OmniZoom: A Universal Plug-and-Play Paradigm for Cross-Device Smooth Zoom Interpolation

• W1: On the fragmentation and innovation of proposed modules

Thank you for the insightful comment. While our framework involves multiple components, they are tightly coupled to address a core challenge in cross-device zoom interpolation: (1) the lack of high-quality, device-agnostic supervision, and (2) the inherent motion ambiguity in standard FI pipelines.

To address (1), we design a device-agnostic data generation pipeline with three coordinated modules. We enhance this process with spatial transition modeling and dynamic color adaptation to decouple geometric and photometric differences between devices. Furthermore, a cross-domain consistency learning scheme ensures the image maintains semantic alignment and photometric fidelity across devices. Together, these modules form a unified pipeline for generating high-quality, device-agnostic supervision. The key advantages of our data generation strategy are summarized in the table below:

To address (2), we propose 3D-TPR to enhance zoom interpolation's motion reasoning and perceptual quality, introducing a 3D trajectory encoder for monocular-disparity-aware correspondence, a texture-focus loss emphasizing high-frequency details, and mask uncertainty penalty suppressing ambiguous predictions in occluded/low-confidence regions, collectively improving robustness and perceptual fidelity. The innovations of our strategies are shown in the following table.

• W2 & W3: Clarification on visual advantages and figure improvements.

We thank the reviewer for this valuable suggestion. We would like to clarify that the visual improvements, particularly ghosting reduction and structure preservation, are already evident in the current results. For example, in Fig. 4, our method produces noticeably clearer outputs in the Huawei power outlet, OPPO grass, and fine lines in the Redmi and iPhone samples. These perceptual gains are also supported by improved no-reference metrics in Table 4.

We agree that enlarging key regions would better highlight these improvements. Due to rebuttal constraints, we cannot update the figures here, but in the camera-ready version, we will revise Figures 3 and 4 by adding zoomed-in patches and more representative examples for clarity.

• W4: Regarding the presentation of limitations.

Thank you for the suggestion. Due to space constraints, we included our limitations in Appendix A.10. We agree that making them more visible improves transparency, and will move the key points into the main paper in the camera-ready version.

Our main limitations are: (1) minor adjustments to training settings (e.g., batch size or loss weights) may be needed for different FI backbones; and (2) while our dataset spans multiple smartphones, it could be further expanded to handle more extreme device variations. That said, our method still generalizes well, as shown by its strong performance on unseen platforms.

• Q1: Regarding the device-agnostic nature of spatial transition modeling.

We thank the reviewer for raising this important point. Our spatial transition module is explicitly designed to be device-agnostic. It models zoom-related camera motion solely based on the relative transformation between the wide and main cameras, without relying on any actual device-specific assumptions such as known baselines, factory calibration data, or distortion priors. We estimate camera intrinsics and extrinsics via COLMAP and interpolate between the two estimated poses to construct a continuous trajectory. As shown in Table 2 and Figure 4, this design leads to strong cross-device generalization.

• Q2: Clarifying t in 3D-TPR framework.

We sincerely thank the reviewer for this thoughtful question.

During training, t = Sim(i, j, k), ∀ (i, j, k). Sim(i, j, k) is derived from ground-truth frames using RAFT and Lite-Mono (see Equation 13).

During inference, the ground-truth frame is not accessible, the exact similarity map cannot be computed. Following the strategy used in prior work such as InterpAny, it is sufficient to provide a t = fixed-timestep map (mentioned in line 237 in main text) in the same manner as conventional time-indexing methods.
However, the semantics of this indexing map have shifted from an uncertain timestep map to a more deterministic motion hint. Physically, this encourages the model to move each object at constant speeds along their trajectories. In practice, this constant-speed assumption serves as a valid approximation for smooth zoom scenarios commonly found in mobile photography.

As shown in Section 4, although this may lead to pixel-level misalignments relative to ground-truth, it still yields significantly improved perceptual quality. This suggests that the model, once trained with motion-aware 3D-TPR supervision, can robustly generalize to uniform motion priors during test time. Indeed, this training-inference strategy is not unique to FI task, but for a wide range of machine learning problems. In some areas, researchers have come up with similar methods [1][2]. We will clarify this design choice in the revised version.

We will clarify this design choice in the revised version.

• Q3: Regarding the clarity of Figure 2.

