Motion Attribution for Video Generation

视频生成的运动归因

Motion Attribution for Video Generation

Xindi Wu Affiliation: NVIDIA Affiliation: Princeton University    Despoina Paschalidou Affiliation: NVIDIA    Jun Gao Affiliation: NVIDIA Affiliation: University of Michigan    Antonio Torralba Affiliation: MIT    Laura Leal-Taixé Affiliation: NVIDIA    Olga Russakovsky Affiliation: Princeton University    Sanja Fidler Affiliation: NVIDIA Affiliation: University of Toronto Affiliation: Vector Institute https://research.nvidia.com/labs/sil/projects/MOTIVE/    Jonathan Lorraine Affiliation: NVIDIA

Abstract

摘要

Despite the rapid progress of video generation models, the role of data in influencing motion is poorly understood. We present Motive (MOTI****on attribution for Video gEneration), a motion-centric, gradient-based data attribution framework that scales to modern, large, high-quality video datasets and models. We use this to study which fine-tuning clips improve or degrade temporal dynamics. Motive isolates temporal dynamics from static appearance via motion-weighted loss masks, yielding efficient and scalable motion-specific influence computation. On text-to-video models, Motive identifies clips that strongly affect motion and guides data curation that improves temporal consistency and physical plausibility. With Motive-selected high-influence data, we improve both motion smoothness and dynamic degree on VBench, achieving a 74.1% human preference win rate compared with the pretrained base model. To our knowledge, this is the first framework to attribute motion rather than visual appearance in video generative models and to use it to curate fine-tuning data.

视频生成模型进展很快，但数据如何影响运动，仍缺少清楚的解释。我们提出 Motive（MOTIon attribution for Video gEneration，视频生成的运动归因）：一套以运动为中心、基于梯度的数据归因框架，能够覆盖当下规模大、质量高的视频数据集与生成模型。我们用它考察：哪些微调片段会提升时序动态，哪些反而会把它做坏。Motive 用运动加权的损失掩码，把时序动态从静态外观里拆出来，从而高效、可扩展地计算运动相关的影响分数。在文本生成视频模型上，Motive 能找出对运动影响很强的片段，并据此筛选数据，提高时间一致性和物理合理性。用它选出的高影响数据微调后，VBench 上的运动平滑度与动态程度一并提升，相对预训练底模的人工偏好胜率为 74.1%。据我们所知，这是第一个把归因对准运动而非视觉外观、并据此筛选微调数据的视频生成框架。

1 Introduction

1 引言

Motion is the defining element of videos. Unlike image generation, which produces a single frame, video generative models capture how objects move, interact, and obey physical constraints (Wiedemer et al., 2025; Kang et al., 2024). Yet even with the rapid progress of video generation, a fundamental question remains:

运动是视频最核心的要素。图像生成只给出一帧，视频生成模型则要刻画物体如何移动、如何互动，以及如何服从物理约束（Wiedemer et al., 2025; Kang et al., 2024）。尽管视频生成进展很快，一个根本问题仍然在：

Why it matters. Diffusion models are data-driven, and their progress has tracked the scaling of data and compute (Saharia et al., 2022; Nichol and Dhariwal, 2021; Ho et al., 2022; Peebles and Xie, 2023). Prior work (Blattmann et al., 2023; Kaplan et al., 2020; Ravishankar et al., 2025) shows that training data shapes key generative properties, including visual quality (Rombach et al., 2022), semantic fidelity (Namekata et al., 2024), and compositionality (Wu et al., 2025; Favero et al., 2025). Motion is no exception. Motion refers to temporal dynamics captured by optical flow, including trajectories, deformations, camera movement, and interactions. If generated motion reflects the data distribution that shaped the model, then attributing motion to influential training clips provides a direct lens on why a model moves the way it does and enables targeted data selection for desired dynamics. High-quality data often matters most in fine-tuning, where large pretraining corpora are inaccessible and carefully selected clips can have an outsized impact. Motion-specific attribution is therefore especially valuable in the fine-tuning regime, where the goal is to identify which clips most influence temporal coherence and physical plausibility.

这件事为什么重要。 扩散模型靠数据驱动，进展也一直跟着数据和算力的扩大走（Saharia et al., 2022; Nichol and Dhariwal, 2021; Ho et al., 2022; Peebles and Xie, 2023）。已有工作（Blattmann et al., 2023; Kaplan et al., 2020; Ravishankar et al., 2025）表明，训练数据会塑造关键的生成性质，包括视觉质量（Rombach et al., 2022）、语义忠实度（Namekata et al., 2024）和组合能力（Wu et al., 2025; Favero et al., 2025）。运动也不例外。这里的运动指光流所刻画的时序动态，包括轨迹、形变、相机运动和物体交互。如果生成出来的运动反映的是塑造模型的数据分布，那么把运动归因到有影响力的训练片段，就能直接看到模型为什么会这样动，也能针对想要的动态去选数据。高质量数据往往在微调阶段最关键：大规模预训练语料通常拿不到，精心挑选的片段却能产生不成比例的影响。因此，专门针对运动的归因，在微调设定里尤其有用——目标就是找出哪些片段最能影响时间连贯性和物理合理性。

Why existing methods are limited for motion. Prior diffusion data attribution focuses on images and explains static content. Extending these methods to videos naïvely collapses motion into appearance, missing the temporal structure that distinguishes videos from images. Three challenges drive this gap: (i) localizing motion so attribution focuses on dynamic regions rather than static backgrounds, (ii) scaling to sequences since gradients must integrate across time, and (iii) capturing temporal relations like velocity, acceleration, and trajectory coherence that single-frame attribution cannot measure. Addressing motion attribution requires methods that explicitly model temporal structure, rather than treating time as an additional spatial axis.

现有方法为什么不适合运动。 先前的扩散数据归因多半面向图像，解释的是静态内容。若把这些方法直接搬到视频，运动会塌进外观里，丢掉把视频和图像区分开的时间结构。缺口主要来自三点：（i）把运动定位到动态区域，而不是静态背景；（ii）序列变长后，梯度必须沿时间积分，计算要能扩得动；（iii）速度、加速度和轨迹连贯性这类时间关系，单帧归因测不到。要做运动归因，必须显式建模时间结构，而不能把时间只当成多出来的空间轴。

Our method. We introduce Motive, a motion attribution framework for video generation models that isolates motion-specific influence. Motive computes gradients with motion-aware masking. As a result, the attribution signal emphasizes dynamic regions rather than static appearance. Efficient approximations make the method practical for large, high-quality datasets and video generative models. Our scores trace generated motion back to training clips, enabling targeted curation and improving motion quality when used to guide fine-tuning. Our key contributions are:

我们的方法。 我们提出 Motive，这是面向视频生成模型的运动归因框架，用来分离出运动本身的影响。Motive 在计算梯度时加入运动感知掩码，归因信号因而更偏向动态区域，而不是静态外观。高效近似使它能在大规模、高质量的数据集和视频生成模型上实际跑起来。得到的分数能把生成运动追溯到训练片段，既可做有针对性的数据筛选，也能在指导微调时提升运动质量。主要贡献如下：

1. Proposing a scalable gradient-based attribution approach for video generation models that is computationally efficient, even at the scale of modern, high-quality data and large generative models (§3.2).

1. 提出一套可扩展的、基于梯度的视频生成归因方法，计算上足够高效，即使面对当下高质量数据和大型生成模型也能使用（§3.2）。

2. Addressing a video-specific bias by correcting frame-length effects in gradient magnitudes, ensuring fair attribution across clips of different durations (§3.3).

2. 纠正视频特有的帧长偏差：对梯度幅度中的帧数效应做校正，使不同时长的片段能够公平比较（§3.3）。

3. Introducing an attribution that emphasizes temporal dynamics to trace which training clips most strongly influence motion quality (§3.4).

3. 引入强调时序动态的归因，追溯哪些训练片段对运动质量影响最大（§3.4）。

4. Showing that we improve motion smoothness and dynamic degree on VBench and in human evaluation (§4), matching, or surpassing, full-dataset fine-tuning performance with only $10\%$ of the data, and outperforming motion-unaware attribution baselines (Tables 1 and 2).

4. 在 VBench 和人工评测中提升运动平滑度与动态程度（§4）；只用 $10\%$ 的数据，即可达到或超过全量微调，并优于不考虑运动的归因基线（表 1 与表 2）。

2 Background

2 背景

Notation and extended related work are in App. §A and §B.

符号约定与扩展相关工作见附录 §A 与 §B。

2.1 Video Generation with Diffusion and Flow-Matching Models

2.1 用扩散与流匹配做视频生成

Diffusion and flow matching in latent space. Let $p_{\bm{\theta}}(\mathbf{v}\mid\mathbf{c})$ be a conditional generator with parameters $\bm{\theta}$, where $\mathbf{v}\in\mathbb{R}^{F\times H\times W\times 3}$ is a clip of height $H$, width $W$, and $F$ frames, and $\mathbf{c}$ denotes conditioning such as text or other multimodal metadata (e.g., fps, depth, pose). We operate in VAE latents: $\mathbf{h}=E(\mathbf{v})$ and train a denoiser or velocity field on noisy latents. A noise scheduler supplies time-dependent coefficients $(\alpha_{t},\sigma_{t})$ controlling signal and noise scales, and the forward noising is:

潜空间中的扩散与流匹配。 记 p_{m{ heta}}(\mathbf{v}\mid\mathbf{c}) 为参数 m{ heta} 的条件生成器，其中 $\mathbf{v}\in\mathbb{R}^{F imes H imes W imes 3}$ 是高 $H$、宽 $W$、$F$ 帧的片段，$\mathbf{c}$ 表示文本或其他多模态元数据（例如帧率、深度、姿态）这类条件。我们在 VAE 潜空间中工作：$\mathbf{h}=E(\mathbf{v})$，并在加噪后的潜变量上训练去噪器或速度场。噪声调度器提供随时间变化的系数 (lpha_{t},\sigma_{t})，分别控制信号与噪声尺度，前向加噪为：

$$\smash{\mathbf{z}(t,\bm{\epsilon})\;=\;\alpha_{t}\,\mathbf{h}\;+\;\sigma_{t}\,\bm{\epsilon},\quad\bm{\epsilon}\sim\mathcal{N}(0,\mathbf{I}),\quad t\in\{1,\ldots,T\}.}\vskip-2.75269pt \tag{1}$$

Denoising diffusion (Ho et al., 2020) trains a network $\bm{\epsilon}_{\bm{\theta}}(\mathbf{z},\mathbf{c},t)$ to predict the injected noise:

去噪扩散（Ho et al., 2020）训练网络 $\bm{\epsilon}_{\bm{\theta}}(\mathbf{z},\mathbf{c},t)$ 去预测注入的噪声：

$$\smash{\mathcal{L}_{\textnormal{diff}}(\bm{\theta};\mathbf{v},\mathbf{c})\;=\;\mathbb{E}_{t,\bm{\epsilon}}\!\left[\left\lVert\bm{\epsilon}_{\bm{\theta}}(\mathbf{z}(t,\bm{\epsilon}),\mathbf{c},t)-\bm{\epsilon}\right\rVert_{2}^{2}\right].}\vskip-2.75269pt \tag{2}$$

Flow matching (Lipman et al., 2022; Albergo et al., 2023) learns a time-dependent vector field $\mathbf{f}_{\bm{\theta}}(\mathbf{z}_{t},\mathbf{c},t)$ that matches the instantaneous velocity $\dot{\mathbf{z}}=\tfrac{\textnormal{d}}{\textnormal{d}t}\mathbf{z}$ induced by a chosen interpolant:

流匹配（Lipman et al., 2022; Albergo et al., 2023）学习一个随时间变化的向量场 $\mathbf{f}_{\bm{\theta}}(\mathbf{z}_{t},\mathbf{c},t)$，去匹配由所选插值路径诱导的瞬时速度 $\dot{\mathbf{z}}=\tfrac{\textnormal{d}}{\textnormal{d}t}\mathbf{z}$：

$$\smash{\mathcal{L}_{\textnormal{flow}}(\bm{\theta};\mathbf{v},\mathbf{c})\;=\;\mathbb{E}_{t,\bm{\epsilon}}\!\left[\left\lVert\mathbf{f}_{\bm{\theta}}(\mathbf{z}(t,\bm{\epsilon}),\mathbf{c},t)-\dot{\mathbf{z}}(t,\bm{\epsilon})\right\rVert_{2}^{2}\right].}\vskip-2.75269pt \tag{3}$$

Both objectives train time-indexed predictors over the latent space by integrating over $t$ and $\bm{\epsilon}$, thus gradient-based methods like attribution share similar challenges.

