Pith. sign in

REVIEW 4 major objections 6 minor 42 references

Reversible training eliminates long-range drift in video interpolation

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 16:53 UTC pith:M4DOZ4FO

load-bearing objection A plausible training-time regularizer for video interpolation, but the central ablation confounds bidirectional supervision with doubled data volume; the main claim needs a matched-data control. the 4 major comments →

arxiv 2604.01700 v2 pith:M4DOZ4FO submitted 2026-04-02 cs.CV cs.MM

Can Video Diffusion Models Predict Past Frames? Bidirectional Cycle Consistency for Reversible Interpolation

Lingyu Liu , Yaxiong Wang , Li Zhu , Zhedong Zheng This is my paper
classification cs.CV cs.MM
keywords video frame interpolationbidirectional trainingcycle consistencydirectional tokenstemporal symmetrycurriculum learninglong-range video generationflow matching
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that video frame interpolation models drift over long sequences because they are trained only to predict forward in time. The authors claim that adding a backward pass—training the same generative backbone to reconstruct the sequence from the end frame back to the start—acts as a regularizer that forces the model to learn reversible, coherent motion. Two learnable tokens tell the shared model which temporal direction to generate, and a curriculum trains on short clips before long ones. The result, they report, is state-of-the-art visual quality, motion smoothness, and long-range consistency on 37- and 73-frame interpolation tasks, with no extra inference cost since the cycle-consistency losses apply only during training.

Core claim

The central claim is that temporal cycle consistency—requiring the same model to return to the original start frame after generating forward and then backward—is a strong self-supervised regularizer for video interpolation. Concretely, the paper shows that appending backward reconstruction losses in both latent and pixel space, plus two learnable directional tokens, to a pretrained rectified-flow video backbone improves frame-level quality metrics and long-range consistency at both 37 and 73 frames while keeping inference identical to the base model. Their ablation attributes the largest benefit to the reverse training: without it, the dynamic-degree metric drops sharply (0.90 to 0.58 on 37

What carries the argument

The load-bearing mechanism is a pair of learnable directional tokens appended to the text conditioning of a shared diffusion-transformer backbone; the choice of token switches the same velocity-field network between forward and backward temporal generation. During training, each ground-truth clip is used twice—once in its original order and once time-reversed—and the total loss sums latent-space and pixel-space reconstruction errors for both directions. This forces the learned velocity field to be consistent under time reversal. A curriculum that first trains on short clips, then on long clips, stabilizes the objective and prevents the model from settling into a static-frame shortcut.

Load-bearing premise

The method assumes real-world video dynamics are approximately reversible—that training the model to generate from the end frame back to the start teaches it something useful; if the true dynamics are stochastic or dissipative, the reversal constraint can bias the model toward smoothed or physically wrong trajectories.

What would settle it

Train the same bidirectional framework on a dataset of irreversible processes (smoke dispersal, splashing water, shattering glass) and compare against a forward-only baseline; if the bidirectional model shows worse FVD/dynamic-degree on those clips, or if its forward-backward reconstruction error on real irreversible videos is substantially higher than on reversible ones, the reversibility assumption fails. More directly, compute the round-trip error: generate forward from A to B, then backward from B to A with the same model, and measure the distance between the reconstructed A and the origin