Thank you for the observation. We agree that the transitions may not be immediately noticeable due to resolution and layout constraints in Figure 2. To clarify, one clear example is the first triplet from the Redmi row. For example, in the first Redmi triplet, the green plant in the background shows a gradual color shift, and the scarf’s position changes consistently across views, indicating a plausible transition path.

In the camera-ready version, we will enlarge key regions and add annotations in Figure 2 to better highlight these transitions.

• Q4: Clarifying visual differences between 2D and 3D-TPR in Fig. 3.

We appreciate this opportunity to clarify: while overall frames may appear similar, 3D-TPR consistently improves challenging regions (fine structures / motion boundaries / occlusions) where ghosting or blurring artifacts occur.

Fig. 3 examples:

• RIFE (row 1): 3D-TPR better preserves tree stripe patterns vs 2D blur.
• IFRNet (row 2): More natural fingers in 3D-TPR vs 2D smearing.
• AMT-S (row 3): Guitar strings remain sharp and aligned in 3D-TPR vs 2D misalignment.
• EMA-VFI (row 4): Clearer bracelet reflections in 3D-TPR vs 2D detail loss.

We also provide additional examples in Appendix Fig. 14. In the revised version, we will clearly highlight these areas to better guide the reader’s attention. These localized enhancements reflect our 3D-aware design's strength, even when full-frame metrics may not show large differences.

• Q5: Complexity and Real-world Applicability

Thank you for this important point. Our method is plug-and-play: all components apply only during training. As shown in Q2, we revert to a uniform map at inference, introducing no extra inference cost and preserving the base model's original inference speed/memory footprint, ensuring full mobile/cloud compatibility.

• Q6: On the impact of optical flow quality on the experimental results.

We thank the reviewer for this valuable question. To analyze optical flow quality's impact on our 3D-TPR framework, we conduct a controlled ablation by downsampling RAFT inputs to different resolutions (1×/2×/4×/8×). As RAFT is resolution-sensitive, this setup offers an effective proxy for flow fidelity variation. All estimated flows are bilinearly upsampled to original resolution for training.

On Vimeo90K with AMT and IFRNet, lower-resolution flows (4×/8×) lead to noticeable degradation (e.g., up to 1.2dB PSNR drop), while 2× performs nearly on par with full-resolution 1× but reduces memory and compute significantly. We thus adopt 2× as a practical efficiency and accuracy trade-off.

Importantly, RAFT is used only during training, with an average cost of 27.5ms per image pair at 2×. No flow estimation is needed during inference. These results confirm 3D-TPR’s robustness: it maintains strong performance even under moderately degraded flow input.

Reference

[1] Y. Wang, D. Stanton, Y. Zhang, R.-S. Ryan, E. Battenberg, J. Shor, Y. Xiao, Y. Jia, F. Ren, and R. A. Saurous, “Style tokens: Unsupervised style modeling, control and transfer in end-to-end speech synthesis,” in International conference on machine learning. PMLR, 2018, pp.5180–5189.

[2] Y. Xu, H. Tan, F. Luan, S. Bi, P. Wang, J. Li, Z. Shi, K. Sunkavalli, G. Wetzstein, Z. Xu et al., “Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model,” arXiv preprint arXiv:2311.09217, 2023.

• W1: Regarding the clarity of relation to ZoomGS and Figure 1.

Thank you for the helpful comment. While our data generation is inspired by ZoomGS’s use of 3DGS for intermediate frame synthesis, this influence is limited to the use of 3DGS as a data generation tool. Our framework is independently designed and introduces three novel components:
(1) spatial transition modeling, (2) color adaptation, and (3) cross-domain consistency learning, which are absent in ZoomGS. Figure 1 illustrates our proposed pipeline. Its current version may not be fully self-explanatory. Due to NeurIPS rebuttal constraints, we will revise this figure in the camera-ready version with clearer module labels and updated captions.

• W2: On the significance of quantitative gains and the need for statistical analysis.

We thank the reviewer for the insightful suggestions. Due to page limits, ablation results were initially in the appendix. We appreciate your attention and will move expanded results to the main paper for better visibility and clarity in the camera-ready version.

