.g., DINO) itself, but how we leverage it within a new E-VFI paradigm. Previous works optimize their model predominantly rely on pixel-level supervision from the ground truth at the output image-level stage. When keyframes and the GT are degraded, this approach forces the model to learn to reproduce these degradations, harming its generalization.

• Our paradigm addresses this by using VFM to lift the problem into a more robust feature space. The role of VFM (such as DINO) is to act as a stable, degradation-insensitive "perceptual anchor". The main contribution of our work is to design a novel framework that operates effectively for dealing with such inspiration, especially with event data assistance. We intentionally keep the VFM with minimal modifications, ensuring that our framework can be easily extended to other domains or tasks by simply swapping in a different, more suitable feature extractor.

Q3:"Unlike prior event-based optical flow methods where events contribute to both motion estimation and image synthesis, the proposed approach uses events only for alignment, without directly enhancing semantic features. The rationale for this design should be discussed."

• This is a deliberate design choice. When keyframe degradation is the primary problem, traditional optical flow methods tend to propagate the degradation from the keyframe to the warped, interpolated frame. This results in an output that, while perhaps spatially aligned, is still perceptually poor in semantic expression. By moving to a feature-level alignment operation, we bypass this issue. We use events for what they excel at—providing precise motion cues for alignment— for facilitating the final synthesis, resulting in high-quality, degradation-insensitive outputs.

Q4:"The ablation study is limited to toggling modules on or off. A deeper analysis is needed, such as replacing DINO with other backbone models or combining other backbones with the proposed supervision to verify the generality of the approach."

• To address this and further verify the generality of our approach, we have conducted a new experiment where we replace the DINO backbone with MoCo v3, another popular self-supervised vision model. After training, the model with MoCo v3 achieved a PSNR of 35.11 and an LPIPS of 0.0079 on the Vimeo90k dataset. While slightly lower than our DINO-based model, these strong results demonstrate that our proposed framework is not tied to a specific backbone and can successfully operate with different feature extractors, confirming the generality of our approach.

Q5:"How is the off-the-shelf reconstruction generator [41] aligned with features extracted by M_f (DINO), especially to support the claim of compatible to multi-level feature input for restoring the image?"

• This compatibility is achieved through our two-stage training process and the data-driven nature of the network. In Stage 1, we specifically train the reconstruction generator (G_r) and the Style Adapter (A_s) to do one thing: reconstruct a high-quality image from the multi-level features provided by DINO. An MLP is used to unify the feature dimensions. This pre-training ensures that the generator learns the exact mapping from the DINO feature distribution to the image distribution. It becomes an "expert" at translating DINO's language into a visual output, ensuring the two are perfectly aligned for the main task in Stage 2.

Q6:"Inconsistently with HS-ERGB, the proposed method achieves better perceptual metrics (e.g., LPIPS) but worse traditional ones (e.g., PSNR) on BS-ERGB and EventAid datasets. An in-depth analysis for this phenomenon is needed, including more significant visual comparison analyses."

Thank you for the comment.

• Our goal is to generate results faithful to the underlying authentic scene, rather than the degraded ground truth.

• The issue you raised is, in fact, a reflection of the success of our method rather than a flaw. The BS-ERGB dataset is well-known for its substantial noise and non-rigid deformations, and both its input frames and ground-truth references are inherently low in quality. The EventAid dataset also contains a significant amount of blur. In such conditions, a robust model should aim to produce results that are visually clean, perceptually aligned with human judgment, rather than replicating the noise and artifacts present in the original data. Traditional pixel-level metrics like PSNR tend to penalize such ‘reasonable deviations’ since they measure absolute pixel differences without considering perceptual quality.

• Our method achieves significant improvements on perceptual metrics such as LPIPS and DISTS, and the qualitative results (Fig.5, Fig.6 and Fig.7) in the paper clearly show that our interpolated frames exhibit much better visual realism than those generated by prior methods. This strongly supports the claim that our model better captures ‘perceptual authenticity’ and is more suited for real-world deployment.

• Moreover, our approach does not suffer from a noticeable disadvantage on traditional metrics like PSNR and SSIM, and in fact achieves competitive results. This further demonstrates the comprehensive effectiveness of our framework. Due to conference requirements, more significant visual comparison analyses will be included in the supplementary material.

Q7:"The paper should clarify differences between training and inference stages and report the inference cost."

