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

Lingyu Liu; Yaxiong Wang; Li Zhu; Zhedong Zheng

cs.CVApr 2, 2026

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

Authors: Lingyu Liu, Yaxiong Wang, Li Zhu, Zhedong Zheng

Organizations: School of Software, Xi’an Jiao-tong University, Xi’an, 710049, China. · School of Computer and Information Science, Hefei University of Technology, Jianghuai Advance Technology Center, Anhui Provincial Key Laboratory of Humanoid Robots. Anhui Provincial Industry Innovation Center of Humanoid Robots, Hefei, 230000, China. · Faculty of Science and Technology, and Institute of Collaborative Innovation, University of Macau, Macau, 999078, China.

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 & tables

Fig. 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 counterpart 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 captures authentic reverse dynamics, producing a coherent backward-walking video that preserves the subject’s orientation, thereby closing the temporal loop faithfully.

Fig. 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 directional tokens. The model is supervised with reconstruction losses in both latent space and pixel space for both directions, which encourages consistent motion modeling under time reversal.

Fig. 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 project website.

Method	Backbone Params	Frames	FVD $\downarrow$	Subject Consistency $\uparrow$	Aesthetic Quality $\uparrow$	Imaging Quality $\uparrow$	Temporal Flickering $\uparrow$	Motion Smoothness $\uparrow$	Dynamic Degree $\uparrow$	VBench Mean $\uparrow$
Ground-Truth	-	37	-	0.884	0.441	0.636	0.949	0.987	0.98	0.81
GI	1.5B	25	698	0.921	0.424	0.593	0.966	0.985	0.63	0.75
ViBidSampler	1.5B	25	693	0.863	0.419	0.540	0.944	0.962	0.96	0.78
FCVG	1.5B	25	992	0.906	0.442	0.554	0.967	0.989	0.63	0.75
Framer	1.5B	25	806	0.882	0.389	0.525	0.961	0.980	0.85	0.76
Wan2.1-Fun (Wan)	1.3B	37	984	0.876	0.451	0.584	0.948	0.965	0.96	0.80

TABLE I: Quantitative Comparisons with Baselines on VidGen-1M. Results are reported using FVD (lower is better) and six VBench metrics (higher is better). FP+finetuned denotes the FramePack backbone fine-tuned on the same data volume as our full model but using only forward videos. Our full models achieve the highest average VBench scores across both 37-frame and 73-frame tasks, closely matching the Ground-Truth distribution and outperforming all baselines in imaging quality and motion smoothness.

Method	Backbone Params	Frames	FVD $\downarrow$	Subject Consistency $\uparrow$	Aesthetic Quality $\uparrow$	Imaging Quality $\uparrow$	Temporal Flickering $\uparrow$	Motion Smoothness $\uparrow$	Dynamic Degree $\uparrow$	VBench Mean $\uparrow$
Ground-Truth	-	37	-	0.936	0.555	0.712	0.980	0.982	0.86	0.84
GI	1.5B	25	674	0.952	0.534	0.699	0.959	0.985	0.43	0.76
ViBidSampler	1.5B	25	693	0.924	0.534	0.657	0.947	0.973	0.75	0.80
FCVG	1.5B	25	789	0.945	0.545	0.677	0.965	0.987	0.36	0.75
Framer	1.5B	25	702	0.933	0.487	0.640	0.969	0.986	0.47	0.75
Wan2.1-Fun (Wan)	1.3B	37	622	0.944	0.531	0.674	0.960	0.985	0.65	0.79

TABLE II: Zero-Shot quantitative evaluation on the unseen UltraVideo dataset. The results highlight the robust generalization of our method across different video distributions. Our methods achieve state-of-the-art scores in Motion Smoothness and Dynamic Degree , confirming their ability to generate coherent and dynamic videos even in zero-shot scenarios.

Fig. 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 outperforms the baselines in VBench score while preserving the same inference efficiency as the original backbones.

Fig. 5: User Study. In both short and long scenarios, our approach is consistently preferred by a large margin, confirming its effectiveness in producing visually appealing and temporally stable videos.