Although improvements on pixel-centric metrics such as PSNR and SSIM appear numerically modest, the qualitative results (e.g., reduced ghosting in Fig. 3-4) demonstrate clear perceptual benefits. We argue that in most real-world applications, the goal of VFI is not to produce pixel-aligned ground truth reconstructions, but to synthesize plausible frames with high perceptual quality. Furthermore, PSNR and SSIM are less sensitive to common VFI artifacts such as blur and ghosting [1], which our design explicitly targets by resolving velocity ambiguity. So, pixel-centric metrics are not always informative in evaluating perceptual quality. To better assess perceptual quality, we evaluate three perceptual metrics: PI, MUSIQ, and FLOLPIPS [2], on Vimeo90K using EMA-VFI. These human-aligned metrics highlight each component’s strengths. Notably, FLOLPIPS is specifically designed for video frame interpolation, and it uniquely incorporates motion consistency priors via flow-guided weighting.

As shown below, our method consistently outperforms 2D baseline across all metrics.
Importantly, our design is not a naive stacking of modules. Each component is intentionally designed to target specific and well-known limitations in Zoom Interpolation: motion ambiguity, texture degradation, and occlusion ambiguities. Their effects are distinct, measurable, and complementary:

• The 3D-TPR encoding (3D-t-m) enhances motion consistency over the 2D baseline, reflected by a clear improvement in FLOLPIPS (0.128 → 0.119). This reduction suggests that incorporating 3D trajectory cues helps mitigate motion ambiguity and improves flow-aligned perceptual coherence.

• The texture-focus strategy (3D-m) achieves the most significant drop in FLOLPIPS (0.128 → 0.113) and the largest gain in MUSIQ (55.968 → 57.362) among individual modules, demonstrating its strong ability to preserve high-frequency details, particularly those poorly captured by 2D baselines.

• The mask uncertainty penalty (3D-t) contributes the most to reducing PI among single components (5.071 → 4.934), indicating its effectiveness in suppressing occlusion artifacts and improving perceptual sharpness.

When integrating all three modules into the complete 3D-TPR design (3D full), the results in the best overall performance across all metrics. This confirms that each module contributes uniquely, and their combination yields a more perceptually robust and accurate interpolation pipeline.

• W3: Regarding the rationale and alternatives of the disparity variation vector.

We thank the reviewer for the thoughtful question. First, we clarify that the disparity variation vector $Var^{z} = \Delta D_{0\rightarrow t} / \Delta D_{0\rightarrow 1}$ is not used during training, nor does it directly affect training or inference. Its role is purely conceptual: to illustrate the notion of normalized motion progression in the depth dimension, complementing the 2D motion ratio $Var^{xy} = V\^{xy-plane}\_{0\rightarrow t} / V\^{xy-plane}\_{0\rightarrow 1}$ in the xy-plane. Our actual 3D-TPR computation does not explicitly use this vector. Instead, we compute cosine similarity between 3D displacement vectors constructed from estimated depth and flow.

To evaluate the robustness of disparity cues in our 3D-TPR encoding, we ablated disparity map sources – the critical factor for similarity computation. Using fixed EMAVFI backbone and RAFT warping, we compared our lightweight Lite-Mono against advanced DepthPro [3].

Note that this ablation does not evaluate the disparity variation vector per se, but instead assesses the source disparity maps used in constructing the 3D trajectory vectors. Vimeo90K results show both 3D-TPR versions consistently outperform the 2D baseline, confirming stable, generalizable motion priors regardless of disparity estimator. Despite sharper DepthPro outputs, its complexity yields only marginal gains. We adopt Lite-Mono for optimal training efficiency-quality trade-off.

• W4 & Q: Clarification on similarity map input at inference.

We sincerely thank the reviewer for this thoughtful question.
During training, the similarity map (replacing fixed-timestep map) is derived from ground-truth frames using RAFT and Lite-Mono (see Equation 13).

During inference, $I_t$ is not accessible, and the exact similarity map cannot be computed. Following the strategy used in prior work such as InterpAny [4], it is sufficient to provide a uniform map (mentioned in line 237 in the main text). In the same manner as conventional time-indexing methods, a fixed-timestep map at time $t$ can be used.
However, the semantics of this indexing map have shifted from an uncertain timestep map to a more deterministic motion hint. Physically, this encourages the model to move each object at constant speeds along their trajectories. In practice, this constant-speed assumption serves as a valid approximation for smooth zoom scenarios commonly found in mobile photography.

As shown in Section 4, although this may lead to pixel-level misalignments relative to ground-truth, it still yields significantly improved perceptual quality. This suggests that the model, once trained with motion-aware 3D-TPR supervision, can robustly generalize to uniform motion priors during test time. Indeed, this training-inference strategy is not unique to FI task, but for a wide range of machine learning problems. In some areas, researchers have come up with similar methods [5][6].