Thank you for the comment. This part will be included in the supplementary material.

• Training stage: The training process is composed of two distinct stages, conducted on a single NVIDIA RTX 3090:

  1. Stage 1: The model was trained for 100 epochs using a single RGB image with a batch size of 28 and image size of 256x256, requiring approximately 24 hours.

  2. Stage 2: The weights of the other parts were frozen, and only the BEGA module was trained for 40 epochs using two RGB images and the corresponding event streams, with a batch size of 32. This stage took about 48 hours to complete.

• Inference Cost and Comparison: In the inference stage, the entire network is used to generate the final frame. The table below details the resource consumption compared to other methods. Our approach achieves a highly competitive runtime and model size, demonstrating an excellent balance between performance and efficiency.

$\dagger$ represents trainable parameters. PSNR and LPIPS scores come from the real dataset HS-ERGB.

Q8:"Does the claim of L82 apply to all existing methods? For example, does EVDI [46] have this problem?"

Thank you for the comment.

• The denoising and deblurring methods mentioned in [35][46], which are currently leading in the field, do exhibit similar issues. They share a similar objective with ours, aiming to alleviate degradation in key frames for E-VFI. However, they take a different approach, opting for a 'denoise-then-interpolate' strategy. While this approach can yield high-quality interpolated frames and clear key frames, the multi-stage process may introduce error accumulation and additional computational complexity, which contributes to their relatively lower performance in E-VFI tasks.
• In contrast, our EPA framework is a single, end-to-end solution for the interpolation task. By operating entirely in a degradation-robust feature space, we circumvent the need for an explicit, separate denoising step. This avoids the potential pitfalls of error accumulation inherent in multi-stage pipelines, offering a more direct solution to the problem.

Thanks for the rebuttal. I find most of my concerns addressed, except that I still believe the uniqueness of events in the framework design is unconvincing. I will raise my rating to borderline accept.

We thank Reviewer pM6N for devoting time to this review and providing valuable comments.

Q1:"The proposed method employs a generator-based architecture to produce the final image, a technique inherently known to boost the perceptual quality of generated outputs. Consequently, it is unclear whether the final improvement in perceptual quality stems more from the generator architecture itself or from the proposed feature-level supervision. The authors should provide more extensive ablation studies or analysis to disentangle these two effects."

Thank you for your comment.

• All components of the training framework we propose are interchangeable. This component primarily serves to synthesize interpolated frames, which is not our main focus. During experiments, we also evaluated U-Net-based [1] and Transformer-based [2] decoders. The U-Net structure increased the number of parameters, while the Transformer-based decoder introduced longer inference latency. Both alternatives yielded similar interpolation quality ($\pm$3%), with larger models offering marginal gains at the cost of efficiency. To strike a balance between performance and runtime, we ultimately adopted a simple yet effective generator.

[1] Chan K C K, Zhou S, Xu X, et al. Basicvsr++: Improving video super-resolution with enhanced propagation and alignment[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022: 5972-5981.

[2] Kim T, Chae Y, Jang H K, et al. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023: 18032-18042.

Q2:"The paper initially discusses "motion blur and image degradation" in broad terms, but the subsequent experiments primarily demonstrate effectiveness on blur and camera noise. Image degradation encompasses a much wider range of phenomena, such as lens distortion, rolling shutter artifacts, compression artifacts, low resolution, rain, and fog. It is questionable whether the proposed method maintains its robustness against these other types of degradation."

Thank you for the comment.

• We have added comparative results for different methods under JPEG compression artifacts (Table 1) and low-resolution degradation (Table 2) in the BS-ERGB pen_03 scene. As shown, our method ensures the best perceptual quality and highly competitive traditional metrics. Although CBMNet is slightly superior in traditional metrics, it suffers a collapse in perceptual quality, with significant color errors in moving objects. TLXNet, being purely based on optical flow, preserves the degradation from key frames to the fullest, resulting in large discrepancies with the non-degraded ground truth and severely affecting its performance. This further supports our viewpoint and validates the effectiveness of our method. Due to conference requirements, the detailed comparison images will be included in the supplementary material.

[Table 1] JPEG compression artifacts:

[Table 2] Low resolution:

Q3:"The proposed method requires a two-stage training process, but its training efficiency is not discussed. Furthermore, the computational efficiency of the method during inference remains unclear, and it is not specified whether it can meet real-time requirements."