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the claim holds, any unidirectional video generator can be turned into an interpolator with reversible dynamics and no change to inference speed.
  • Bidirectional cycle-consistency can act as a regularization principle for long-horizon video generation, not just interpolation, anchoring generated trajectories to their endpoints.
  • Directional tokens give explicit temporal-orientation control, which could support applications such as consistent reverse playback or time-editing.
  • The short-to-long curriculum is a practical training recipe that avoids the still-video collapse common when training long-sequence video models from scratch.
  • Because the constraints are training-only, existing efficient autoregressive or flow-based samplers remain untouched at deployment.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The reversibility assumption limits the method to near-deterministic, entropy-preserving dynamics; for irreversible processes (smoke, splashes, shattering), the paper's own limitation note suggests cycle-consistency may bias outputs toward smoothed, physically wrong trajectories—an empirically testable prediction.
  • The same bidirectional regularizer could transfer to audio or time-series interpolation, where time-reversal symmetry is a better approximation than in complex video scenes.
  • Directional tokens might learn an implicit 'arrow-of-time' representation; probing them could show whether the model truly encodes temporal direction or merely memorizes a label.
  • A direct stress test is measuring the forward-backward round-trip error on videos with irreversible dynamics—if the model fails to return to the start frame, the reversibility prior is not doing the claimed work.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes a bidirectional cycle-consistent training framework for text-guided video frame interpolation. Built on pre-trained long-video diffusion/flow backbones (FramePack and Wan2.1-Fun), it introduces learnable directional tokens (τ_fwd, τ_bwd) to condition a shared backbone on temporal orientation, and trains the model jointly on forward and time-reversed sequences with latent- and pixel-space reconstruction losses. A curriculum schedule first trains on 37-frame clips then fine-tunes on 73-frame clips. At inference, only a single forward pass is used, so the method adds no computational overhead relative to its backbone. Experiments report FVD and VBench metrics on 37- and 73-frame interpolation, claim state-of-the-art quality, and include ablations removing reverse training, directional tokens, and the curriculum.

Significance. If the central claim is validated, the proposal is a practical and low-cost regularizer: cycle-consistent bidirectional training improves long-range interpolation quality without increasing inference cost. The idea of using temporal symmetry as a training-only constraint is well motivated, and the paper provides a clean architectural mechanism (directional tokens) and a clear experimental protocol. The manuscript also includes qualitative demonstrations that the full model produces more coherent motion than unidirectional baselines. However, the empirical evidence for the central attribution — that cycle consistency, rather than additional training data or length mismatch, drives the gains — is currently inconclusive due to a confounded ablation, a small evaluation set without significance testing, and ambiguous train/evaluation disjointness.

major comments (4)
  1. [§4.3, Table 2] The 'w/o Reverse Training' ablation is confounded with data volume. In the full model, each ground-truth video contributes two training samples (forward and backward), while the forward-only ablation contributes one. Under the stated 'all other settings remain identical' protocol (same 4 epochs, same batch size), the full model sees roughly twice as many gradient updates. The large FVD/dynamic-degree gaps (e.g., 885 vs. 937 at 37 frames; 0.90 vs. 0.58 dynamic degree) could therefore reflect additional training samples rather than the cycle-consistency constraint. A matched-data control is needed — e.g., training forward-only on the same number of samples (by duplicating forward sequences or sampling additional videos) — or a variant that trains on reversed sequences without the bidirectional consistency objective. Without such a control, the paper's central attribution of gains to cycle-
  2. [§4.1, §4.2, Tables 1–2] All quantitative claims rest on a single evaluation set of only 100 videos, with no error bars, no multiple random seeds, and no statistical significance tests. Many reported differences are small in absolute terms (e.g., Table 1: FVD 885 vs. 901, VBench imaging quality 0.6024 vs. 0.5941; Table 2: FVD 885 vs. 901, temporal flickering 0.9877 vs. 0.9774). Given the noise typical of generative video evaluation, the claimed state-of-the-art improvements may not be distinguishable from run-to-run variation. The authors should report standard deviations across seeds, bootstrap confidence intervals, or per-video paired significance tests. The 100-video evaluation also limits the generalizability of the conclusions.
  3. [§4.1, Datasets] Training and evaluation sets are both sampled from VidGen-1M, but the manuscript does not explicitly state that the 100 evaluation videos are disjoint from the 5,000 training videos. The phrase 'we further sample additional 100 high-quality videos from VidGen-1M' suggests a separate sample, but it is not guaranteed to be non-overlapping. If any evaluation videos overlap with training, the reported FVD and VBench numbers are inflated and the comparison to baselines is invalid. The authors must state explicitly that the evaluation set is disjoint from the training set and, ideally, verify no near-duplicate content.
  4. [§4.2, Table 1] The baseline comparison is complicated by mismatched sequence lengths: GI is evaluated at 25 frames, SFI at 33/65 frames, while the proposed methods are evaluated at 37/73 frames. FVD and VBench metrics, especially dynamic degree and motion smoothness, are sensitive to sequence length and frame rate. The paper acknowledges these length differences in the text but does not control for their effect. For a fair 'state-of-the-art' claim, the authors should either evaluate all methods at matched output lengths (e.g., generate 37/73 frames from GI and SFI where possible) or provide a length-controlled comparison that isolates the effect of the proposed training objective from the effect of generating longer sequences.
minor comments (6)
  1. [§3.3, Eq. (3)] The ground-truth latent for the backward direction, x^*_bwd, is used in Eq. (3) but never defined. It should be defined as the latent encoding of Reverse(V^*_fwd), consistent with the pixel-space definition V^*_bwd = Reverse(V^*_fwd).
  2. [§4.3, Table 2 caption] The caption says 'five VBench metrics' but the table lists six metrics (Subject Consistency, Aesthetic Quality, Imaging Quality, Temporal Flickering, Motion Smoothness, Dynamic Degree). Please correct the caption or the table.
  3. [§4.2, Figure 4] The text says the VBench score in Figure 4 is the average of six perceptual metrics, but the ablation table refers to five. Please ensure consistent counting of the VBench dimensions.
  4. [§3.4, Multi-Rate Temporal Sampling] The description says the 37-frame and 73-frame variants of the same clip 'share identical start and end frames.' If the same clip is temporally resampled to different lengths, it is unclear how both can have literally identical end frames unless the end frame is defined as the last frame of the resampled clip. Please clarify the construction.
  5. [§2, Related Work] Reference [28] (Framer) is cited in the introduction and related work but has no publication venue in the reference list. Also, the claim that 'our work is the first to adapt such a long video generator for bidirectional interpolation' is strong given that [15,29,34] also use backward/time-reversed generation; please soften or more precisely distinguish the contribution.
  6. [§4.2, Baselines] Table 1 reports parameter counts and frame lengths, but not inference times in the table. Figure 4 provides a trade-off plot, but no numerical latency values are given; adding a table with mean inference time and standard deviation would strengthen the 'no extra computational cost' claim.