We will clarify this design choice in the revised version. Moreover, we have observed that if exact similarity map is used as input for inference(Appendix A.8), frame interpolation can yield gains. Therefore, we also aim to employ more lightweight methods in our subsequent work to obtain a coarse map to enhance performance.

• W5: Ablation on ColorNet.

Thank you for the helpful suggestion. We conducted a no-reference image quality comparison between our ColorNet and common photometric calibration baselines:
(1) No Calibration, (2) Mean-Std Normalization [7], and (3) Reinhard Color Transfer [8].
Evaluations were performed on rendered intermediate frames using NIQE, PI, CLIP-IQA, and MUSIQ.

As shown below, ColorNet consistently outperforms all baselines, highlighting its advantage in modeling fine-grained photometric variations that simpler methods cannot capture. We will include these ablations in the revised manuscript to support our design choice.

• Limitations: Clarification on limitations' placement and content:

We thank the reviewer for pointing this out. Due to the page limit of the main paper, we included the limitations in Appendix A. We will integrate them into the main paper in the camera-ready version.

Our main limitations are:
(1) while OmniZoom supports plug-and-play usage across different FI backbones, mild adjustments to training schedules may improve results in practice; and (2) our dataset can be further enriched to improve robustness under extreme device variations.

Reference

[1] The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition

[2] Flolpips: A bespoke video quality metric for frame interpolation.

[3] Depth pro: Sharp monocular metric depth in less than a second. In International Conference on Learning Representations.

[4] Clearer frames, anytime: Resolving velocity ambiguity in video frame interpolation. In European Conference on Computer Vision.

[5] Style tokens: Unsupervised style modeling, control and transfer in end-to-end speech synthesis, in International conference on machine learning.

[6] Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model.

[7] Perceptual losses for real-time style transfer and super-resolution. In European Conference on Computer Vision.

[8] Color transfer between images. IEEE Computer Graphics and Applications.

• W1 & Q1: Clarification of "Motion Ambiguity" and Novelty Compared to InterpAny.

We thank the reviewer for this important question. While both our method and InterpAny [1] adopt pixel-wise similarity maps, our core idea is to reformulate motion progress estimation in the full 3D space by leveraging monocular depth, offering a more physically grounded and geometrically discriminative alternative to tackle velocity ambiguity. We clarify the terminology and highlight the key differences below.

1. Clarifying Terminology:

Fixed-timestep map: The fixed-timestep map is typically used as the temporal input for interpolation models, where the start and end frames are assigned time values of 0 and 1, respectively. It represents the specific time value for the frame to be inserted. For example, interpolating the middle frame corresponds to a map filled with 0.5. This 1D time prior, however, lacks object-specific motion awareness.

Velocity ambiguity: As defined in InterpAny, velocity ambiguity refers to the inherent uncertainty in object motion trajectories between two frames, due to the absence of intermediate observations. A given pixel may follow multiple plausible paths (e.g., curved vs. linear motion) or exhibit non-uniform speeds (e.g., due to foreshortening). Since fixed-timestep maps provide uniform temporal priors, interpolation models cannot resolve such ambiguity per object.

2. Comparison with InterpAny

InterpAny's approach: InterpAny tackles velocity ambiguity by replacing the fixed-timestep map with a 2D distance index map, computed from the similarity between optical flows. This encourages the model to infer spatially adaptive motion progress during training.
However, this formulation is limited to motion in the xy-plane. As shown in Table 2 and Figure 3 in the main text, relying solely on 2D optical flow leads to indistinguishable distance indices for geometrically distinct 3D motions, thereby failing to resolve velocity ambiguity in realistic scenes.

For instance, in zoom scenarios where objects exhibit foreshortened motion (i.e., moving toward or away from the camera), InterpAny’s 2D similarity map fails to differentiate fronto-parallel versus depth-variant trajectories, leading to ghosting or temporal drift. Our 3D-TPR explicitly addresses this case.