Thank you for the question. This part will be included in the supplementary material.

• Training stage: The training process is composed of two distinct stages, conducted on a single NVIDIA RTX 3090:

  1. Stage 1: The model was trained for 100 epochs using a single RGB image with a batch size of 28 and image size of 256x256, requiring approximately 24 hours.

  2. Stage 2: The weights of the other parts were frozen, and only the BEGA module was trained for 40 epochs using two RGB images and the corresponding event streams, with a batch size of 32. This stage took about 48 hours to complete.

• Inference Cost and Comparison: The table below details the resource consumption compared to other methods. Our approach achieves a highly competitive runtime and model size, demonstrating an excellent balance between performance and efficiency.

$\dagger$ represents trainable parameters. PSNR and LPIPS scores come from the real dataset HS-ERGB.

Q4:"The robustness of DINO features is not guaranteed in specialized domains, such as medical microscopy videos or materials analysis. Since the proposed method relies heavily on DINO features for semantic alignment, its broad applicability and performance in such out-of-distribution scenarios are uncertain."

Thank you for the comment.

• Our work, as you mentioned, is a framework-based approach. The areas you highlighted are indeed critical application scenarios. We believe that by replacing DINO with domain-specific vision models, a competitive application performance can be achieved. However, due to limitations in available hardware and the scarcity of event-based datasets, we regret that we are currently unable to validate our method in specific domains, such as medical microscopy videos or material analysis. This is, however, a direction for future work.

Dear Reviewer pM6N,

Thank you again for your time and thoughtful feedback on our submission. We would like to kindly remind you that we have responded to your comments. If you have any further questions or concerns, we would be more than happy to address them.

We sincerely appreciate the effort you’ve put into reviewing our work—your insights have been immensely helpful in improving our paper.

Please don’t hesitate to reach out if there’s anything else we can clarify. Thank you once again for your valuable contribution.

Best regards,

The Authors

We thank Reviewer tAbV for devoting time to this review and providing valuable comments.

Q1:"The core ideas underpinning EPA are not new. The use of feature-space supervision via pre-trained networks is the principle behind perceptual losses (e.g., LPIPS ), which have been a staple in image synthesis for years. The paper attempts to distinguish its work by performing explicit feature alignment and generation in this space, rather than just using it as a loss function. However, this is more of an architectural choice than a paradigm shift. The components themselves—a DINO feature extractor, deformable convolutions for alignment , and a flow-based generator —are all existing technologies. The contribution amounts to a novel combination of these parts, which is an engineering contribution."

Thank you for your insightful analysis of our architecture.

• Our core contribution lies in proposing a novel framework that fundamentally shifts how event-based video frame interpolation (E-VFI) is approached, especially in the presence of real-world image degradation. Our novelty is threefold:

  1. Addressing a Critical, Under-explored Problem: We attempt to address a common but often overlooked issue in E-VFI—degraded keyframes—in a single-stage manner, moving beyond the idealistic assumption of clean inputs prevalent in prior work.

  2. A True Feature-Space Paradigm: Unlike methods that simply use LPIPS as a final loss function, our EPA framework performs the core operations of motion alignment and interpolation directly within the semantic-perceptual feature space. The primary learning objective (L_bf in Eq. 3) is driven by feature alignment, not pixel-level errors. This is a fundamental architectural and conceptual shift.

  3. A Novel, Purpose-Built Module: The Bidirectional Event-Guided Alignment (BEGA) module is specifically designed for this new paradigm. It leverages event data to guide the alignment of semantic features, which is fundamentally different from traditional optical flow estimation on pixels.

• The significant perception performance gains, particularly on challenging real-world datasets (Tables 2, 3), validate that our contribution is more than a novel combination of parts; it is a new, effective paradigm for robust E-VFI.

Q2:"Could you please elaborate on the rationale for re-injecting low-level features from the input keyframes via the Style Adapter? Given that the paper's primary motivation is to build resilience to defects in these keyframes, this appears counter-intuitive. Is there a mechanism that prevents the adapter from simply passing on the input noise or blur, and if so, how was it validated?"

