acceptodds
Under review as a conference paper at ICLR 2027

STRIDE: Scene Transition via Interpolative Diffusion Estimation

Abstract

Video transition generation, the synthesis of coherent intermediate frames bridging a given start and end frame, is a core problem in generative video modeling. Existing diffusion and flow methods are not only computationally heavy but also rely on large pretrained image/video backbones. We present STRIDE (cene ansition via nterpolative iffusion stimation), an inference-efficient bidirectional latent-diffusion framework whose 3D U-Net denoiser (1.54B parameters) is customized and trained as generative-prior-free, that carries no transition or motion prior from any pretrained generator (a frozen VAE and CLIP serve only as fixed interfaces). STRIDE couples position-aware progressive masking, coupled bidirectional temporal attention with a forward/backward consistency loss, and multi-anchor CLIP conditioning, is trained under a four-phase motion-filtered curriculum with an annealed-fusion inference schedule. Operating in a compact pixel and latent space, it infers to faster than DynamiCrafter, SEINE, TVG, and Wan2.1-VACE-1.3B, at to lower per-clip compute than DynamiCrafter, SEINE, and TVG. Further, it attains the best inter-frame structural consistency (VSI, FSIM, and tMS-SSIM on all three benchmark datasets i.e., Morph-Bench, TC-Bench, and TransitBench and tPSNR on one of the three). Although SVD-KFI (Generative Inbetweening) wins the four perceptual/appearance-similarity metrics (CLIPSim, DreamSIM, LPIPS, EMD) on all three benchmarks, its inference is slower than STRIDE, with Wan and Framer occupying intermediate positions. Because all eight metrics are computed between consecutive frames, they measure inter-frame smoothness rather than motion fidelity. This makes the resulting frontier an operating-point trade-off between inter-frame structural consistency and perceptual/appearance similarity, rather than a quality ranking. In our human study (5 models 48 clips 3 raters ratings), we observe high inter-rater agreement (Krippendorff's –). The Friedman omnibus test reveals significant differences in transition-smoothness ratings across models (). Furthermore, Holm-corrected pairwise Wilcoxon tests confirm that STRIDE is significantly preferred over all four pretrained-prior baselines (DynamiCrafter, SEINE, TVG, Wan2.1-VACE-1.3B). Despite using no pretrained diffusion backbone, STRIDE achieves significantly smoother transitions than pretrained-prior baselines, at the cost of occasional reductions in intermediate-frame visual quality near the transition midpoint.

Primary area:
generative models
Keywords:
Scene-transition, bidirectional fusion, diffusion, Video frame interpolation, Video infilling

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_uxciOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper focuses on video transition generation from two boundary images. A shared denoiser predicts noise under forward and backward temporal-attention masks, and the predictions are combined using position-dependent weights. Its training uses progressive frame masking and a motion-filtered curriculum.

Critical strengths
  1. The experiments show strong smoothness and latency. The proposed method achieves a human smoothness score of 3.92, which is the best over all baselines. Several consecutive-frame structural-consistency metrics support this result. The generation latency is lower than WAN.

  2. The paper reports important failure modes. The ground-truth fidelity evaluation, loss interventions, and unsuccessful ghosting-proxy validation are informative. I appreciate that the authors acknowledge the limits of consecutive-frame metrics and disclose the smoothness–fidelity trade-off rather than claiming universal quality superiority.

Critical weaknesses
  1. Unconvincing results. In Tables 6 and 7, the proposed method does not show strong performance. Most evaluation metrics are not SOTA. Moreover, in Tab. 2, the proposed method achieves the best smoothness, but other metrics are significantly weaker than other baselines. Those results show that the proposed method fails to generate high quality videos. There is also very limited visual examples (qualitative results).

  2. The comparison is not fair. STRIDE runs at 256², whereas DynamiCrafter and TVG run at 768² and SEINE at 784², despite the paper describing the latter checkpoints' native resolution as 320×512 (Appendix A.1 and Table 5). I do not believe this is fair. The training data should be the same unless the authors claim a new dataset contribution or their training data is not obtainable.

  3. The presentation makes the method and evidence unnecessarily difficult to assess. All tables are deferred to the appendix: the human-study summary is on p.14, efficiency on p.19, and the main benchmark tables on pp.20–21. Figure 1 combines training and inference through dense text and equations. The main qualitative comparison is deferred to Figure 4 on p.17.

Questions for authors

I do not have any reason to accept this paper, regardless of the rebuttal. The paper has to be largely reformatted, and the evaluation metrics should be more standardized. The goal should be clear.

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

The paper presents a diffusion framework for generating intermediate frames between two input frames. The method trains a 3D U-Net from scratch without relying on a pretrained generative backbone, while using frozen VAE and CLIP encoders. It combines progressive masking, bidirectional temporal attention, multi-anchor conditioning, and an annealed fusion strategy.

The authors evaluate STRIDE on transtion benchmarks and report improvements in temporal smoothness and inference efficiency compared to several pretrained video generation methods. They also conduct a human evaluation and analyze the degradation of intermediate-frame quality near the midpoint.

Critical strengths

None

Critical weaknesses

1. Unclear novelty and poor presentation.

The paper is very difficult to follow, even after careful reading. The method combines many components, but it is not clear what the main technical novelty is compared to existing VFI and generative inbetweening methods. The motivation for each component is also not well explained, and there are no sufficient archtectural ablations to justify the proposed design. As a result, it is difficult to understand which part of the proposed method is actually responsible for the reported improvements.