Our 3D-TPR approach: We extend the 2D distance index into 3D space by incorporating depth variation. Specifically, we estimate monocular disparity map and compute the 3D trajectory vectors between the target and end frames. Their cosine similarity yields a 3D trajectory progress ratio (3D-TPR), which better reflects true physical motion and resolves velocity ambiguity in both spatial and depth dimensions. This 3D similarity map replaces the fixed-timestep map, just as in InterpAny, but provides richer geometric priors. The formulation is fully differentiable and can be plugged into existing VFI models without architectural changes, applicable to existing VFI models as a plug-in.

3. Additional Contributions Beyond InterpAny

• We propose a gradient-coupled texture-focus strategy that enhances sharpness in high-frequency regions by reweighting the loss using image gradients. This design allows texture regions to receive stronger supervision without relying on pre-trained perceptual networks (Appendix A.2.2).

• We introduce a mask uncertainty penalty to regularize ambiguous fusion masks in bidirectional interpolation. This leads to improved sharpness near motion boundaries and promotes faster and more stable convergence during training (Appendix A.2.3).

4. Summary of Differences

The table below compares InterpAny and our 3D-TPR framework across major aspects. Our method introduces three main improvements:
(1) We extend motion modeling into 3D space using 3D-TPR encoding, which better distinguishes depth-related motions.
(2) A texture-focus strategy that enhances detail preservation in high-frequency regions.
(3) A mask uncertainty penalty to incentivize perceptually sharper frame synthesis.

• W2 & Q2: On the necessity of all components and the significance of overall performance gain.

We sincerely thank the reviewer for these valuable comments. Although improvements on pixel-centric metrics such as PSNR and SSIM appear numerically modest, the qualitative results (e.g., reduced ghosting and enhanced sharpness in Fig. 3 and Fig. 4) demonstrate clear perceptual benefits.
We argue that in most real-world applications, the goal of video frame interpolation (VFI) is not to produce pixel-aligned ground truth reconstructions, but to synthesize plausible intermediate frames with high perceptual quality. Furthermore, PSNR and SSIM are less sensitive to common VFI artifacts such as blur and ghosting [2], which our design explicitly targets by resolving velocity ambiguity. As such, pixel-centric metrics are not always informative in evaluating perceptual quality.

To better reflect perceptual quality, we evaluate three additional perceptual metrics: PI, MUSIQ, and FLOLPIPS [3] on the Vimeo90K benchmark using EMA-VFI. Due to space constraints, we include results from one representative network here, while full results will be provided in the revised paper. These metrics are more aligned with human judgment and highlight the strengths of each component.
Notably, FLOLPIPS is specifically designed for video frame interpolation, and it uniquely incorporates motion consistency priors via flow-guided weighting. This makes it particularly appropriate for evaluating perceptual quality and temporal coherence in smooth zoom interpolation.

As shown below, our method consistently outperforms InterpAny (2D baseline) across all metrics.
Importantly, our design is not a naive stacking of modules. Each component is intentionally designed to target specific and well-known limitations in Zoom Interpolation: motion ambiguity, texture degradation, and occlusion ambiguities. Their effects are distinct, measurable, and complementary:

• The 3D-TPR encoding (3D-t-m) enhances motion consistency over the 2D baseline, reflected by a clear improvement in FLOLPIPS (0.128 → 0.119). This reduction suggests that incorporating 3D trajectory cues helps mitigate motion ambiguity and improves flow-aligned perceptual coherence.

• The texture-focus strategy (3D-m) achieves the most significant drop in FLOLPIPS (0.128 → 0.113) and the largest gain in MUSIQ (55.968 → 57.362) among individual modules, demonstrating its strong ability to preserve high-frequency details, particularly those poorly captured by 2D baselines.

• The mask uncertainty penalty (3D-t) contributes the most to reducing PI among single components (5.071 → 4.934), indicating its effectiveness in suppressing occlusion artifacts and improving perceptual sharpness.

When integrating all three modules into the complete 3D-TPR design (3D full), the results in the best overall performance across all metrics. This confirms that each module contributes uniquely, and their combination yields a more perceptually robust and accurate interpolation pipeline. These improvements are statistically significant ($p < 0.05$), as verified by paired $t$-tests over all test samples.

• W3 & Q3: Explanation of Key Terms:

Velocity Ambiguity and Fixed-timestep Map explanation see in W1 & Q1.