• Our design takes into account the issue you mentioned. Introducing a style adapter to reintroduce low-level features from key frames was a carefully considered decision.
• Initially, our method was designed without an adapter. However, during experimentation, we observed that the VFM model’s disregard for motion backgrounds led to color shift issues. As you pointed out, we needed an adapter to filter and transfer low-level information from the key frames. In this regard, we explored several off-the-shelf strategies and found that a simple gating module was sufficient to achieve our goal. We believe this could be due to the semantic-detail separation learning approach, which allows for faster convergence even with fewer data.
• Specifically, after applying the gating module, our PSNR on Vimeo90k improved by 4.32 and LPIPS improved by 0.02, effectively preventing the propagation of degraded information. Since this is not the main focus of our paper, we did not include related experiments in the main text, but we will add them in the revised version.

Q3:"How do you interpret the significant drop in PSNR/SSIM on the noisy BS-ERGB dataset? Does this suggest that in the presence of heavy degradation, the feature-level supervision prioritizes perceptual "plausibility" to such an extent that it sacrifices structural fidelity to the ground truth? How can a user be confident that the generated output is faithful to the scene's motion and not a well-textured hallucination?"

• This is an excellent point highlighting the classical trade-off between fidelity and perceptual quality. Our goal is to generate results faithful to the underlying authentic scene, rather than the degraded ground truth.

• The issue you raised is, in fact, a reflection of the success of our method rather than a flaw. The BS-ERGB dataset is well-known for its substantial noise and non-rigid deformations, and both its input frames and ground-truth references are inherently low in quality. In such conditions, a robust model should aim to produce results that are visually clean, perceptually aligned with human judgment, rather than replicating the noise and artifacts present in the original data. Traditional pixel-level metrics like PSNR tend to penalize such ‘reasonable deviations’ since they measure absolute pixel differences without considering perceptual quality.

• Our method achieves significant improvements on perceptual metrics such as LPIPS and DISTS, and the qualitative results (Fig.5, Fig.6 and Fig.7) in the paper clearly show that our interpolated frames exhibit much better visual realism than those generated by prior methods. This strongly supports the claim that our model better captures ‘perceptual authenticity’ and is more suited for real-world deployment.

• Moreover, our approach does not suffer from a noticeable disadvantage on traditional metrics like PSNR and SSIM, and in fact achieves competitive results. This further demonstrates the comprehensive effectiveness of our framework.

Q4:"The paper claims to move away from "direct image-level supervision" , yet the final training stage employs the L_cg loss, which includes LPIPS and Laplacian loss on the output image. Could you clarify why this is not considered direct image-level supervision?"

Thank you for the comment.

• The phrasing could indeed lead to some ambiguity, and we have corrected it. By 'direct image-level supervision,' we refer to pixel-level error calculations, such as MSE loss and L1 loss, which can introduce degradation patterns into the model when the ground truth is degraded. Previous works typically combine pixel-level loss (e.g., L1 loss, MSE loss) with perceptual loss (e.g., LPIPS, Laplacian loss), which reduces robustness in the face of various degradation scenarios. Our training approach avoids this issue.

• Specifically, our EPA framework employs a decoupled, dual-supervision mechanism. The primary driver of the interpolation process is the feature-level alignment loss L_bf , through which the model learns to perform interpolation within a degradation-robust semantic space. This process is safeguarded from the influence of flawed ground truth pixels, thereby shifting the core learning task away from the fragile image domain, which validates our claim.

Q5:"Beyond the L2 distance comparison in Figure 8, have you performed any other analysis to more rigorously quantify the "degradation-insensitivity" of DINO features? For example, how does the signal-to-noise ratio of the features compare to that of the input images under varying levels of degradation?"

Thank you for the suggestion.

• To provide a more rigorous analysis beyond the L2 distance comparison in Figure 8, we have calculated the signal-to-noise ratio (SNR) for both the image level and feature levels across various degradation types. As the table below clearly demonstrates, the SNR is significantly more stable in the DINO feature space, especially in deeper layers, quantitatively validating their superior robustness to degradation.

Dear Reviewer tAbV,

Thank you again for your time and thoughtful feedback on our submission. We would like to kindly remind you that we have responded to your comments. If you have any further questions or concerns, we would be more than happy to address them.

We sincerely appreciate the effort you’ve put into reviewing our work—your insights have been immensely helpful in improving our paper.

Please don’t hesitate to reach out if there’s anything else we can clarify. Thank you once again for your valuable contribution.

Best regards,

The Authors