Frames	Setting	FVD $\downarrow$	Subject Consistency $\uparrow$	Aesthetic Quality $\uparrow$	Imaging Quality $\uparrow$	Temporal Flickering $\uparrow$	Motion Smoothness $\uparrow$	Dynamic Degree $\uparrow$	VBench Mean $\uparrow$
37	w/o Reverse Training	937	0.876	0.466	0.588	0.987	0.992	0.58	0.75
	Forward Data Augmentation	932	0.880	0.461	0.553	0.988	0.991	0.59	0.74
	w/o Directional Tokens	1018	0.899	0.470	0.590	0.975	0.990	0.67	0.77
	w/o Pixel Loss	897	0.866	0.462	0.548	0.964	0.982	0.89	0.79
	Mixed Training Length	901	0.883	0.466	0.594	0.977	0.988	0.87	0.80
	Full model	885	0.885	0.465	0.602	0.988	0.992	0.90	0.81

TABLE III: Quantitative ablation study of key components in our framework. We evaluate all variants on 37-frame and 73-frame interpolation task using FVD (lower is better) and five VBench metrics (higher is better). Our full model achieves the best overall performance, with consistent improvements in imaging quality and dynamic degree.

Fig. 6: 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 project website.

Frames	Methods	Videos	VidGen-1M			UltraVideo
			FVD $\downarrow$	VBench $\uparrow$	$\mathcal{C}_{\ell_{1}}\downarrow$	FVD $\downarrow$	VBench $\uparrow$	$\mathcal{C}_{\ell_{1}}\downarrow$
37	Baseline	$\hat{\mathcal{V}}_{\mathrm{fwd}}$	1049	0.79	0.141	878	0.80	0.094
		$\mathrm{Reverse}(\hat{\mathcal{V}}_{\mathrm{bwd}})$	1075	0.76		910	0.79
	Ours	$\hat{\mathcal{V}}_{\mathrm{fwd}}$	885	0.81	0.108	477	0.83	0.057
		$\mathrm{Reverse}(\hat{\mathcal{V}}_{\mathrm{bwd}})$	987	0.77		625	0.83
73	Baseline	$\hat{\mathcal{V}}_{\mathrm{fwd}}$	686	0.79	0.132	565	0.80	0.086

TABLE IV: Evaluation of forward and backward generation.

\hat{\mathcal{V}}_{\mathrm{fwd}}

and

\hat{\mathcal{V}}_{\mathrm{bwd}}

denote the generated forward and backward sequences; backward videos are temporally reversed before evaluation against the forward ground truth.

\mathcal{C}_{\ell_{1}}

is the per-pixel mean absolute error between

\hat{\mathcal{V}}_{\mathrm{fwd}}

and

\mathrm{Reverse}(\hat{\mathcal{V}}_{\mathrm{bwd}})

of the same input pair (pixel range

[0,1]

), serving as an empirical measure of the cycle discrepancy in Eq. 4 . For the baseline, backward generation is obtained by swapping the endpoint conditions.

Fig. 7: T-SNE visualization of directional tokens. The forward and backward tokens appear as two isolated points, located far from the dense cluster of standard text embeddings. Their near-orthogonal relationship (cosine similarity $=0.0221$ ) confirms that our model decouples temporal direction from semantic content.

Fig. 8: Effect of $\lambda_{\text{rev}}$ on generation quality. $\lambda_{\text{rev}}=0$ yields static videos, $\lambda_{\text{rev}}=1$ achieves the best dynamic degree (0.90), while $\lambda_{\text{rev}}=10$ over-constrains and degrades performance.

Video Frame Interpolation

Video frame interpolation estimates intermediate frames between observed video frames to create smoother playback and reconstruct temporal detail.

See all 26 papers

Mar 5, 2026

FC-VFI: Faithful and Consistent Video Frame Interpolation for High-FPS Slow Motion Video Generation

Large pre-trained video diffusion models excel in video frame interpolation but struggle to generate high fidelity frames due to reliance on intrinsic generative priors, limiting detail preservation from start and end frames. Existing methods often depend on motion control for temporal consistency, yet dense optical flow is error-prone, and sparse points lack structural context. In this paper, we propose FC-VFI for faithful and consistent video frame interpolation, supporting

4\times

x and

8\times

interpolation, boosting frame rates from 30 FPS to 120 and 240 FPS at

2560\times 1440

resolution while preserving visual fidelity and motion consistency. We introduce a temporal modeling strategy on the latent sequences to inherit fidelity cues from start and end frames and leverage semantic matching lines for structure-aware motion guidance, improving motion consistency. Furthermore, we propose a temporal difference loss to mitigate temporal inconsistencies. Extensive experiments show FC-VFI achieves high performance and structural integrity across diverse scenarios.

