TCVBM: Time-Correlated Video Bridge Matching

• TCVBM is a framework that constructs stochastic bridges with explicit temporal correlations to synthesize and manipulate coherent video sequences.
• It employs time-correlated stochastic differential equations to jointly model inter-frame dependencies, ensuring smooth transitions in video tasks.
• Empirical results show TCVBM outperforms traditional diffusion and bridge-matching models in metrics like FVD, PSNR, and SSIM, demonstrating enhanced temporal realism.

Time-Correlated Video Bridge Matching (TCVBM) is a framework for modeling, manipulating, and generating temporally coherent video data by constructing stochastic bridges between video distributions with explicit temporal correlation modeling. Unlike classical diffusion or naïve bridge matching models, TCVBM directly incorporates inter-frame dependencies both in the prior process and in the dynamics of the video transformation, enabling high-fidelity and temporally smooth outcomes across canonical video tasks such as interpolation, image-to-video synthesis, and super-resolution (Vasilev et al., 14 Oct 2025).

1. Motivation and Problem Scope

Video generation and manipulation require not just high-quality spatial artifacts but also rigorous temporal coherence. Diffusion models such as DDPM and DDIM specialize in mapping noise to data distributions and excel at unconditional "imagination from scratch"; however, they are not tailored for structured translation tasks between complex data distributions (e.g., transforming a low-resolution video to high resolution, or filling in missing frames). Standard Bridge Matching (BM) methods generalize diffusion to data-to-data translations by constructing stochastic bridges, typically formulated between initial $p(x_0)$ and terminal $p(x_T|x_0)$ distributions. Naïve BM applied to video fails to enforce frame-to-frame consistency, leading to temporal flicker and incoherent motion.

TCVBM addresses this by formulating the translation task over the joint space of full video sequences, modeling the generative process as a single stochastic bridge endowed with explicit temporal correlation priors. This approach results in solutions that preserve inter-frame continuity, yielding outputs suitable for video-centric applications requiring both high spatial detail and stable motion dynamics (Vasilev et al., 14 Oct 2025).

2. Mathematical Foundations

2.1. Time-Correlated Stochastic Differential Equation (SDE)

A video of $N$ frames is denoted as $X = (x^1, \dots, x^N) \in \mathbb{R}^{N \times D}$, where each $x^n \in \mathbb{R}^D$. Temporal smoothness is encoded by evolving each feature-dimension $x_t^{(d)} \in \mathbb{R}^N$ under

$$d\,x_t^{(d)} = (A\,x_t^{(d)} + b^{(d)})\,dt + g(t)\,dW_t^{(d)},$$

where

• $A \in \mathbb{R}^{N \times N}$: symmetric tridiagonal matrix ($-2$ diagonal, $+1$ off-diagonal), enforcing temporal regularization so each frame moves toward the mean of its neighbors,
• $p(x_T|x_0)$0: boundary conditions,
• $p(x_T|x_0)$1: scalar (commonly constant),
• $p(x_T|x_0)$2: independent Wiener processes.

The system evolves in compact form as $p(x_T|x_0)$3, resulting in joint Gaussian marginals due to linearity.

2.2. Bridge Distribution

For $p(x_T|x_0)$4 fixed under the prior SDE, $p(x_T|x_0)$5, with

$p(x_T|x_0)$6

Conditioning further on $p(x_T|x_0)$7 yields the posterior bridge:

$p(x_T|x_0)$8

with explicit closed-form solutions for the mean and covariance, integrating inter-frame covariances through the structure of $p(x_T|x_0)$9 and the resulting $N$0.

2.3. Learning Objective

Given clean video $N$1 and a corresponding degraded (noisy) version $N$2, a bridge sample $N$3 is drawn for a random $N$4. The drift-predictor $N$5 is trained to match the score:

$N$6

By reparameterization (Proposition 3), this is equivalent to mean-square regression on a predicted clean frame $N$7:

$N$8

3. Algorithms and Model Architecture

3.1. Training and Inference Procedures

Training (Algorithm 1):

1. Sample $N$9, $X = (x^1, \dots, x^N) \in \mathbb{R}^{N \times D}$0.
2. Sample $X = (x^1, \dots, x^N) \in \mathbb{R}^{N \times D}$1.
3. Draw $X = (x^1, \dots, x^N) \in \mathbb{R}^{N \times D}$2.
4. Take a gradient step on $X = (x^1, \dots, x^N) \in \mathbb{R}^{N \times D}$3.