Circularity Check

0 steps flagged

No circular derivation: benchmark-validated training recipe with no fitted-then-predicted quantities.

full rationale

The paper's derivation chain is an empirical training recipe. Section 3.3 defines a four-term loss (Eq. 3) that supervises forward and backward reconstructions against ground-truth latent and pixel sequences (V*_bwd = Reverse(V*_fwd)); the directional tokens in Eq. (1)-(2) are learned parameters, not fitted constants that are later reported as predictions. The central claims are validated by Table 1's FVD/VBench comparisons on separately sampled evaluation videos from VidGen-1M, and by the Table 2 ablation. No equation in the paper reduces to its own inputs: nothing is fit to the evaluation metrics and then 'predicted'; no uniqueness theorem or prior result by the authors is invoked to force the architecture. The citation to cycle-consistency of time [30] is used as inspiration, not as load-bearing evidence, and the only overlapping-author citation (Open-SORA [41]) appears in related work without argumentative weight. The acknowledged reversibility assumption in the Limitations is an honest boundary condition, not circularity. The skeptic's concern that the 'w/o Reverse Training' ablation also halves the number of training samples is a real experimental confound, but it concerns attribution of the improvement to cycle-consistency vs. data volume; it is not a case where a prediction is equivalent by construction to a fitted input. Therefore no circular step is identifiable.

Axiom & Free-Parameter Ledger

5 free parameters · 3 axioms · 1 invented entities

The framework rests on hyperparameters chosen by hand (LoRA rank, learning rates, curriculum lengths, dataset subsampling) and on the domain assumption that video dynamics are approximately reversible. The learnable directional tokens are internal model parameters, not physically grounded entities. The central empirical claim depends on the evaluation set overlapping in distribution with the training set, which is an unverified assumption about representativeness.