Ganggui Ding, Hao Chen, Xiaogang Xu

⋆Zhejiang University · † Chinese University of Hong Kong

Video Frame Interpolation Video Diffusion Models Temporal Consistency in Video Generation

Ask PDF

Jul 13, 2026

Cycle-World: Mitigating Error Accumulation in Long-term Video World Models via Reverse-Prediction Cycle Consistency

Autoregressive diffusion models have enabled high-quality video generation, yet their sequential nature inherently suffers from error accumulation. In long-horizon video synthesis, minor prediction deviations compound over time, inevitably leading to unconstrained generative drift, structural collapse, and severe visual degradation. To address this, we propose Cycle-World, a novel framework designed for stable and temporally consistent long-video generation. Our approach tackles error drift by enforcing strict temporal reversibility across both the training and inference phases. Theoretically, we demonstrate that forward generative drift can be strictly bottlenecked by a cycle-consistency objective. During training, we integrate an efficient reverse-prediction model to implicitly embed causal constraints into the forward generator, compelling it to produce reversible sequences that tightly adhere to the natural video manifold. At inference time, we repurpose this frozen reverse model as a runtime corrector. Through gradient-based cycle guidance, it iteratively refines the generated latent representations, actively suppressing accumulated errors before they are committed to the historical context. Extensive experiments on the VBench benchmark demonstrate that Cycle-World's dual-phase synergy significantly mitigates error drift, achieving state-of-the-art overall generation quality and long-horizon temporal consistency in 60-second synthesis.

Zihan Su, Teng Hu, Jiangning Zhang +4

School of Computer Science, Shanghai Jiao Tong University, Shanghai, China · Institute of Cyber-Systems and Control, Zhejiang University, Hangzhou, China · Nanyang Technological University, Singapore

Video Diffusion Models Long-Horizon Video Generation Autoregressive Video Generation Temporal Consistency in Video Generation

Ask PDF

Aug 13, 2026

SNM-VFI: Symmetric Nonlinear Motion-Guided Generative Video Frame Interpolation

We propose Symmetric Nonlinear Motion-guided Generative Video Frame Interpolation (SNM-VFI), a training-free framework for motion-controllable generative video frame interpolation with pre-trained optical flow and video diffusion models. Unlike conventional diffusion-based VFI methods that synthesize intermediate frames from random noise, SNM-VFI guides the generative process with correspondence-aware frames produced by a symmetric nonlinear motion model. Specifically, we first utilize a pre-trained optical flow model to construct multi-frame nonlinear flow-based intermediate frames and confidence maps. These flow-guided frames are then encoded as latent priors to initialize and iteratively guide a pre-trained Video Diffusion model, enabling the diffusion model to preserve dense motion correspondence while improving perceptual realism. To further enhance output quality, we employ confidence maps to fuse structurally reliable flow-based predictions with diffusion-generated details in uncertain regions such as occlusions and object boundaries. Extensive evaluations on challenging benchmarks, including DAVIS, Sintel, and KITTI, demonstrate that SNM-VFI achieves strong perceptual quality, competitive reconstruction accuracy, and robust temporal coherence across diverse motion scenarios.

Jisoo Jeong, Hong Cai, Jamie Menjay Lin +5

Qualcomm AI Research · Qualcomm Technologies, Inc.

Video Diffusion Models Diffusion Guidance Video Frame Interpolation

Ask PDF

Abstract

Video Frame Interpolation

Explore similar work

FC-VFI: Faithful and Consistent Video Frame Interpolation for High-FPS Slow Motion Video Generation

Cycle-World: Mitigating Error Accumulation in Long-term Video World Models via Reverse-Prediction Cycle Consistency

SNM-VFI: Symmetric Nonlinear Motion-Guided Generative Video Frame Interpolation