2. Insufficient experimental validation

I am not convinced that the experiments demonstrate the effectiveness of the proposed mehod as a VFI method. The main evaluation focuses on temporal smoothness, but this does not necessarily mean that the generated intermediate frames are meaningful or accurate. Comparisons with conventional VFI methods such as FILM and inbetweening methods such as TRF are missing. Although the authors evaluate ground-truth fidelity on Vimeo-90K, they do not compare against existing VFI methods.

The qualitative comparisons are also limited, and some results in the supplementary videos appear closer to smooth crossfading than actual intermediate motion. In addition, the FVD results in Appendix A6 raise concerns about whether the proposed method provides a clear advantage over simple interpolation. I do not think the current experiments provide sufficient evidence that the proposed method offers a clear advantage over existing interpolation approaches.

Questions for authors

Q1. I am not sure what the main novelty of the proposed method is. Bidirectional diffusion and forward/backward fusion have already been studied in previous works. Could the authors explain more clearly what is technically new in the proposed architecture and fusion strategy compared to existing video interpolation methods?

Q2. The paper introduces several components, but it is difficult to understand which component actually contributes to the performance. Could the authors provide ablation studies with unidirectional temporal attention, standard random masking, and simpler forward/backward fusion? These experiments are necessary to show why the proposed design is needed.

Q3. Why are conventional video frame interpolation methods such as FILM not included in the experiments? Since the method generates intermediate frames between two given frames, I think comparisons with these methods are important. Also, the experiments mainly focus on large scene transitions, often between two frames with completely unrelated content. I would like to see how the proposed method performs on standard VFI tasks, where the input frames are temporally close and belong to the same scene. Although the authors report ground-truth fidelity on Vimeo-90K, there is no comparison with existing VFI methods. I think such comparisons are necessary. In particular, I would like to see whether the proposed method can reconstruct actual intermediate motion better than existing interpolation methods, rather than only achieving high temporal smoothness.

Q4. Could the authors compare the method with simple crossfade and latent interpolation on the main benchmarks? The FVD results in Appendix A6 are also concerning, as simple crossfade achieves a much lower FVD than the proposed method on Vimeo-90K. Also, aftre watching the supplementary videos, I don't think the proposed method tends to produce smooth or meaningful transitions between two ekeypoints. I think comparisons with simple interpolation baselines would help clarify this point.

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

This paper is about video transitioning via diffusion: Given start and end frames of a sequences, the task of the model is to estimate frames that generate a transition from the start image to the end image of the sequence. Whereas the standard image-to-video task uses a start image along with a prompt and generates some sequence that could follow the start image, this paper assumes that there are 2 images given, so the end of the sequence is also fixed.

The proposed network is based on a 3D-Unet. The model starts generation from both the first and the last image which leads to two different denoised samples at each frame at each denoising step. The authors propose a linear blend between both sources which appears both reasonable and novel.

There are extensive quantitative results in the appendix.

Critical strengths

The problem considered by the authors is interesting as indicated by several prior works for instance the Ant team (Wang et al., Framer) and Google Deepmind (Wang et al., Generative inbetweening). The proposed solution to linearly interpolate the denoised samples from the forward and backward generation is simple but effective. There is a comprehensive quantitative evaluation in the appendix.

Critical weaknesses

efficiency comparison

  • Sec 4.6 compares the proposed with with a resolution of 256^2 with prior works that use higher resolutions (eg a Wan baseline with resolution 614^2). This is not a fair comparison! The authors state that their algorithm is about 4 times faster than the baseline which generates about 5 times more pixels per image. Effectively, this result is worse.

qualitative examples

  • The only qualitative results of this paper are shown in fig 4. Why did the authors not provide more examples from more general tasks than interpolation between faces?

structure of the paper

  • In my opinion, the structure of the paper is a bit odd. The method section is about 4 pages long which is too long. Likewise, fig 1 add little or no value and can be removed. The extra space can then be used to move some of the experimental results to the main paper.

Results

  • At the moment, the main paper only contains fig 2 which is some sort of smoothness evaluation. What I can infer from that figure is that the proposed algorithm generates significantly smoother frames than prior works. Can this be achieved by taking a result from any of the baselines and running a smoothness filter over it? I think measures such as FID, etc, are more interesting.
  • According to tables 6-11, the proposed algorithm performs about same as good as the best competitor.
Questions for authors

Please add more qualitative examples from other types of transition.

Comments on the paper

1 comment
Area_Chair_zCsZOpenReview ↗

Closing discussion after the rebuttal

Thank you to the authors for the rebuttal and to the reviewers for engaging in the discussion. As we wrap up, I’d appreciate a brief final assessment from each reviewer. After reading the rebuttal and the discussion so far, what is the main reason you lean toward acceptance or rejection, and what evidence supports that view? Please note which concerns the rebuttal resolved and which remaining concerns, if any, are decisive for your recommendation.

If your assessment differs from another reviewer’s, please point to the specific claim or evidence you interpret differently. Authors are welcome to clarify any factual points that remain open. This will help us identify where we agree and understand the reasons for any remaining disagreement.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.