Justification for Geometric–Photometric Decoupling

Our decoupling of geometric and photometric characteristics is a central component for enabling cross-device generalization in zoom interpolation. Below, we clarify how this design removes device-dependent parameters:

Geometric decoupling
Our spatial transition module operates entirely on relative camera transformations, using estimated intrinsics and extrinsics from the input wide and main images. It does not assume any actual device-specific parameters such as fixed baselines, known calibration, or distortion models. Thus, geometry is modeled purely through camera pose interpolation, independent of the device configuration. This ensures generalizability across devices with varying hardware setups.

Photometric decoupling
Instead of relying on handcrafted calibration curves or device-specific tone mappings, our ColorNet module learns content-adaptive color corrections from the input pair. Each image is encoded into a latent embedding, and ColorNet predicts per-Gaussian color offsets to align appearance across views. This data-driven design allows the system to adaptively handle color differences without assuming any pre-defined device behavior.

Summary
By disentangling geometric transitions from hardware configuration, and photometric alignment from device-specific color priors, our framework avoids hardcoded assumptions and achieves device-agnostic generalization (see Table 2 and Figure 4). We will further clarify this rationale in the revised manuscript.

[1] Clearer frames, anytime: Resolving velocity ambiguity in video frame interpolation.

[2] The unreasonable effectiveness of deep features as a perceptual metric.

[3] Flolpips: A bespoke video quality metric for frame interpolation.

• Questions:

Thank you very much for this insightful question. We sincerely appreciate your attention to detail and your concern about the resolution settings, which are indeed crucial for evaluating the practicality and scalability of our framework in real-world mobile photography scenarios.

We would like to clarify that during pretraining, we used the Vimeo-90K septuplet dataset with a native resolution of 448×256. During finetuning, our synthetic Zoom Interpolation dataset contains higher-resolution samples: 1632×1224 for Redmi and 2133×1600 for Huawei. However, we adopt a patch-based training strategy in both stages, consistent with the original design of each FI model: we use 256×256 patches for EMAVFI, and 224×224 for RIFE, AMT, and IFRNet, identical to their official training protocols. Therefore, there is no resolution mismatch between the pretraining and finetuning stages.

At inference time, OmniZoom supports arbitrary-resolution inputs and can seamlessly adapt to the native resolutions of real mobile devices, including 1632×1224 (Redmi), 2133×1600 (Huawei), 2016×1512 (iPhone), and 2048×1536 (OPPO). Although these resolutions are higher than the training patches, we observe no noticeable degradation in performance or visual quality for two key reasons:

• All FI models used are fully convolutional, and therefore generalize naturally to higher resolution inputs without any structural modification or performance degradation.

• The Zoom Interpolation task exhibits continuous and smooth camera motion, with predominantly local spatial variations. As such, models trained on moderate-sized patches are capable of capturing high-resolution local correspondences effectively at test time.

We greatly appreciate this question as it allowed us to better explain this design detail, and we will revise the paper accordingly to ensure clarity for readers.

---

• Limitations:

We sincerely thank the reviewer for raising this important point regarding data scale and its role in enabling cross-device smooth zoom interpolation.

While our dataset consists of 205 synthetic zoom sequences (16 frames each) captured from two devices, we emphasize that its design prioritizes structural diversity and cross-device generalizability over raw volume. Specifically, we focus on capturing geometric and photometric variations that commonly arise in real-world zoom operations across devices.

Our device-agnostic supervision pipeline includes:

• Device-agnostic spatial transition modeling: We interpolate both intrinsic and extrinsic parameters between wide and main camera views to simulate smooth and geometrically coherent virtual viewpoints, independent of device-specific assumptions.

• Photometric decoupling via color adaptation: A learnable ColorNet predicts per-Gaussian color offsets to adapt to photometric shifts between devices, ensuring robust rendering across camera styles.

• Cross-domain consistency constraints: We enforce semantic and textural consistency between rendered and real images, enhancing realism and semantic alignment.

• Extensive data augmentation: We apply random resizing, cropping, horizontal/vertical flipping, rotation, temporal reversal, channel permutation, and pose jittering to enrich the diversity of spatial and photometric variations. These augmentations empirically improve the model’s robustness under diverse conditions.

In addition, all models are initialized from strong pretrained FI backbones (e.g., RIFE, EMAVFI) and finetuned with our synthetic supervision. This significantly reduces the demand for large-scale data while ensuring effective adaptation to zoom-specific motion.

As shown in Table 2 and Figure 4, our models generalize well to unseen devices (e.g., iPhone, OPPO), validating the transferability of our supervision design. In future work, we plan to expand the dataset to include a wider variety of devices and optical settings.