free parameters (5)
  • LoRA rank = 64
    Hyperparameter controlling adapter capacity; chosen by hand without reported sweep, and central to the fine-tuning claim.
  • Directional token learning rate = 2e-3
    Hand-chosen learning rate for τ_fwd and τ_bwd; part of the training recipe but no sensitivity analysis is given.
  • Curriculum sequence lengths = 37 frames then 73 frames
    Fixed durations matching FramePack's packing; the curriculum design is a central component and the specific lengths are chosen, not derived.
  • Training set size = 5,000 videos from VidGen-1M
    Subsampling based on three VBench metrics; exact thresholds are not given, making the training-set construction a hand-chosen free parameter.
  • Training epochs / batch size = 4 epochs, batch size 6
    Training hyperparameters; no ablation is reported for them.
axioms (3)
  • domain assumption Approximate reversibility of real-world video dynamics
    The cycle-consistency objective assumes a forward trajectory can be reversed to recover the start; the Limitations section explicitly acknowledges this may fail for stochastic or irreversible processes.
  • domain assumption Time-reversed videos are valid learning signals
    Training on reversed ground-truth clips assumes that reversed dynamics are a plausible target; many real-world clips (e.g., gravity-driven falls) are physically implausible when reversed.
  • domain assumption Evaluation videos from the same dataset as training are representative
    The 100 evaluation videos are drawn from VidGen-1M, and overlap with the 5,000 training videos is not ruled out; in-distribution performance may overstate generalization.
invented entities (1)
  • Learnable directional tokens τ_fwd and τ_bwd no independent evidence
    purpose: Condition the shared backbone on temporal orientation (forward vs. backward) during training and, at inference, select the forward direction.
    These are newly introduced learned embeddings with no external falsifiable handle; their only evidence is the paper's in-house experiments on sampled data.

pith-pipeline@v1.3.0-alltime-deepseek · 11154 in / 11060 out tokens · 109614 ms · 2026-08-02T16:53:12.132930+00:00 · methodology

0 comments
read the original abstract

Video frame interpolation aims to synthesize realistic intermediate frames between given endpoints while adhering to specific motion semantics. While recent generative models have improved visual fidelity, they predominantly operate in a unidirectional manner, lacking mechanisms to self-verify temporal consistency. This often leads to motion drift, directional ambiguity, and boundary misalignment, especially in long-range sequences. Inspired by the principle of temporal cycle-consistency in self-supervised learning, we propose a novel bidirectional framework that enforces symmetry between forward and backward generation trajectories. Our approach introduces learnable directional tokens to explicitly condition a shared backbone on temporal orientation, enabling the model to jointly optimize forward synthesis and backward reconstruction within a single unified architecture. This cycle-consistent supervision acts as a powerful regularizer, ensuring that generated motion paths are logically reversible. Furthermore, we employ a curriculum learning strategy that progressively trains the model from short to long sequences, stabilizing dynamics across varying durations. Crucially, our cyclic constraints are applied only during training; inference requires a single forward pass, maintaining the high efficiency of the base model. Extensive experiments show that our method achieves state-of-the-art performance in imaging quality, motion smoothness, and dynamic control on both 37-frame and 73-frame tasks, outperforming strong baselines while incurring no additional computational overhead.

Figures

Figures reproduced from arXiv: 2604.01700 by Lingyu Liu, Li Zhu, Yaxiong Wang, Zhedong Zheng.

Figure 1
Figure 1. Figure 1: Cycle-consistency of Time. Given identical start/end frames, we test temporal symmetry by generating (top) a forward sequence and (bottom) its time-reversed coun￾terpart via swapped endpoints. The baseline fails to synthesize true backward motion and instead resolves the constraint via a directional flip, where the dog re-orients to walk forward. In contrast, our model achieves robust cycle-consistency. It… view at source ↗
Figure 2
Figure 2. Figure 2: A brief overview of our framework. During training, each ground-truth video is used to construct two samples. The forward sample interpolates from the original start frame to the original end frame. The backward sample interpolates in the reverse temporal direction, starting from the original end frame and ending at the original start frame. These two directions are controlled by distinct learnable directi… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative Comparisons with Baselines. Our methods (Wan+Ours and FP+Ours) achieve significantly smoother trajectories and coherent temporal dynamics on both short videos (37 frames) and long videos (73 frames). Videos can be viewed in our supplementary material. than baselines, resulting in videos that are not only visually high-quality but also temporally stable. Efficiency [PITH_FULL_IMAGE:figures/full… view at source ↗
Figure 4
Figure 4. Figure 4: Efficiency vs. Performance. Models closer to the top-left corner exhibit faster inference and higher quality. Circle area indicates model parameter size. Our method is complementary to the existing video interpolation models, outperforming baselines in terms of VBench score and achieving a higher average quality across the evaluated metrics, while maintaining the same inference time as their backbones. and… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative ablation study of key components in our framework. Our full model generates a fluid motion sequence that naturally evolves from a “sliding” preparation into a full “jump”, ensuring high temporal coherence and dynamic realism. Videos can be viewed in our supplementary material. successfully synthesizes a coherent trajectory, seamlessly transitioning from the initial “sliding” phase into the subs… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