两个目标都是在潜空间上训练随时间索引的预测器，并对 $t$ 与 $\bm{\epsilon}$ 积分，因此基于梯度的方法（包括归因）会面对相近的挑战。

From images to video for generation. Adding a temporal axis materially changes modeling and training. Generation must capture spatial appearance and temporal dynamics such as object and camera motion, deformations, and interactions. Modern systems extend image backbones with temporal capacity, for example, 3D U-Nets or 2D U-Nets augmented with temporal attention, causal or sliding-window context, and factorized space-time blocks, often trained in a latent-video VAE that compresses frames while preserving temporal cues. Training departs from images along several axes, which we address in §3: (i) Compute and storage. Longer sequences multiply the cost of sampling timesteps, noise draws, and frames, motivating fixed-timestep or small-subset estimators that reduce variance without prohibitive cost (§3.2). (ii) Variable horizon. Clips vary in $F$ and frame rate (§3.3). (iii) Time-specific failure modes. Typical artifacts include inconsistent trajectories, temporal flicker, identity drift, and physically implausible dynamics despite sharp individual frames (§3.4).

从图像走到视频生成。 加上时间轴后，建模和训练都会实质性改变。生成既要抓住空间外观，也要抓住物体与相机运动、形变和交互这类时序动态。现代系统会在图像主干上补时间能力，例如 3D U-Net，或在 2D U-Net 上加入时间注意力、因果或滑窗上下文，以及空间–时间分解模块；训练往往在潜空间视频 VAE 里进行，压缩帧的同时保留时间线索。训练相对图像还沿几个轴分开，我们在 §3 处理：（i）计算与存储。 序列变长后，采样时间步、噪声抽取和帧数的代价成倍增加，因而需要固定时间步或小子集估计器，在代价可接受的前提下压低方差（§3.2）。（ii）可变时长。 片段的 $F$ 和帧率并不固定（§3.3）。（iii）时间上特有的失败模式。 常见伪影包括轨迹不一致、时间闪烁、身份漂移，以及单帧很锐利但动态在物理上不合理（§3.4）。

Motion representations in videos. We denote our video as $\mathbf{v}=[\mathbf{f}_{f}]_{f=1}^{F}$ with $\mathbf{f}_{f}\in\mathbb{R}^{H\times W\times 3}$ being the $f$-th frame. We represent motion via optical flow between consecutive frames: $\mathbf{F}_{f}:\{1,\ldots,H\}\times\{1,\ldots,W\}\to\mathbb{R}^{2}$, where each flow vector in $\mathbb{R}^{2}$ encodes the horizontal displacement $\mathrm{d}w$ and vertical displacement $\mathrm{d}h$ of a pixel. The motion magnitude is $M_{\!f}(h,w)=\|\mathbf{F}_{f}(h,w)\|_{2}$. The $M_{\!f}$ over frames $f$ and pixels $h,w$ summarizes the amount and spatial layout of motion in a clip, which we use to provide masks in our motion-weighted loss in §3.

视频中的运动表示。 我们把视频记为 $\mathbf{v}=[\mathbf{f}_{f}]_{f=1}^{F}$，其中第 $f$ 帧 $\mathbf{f}_{f}\in\mathbb{R}^{H\times W\times 3}$。运动用相邻帧之间的光流表示：$\mathbf{F}_{f}:\{1,\ldots,H\}\times\{1,\ldots,W\}\to\mathbb{R}^{2}$，每个 $\mathbb{R}^{2}$ 中的流向量编码一个像素的水平位移 $\mathrm{d}w$ 与垂直位移 $\mathrm{d}h$。运动幅度为 $M_{\!f}(h,w)=\|\mathbf{F}_{f}(h,w)\|_{2}$。对帧 $f$ 和像素 $h,w$ 上的 $M_{\!f}$ 做汇总，可以得到片段里运动的多少与空间分布，我们在 §3 用它为运动加权损失提供掩码。

2.2 Data Attribution

2.2 数据归因

Data attribution measures how individual training samples affect a model’s predictions (Bae et al., 2024). A classic approach to data attribution is to use influence functions (Koh and Liang, 2017). Intuitively, the influence of a training sample measures: if we upweight this training example, how much would the model’s prediction on a test datum change? Consider a loss function $\mathcal{L}(\bm{\theta};\mathbf{x})$ and a test sample $\mathbf{x}_{\textnormal{test}}$, the influence of a training point $\mathbf{x}_{n}$ can be quantified as:

数据归因衡量单个训练样本如何影响模型的预测（Bae et al., 2024）。经典做法是影响函数（Koh and Liang, 2017）。直观地说，一个训练样本的影响问的是：如果把这个训练例子的权重加大，模型在某个测试样本上的预测会变多少？给定损失函数 $\mathcal{L}(\bm{\theta};\mathbf{x})$ 和测试样本 $\mathbf{x}_{\textnormal{test}}$，训练点 $\mathbf{x}_{n}$ 的影响可以写成：

$$\displaystyle I(\mathbf{x}_{n},\mathbf{x}_{\textnormal{test}}) \\ \displaystyle\penalty\ =\penalty\ -\nabla_{\bm{\theta}}\mathcal{L}(\bm{\theta};\mathbf{x}_{\textnormal{test}})^{\!\top}\,\mathbf{H}_{\bm{\theta}}^{-1}\,\nabla_{\bm{\theta}}\mathcal{L}(\bm{\theta};\mathbf{x}_{n}),\quad\mathbf{H}_{\bm{\theta}}=\tfrac{1}{N}\sum\nolimits_{n=1}^{N}\nabla_{\bm{\theta}}^{2}\mathcal{L}(\bm{\theta};\mathbf{x}_{n}), \tag{4}$$

where the inverse Hessian captures the curvature of the loss landscape, yet computing or storing it is infeasible at modern model and dataset scales. Thus, practical methods (e.g., TracIn (Pruthi et al., 2020) and TRAK (Park et al., 2023)) approximate influence via gradient inner products or gradient feature projections.

其中逆 Hessian 刻画损失曲面的曲率，但在当下的模型和数据规模下，计算或存储它都不可行。因此，实用方法（例如 TracIn（Pruthi et al., 2020）和 TRAK（Park et al., 2023））用梯度内积或梯度特征投影来近似影响。

Attribution in diffusion models. Diffusion training aggregates gradients over timesteps $t$ and noise draws $\bm{\epsilon}$, where gradient norms vary systematically with $t$, producing a timestep bias where examples aligned with large-norm timesteps appear spuriously influential. Diffusion-ReTrac (Xie et al., 2024) reduces this bias by normalizing gradients and sub-sampling $t$ and $\bm{\epsilon}$ for influence. Let $\mathcal{L}_{\textnormal{diff}}$ denote the diffusion loss, and with the sampled-timestep-and-noise set $\mathcal{T}$, we compute a cosine-style score for normalized test and train gradients:

扩散模型中的归因。 扩散训练会在时间步 $t$ 和噪声抽取 $\bm{\epsilon}$ 上聚合梯度，而梯度范数会随 $t$ 系统性变化，从而产生时间步偏差：与大范数时间步对齐的样本，会看起来格外有影响力。Diffusion-ReTrac（Xie et al., 2024）通过对梯度做归一化、并对 $t$ 与 $\bm{\epsilon}$ 做子采样，来减轻这种偏差。记 $\mathcal{L}_{\textnormal{diff}}$ 为扩散损失，采样得到的时间步–噪声集合为 $\mathcal{T}$，我们对归一化后的测试梯度与训练梯度计算余弦风格的分数：

$$I_{\textnormal{diff}}(\mathbf{x}_{n},\mathbf{x}_{\textnormal{test}})=\left(\frac{1}{|\mathcal{T}_{\textnormal{test}}|}\sum_{t,\bm{\epsilon}\in\mathcal{T}_{\textnormal{test}}}\frac{\nabla_{\bm{\theta}}\mathcal{L}_{\textnormal{diff}}(\bm{\theta};\mathbf{x}_{\textnormal{test}},t,\bm{\epsilon})}{\big\|\nabla_{\bm{\theta}}\mathcal{L}_{\textnormal{diff}}(\bm{\theta};\mathbf{x}_{\textnormal{test}},t,\bm{\epsilon})\big\|}\right)^{\!\top}\left(\frac{1}{|\mathcal{T}_{n}|}\sum_{t,\bm{\epsilon}\in\mathcal{T}_{n}}\frac{\nabla_{\bm{\theta}}\mathcal{L}_{\textnormal{diff}}(\bm{\theta};\mathbf{x}_{n},t,\bm{\epsilon})}{\big\|\nabla_{\bm{\theta}}\mathcal{L}_{\textnormal{diff}}(\bm{\theta};\mathbf{x}_{n},t,\bm{\epsilon})\big\|}\right) \tag{5}$$

Averaging gradients over $(t,\bm{\epsilon})$ stabilizes estimates, and normalization mitigates timestep-induced scale effects. Attribution quality is also sensitive to the measurement function used to score examples, such as denoising loss versus likelihood proxies (Zheng et al., 2023).

在 $(t,\bm{\epsilon})$ 上对梯度取平均，可以使估计更稳；归一化则减轻时间步带来的尺度效应。归因质量还对用来给样本打分的度量函数敏感，例如去噪损失相对似然代理（Zheng et al., 2023）。

Why vanilla attribution is insufficient for videos. Naïvely applying gradient-based attribution to video diffusion risks treating appearance and motion alike, overemphasizing appearance (objects, textures, backgrounds) while overlooking dynamics (Park et al., 2025; Tulyakov et al., 2018). Its cost grows with clip length, sampled timesteps, noise draws, and gradient dimensionality, making naïve methods impractical at modern video scales. To improve motion, we need attribution that suppresses static appearance, emphasizes motion-specific signals, and remains efficient (§3). Motion is distributed across frames and entangled with static cues, thus influence cannot be assigned frame-independently.

为什么把归因原样用到视频不够。 若把基于梯度的归因直接套到视频扩散上，外观和运动容易被一视同仁，结果是物体、纹理、背景这类外观被过分强调，动态反而被忽略（Park et al., 2025; Tulyakov et al., 2018）。代价还会随片段长度、采样时间步、噪声抽取和梯度维数一起涨，在当下的视频规模下几乎跑不动。要改善运动，归因必须压住静态外观、突出运动信号，同时还要足够高效（§3）。运动分布在多帧里，又和静态线索缠在一起，因此不能按帧独立分配影响。

3 Method

3 方法

We formalize the problem in §3.1 and develop a practical framework for motion attribution in video diffusion models with four components: scalable gradient computation (§3.2), frame-length bias fix (§3.3), motion-aware weighting (§3.4), and target data selection (§3.5). We provide efficiency analysis (§3.6) demonstrating scalability to billion-parameter models and large-scale video datasets.