Inference (Algorithm 2):

Given $X = (x^1, \dots, x^N) \in \mathbb{R}^{N \times D}$4, backward-sample along the discretized $X = (x^1, \dots, x^N) \in \mathbb{R}^{N \times D}$5, recursively predicting $X = (x^1, \dots, x^N) \in \mathbb{R}^{N \times D}$6 and sampling $X = (x^1, \dots, x^N) \in \mathbb{R}^{N \times D}$7 from the conditional $X = (x^1, \dots, x^N) \in \mathbb{R}^{N \times D}$8 until reaching $X = (x^1, \dots, x^N) \in \mathbb{R}^{N \times D}$9.

3.2. Network Design

Standard 2D U-Nets are replaced or augmented to enable joint modeling across frames. Implementations include:

• 3D convolutions over indices $x^n \in \mathbb{R}^D$0,
• 2D conv blocks with cross-frame attention,
• Transformer-style self-attention on patch embeddings along the temporal axis.

Empirically, TCVBM uses a lightweight U-Net concatenating $x^n \in \mathbb{R}^D$1 frames as $x^n \in \mathbb{R}^D$2 channels, but more sophisticated spatiotemporal modules can target higher-complexity video tasks (Vasilev et al., 14 Oct 2025).

4. Experimental Evaluation and Metrics

TCVBM is empirically validated on the MovingMNIST dataset, using 10-frame subclips. The following video generation tasks are considered:

1. Frame Interpolation: generate middle frames given endpoints.
2. Image-to-Video: generate full sequence given initial frame.
3. Super-Resolution: generate high-resolution frames from low-res input.

Evaluation metrics:

• PSNR (↑): Per-frame reconstruction fidelity.
• SSIM (↑): Structural similarity index.
• LPIPS (↓): Learned perceptual image similarity.
• FVD (↓): Frechet Video Distance, measuring overall spatial/temporal realism.

All compared methods (TCVBM, DDPM, DDIM, BM) use the same training settings and architectural backbone; differences arise from loss formulation and sampling algorithm.

Representative Results

Metric	DDIM	DDPM	BM	TCVBM
Frame Interpolation (FVD↓)	33.6	32.4	34.3	30.5
Image-to-Video (FVD↓)	77.7	75.9	49.3	44.9
Video Super-Resolution (FVD↓)	334.7	607.9	53.5	59.5

TCVBM achieves improvements in both temporal realism (FVD, LPIPS) and per-frame accuracy (PSNR, SSIM), with the most substantial gains observed in FVD and LPIPS metrics (Vasilev et al., 14 Oct 2025).

5. Ablation Studies and Impact of Temporal Correlation Modeling

TCVBM was evaluated under multiple ablations:

• Different prior initializations (Gaussian, linear, static) confirm that noise-based initialization is robust.
• Hyperparameter sweeps show optimal trade-off with $x^n \in \mathbb{R}^D$3 and $x^n \in \mathbb{R}^D$4 scaling $x^n \in \mathbb{R}^D$5.
• Alternative time-varying correlation coefficients $x^n \in \mathbb{R}^D$6 provided no clear improvement over the baseline constant-$x^n \in \mathbb{R}^D$7 prior.

Visualizations show perceptually smoother motion, sharper details, and less flicker compared with all baselines, validating the utility of explicit temporal correlation in the bridge process.

6. Broader Context and Related Approaches

Alternative approaches to modeling time-correlated video data include multi-marginal Schrödinger bridges, such as in MMtSBM (Gravier et al., 2 Oct 2025), which construct dynamic transport plans matching distributions at multiple temporal marginals. MMtSBM operates by alternately fitting Markovian and reciprocal (bridge-filled) processes across all timepoints, but with more combinatorial complexity and different mathematical guarantees than TCVBM. Classical BM methods and standard diffusion models are limited by their independent, frame-wise generation and do not enforce sequence-level consistency (Vasilev et al., 14 Oct 2025).

TCVBM's contributions are thus situated at the intersection of score-based generative modeling, optimal transport, and temporal modeling—bridging the gap for temporally coherent video generation, interpolation, and super-resolution by explicitly modeling spatiotemporal statistical dependencies.

7. Limitations and Prospects

TCVBM, by design, targets translation tasks between distributions of entire video sequences; it inherits simplicity and efficiency from mean-squared regression training and backward-sampling inference. Its empirical validation currently centers on synthetic data (MovingMNIST); application to natural video would require scaling architectural capacity and memory-handling. The framework demonstrates robustness to initialization strategies and hyper-parameter selection. Future extensions may integrate advanced spatiotemporal architectures or adapt temporal correlation structures for higher realism in natural video domains (Vasilev et al., 14 Oct 2025).