42 extracted references · 7 linked inside Pith

  1. [1]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Bao, W., Lai, W.S., Ma, C., Zhang, X., Gao, Z., Yang, M.H.: Depth-aware video frame interpolation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3703–3712 (2019)

    2019

  2. [2]

    arXiv preprint arXiv:2311.15127 (2023)

    Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023)

    Pith/arXiv arXiv 2023

  3. [3]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion mod- els. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 22563–22575 (2023)

    2023

  4. [4]

    arXiv e-prints pp

    Chen, L., Cun, X., Li, X., He, X., Yuan, S., Chen, J., Shan, Y., Yuan, L.: Sci-fi: Symmetric constraint for frame inbetweening. arXiv e-prints pp. arXiv–2505 (2025)

    2025

  5. [5]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Danier, D., Zhang, F., Bull, D.: Ldmvfi: Video frame interpolation with latent diffusion models. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 1472–1480 (2024)

    2024

  6. [6]

    IEEE Transactions on Multimedia26, 9187–9202 (2024)

    Ding, C., Lin, M., Zhang, H., Liu, J., Yu, L.: Video frame interpolation with stereo event and intensity cameras. IEEE Transactions on Multimedia26, 9187–9202 (2024)

    2024

  7. [7]

    In: European Conference on Computer Vision

    Feng, H., Ding, Z., Xia, Z., Niklaus, S., Abrevaya, V., Black, M.J., Zhang, X.: Explorative inbetweening of time and space. In: European Conference on Computer Vision. pp. 378–395. Springer (2024)

    2024

  8. [8]

    Advances in neural information processing systems35, 8633– 8646 (2022)

    Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models. Advances in neural information processing systems35, 8633– 8646 (2022)

    2022

  9. [9]

    arXiv preprint arXiv:2507.05173 (2025)

    Hong, Y., Zhang, J., Yi, R., Wang, Y., Cao, W., Hu, X., Xue, Z., Wang, Y., Wang, C., Ma, L.: Semantic frame interpolation. arXiv preprint arXiv:2507.05173 (2025)

    Pith/arXiv arXiv 2025

  10. [10]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Hu, M., Jiang, K., Zhong, Z., Wang, Z., Zheng, Y.: Iq-vfi: Implicit quadratic mo- tion estimation for video frame interpolation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6410–6419 (2024)

    2024

  11. [11]

    In: European conference on computer vision

    Huang, Z., Zhang, T., Heng, W., Shi, B., Zhou, S.: Real-time intermediate flow estimation for video frame interpolation. In: European conference on computer vision. pp. 624–642. Springer (2022)

    2022

  12. [12]

    In: Proceedings of the 32nd ACM International Conference on Multimedia

    Huang, Z., Yu, Y., Yang, L., Qin, C., Zheng, B., Zheng, X., Zhou, Z., Wang, Y., Yang, W.: Motion-aware latent diffusion models for video frame interpolation. In: Proceedings of the 32nd ACM International Conference on Multimedia. pp. 1043– 1052 (2024) 16 Liu et al

    2024

  13. [13]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)

    Huang, Z., He, Y., Yu, J., Zhang, F., Si, C., Jiang, Y., Zhang, Y., Wu, T., Jin, Q., Chanpaisit, N., Wang, Y., Chen, X., Wang, L., Lin, D., Qiao, Y., Liu, Z.: VBench: Comprehensive benchmark suite for video generative models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)

    2024

  14. [14]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Jain, S., Watson, D., Tabellion, E., Poole, B., Kontkanen, J., et al.: Video inter- polation with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7341–7351 (2024)