我们在 §3.1 形式化问题，并给出一套可实际使用的视频扩散运动归因框架，包含四个部分：可扩展的梯度计算（§3.2）、帧长偏差校正（§3.3）、运动感知加权（§3.4），以及目标数据选择（§3.5）。§3.6 给出效率分析，说明它可以扩到十亿参数级模型和大规模视频数据集。

Figure 1: Motive. Top. Motion-gradient computation (§3.4) has three steps: (1) detect motion with AllTracker, (2) compute motion-magnitude patches, (3) apply loss-space motion masks to focus gradients on dynamic regions. Bottom. Our method (§3.2) is made scalable via a single-sample variant with common randomness and a projection, computed for each pair of training and query data, aggregated (§3.5) for a final ranking, and eventually used to select fine-tuning subsets.

图 1：Motive。 上。 运动梯度计算（§3.4）分三步：（1）用 AllTracker 检测运动，（2）计算运动幅度图块，（3）在损失空间施加运动掩码，让梯度集中在动态区域。下。 我们的方法（§3.2）靠单样本变体、共享随机性和投影来扩规模；对每一对训练数据与查询数据计算分数，再按 §3.5 汇总得到最终排序，并据此选出微调子集。

3.1 Problem Formulation

3.1 问题形式化

We study data attribution for motion in the fine-tuning setting. Let $\mathcal{D}_{\textnormal{ft}}=\{(\mathbf{v}_{n},\mathbf{c}_{n})\}_{n=1}^{N}$ be the fine-tuning corpus. Given a query video $(\hat{\mathbf{v}},\hat{\mathbf{c}})$, we assign to each training clip $(\mathbf{v}_{n},\mathbf{c}_{n})$ a motion-aware influence score $I(\mathbf{v}_{n},\hat{\mathbf{v}};\bm{\theta})$ that explains how it contributes to the dynamics observed in $\hat{\mathbf{v}}$. The score should satisfy: (i) predictivity, rankings correlate with observed changes from fine-tuning on the most influential subsets; (ii) efficiency, scales to modern video generators, such as forgoing explicit Hessian inversion, expensive per-data integration, or prohibitive storage. To do this, we augment the influence target defined in Eq. 5 to be (a) lower variance for stable rankings with feasible levels of compute, (b) more scalable to store, and (c) motion-centric.

我们在微调设定下研究运动的数据归因。记微调语料为 $\mathcal{D}_{\textnormal{ft}}=\{(\mathbf{v}_{n},\mathbf{c}_{n})\}_{n=1}^{N}$。给定查询视频 $(\hat{\mathbf{v}},\hat{\mathbf{c}})$，我们为每个训练片段 $(\mathbf{v}_{n},\mathbf{c}_{n})$ 赋予运动感知的影响分数 $I(\mathbf{v}_{n},\hat{\mathbf{v}};\bm{\theta})$，用来解释它如何贡献于 $\hat{\mathbf{v}}$ 中观察到的动态。分数应满足：（i）可预测性，排序应与在最有影响的子集上微调后观察到的变化相关；（ii）效率，能够扩到现代视频生成器，例如不必显式求逆 Hessian、不必对每条数据做昂贵积分，也不需要难以承受的存储。为此，我们把式 5 中的影响目标改造成：（a）方差更低，从而在可行的计算量下得到稳定排序；（b）更省存储；（c）以运动为中心。

Fine-tuning Subset Selection. For a budget $K\ll N$, we get a motion-influential subset by ranking scores and taking the top-$K$ examples. For multiple query motions, we combine selections as described in §3.5. The resulting subsets serve as candidates for motion-centric fine-tuning.

微调子集选择。 在预算 $K\ll N$ 下，我们按分数排序并取前 $K$ 个样本，得到对运动有影响的子集。若有多种查询运动，则按 §3.5 所述合并选择结果。得到的子集作为以运动为中心的微调候选。

3.2 Scalable Gradient-based Attribution

3.2 可扩展的基于梯度的归因

We make attribution scalable for modern, large video datasets and models via inverse-Hessian approximations, lower-variance gradient-similarity estimators, low-cost single-sample estimators, and a Fastfood projection for tractable storage.

为了在现代大规模视频数据集和模型上做归因，我们采用逆 Hessian 近似、更低方差的梯度相似度估计、低成本的单样本估计器，以及用于可控存储的 Fastfood 投影。

Approximating the inverse-Hessian. Computing exact inverse-Hessian-vector products is infeasible for modern neural networks. We estimate influence via gradient similarity, using an identity preconditioner for the inverse Hessian (Koh and Liang, 2017; Pruthi et al., 2020; Park et al., 2023).

近似逆 Hessian。 对现代神经网络来说，精确计算逆 Hessian–向量积并不可行。我们用梯度相似度来估计影响，并把逆 Hessian 的预条件子取成单位阵（Koh and Liang, 2017; Pruthi et al., 2020; Park et al., 2023）。

Common randomness for stable rankings. To reduce variance without changing the target, we evaluate train and test gradients under the same $(t,\bm{\epsilon})$ pairs and average over a small set $\mathcal{T}$ (Xie et al., 2024; Lin et al., 2024). This paired averaging stabilizes rankings compared to independent draws:

用共享随机性稳住排序。 为了在不改目标的前提下降低方差，我们在同一组 $(t,\bm{\epsilon})$ 上计算训练梯度与测试梯度，并在一个较小的集合 $\mathcal{T}$ 上取平均（Xie et al., 2024; Lin et al., 2024）。这种成对平均，比独立抽样更能稳住排序：

$$I_{\textnormal{diff}}^{1}(\mathbf{x}_{n},\mathbf{x}_{\textnormal{test}})=\frac{1}{|\mathcal{T}|}\sum_{t,\bm{\epsilon}\in\mathcal{T}}\frac{\nabla_{\bm{\theta}}\mathcal{L}_{\textnormal{diff}}(\bm{\theta};\,\mathbf{x}_{\textnormal{test}},t,\bm{\epsilon})}{\big\|\nabla_{\bm{\theta}}\mathcal{L}_{\textnormal{diff}}(\bm{\theta};\,\mathbf{x}_{\textnormal{test}},t,\bm{\epsilon})\big\|}^{\!\top}\frac{\nabla_{\bm{\theta}}\mathcal{L}_{\textnormal{diff}}(\bm{\theta};\,\mathbf{x}_{n},t,\bm{\epsilon})}{\big\|\nabla_{\bm{\theta}}\mathcal{L}_{\textnormal{diff}}(\bm{\theta};\,\mathbf{x}_{n},t,\bm{\epsilon})\big\|} \tag{6}$$

Single-sample variant for reduced compute. We fix a single $t_{\textnormal{fix}}$ and a shared draw $\bm{\epsilon}_{\textnormal{fix}}\sim\mathcal{N}(0,\mathbf{I})$ for all train–test pairs. Sharing $(t_{\textnormal{fix}},\bm{\epsilon}_{\textnormal{fix}})$ allows low enough variance for the efficient single-sample estimator to maintain relative ordering (Xie et al., 2024; Lin et al., 2024). The estimator becomes:

用于降低计算量的单样本变体。 我们固定单个 $t_{\textnormal{fix}}$，并为所有训练–测试对共享同一次抽取 $\bm{\epsilon}_{\textnormal{fix}}\sim\mathcal{N}(0,\mathbf{I})$。共享 $(t_{\textnormal{fix}},\bm{\epsilon}_{\textnormal{fix}})$ 可以使方差足够低，让高效的单样本估计器仍能保持相对排序（Xie et al., 2024; Lin et al., 2024）。估计器变为：

$$I_{\textnormal{diff}}^{2}(\mathbf{x}_{n},\mathbf{x}_{\textnormal{test}})=\frac{\nabla_{\bm{\theta}}\mathcal{L}_{\textnormal{diff}}(\bm{\theta};\,\mathbf{x}_{\textnormal{test}},t_{\textnormal{fix}},\bm{\epsilon}_{\textnormal{fix}})}{\big\|\nabla_{\bm{\theta}}\mathcal{L}_{\textnormal{diff}}(\bm{\theta};\,\mathbf{x}_{\textnormal{test}},t_{\textnormal{fix}},\bm{\epsilon}_{\textnormal{fix}})\big\|}^{\!\top}\frac{\nabla_{\bm{\theta}}\mathcal{L}_{\textnormal{diff}}(\bm{\theta};\,\mathbf{x}_{n},t_{\textnormal{fix}},\bm{\epsilon}_{\textnormal{fix}})}{\big\|\nabla_{\bm{\theta}}\mathcal{L}_{\textnormal{diff}}(\bm{\theta};\,\mathbf{x}_{n},t_{\textnormal{fix}},\bm{\epsilon}_{\textnormal{fix}})\big\|} \tag{7}$$

Structured projection for reduced storage. To operate at model scale, we apply a Johnson–Lindenstrauss projection via Fastfood (Le et al., 2014) and then normalize. Let

用结构化投影降低存储。 为了在模型规模上运行，我们先用 Fastfood（Le et al., 2014）做 Johnson–Lindenstrauss 投影，再做归一化。令

$$\mathbf{P}\in\mathbb{R}^{D^{\prime}\times D}\;\;\text{implemented as}\;\;\mathbf{P}:=\tfrac{1}{\xi\sqrt{D^{\prime}}}\,\mathbf{S}\mathbf{Q}\mathbf{G}\bm{\Pi}\mathbf{Q}\mathbf{B},\vskip-5.0pt \tag{8}$$

where $\mathbf{Q}$ is the Walsh–Hadamard matrix, $\mathbf{B}$ is a diagonal Rademacher matrix, $\bm{\Pi}$ is a random permutation, $\mathbf{G}$ is a diagonal Gaussian scaling, and $\mathbf{S}$ is a diagonal rescaling, and $\xi$ normalizes the variance. The projected, normalized gradient is:

其中 $\mathbf{Q}$ 是 Walsh–Hadamard 矩阵，$\mathbf{B}$ 是对角 Rademacher 矩阵，$\bm{\Pi}$ 是随机置换，$\mathbf{G}$ 是对角高斯缩放，$\mathbf{S}$ 是对角再缩放，$\xi$ 用来归一化方差。投影并归一化后的梯度为：

$$\tilde{\mathbf{g}}\big(\bm{\theta},\mathbf{x}\big)\penalty\ :=\penalty\ \frac{\mathbf{P}\,\nabla_{\bm{\theta}}\mathcal{L}_{\textnormal{diff}}(\bm{\theta},\mathbf{x},t_{\textnormal{fix}},\bm{\epsilon}_{\textnormal{fix}})}{\|\mathbf{P}\,\nabla_{\bm{\theta}}\mathcal{L}_{\textnormal{diff}}(\bm{\theta},\mathbf{x},t_{\textnormal{fix}},\bm{\epsilon}_{\textnormal{fix}})\|}. \tag{9}$$

Then the influence score is the dot product of normalized projected gradients (i.e., cosine similarity) in $\mathbb{R}^{D^{\prime}}$:

于是影响分数就是 $\mathbb{R}^{D^{\prime}}$ 中归一化投影梯度的内积（也即余弦相似度）：

$$I_{\textnormal{diff}}^{3}(\mathbf{x}_{n},\mathbf{x}_{\textnormal{test}})=\tilde{\mathbf{g}}(\bm{\theta},\mathbf{x}_{\textnormal{test}})^{\top}\tilde{\mathbf{g}}(\bm{\theta},\mathbf{x}_{n}) \tag{10}$$

This keeps compute $\mathcal{O}(D^{\prime}\log D^{\prime})$ for projection and $\mathcal{O}(D^{\prime})$ per dot product, with storage $\mathcal{O}(|\mathcal{D}|\,D^{\prime})$, while staying close to the ranking behavior of full-gradient cosine similarity (Park et al., 2023).