    2024

  15. [15]

    In: The Fourteenth International Conference on Learning Representations (2026)

    Jeon, W., Shin, S., Shin, D., Jeon, H.G.: Motion prior distillation in time reversal sampling for generative inbetweening. In: The Fourteenth International Conference on Learning Representations (2026)

    2026

  16. [16]

    arXiv preprint arXiv:2412.03603 (2024)

    Kong, W., Tian, Q., Zhang, Z., Min, R., Dai, Z., Zhou, J., Xiong, J., Li, X., Wu, B., Zhang, J., et al.: Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603 (2024)

    Pith/arXiv arXiv 2024

  17. [17]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Li, Z., Zhu, Z.L., Han, L.H., Hou, Q., Guo, C.L., Cheng, M.M.: Amt: All-pairs multi-field transforms for efficient frame interpolation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9801– 9810 (2023)

    2023

  18. [18]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Liu, C., Zhang, G., Zhao, R., Wang, L.: Sparse global matching for video frame interpolation with large motion. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 19125–19134 (2024)

    2024

  19. [19]

    In: Proceedings of the IEEE international conference on computer vision

    Niklaus, S., Mai, L., Liu, F.: Video frame interpolation via adaptive separable convolution. In: Proceedings of the IEEE international conference on computer vision. pp. 261–270 (2017)

    2017

  20. [20]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Pardo, A., Pizzati, F., Zhang, T., Pondaven, A., Torr, P., Perez, J.C., Ghanem, B.: Matchdiffusion: Training-free generation of match-cuts. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14973–14982 (2025)

    2025

  21. [21]

    In: 2024 International Conference on Knowledge Engi- neering and Communication Systems (ICKECS)

    Prasanna, B., Niranjan, S., et al.: Video frame interpolation using real-time inter- mediate flow estimation. In: 2024 International Conference on Knowledge Engi- neering and Communication Systems (ICKECS). vol. 1, pp. 1–5. IEEE (2024)

    2024

  22. [22]

    In: European Conference on Computer Vi- sion

    Reda, F., Kontkanen, J., Tabellion, E., Sun, D., Pantofaru, C., Curless, B.: Film: Frame interpolation for large motion. In: European Conference on Computer Vi- sion. pp. 250–266. Springer (2022)

    2022

  23. [23]

    In: European Conference on Computer Vision

    Shen, L., Liu, T., Sun, H., Ye, X., Li, B., Zhang, J., Cao, Z.: Dreammover: Lever- aging the prior of diffusion models for image interpolation with large motion. In: European Conference on Computer Vision. pp. 336–353. Springer (2024)

    2024

  24. [24]

    arXiv preprint arXiv:2408.02629 (2024)

    Tan, Z., Yang, X., Qin, L., Li, H.: Vidgen-1m: A large-scale dataset for text-to- video generation. arXiv preprint arXiv:2408.02629 (2024)

    Pith/arXiv arXiv 2024

  25. [25]

    Unterthiner, T., Van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Fvd: A new metric for video generation (2019)

    2019

  26. [26]

    Advances in neural information processing systems35, 23371–23385 (2022)

    Voleti, V., Jolicoeur-Martineau, A., Pal, C.: Mcvd-masked conditional video diffu- sion for prediction, generation, and interpolation. Advances in neural information processing systems35, 23371–23385 (2022)

    2022

  27. [27]

    arXiv preprint arXiv:2503.20314 (2025)

    Wan, T., Wang, A., Ai, B., Wen, B., Mao, C., Xie, C.W., Chen, D., Yu, F., Zhao, H., Yang, J., et al.: Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314 (2025)

    Pith/arXiv arXiv 2025

  28. [28]

    In: The Thirteenth International Conference on Learning Representations

    Wang, W., Wang, Q., Zheng, K., OUYANG, H., Chen, Z., Gong, B., Chen, H., Shen, Y., Shen, C.: Framer: Interactive frame interpolation. In: The Thirteenth International Conference on Learning Representations

  29. [29]

    arXiv preprint arXiv:2408.15239 (2024) Title Suppressed Due to Excessive Length 17