这样，投影的计算量为 $\mathcal{O}(D^{\prime}\log D^{\prime})$，每次内积为 $\mathcal{O}(D^{\prime})$，存储为 $\mathcal{O}(|\mathcal{D}|\,D^{\prime})$，同时排序行为仍接近全梯度余弦相似度（Park et al., 2023）。

3.3 Video-specific Frame-length Bias Fix

3.3 视频特有的帧长偏差校正

Raw gradient magnitudes depend on the number of frames $F$ in the video $\mathbf{v}$, thereby biasing scores toward longer videos. We correct this by normalizing for frame count before the projection–normalization step:

原始梯度幅度依赖于视频 $\mathbf{v}$ 的帧数 $F$，分数会因此偏向更长的视频。我们在投影–归一化之前按帧数做一次归一化来纠正这一点：

$$\smash{\nabla_{\bm{\theta}}\mathcal{L}_{\textnormal{diff}}(\bm{\theta};\,\mathbf{v},t_{\textnormal{fix}},\bm{\epsilon}_{\textnormal{fix}})\;\leftarrow\;\frac{1}{F}\,\nabla_{\bm{\theta}}\mathcal{L}_{\textnormal{diff}}(\bm{\theta};\,\mathbf{v},t_{\textnormal{fix}},\bm{\epsilon}_{\textnormal{fix}}).} \tag{11}$$

We still apply $\ell_{2}$ normalization in Eq. 10, further stabilizing scales across examples. Together, single-timestep, common randomness, projection, and frame-length correction form a compact, scalable estimator that we use throughout. However, naïve video-level attribution conflates appearance with motion, often ranking clips high due to shared backgrounds or objects, while offering little insight into dynamics.

我们仍在式 10 中施加 $\ell_{2}$ 归一化，进一步稳住不同样本之间的尺度。单时间步、共享随机性、投影和帧长校正合在一起，构成我们全程使用的紧凑、可扩展估计器。不过，若只做整段视频级的朴素归因，外观会和运动缠在一起：片段常常因为共享背景或物体而排到前面，对动态却几乎给不出信息。

3.4 Motion Attribution

3.4 运动归因

To move beyond whole-video influence, we introduce motion attribution, which isolates the contribution of training data to temporal dynamics. Unlike video-level attribution, which treats each clip as a single unit and conflates appearance with motion, motion attribution reweights per-location gradients using motion masks, assigning influence via dynamic behavior rather than static content.

为了超越整段视频的影响分数，我们引入运动归因，把训练数据对时序动态的贡献单独拆出来。整段归因把每条片段当成一个整体，外观与运动混在一起；运动归因则用运动掩码对各个位置的梯度重新加权，按动态行为而不是静态内容来分配影响。

Motion Masking Attribution. Motion is what distinguishes video diffusion from image diffusion. Our goal is to understand how training data shapes motion in video diffusion models. Prior work has emphasized architectural or algorithmic changes for motion modeling (Peebles and Xie, 2023; Blattmann et al., 2023; Guo et al., 2023), many of the largest generative gains have instead come from scaling and curating massive video corpora, which in turn enable impressive motion synthesis results in video diffusion models (Ho et al., 2022; Wan et al., 2025; Tan et al., 2024; Yang et al., 2024). Yet we lack tools that quantify how specific training clips shape particular motion patterns. We address this by attributing motion back to data via motion-weighted gradients, which yields actionable signals for targeted data selection, artifact diagnosis, and selective fine-tuning.

运动掩码归因。 运动正是视频扩散区别于图像扩散的地方。我们的目标，是理解训练数据如何塑造视频扩散模型中的运动。已有工作更强调用架构或算法改进来建模运动（Peebles and Xie, 2023; Blattmann et al., 2023; Guo et al., 2023）；但生成质量上许多最大的提升，其实来自大规模视频语料的扩展与筛选，这些语料也让视频扩散能够合成令人印象深刻的运动（Ho et al., 2022; Wan et al., 2025; Tan et al., 2024; Yang et al., 2024）。我们却缺少工具，去量化特定训练片段如何塑造特定的运动模式。我们的做法是：用运动加权梯度把运动追溯回数据，从而为有针对性的数据选择、伪影诊断和选择性微调提供可操作的信号。

Motion Detection and Latent Space Mapping. Given a video $\mathbf{v}\in\mathbb{R}^{F\times H\times W\times 3}$ with $F$ frames of resolution $H\times W$, we first encode it into the VAE latent space as $\mathbf{h}=E(\mathbf{v})\in\mathbb{R}^{F\times\nicefrac{{H}}{{s}}\times\nicefrac{{W}}{{s}}\times C}$, with downsampling factor $s=8$ and $C=16$ following the Wan2.1 backbone used in our experiments.

运动检测与潜空间映射。 给定视频 $\mathbf{v}\in\mathbb{R}^{F\times H\times W\times 3}$，共 $F$ 帧、分辨率为 $H\times W$，我们先把它编码到 VAE 潜空间：$\mathbf{h}=E(\mathbf{v})\in\mathbb{R}^{F\times\nicefrac{{H}}{{s}}\times\nicefrac{{W}}{{s}}\times C}$。按实验中使用的 Wan2.1 主干，下采样因子 $s=8$，通道数 $C=16$。

For motion computation, we use AllTracker (Harley et al., 2025) to extract motion information in pixel space: $A=\mathcal{A}(\mathbf{v})\in\mathbb{R}^{F\times H\times W\times 4}$, where the first two channels contain optical flow maps $A_{:,:,:,0:2}$ indicating pixel displacement between frames, and the remaining channels $A_{:,:,:,2:4}$ encode visibility and confidence scores. We extract displacement vectors at each pixel location as:

计算运动时，我们用 AllTracker（Harley et al., 2025）在像素空间提取运动信息：$A=\mathcal{A}(\mathbf{v})\in\mathbb{R}^{F\times H\times W\times 4}$，前两个通道是光流图 $A_{:,:,:,0:2}$，表示帧间像素位移，其余通道 $A_{:,:,:,2:4}$ 编码可见性与置信度。每个像素位置的位移向量取为：

$$\smash{\mathbf{D}_{\!f}(h,w)=(A_{f,h,w,0},A_{f,h,w,1})=(\mathrm{d}w,\mathrm{d}h).}\vskip-5.49698pt \tag{12}$$

We then bilinearly downsample motion quantities from $(H,W)$ to the latent grid $\big(\tfrac{H}{s},\tfrac{W}{s}\big)$ so that our masking lives where gradients are computed.

随后把运动量从 $(H,W)$ 双线性下采样到潜空间网格 $\big(\tfrac{H}{s},\tfrac{W}{s}\big)$，使掩码落在实际计算梯度的位置。

Motion-Weighted Gradient Computation. We define the motion magnitude at each location as: $M_{\!f}(h,w)\!=\!\|\mathbf{D}_{\!f}(h,w)\|_{2}$. To obtain comparable motion weights across frames and pixels, we min–max normalize over all frames and pixels, ensuring values lie in $[0,1]$:

运动加权的梯度计算。 我们把每个位置的运动幅度定义为 $M_{\!f}(h,w)\!=\!\|\mathbf{D}_{\!f}(h,w)\|_{2}$。为了在帧与像素之间得到可比较的运动权重，我们在全部帧和像素上做最小–最大归一化，使取值落在 $[0,1]$：

$$\mathbf{W}(f,h,w)=\frac{M_{\!f}(h,w)-m_{\min}}{m_{\max}-m_{\min}+\zeta}, \tag{13}$$

where $m_{\min}=\min_{f^{\prime},h^{\prime},w^{\prime}}M_{\!f}(h^{\prime},w^{\prime})$, $m_{\max}=\max_{f^{\prime},h^{\prime},w^{\prime}}M_{\!f}(h^{\prime},w^{\prime})$, and $\zeta=10^{-6}$ ensures a positive denominator. This normalization mitigates bias from absolute motion scale, yielding weights that emphasize relative motion saliency rather than raw magnitude, following prior practice in video saliency detection (Fang et al., 2013). Let $(\tilde{h},\tilde{w})$ index the latent grid. We obtain latent-aligned weights by bilinear downsampling:

其中 $m_{\min}=\min_{f^{\prime},h^{\prime},w^{\prime}}M_{\!f}(h^{\prime},w^{\prime})$，$m_{\max}=\max_{f^{\prime},h^{\prime},w^{\prime}}M_{\!f}(h^{\prime},w^{\prime})$，$\zeta=10^{-6}$ 保证分母为正。这一归一化减轻绝对运动尺度带来的偏差，使权重强调相对运动显著性，而不是原始幅度，这也符合视频显著性检测中的既有做法（Fang et al., 2013）。用 $(\tilde{h},\tilde{w})$ 索引潜空间网格，再经双线性下采样得到与潜空间对齐的权重：

$$\smash{\tilde{\mathbf{W}}(f,\tilde{h},\tilde{w})\penalty\ =\penalty\ \operatorname{Bilinear}\!\left(\mathbf{W}(\cdot,\cdot,\cdot),F,\penalty\ \tfrac{H}{s},\penalty\ \tfrac{W}{s}\right).} \tag{14}$$

We compute per-location squared error at fixed $(t_{\textnormal{fix}},\bm{\epsilon}_{\textnormal{fix}})$ at each frame $f$ and “latent pixel” $(\tilde{h},\tilde{w})$:

我们在固定的 $(t_{\textnormal{fix}},\bm{\epsilon}_{\textnormal{fix}})$ 下，对每一帧 $f$ 和每个「潜空间像素」$(\tilde{h},\tilde{w})$ 计算逐位置平方误差：

$$\displaystyle\tilde{\mathcal{L}}_{\bm{\theta},\mathbf{v},\mathbf{c}}(f,\tilde{h},\tilde{w})\penalty\ =\penalty\ \Big( \\ \displaystyle[\bm{\epsilon}_{\bm{\theta}}(\mathbf{z}(\mathbf{v},t_{\textnormal{fix}},\bm{\epsilon}_{\textnormal{fix}}),t_{\textnormal{fix}},\mathbf{c})]_{f,\tilde{h},\tilde{w}}-[\bm{\epsilon}_{\textnormal{target}}(t_{\textnormal{fix}},\bm{\epsilon}_{\textnormal{fix}})]_{f,\tilde{h},\tilde{w}}\,\Big)^{2},$$

and define the motion-weighted loss by averaging over frames and latent spatial locations:

再对帧和潜空间位置取平均，定义运动加权损失：

$$\mathcal{L}_{\textnormal{mot}}(\bm{\theta};\mathbf{v},\mathbf{c})\penalty\ =\penalty\ \frac{1}{F_{\mathbf{v}}}\operatorname*{mean}_{f,\tilde{h},\tilde{w}}\Big[\tilde{\mathbf{W}}_{\mathbf{v},\mathbf{c}}(f,\tilde{h},\tilde{w})\cdot\tilde{\mathcal{L}}_{\bm{\theta},\mathbf{v},\mathbf{c}}(f,\tilde{h},\tilde{w})\Big]. \tag{16}$$

Notably, when $\tilde{\mathbf{W}}$ is all ones, this recovers the standard objective with no motion emphasis. The $\nicefrac{{1}}{{F_{\mathbf{v}}}}$ factor corrects for frame-length bias and $F_{\mathbf{v}}$ signifies how the number of frames may be video-dependent. The corresponding motion-weighted gradient for attribution is:

值得注意的是，当 $\tilde{\mathbf{W}}$ 全为 1 时，它会退回没有运动强调的标准目标。因子 $\nicefrac{{1}}{{F_{\mathbf{v}}}}$ 用来校正帧长偏差，$F_{\mathbf{v}}$ 标明帧数可以随视频变化。相应的、用于归因的运动加权梯度为：

$$\displaystyle I_{\textnormal{mot}}\!(\mathbf{v}_{n},\!\hat{\mathbf{v}}) \\ \displaystyle\!=\!\tilde{\mathbf{g}}_{\textnormal{mot}}\!(\bm{\theta}\!,\!\hat{\mathbf{v}})^{\!\top}\!\tilde{\mathbf{g}}_{\textnormal{mot}}\!(\bm{\theta}\!,\!\mathbf{v}_{n}),\quad\textnormal{where }\mathbf{g}_{\textnormal{mot}}\!:=\!\nabla_{\!\bm{\theta}}\mathcal{L}_{\textnormal{mot}}\textnormal{ and }\tilde{\mathbf{g}}_{\textnormal{mot}}\!(\bm{\theta}\!,\!\mathbf{v})\!:=\!\frac{\mathbf{P}\mathbf{g}_{\textnormal{mot}}\!(\bm{\theta}\!,\!\mathbf{v}\!,\!t_{\textnormal{fix}},\!\bm{\epsilon}_{\textnormal{fix}})}{\|\mathbf{P}\mathbf{g}_{\textnormal{mot}}\!(\bm{\theta}\!,\!\mathbf{v}\!,\!t_{\textnormal{fix}},\!\bm{\epsilon}_{\textnormal{fix}})\|}. \tag{17}$$

Loss-space masking leaves forward noising and generation unchanged and reweights only attribution, avoiding interactions between motion weighting and noise injection. In contrast, our motion-aware attribution emphasizes dynamic regions and de-emphasizes static backgrounds, so rankings identify training clips that most strongly shape the model’s motion rather than appearance.

在损失空间做掩码，前向加噪和生成过程保持不变，加权只作用于归因，从而避免运动加权与噪声注入之间的相互干扰。相比之下，运动感知归因会强调动态区域、压低静态背景，排序因此能找出最能塑造模型运动（而不是外观）的训练片段。

3.5 Most Influential Fine-tuning Subset Selection

3.5 选择最有影响的微调子集

Goal. Given a query clip $(\hat{\mathbf{v}},\hat{\mathbf{c}})$, we compute a motion-aware attribution value for each fine-tuning sample $(\mathbf{v}_{n},\mathbf{c}_{n})\in\mathcal{D}_{\textnormal{ft}}$ using $I_{\textnormal{mot}}\!\left(\mathbf{v}_{n},\,\hat{\mathbf{v}}\right)$ (Eq. 17). We construct a fine-tuning set $\mathcal{S}$ for one or many query videos $\hat{\mathbf{v}}$.

目标。 给定查询片段 $(\hat{\mathbf{v}},\hat{\mathbf{c}})$，我们用 $I_{\textnormal{mot}}\!\left(\mathbf{v}_{n},\,\hat{\mathbf{v}}\right)$（式 17）为微调集合 $\mathcal{D}_{\textnormal{ft}}$ 中的每个样本 $(\mathbf{v}_{n},\mathbf{c}_{n})$ 计算运动感知的归因值。我们为一条或多条查询视频 $\hat{\mathbf{v}}$ 构造微调集合 $\mathcal{S}$。

Single-query-point fine-tuning selection. For a budget of $K$ data points, we select the $K$ highest-scoring examples. In practice, $K$ is chosen as a percentile of the dataset size (e.g., top 1–10%), ensuring the subset scales consistently across datasets.

单查询点的微调选择。 在预算为 $K$ 个数据点时，我们选取分数最高的 $K$ 个样本。实践中，$K$ 取数据集大小的百分位（例如前 1–10%），使子集规模在不同数据集上保持一致。

Multi-query-point fine-tuning selection: aggregating attribution scores. For $Q$ queries, we adopt the majority voting approach from ICONS (Wu et al., 2024) and aggregate motion-aware influence scores across queries by percentile thresholding and voting. A sample receives a vote if the score is above the percentile cutoff $\tau$ for that query. The consensus score of a candidate $\mathbf{v}_{n}$ is the total number of queries that vote for it. We then rank all training samples by $\operatorname{MajVote}(\mathbf{v}_{n})$ and select the top-$K$ to form the fine-tuning subset. This formulation emphasizes samples that are consistently influential across multiple queries, without requiring cross-query calibration of raw scores:

多查询点的微调选择：汇总归因分数。 对于 $Q$ 个查询，我们采用 ICONS（Wu et al., 2024）中的多数投票，用百分位阈值和投票，把各查询上的运动感知影响分数汇总起来。若某个样本在该查询上的分数高于百分位截断 $\tau$，就得到一票。候选 $\mathbf{v}_{n}$ 的共识分数，是为它投票的查询总数。随后按 $\operatorname{MajVote}(\mathbf{v}_{n})$ 对全部训练样本排序，并取前 $K$ 个组成微调子集。这一写法强调在多个查询上持续有影响的样本，而不需要对原始分数做跨查询校准：

$$\displaystyle\operatorname{MajVote}_{n} \\ \displaystyle=\!\sum\nolimits_{q=1}^{Q}\!\mathbb{I}\!\big[\,I_{\textnormal{mot}}(\mathbf{v}_{n},\hat{\mathbf{v}}_{q})\!>\!\tau\,\big],\quad\mathcal{S}_{\textnormal{vote}}(K)\!=\!\bigl\{\mathbf{v}_{n}|\mathbf{v}_{n}\text{ in top-}K\text{ by }\operatorname{MajVote}\bigr\}.$$

3.6 Computational Efficiency Analysis

3.6 计算效率分析

Gradient Compute. Naïvely averaging over timesteps and noise for every example costs $\mathcal{O}(|\mathcal{D}|\,|\mathcal{T}|\,B)$, where $B$ is a single forward+backward cost and $|\mathcal{T}|$ is the number of sampled $t,\bm{\epsilon}$ per data. Using a single sample reduces this to $\mathcal{O}(|\mathcal{D}|\,B)$, which is key to keeping the cost reasonable for modern video datasets and models, while reusing a sample across data allows low enough variance for stable rankings. Projection adds $\mathcal{O}(D^{\prime}\log D^{\prime})$ per example using Fastfood (Le et al., 2014), negligible relative to a backward pass.

梯度计算。 若对每个样本在时间步和噪声上朴素取平均，代价是 $\mathcal{O}(|\mathcal{D}|\,|\mathcal{T}|\,B)$，其中 $B$ 是一次前向加反向的代价，$|\mathcal{T}|$ 是每条数据采样的 $t,\bm{\epsilon}$ 数目。改用单次采样后降为 $\mathcal{O}(|\mathcal{D}|\,B)$，这是在现代视频数据集和模型上把代价控制住的关键；同一组采样在数据间复用，也能把方差压到足以稳住排序。投影对每个样本额外增加 $\mathcal{O}(D^{\prime}\log D^{\prime})$（Fastfood，Le et al., 2014），相对一次反向可以忽略。

Gradient Storage. Storing full gradients is $\mathcal{O}(|\mathcal{D}|\,D)$. We instead store only projected vectors, $\mathcal{O}(|\mathcal{D}|\,D^{\prime})$, plus the structured Fastfood state, $\mathcal{O}(D)$. Since $D^{\prime}$ is typically orders of magnitude smaller than $D$, this transformation makes storage tractable for billion-parameter models.

梯度存储。 存储完整梯度是 $\mathcal{O}(|\mathcal{D}|\,D)$。我们改为只存投影向量，即 $\mathcal{O}(|\mathcal{D}|\,D^{\prime})$，再加上结构化的 Fastfood 状态 $\mathcal{O}(D)$。由于 $D^{\prime}$ 通常比 $D$ 小几个数量级，这一变换使十亿参数级模型的存储变得可行。

Data Ranking Compute. Influence computation in Eq. 10 is an inner product in $\mathbb{R}^{D^{\prime}}$, so evaluating all train examples against a query is $\mathcal{O}(|\mathcal{D}|\,D^{\prime})$, and sorting is $\mathcal{O}(|\mathcal{D}|\log|\mathcal{D}|)$.

数据排序的计算。 式 10 中的影响计算是 $\mathbb{R}^{D^{\prime}}$ 中的内积，因此用一条查询评估全部训练样本的代价是 $\mathcal{O}(|\mathcal{D}|\,D^{\prime})$，排序为 $\mathcal{O}(|\mathcal{D}|\log|\mathcal{D}|)$。

Figure 2: Motion Attribution Samples with Wan2.1-T2V-1.3B. Top: Query clips showing float (left) and roll (right) motions. Middle: Top-ranked positive training samples identified by Motive with high influence scores. Bottom: Negative influence samples with minimal, camera-only motion, or cartoon-style content that conflict with target motions. Videos are included in the supplementary material.

图 2：用 Wan2.1-T2V-1.3B 做运动归因的样本。 上：展示漂浮（左）和滚动（右）运动的查询片段。中：Motive 找出的、影响分数很高的正向前排训练样本。下：负影响样本，运动极少、只有镜头运动，或是与目标运动冲突的卡通风格内容。视频见补充材料。

Additional Motion-Emphasis Compute. Motion-specific overhead primarily stems from AllTracker mask extraction with complexity $\mathcal{O}(|\mathcal{D}|\cdot H\cdot W\cdot F)$ for clip length $F$ and frame resolution $H\times W$. Masks are extracted once, cached, and negligible relative to gradient cost. We provide a detailed runtime breakdown in App. G.1.

额外的运动强调计算。 运动相关的额外开销，主要来自 AllTracker 的掩码提取，复杂度为 $\mathcal{O}(|\mathcal{D}|\cdot H\cdot W\cdot F)$，其中 $F$ 是片段长度，$H\times W$ 是帧分辨率。掩码只需提取一次并缓存，相对梯度代价可以忽略。更细的运行时间分解见附录 G.1。

4 Experiment

4 实验

4.1 Setup

4.1 设置

Fine-tuning Datasets. We evaluate our motion attribution framework on two large-scale video datasets: VIDGEN-1M (Tan et al., 2024) and 4DNeX-10M (Chen et al., 2025), both of which offer diverse motion patterns, rich temporal dynamics, and complex scenes. For our experiments, we use 10k videos from both datasets, which provide sufficient scale and diversity to thoroughly evaluate motion attribution methods across different temporal patterns and video generation scenarios.

微调数据集。 我们在两个大规模视频数据集上评估运动归因框架：VIDGEN-1M（Tan et al., 2024）和 4DNeX-10M（Chen et al., 2025）。两者都具有多样的运动模式、丰富的时序动态和复杂场景。实验中，我们各取 1 万段视频，这一规模和多样性足以在不同时间模式和视频生成场景下，较充分地评估运动归因方法。

Motion Query Data. To evaluate our motion attribution, we curate a set of query videos representing distinct motion patterns and scenarios. Our query dataset consists of videos spanning $10$ motion categories, with a focus on object dynamics: compress, bounce, roll, explode, float, free fall, slide, spin, stretch, swing. Five videos, totaling $50$ queries, represent each motion type. These videos are chosen for their clear, isolated motions, serving as a basis for evaluating attribution quality and downstream motion generation. Details are in App. F.2.

运动查询数据。 为评估运动归因，我们整理了一组代表不同运动模式和场景的查询视频。查询集覆盖 $10$ 类运动，重点放在物体动态：压缩、弹跳、滚动、爆炸、漂浮、自由落体、滑动、旋转、拉伸、摆动。每类运动有五段视频，共 $50$ 条查询。这些视频运动清楚、相对独立，用作评估归因质量以及下游运动生成的基础。细节见附录 F.2。

Model & Baselines. We use pretrained models Wan2.1-T2V-1.3B and Wan2.2-TI2V-5B, with additional results on LTX-2B in App. C. Our baselines: Base model: pretrained, no fine-tuning; Full fine-tuning: approximate upper bound using the complete dataset; Random selection: uniform sampling; Motion magnitude: selects videos with the highest average motion magnitude; V-JEPA embeddings: selects most representative videos of motion patterns using V-JEPA (Assran et al., 2025) features, capturing high-level motion semantics; and Ours w/o motion masking: influence of the entire video level without motion-specific masking.

模型与基线。 我们使用预训练模型 Wan2.1-T2V-1.3B 与 Wan2.2-TI2V-5B，附录 C 另给出 LTX-2B 上的结果。基线包括：底模，即不做微调的预训练模型；全量微调，用全部数据近似上界；随机选取，均匀抽样；运动幅度，选取平均运动幅度最高的视频；V-JEPA 嵌入，用 V-JEPA（Assran et al., 2025）特征选出最能代表运动模式的视频，以捕捉高层运动语义；以及本文去掉运动掩码的变体，即在整段视频上计算影响、不做运动专用掩码。

Figure 3: Qualitative Comparisons. We compare four motion scenarios across the base model (Wan2.1-T2V-1.3B), or fine-tuned with random selection and Motive. Our approach produces more realistic motion. Videos are included in the supplementary material.

图 3：定性比较。 我们在四种运动场景下比较底模（Wan2.1-T2V-1.3B）、用随机选取微调，以及用 Motive 微调。我们的方法生成的运动更接近真实。视频见补充材料。

	Wan2.1-T2V-1.3B	Wan2.2-TI2V-5B
Method	Subj.	Bg.	Mot.	Dyn.	Aesth.	Img.	Subj.	Bg.	Mot.	Dyn.	Aesth.	Img.
Base	95.3	96.4	96.3	39.6	45.3	65.7	94.9	96.4	97.5	42.0	44.4	65.5
Full FT	95.9	96.6	96.3	42.0	45.0	63.9	95.3	96.5	97.5	45.3	44.8	66.2
Random	95.3	96.6	96.3	41.3	45.7	65.1	94.7	96.2	97.3	41.6	44.6	65.2
Motion mag.	95.6	96.2	95.7	40.1	45.1	63.2	95.0	96.3	97.4	44.9	45.0	65.1
V-JEPA	95.7	96.0	95.6	41.6	44.9	62.7	95.2	96.4	97.3	45.6	44.9	64.8
Ours w/o MM	95.4	96.1	96.3	43.8	45.7	63.2	94.9	96.5	97.4	43.8	45.2	64.8
Ours (Motive)	96.3	96.1	96.3	47.6	46.0	64.6	95.1	96.6	97.6	48.3	45.6	65.5

Table 1: VBench Evaluation. Performance comparison on VBench (Huang et al., 2024) across different baselines (all values in $\%$, higher is better). All selection methods use $10\%$ of training data; our method uses majority vote aggregation (§3.5) across motion queries. MM: motion masking. Subj.: Subject Consistency, Bg.: Background Consistency, Mot.: Motion Smoothness, Dyn.: Dynamic Degree, Aesth.: Aesthetic Quality, Img.: Imaging Quality.

	Wan2.1-T2V-1.3B	Wan2.2-TI2V-5B
方法	主体	背景	运动	动态	美学	成像	主体	背景	运动	动态	美学	成像
底模	95.3	96.4	96.3	39.6	45.3	65.7	94.9	96.4	97.5	42.0	44.4	65.5
全量微调	95.9	96.6	96.3	42.0	45.0	63.9	95.3	96.5	97.5	45.3	44.8	66.2
随机选取	95.3	96.6	96.3	41.3	45.7	65.1	94.7	96.2	97.3	41.6	44.6	65.2
运动幅度	95.6	96.2	95.7	40.1	45.1	63.2	95.0	96.3	97.4	44.9	45.0	65.1
V-JEPA	95.7	96.0	95.6	41.6	44.9	62.7	95.2	96.4	97.3	45.6	44.9	64.8
本文（无运动掩码）	95.4	96.1	96.3	43.8	45.7	63.2	94.9	96.5	97.4	43.8	45.2	64.8
本文（Motive）	96.3	96.1	96.3	47.6	46.0	64.6	95.1	96.6	97.6	48.3	45.6	65.5

表 1：VBench 评测。 在 VBench（Huang et al., 2024）上与各基线比较（数值均为 $\%$，越高越好）。所有选取方法都只用 $10\%$ 训练数据；我们的方法按 §3.5 在各运动查询上做多数投票汇总。MM：运动掩码。主体：主体一致性；背景：背景一致性；运动：运动平滑度；动态：动态程度；美学：美学质量；成像：成像质量。

Benchmark. We evaluate our attribution using VBench (Huang et al., 2024) metrics across six dimensions: subject consistency, background consistency, motion smoothness, dynamic degree, aesthetic quality, and imaging quality. Motion smoothness and dynamic degree are our primary targets for temporal dynamics, while other metrics help maintain visual quality. We use custom evaluation prompts, following VBench’s descriptive style, for 10 motion types, with $5$ prompts each, to assess our framework’s effectiveness on target motions.

基准。 我们用 VBench（Huang et al., 2024）的六个维度评估归因：主体一致性、背景一致性、运动平滑度、动态程度、美学质量和成像质量。运动平滑度与动态程度是时序动态的主要目标，其余指标用来盯住视觉质量。我们按 VBench 的描述风格，为 10 类运动各写 $5$ 条定制评测提示，以检验框架在目标运动上的效果。

Implementation Details. We finetune base models with our Motive-selected high-quality video data following the official and DiffSynth-Studio implementation. During fine-tuning, we update only the DiT backbone while freezing the T5 text encoder and VAE. All models are trained at a resolution of $480\times 832$ with a learning rate of $1\times 10^{-5}$. Specialist models are trained on single motion category selected data, while generalist models use aggregated selections (both with top $10\%$ selection from VIDGEN-1M (Tan et al., 2024) or 4DNeX (Chen et al., 2025)). Compute and runtime details are provided in App. G.1.

实现细节。 我们按官方实现和 DiffSynth-Studio，用 Motive 选出的高质量视频数据微调底模。微调时只更新 DiT 主干，冻结 T5 文本编码器和 VAE。所有模型在 $480\times 832$ 分辨率上训练，学习率为 $1\times 10^{-5}$。专模型用单一运动类别选出的数据训练，通模型用汇总后的选择结果（二者都从 VIDGEN-1M（Tan et al., 2024）或 4DNeX（Chen et al., 2025）中取前 $10\%$）。计算与运行时间细节见附录 G.1。

4.2 Main Results

4.2 主要结果

High-influence selection and negative filtering. Fig. 2 shows our motion-aware attribution ranks clips with clear, physically grounded dynamics and downranks those with little transferable motion. For rolling and floating, positives show continuous trajectories and smooth temporal evolution (turbulent water, planetary rotation). Negatives are mostly static footage, camera-only motion, or cartoons whose simplified kinematics do not transfer. This pattern holds across motion categories and aligns with the quantitative gains below.

高影响选择与负向过滤。 图 2 显示，运动感知归因会把动态清楚、有物理依据的片段排到前面，并把几乎没有可迁移运动的片段压下去。对滚动和漂浮来说，正样本呈现连续轨迹和平滑的时间演化（湍急水面、行星自转）。负样本多半是几乎静止的镜头、只有相机在动，或运动学被简化、难以迁移的卡通。这一模式在各类运动上都成立，也和下面的定量提升一致。

Qualitative improvements across motion types. Fig. 3 compares the base Wan2.1-T2V-1.3B model, finetuned with random selection or Motive with equal data budgets, across $4$ scenarios. Our method yields higher motion fidelity and temporal consistency than baselines, especially for complex deformation and physics-driven motion.

各类运动上的定性改进。 图 3 在 $4$ 种场景下比较底模 Wan2.1-T2V-1.3B、用随机选取微调，以及用同等数据预算的 Motive。我们的方法在运动保真度和时间一致性上高于各基线，尤其是在复杂形变和由物理驱动的运动上。

Quantitative Results. We evaluate our approach using VBench (Huang et al., 2024), demonstrating consistent improvements in motion when fine-tuning with attribution-selected data. As shown in Tab. 1, Motive consistently achieves the highest dynamic degree scores across both models: $47.6\%$ on Wan2.1-T2V-1.3B and $48.3\%$ on Wan2.2-TI2V-5B, significantly outperforming random selection ($41.3\%$ and $41.6\%$) and whole video attribution ($43.8\%$ for both models). Our method also excels in aesthetic quality ($46.0\%$ and $45.6\%$), while maintaining competitive motion smoothness ($96.3\%$ and $97.6\%$). Notably, using only $10\%$ of training data, our approach surpasses the full fine-tuning on dynamic degree ($42.0\%$ and $45.3\%$) on both models, demonstrating superior performance of motion-specific attribution for targeted fine-tuning. Additional results on LTX-2B are in App. C; motion magnitude distribution analysis is in App. D.

定量结果。 我们用 VBench（Huang et al., 2024）评估方法，表明用归因选出的数据微调后，运动指标会稳定提升。如表 1 所示，Motive 在两个模型上都拿到最高的动态程度：$47.6\%$（Wan2.1-T2V-1.3B）和 $48.3\%$（Wan2.2-TI2V-5B），明显高于随机选取（$41.3\%$ 与 $41.6\%$）和整段视频归因（两个模型均为 $43.8\%$）。美学质量同样最高（$46.0\%$ 与 $45.6\%$），运动平滑度仍保持在有竞争力的水平（$96.3\%$ 与 $97.6\%$）。尤其是，只用 $10\%$ 训练数据，我们的方法在两个模型上的动态程度都超过全量微调（$42.0\%$ 与 $45.3\%$），说明针对运动的归因更适合做有目标的微调。LTX-2B 上的补充结果见附录 C，运动幅度分布分析见附录 D。

Figure 4: Impact of Frame-Length Normalization on Motion Attribution. Comparison of top-ranked samples for floating motion query. Left: With proper frame-length normalization, top samples consistently exhibit floating motion (waves, floating objects, surfing). Right: Without normalization, rankings are biased by video length, resulting in no coherent patterns among top samples.

图 4：帧长归一化对运动归因的影响。 比较漂浮运动查询下排在最前面的样本。左：做了恰当的帧长归一化后，前列样本稳定呈现漂浮运动（波浪、漂浮物、冲浪）。右：不做归一化时，排序被视频长度带走，前列样本之间看不到连贯模式。

4.3 Human Evaluation

4.3 人工评测

Method	Win ($\%$)	Tie ($\%$)	Loss ($\%$)
Ours vs. Base	$74.1$	$12.3$	$13.6$
Ours vs. Random	$58.9$	$12.1$	$29.0$
Ours vs. Full FT	$53.1$	$14.8$	$32.1$
Ours vs. w/o MM	$46.9$	$20.0$	$33.1$

Table 2: Human evaluation. Pairwise comparisons across $50$ videos with $17$ participants ($850$ total). Win, tie, and loss rates show where our method is preferred, rated equal, or outperformed.