    Wang, X., Zhou, B., Curless, B., Kemelmacher-Shlizerman, I., Holynski, A., Seitz, S.M.: Generative inbetweening: Adapting image-to-video models for keyframe in- terpolation. arXiv preprint arXiv:2408.15239 (2024) Title Suppressed Due to Excessive Length 17

    Pith/arXiv arXiv 2024

  30. [30]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Wang, X., Jabri, A., Efros, A.A.: Learning correspondence from the cycle- consistency of time. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2566–2576 (2019)

    2019

  31. [31]

    ACM Transactions on Graphics (TOG)43(6), 1–11 (2024)

    Xing, J., Liu, H., Xia, M., Zhang, Y., Wang, X., Shan, Y., Wong, T.T.: Tooncrafter: Generative cartoon interpolation. ACM Transactions on Graphics (TOG)43(6), 1–11 (2024)

    2024

  32. [32]

    In: European Conference on Computer Vision

    Xing, J., Xia, M., Zhang, Y., Chen, H., Yu, W., Liu, H., Liu, G., Wang, X., Shan, Y., Wong, T.T.: Dynamicrafter: Animating open-domain images with video diffusion priors. In: European Conference on Computer Vision. pp. 399–417. Springer (2024)

    2024

  33. [33]

    ACM computing surveys56(4), 1–39 (2023)

    Yang, L., Zhang, Z., Song, Y., Hong, S., Xu, R., Zhao, Y., Zhang, W., Cui, B., Yang, M.H.: Diffusion models: A comprehensive survey of methods and applications. ACM computing surveys56(4), 1–39 (2023)

    2023

  34. [34]

    In: The Thirteenth International Conference on Learning Representations

    Yang, S., Kwon, T., Ye, J.C.: Vibidsampler: Enhancing video interpolation using bidirectional diffusion sampler. In: The Thirteenth International Conference on Learning Representations

  35. [35]

    In: The Thirteenth International Conference on Learning Representations

    Yang, Z., Teng, J., Zheng, W., Ding, M., Huang, S., Xu, J., Yang, Y., Hong, W., Zhang, X., Feng, G., et al.: Cogvideox: Text-to-video diffusion models with an expert transformer. In: The Thirteenth International Conference on Learning Representations

  36. [36]

    In: Proceedings of the IEEE/CVF International Confer- ence on Computer Vision

    Yang, Z., Zhang, J., Yu, Y., Lu, S., Bai, S.: Versatile transition generation with image-to-video diffusion. In: Proceedings of the IEEE/CVF International Confer- ence on Computer Vision. pp. 16981–16990 (2025)

    2025

  37. [37]

    In: The Fourteenth International Conference on Learning Representations (2026)

    Zhang, G., Wang, H., Wang, C., Zhou, Y., Lu, Q., Wang, L.: Arbitrary genera- tive video interpolation. In: The Fourteenth International Conference on Learning Representations (2026)

    2026

  38. [38]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Zhang, G., Zhu, Y., Wang, H., Chen, Y., Wu, G., Wang, L.: Extracting motion and appearance via inter-frame attention for efficient video frame interpolation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5682–5692 (2023)

    2023

  39. [39]

    arXiv preprint arXiv:2504.12626 (2025)

    Zhang, L., Agrawala, M.: Packing input frame context in next-frame prediction models for video generation. arXiv preprint arXiv:2504.12626 (2025)

    arXiv 2025

  40. [40]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Zhang, Z., Chen, H., Zhao, H., Lu, G., Fu, Y., Xu, H., Wu, Z.: Eden: Enhanced diffusion for high-quality large-motion video frame interpolation. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 2105–2115 (2025)

    2025

  41. [41]

    arXiv preprint arXiv:2412.20404 (2024)

    Zheng, Z., Peng, X., Yang, T., Shen, C., Li, S., Liu, H., Zhou, Y., Li, T., You, Y.: Open-sora: Democratizing efficient video production for all. arXiv preprint arXiv:2412.20404 (2024)

    Pith/arXiv arXiv 2024

  42. [42]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Zhu, T., Ren, D., Wang, Q., Wu, X., Zuo, W.: Generative inbetweening through frame-wise conditions-driven video generation. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 27968–27978 (2025)

    2025