方法	胜（$\%$）	平（$\%$）	负（$\%$）
本文 vs. 底模	$74.1$	$12.3$	$13.6$
本文 vs. 随机选取	$58.9$	$12.1$	$29.0$
本文 vs. 全量微调	$53.1$	$14.8$	$32.1$
本文 vs. 无运动掩码	$46.9$	$20.0$	$33.1$

表 2：人工评测。 对 $50$ 条视频、由 $17$ 名参与者做两两比较（共 $850$ 次判断）。胜、平、负分别表示更偏好我们的方法、两者相当，或认为基线更好。

Automated scores can miss perceptual motion quality, so we run a human evaluation pairwise comparison protocol: participants view two generated videos and choose which shows better motion. We recruit $17$ annotators and evaluate $10$ motion categories. For each category, we prepare $5$ test cases and compare our method to baselines across three pairings, yielding a balanced set of judgments. Presentation order is randomized, and ties are allowed. We report win rate (fraction our method is preferred), tie rate, and overall preference. As shown in Tab. 2, annotators favor our attribution-guided selection: $74.1\%$ win rate vs. the base model and $53.1\%$ vs. the full fine-tuned model, showing perceptually meaningful motion improvements.

自动分数可能漏掉可感知的运动质量，因此我们做人工两两比较：参与者同时看两段生成视频，判断哪一段运动更好。我们招募 $17$ 名标注者，评测 $10$ 类运动。每类准备 $5$ 个测试用例，并把我们的方法与各基线按三种配对比较，得到一组均衡的判断。呈现顺序随机，允许打平。我们报告胜率（更偏好我们方法的比例）、平局率和总体偏好。如表 2 所示，标注者更认可由归因指导的选择：相对底模的胜率为 $74.1\%$，相对全量微调模型为 $53.1\%$，说明运动上的改进在感知上是有意义的。

4.4 Ablations

4.4 消融

Figure 5: Projection dimension analysis. Spearman correlation between projected and full gradients shows rapid improvement with projection dimension, with $512$ providing a strong trade-off between accuracy and efficiency.

图 5：投影维数分析。 投影梯度与全梯度之间的 Spearman 相关随投影维数迅速上升，$512$ 维在精度和效率之间给出很好的折中。

Projected Gradients Preserve Influence Rankings. Comparing full gradients for attribution is infeasible at a billion-parameter scale. We reduce dimensionality with structured random projections that preserve influence geometry, ablating $D^{\prime}\!\in\!\{128,\dots,2048\}$ against the full-gradient baseline. We assess ranking preservation via Spearman correlation with unprojected scores (Fig. 5). Small projections preserve rankings poorly: $D^{\prime}\!=\!128$ yields $\rho\!=\!46.9\%$. Preservation improves with size: $D^{\prime}\!=\!512$ reaches $\rho\!=\!74.7\%$. Beyond that, gains are marginal while cost rises: $D^{\prime}\!=\!1024$ ($\rho\!=\!75.7\%$) and $D^{\prime}\!=\!2048$ ($\rho\!=\!76.1\%$). Thus, $D^{\prime}\!=\!512$ offers the best trade-off, scaling to large models while maintaining quality.

投影梯度能够保住影响排序。 在十亿参数规模上，直接比较完整梯度做归因并不可行。我们用结构化随机投影降低维数，同时尽量保住影响几何，并在 $D^{\prime}\!\in\!\{128,\dots,2048\}$ 上相对全梯度基线做消融。我们用与未投影分数的 Spearman 相关来衡量排序是否保住（图 5）。投影过小时，排序保得不好：$D^{\prime}\!=\!128$ 时 $\rho\!=\!46.9\%$。维数加大后会改善：$D^{\prime}\!=\!512$ 时 $\rho\!=\!74.7\%$。再往上，收益有限而代价继续上升：$D^{\prime}\!=\!1024$（$\rho\!=\!75.7\%$）、$D^{\prime}\!=\!2048$（$\rho\!=\!76.1\%$）。因此 $D^{\prime}\!=\!512$ 是最好的折中，既能扩到大模型，又能保住质量。

Single-timestep attribution. Using a single timestep avoids the cost of averaging across timesteps while closely matching the multi-timestep baseline. With fixed $t\!=\!751$ (midpoint of the $1000$-step trajectory), we obtain $\rho\!=\!66\%$ agreement with ground truth computed using $10$ evenly-spaced timesteps based on flow matching schedule. High timesteps (early denoising) heavily obscure motion with noise; low timesteps (late denoising) operate on nearly formed videos, where gradients reflect fine details. $t\!=\!751$ (mid-denoising) balances influence ranking correlation and compute efficiency. Averaging multiple timesteps yields minimal gains, a single fixed timestep is therefore sufficient for variance-reduced, scalable attribution.

单时间步归因。 只用一个时间步，可以避开在多个时间步上取平均的代价，同时仍接近多时间步基线。把 $t\!=\!751$ 固定在 $1000$ 步轨迹的中点时，我们与按流匹配调度均匀选取的 $10$ 个时间步算出的参考结果，相关达到 $\rho\!=\!66\%$。高时间步（去噪早期）噪声很重，运动会几乎被盖住；低时间步（去噪晚期）面对的已是接近成形的视频，梯度更多反映细部。$t\!=\!751$（去噪中段）在影响排序相关和计算效率之间较为平衡。对多个时间步再取平均，收益很小，因此单个固定时间步已足以做降方差、可扩展的归因。

Frame-Length Normalization. As in the Wan training protocol, we standardize all videos to 81 frames at 16 fps (satisfying the $4n{+}1$ constraint) for fair attribution across clips of different raw lengths. Without standardization, gradient-based scores correlate strongly with video length rather than motion quality ($\rho\!=\!78.0\%$), leading to longer clips ranking higher regardless of dynamics. Standardizing frames reduces spurious length correlations by $54.0\%$ while preserving motion-based correlation, so rankings reflect motion rather than duration. As in Fig. 4, normalization clarifies motion-specific patterns. For floating queries with frame-length normalization (left), top-ranked samples consistently show wave dynamics, floating objects, and surfing, all matching the target motion. Without normalization (right), top samples lack coherent similarity because rankings are driven by clip length, harming motion-relevant training example identification.

帧长归一化。 按 Wan 的训练协议，我们把所有视频统一成 16 fps 下的 81 帧（满足 $4n{+}1$ 约束），以便在原始长度不同的片段之间公平归因。若不做标准化，基于梯度的分数会与视频长度强相关，而不是与运动质量相关（$\rho\!=\!78.0\%$），更长的片段会不管动态如何都排到前面。把帧数标准化后，这种虚假的长度相关下降 $54.0\%$，同时保住基于运动的相关，排序因而反映运动而不是时长。如图 4 所示，归一化会把运动特有的模式看得更清楚。对漂浮查询，做了帧长归一化后（左），前列样本稳定出现波浪动态、漂浮物和冲浪，都与目标运动相符。不做归一化时（右），前列样本缺少连贯的相似性，因为排序被片段长度带走，不利于找出与运动相关的训练例子。

5 Conclusion

5 结论

We address a central and underexplored question in video diffusion: which training clips influence the motion in generated videos? We introduced Motive, a motion-aware data attribution framework for video diffusion that isolates temporal dynamics from static appearance. By tracing generated motion back to influential training clips, our method enables targeted data curation that improves motion quality with a fraction of the data. As video models scale, such data-level understanding will be key to diagnosing failure modes and building more controllable generative systems.

我们面对视频扩散里一个关键却讨论不多的问题：哪些训练片段影响了生成视频里的运动？我们提出 Motive，这是面向视频扩散的运动感知数据归因框架，用来把时序动态从静态外观中拆出来。通过把生成运动追溯到有影响的训练片段，该方法能够做有针对性的数据筛选，并用更少的数据提升运动质量。随着视频模型继续变大，这种数据层面的理解，对诊断失败模式、以及构建更可控的生成系统，都会越来越重要。

Acknowledgements

致谢

We thank the following people (listed alphabetically by last name) for their helpful discussions, feedback, or participation in human studies: Allison Chen, Sanghyuk Chun, Zhiwei Deng, Amaya Dharmasiri, Xingyu Fu, Will Hwang, Yifeng Jiang, Amogh Joshi, Pang Wei Koh, Chen-Hsuan Lin, Huan Ling, Tiffany Ling, Shaowei Liu, Zhengyi Luo, Rafid Mahmood, Kaleb S. Newman, Julian Ost, Zeeshan Patel, Davis Rempe, Rulin Shao, Esin Tureci, Anya Tsvetkov, Jiachen T. Wang, Sheng-Yu Wang, Tingwu Wang, Zian Wang, Hongyu Wen, Jon Williams, Mengzhou Xia, Donglai Xiang, Yilun Xu, William Yang, and Haotian Zhang.

我们感谢以下人士（按姓氏字母顺序排列）在讨论、反馈或人工评测中给予的帮助：Allison Chen, Sanghyuk Chun, Zhiwei Deng, Amaya Dharmasiri, Xingyu Fu, Will Hwang, Yifeng Jiang, Amogh Joshi, Pang Wei Koh, Chen-Hsuan Lin, Huan Ling, Tiffany Ling, Shaowei Liu, Zhengyi Luo, Rafid Mahmood, Kaleb S. Newman, Julian Ost, Zeeshan Patel, Davis Rempe, Rulin Shao, Esin Tureci, Anya Tsvetkov, Jiachen T. Wang, Sheng-Yu Wang, Tingwu Wang, Zian Wang, Hongyu Wen, Jon Williams, Mengzhou Xia, Donglai Xiang, Yilun Xu, William Yang, and Haotian Zhang.

References

参考文献

Agarwal et al. [2025] Niket Agarwal, Arslan Ali, Maciej Bala, Yogesh Balaji, Erik Barker, Tiffany Cai, Prithvijit Chattopadhyay, Yongxin Chen, Yin Cui, Yifan Ding, et al. Cosmos world foundation model platform for physical ai. arXiv preprint arXiv:2501.03575 , 2025.

Albergo et al. [2023] Michael S Albergo, Nicholas M Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions. arXiv preprint arXiv:2303.08797 , 2023.

Assran et al. [2025] Mido Assran, Adrien Bardes, David Fan, Quentin Garrido, Russell Howes, Matthew Muckley, Ammar Rizvi, Claire Roberts, Koustuv Sinha, Artem Zholus, et al. V-jepa 2: Self-supervised video models enable understanding, prediction and planning. arXiv preprint arXiv:2506.09985 , 2025.

Bae et al. [2024] Juhan Bae, Wu Lin, Jonathan Lorraine, and Roger B Grosse. Training data attribution via approximate unrolling. Advances in Neural Information Processing Systems , 37:66647–66686, 2024.

Balduzzi et al. [2018] David Balduzzi, Sebastien Racaniere, James Martens, Jakob Foerster, Karl Tuyls, and Thore Graepel. The mechanics of n-player differentiable games. In International Conference on Machine Learning , pages 354–363. PMLR, 2018.

Bao et al. [2023] Yuxiang Bao, Di Qiu, Guoliang Kang, Baochang Zhang, Bo Jin, Kaiye Wang, and Pengfei Yan. Latentwarp: Consistent diffusion latents for zero-shot video-to-video translation. arXiv preprint arXiv:2311.00353 , 2023.

Bettencourt et al. [2026] Jesse Bettencourt, Xindi Wu, Matan Atzmon, James Lucas, and Jonathan Lorraine. Variance reduction for expectations with diffusion teachers. arXiv preprint arXiv:2605.21489 , 2026.

Blattmann et al. [2023] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 , 2023.

Brokman et al. [2024] Jonathan Brokman, Omer Hofman, Roman Vainshtein, Amit Giloni, Toshiya Shimizu, Inderjeet Singh, Oren Rachmil, Alon Zolfi, Asaf Shabtai, Yuki Unno, et al. Montrage: Monitoring training for attribution of generative diffusion models. In European Conference on Computer Vision , pages 1–17. Springer, 2024.

Chen et al. [2023] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video diffusion models with motion prior and reward feedback learning. arXiv preprint arXiv:2305.13840 , 2023.

Chen et al. [2025] Zhaoxi Chen, Tianqi Liu, Long Zhuo, Jiawei Ren, Zeng Tao, He Zhu, Fangzhou Hong, Liang Pan, and Ziwei Liu. 4dnex: Feed-forward 4d generative modeling made easy. arXiv preprint arXiv:2508.13154 , 2025.

Cook et al. [1982] RDWS Cook et al. Residuals and influence in regression. 1982.

Evans et al. [2025] Zach Evans, Julian D Parker, CJ Carr, Zack Zukowski, Josiah Taylor, and Jordi Pons. Stable audio open. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages 1–5. IEEE, 2025.

Fang et al. [2013] Yuming Fang, Weisi Lin, Zhenzhong Chen, Chia-Ming Tsai, and Chia-Wen Lin. A video saliency detection model in compressed domain. IEEE transactions on circuits and systems for video technology , 24(1):27–38, 2013.

Favero et al. [2025] Alessandro Favero, Antonio Sclocchi, Francesco Cagnetta, Pascal Frossard, and Matthieu Wyart. How compositional generalization and creativity improve as diffusion models are trained. arXiv preprint arXiv:2502.12089 , 2025.

Feldman and Zhang [2020] Vitaly Feldman and Chiyuan Zhang. What neural networks memorize and why: Discovering the long tail via influence estimation. Advances in Neural Information Processing Systems , 33:2881–2891, 2020.

Georgiev et al. [2023] Kristian Georgiev, Joshua Vendrow, Hadi Salman, Sung Min Park, and Aleksander Madry. The journey, not the destination: How data guides diffusion models. arXiv preprint arXiv:2312.06205 , 2023.

Geyer et al. [2023] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373 , 2023.

Google DeepMind [2025] Google DeepMind. Veo-3: Advancing controllable and physically plausible video generation. Technical report, Google DeepMind, 2025. URL https://storage.googleapis.com/deepmind-media/veo/Veo-3-Tech-Report.pdf .

Guo et al. [2023] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725 , 2023.

HaCohen et al. [2024] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103 , 2024.

Hammoudeh and Lowd [2024] Zayd Hammoudeh and Daniel Lowd. Training data influence analysis and estimation: A survey. Machine Learning , 113(5):2351–2403, 2024.

Harley et al. [2025] Adam W Harley, Yang You, Xinglong Sun, Yang Zheng, Nikhil Raghuraman, Yunqi Gu, Sheldon Liang, Wen-Hsuan Chu, Achal Dave, Pavel Tokmakov, et al. Alltracker: Efficient dense point tracking at high resolution. arXiv preprint arXiv:2506.07310 , 2025.

Ho et al. [2020] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems , 33:6840–6851, 2020.

Ho et al. [2022] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in neural information processing systems , 35:8633–8646, 2022.

Horn and Schunck [1981] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial intelligence , 17(1-3):185–203, 1981.

Huang et al. [2024] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 21807–21818, 2024.

Jia et al. [2021] Ruoxi Jia, Fan Wu, Xuehui Sun, Jiacen Xu, David Dao, Bhavya Kailkhura, Ce Zhang, Bo Li, and Dawn Song. Scalability vs. utility: Do we have to sacrifice one for the other in data importance quantification? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021.

Kang et al. [2024] Bingyi Kang, Yang Yue, Rui Lu, Zhijie Lin, Yang Zhao, Kaixin Wang, Gao Huang, and Jiashi Feng. How far is video generation from world model: A physical law perspective. arXiv preprint arXiv:2411.02385 , 2024.

Kaplan et al. [2020] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361 , 2020.

Koh and Liang [2017] Pang Wei Koh and Percy Liang. Understanding black-box predictions via influence functions. In International Conference on Machine Learning , pages 1885–1894. PMLR, 2017.

Kwon et al. [2023] Yongchan Kwon, Eric Wu, Kevin Wu, and James Zou. Datainf: Efficiently estimating data influence in lora-tuned llms and diffusion models. arXiv preprint arXiv:2310.00902 , 2023.

Le et al. [2014] Quoc Viet Le, Tamás Sarlós, and Alexander Johannes Smola. Fastfood: Approximate kernel expansions in loglinear time. arXiv preprint arXiv:1408.3060 , 2014.

Lin et al. [2024] Jinxu Lin, Linwei Tao, Minjing Dong, and Chang Xu. Diffusion attribution score: Evaluating training data influence in diffusion models. arXiv preprint arXiv:2410.18639 , 2024.

Lipman et al. [2022] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747 , 2022.

Lorraine and Duvenaud [2018] Jonathan Lorraine and David Duvenaud. Stochastic hyperparameter optimization through hypernetworks. arXiv preprint arXiv:1802.09419 , 2018.

Lorraine and Hossain [2024] Jonathan Lorraine and Safwan Hossain. Jacnet: Learning functions with structured jacobians. arXiv preprint arXiv:2408.13237 , 2024.

Lorraine et al. [2020] Jonathan Lorraine, Paul Vicol, and David Duvenaud. Optimizing millions of hyperparameters by implicit differentiation. In International conference on artificial intelligence and statistics , pages 1540–1552. PMLR, 2020.

Lorraine et al. [2021] Jonathan Lorraine, Paul Vicol, Jack Parker-Holder, Tal Kachman, Luke Metz, and Jakob Foerster. Lyapunov exponents for diversity in differentiable games. arXiv preprint arXiv:2112.14570 , 2021.

Lorraine et al. [2022a] Jonathan Lorraine, Nihesh Anderson, Chansoo Lee, Quentin De Laroussilhe, and Mehadi Hassen. Task selection for automl system evaluation. arXiv preprint arXiv:2208.12754 , 2022a.

Lorraine et al. [2022b] Jonathan P Lorraine, David Acuna, Paul Vicol, and David Duvenaud. Complex momentum for optimization in games. In International Conference on Artificial Intelligence and Statistics , pages 7742–7765. PMLR, 2022b.

Lorraine [2024] Jonathan Peter Lorraine. Scalable nested optimization for deep learning . PhD thesis, University of Toronto (Canada), 2024.

Loshchilov and Hutter [2017] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 , 2017.

Lucas and Kanade [1981] Bruce D Lucas and Takeo Kanade. An iterative image registration technique with an application to stereo vision. In IJCAI’81: 7th international joint conference on Artificial intelligence , volume 2, pages 674–679, 1981.

Mehta et al. [2024] Nikhil Mehta, Jonathan Lorraine, Steve Masson, Ramanathan Arunachalam, Zaid Pervaiz Bhat, James Lucas, and Arun George Zachariah. Improving hyperparameter optimization with checkpointed model weights. In European Conference on Computer Vision , pages 75–96. Springer, 2024.

Mlodozeniec et al. [2024] Bruno Mlodozeniec, Runa Eschenhagen, Juhan Bae, Alexander Immer, David Krueger, and Richard Turner. Influence functions for scalable data attribution in diffusion models. arXiv preprint arXiv:2410.13850 , 2024.

Namekata et al. [2024] Koichi Namekata, Amirmojtaba Sabour, Sanja Fidler, and Seung Wook Kim. Emerdiff: Emerging pixel-level semantic knowledge in diffusion models. arXiv preprint arXiv:2401.11739 , 2024.

Nichol and Dhariwal [2021] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning , pages 8162–8171. PMLR, 2021.

Park et al. [2023] Sung Min Park, Kristian Georgiev, Andrew Ilyas, Guillaume Leclerc, and Aleksander Madry. Trak: Attributing model behavior at scale. arXiv preprint arXiv:2303.14186 , 2023.

Park et al. [2025] Yonghyun Park, Chieh-Hsin Lai, Satoshi Hayakawa, Yuhta Takida, Naoki Murata, Wei-Hsiang Liao, Woosung Choi, Kin Wai Cheuk, Junghyun Koo, and Yuki Mitsufuji. Concept-trak: Understanding how diffusion models learn concepts through concept-level attribution. arXiv preprint arXiv:2507.06547 , 2025.

Peebles and Xie [2023] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision , pages 4195–4205, 2023.

Pruthi et al. [2020] Garima Pruthi, Frederick Liu, Satyen Kale, and Mukund Sundararajan. Estimating training data influence by tracing gradient descent. Advances in Neural Information Processing Systems , 33, 2020.

Raffel et al. [2020] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research , 21(140):1–67, 2020.

Raghu et al. [2021] Aniruddh Raghu, Jonathan Lorraine, Simon Kornblith, Matthew McDermott, and David K Duvenaud. Meta-learning to improve pre-training. Advances in Neural Information Processing Systems , 34:23231–23244, 2021.

Ravishankar et al. [2025] Rahul Ravishankar, Zeeshan Patel, Jathushan Rajasegaran, and Jitendra Malik. Scaling properties of diffusion models for perceptual tasks. In Proceedings of the Computer Vision and Pattern Recognition Conference , 2025.

Richter-Powell et al. [2025] Jessie Richter-Powell, Antonio Torralba, and Jonathan Lorraine. Score distillation sampling for audio: Source separation, synthesis, and beyond. arXiv preprint arXiv:2505.04621 , 2025.

Rombach et al. [2022] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 10684–10695, 2022.

Saharia et al. [2022] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems , 35, 2022.

Shi et al. [2025] Qingyu Shi, Jianzong Wu, Jinbin Bai, Jiangning Zhang, Lu Qi, Yunhai Tong, and Xiangtai Li. Decouple and track: Benchmarking and improving video diffusion transformers for motion transfer. arXiv preprint arXiv:2503.17350 , 2025.

Song et al. [2024] Yanke Song, Jonathan Lorraine, Weili Nie, Karsten Kreis, and James Lucas. Multi-student diffusion distillation for better one-step generators. arXiv preprint arXiv:2410.23274 , 2024.

Tan et al. [2024] Zhiyu Tan, Xiaomeng Yang, Luozheng Qin, and Hao Li. Vidgen-1m: A large-scale dataset for text-to-video generation. arXiv preprint arXiv:2408.02629 , 2024.

Teed and Deng [2020] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision , pages 402–419. Springer, 2020.

Tulyakov et al. [2018] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 1526–1535, 2018.

Vicol et al. [2022] Paul Vicol, Jonathan P Lorraine, Fabian Pedregosa, David Duvenaud, and Roger B Grosse. On implicit bias in overparameterized bilevel optimization. In International Conference on Machine Learning , pages 22234–22259. PMLR, 2022.

Wan et al. [2025] Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, Jianyuan Zeng, Jiayu Wang, Jingfeng Zhang, Jingren Zhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, Keyu Yan, Lianghua Huang, Mengyang Feng, Ningyi Zhang, Pandeng Li, Pingyu Wu, Ruihang Chu, Ruili Feng, Shiwei Zhang, Siyang Sun, Tao Fang, Tianxing Wang, Tianyi Gui, Tingyu Weng, Tong Shen, Wei Lin, Wei Wang, Wei Wang, Wenmeng Zhou, Wente Wang, Wenting Shen, Wenyuan Yu, Xianzhong Shi, Xiaoming Huang, Xin Xu, Yan Kou, Yangyu Lv, Yifei Li, Yijing Liu, Yiming Wang, Yingya Zhang, Yitong Huang, Yong Li, You Wu, Yu Liu, Yulin Pan, Yun Zheng, Yuntao Hong, Yupeng Shi, Yutong Feng, Zeyinzi Jiang, Zhen Han, Zhi-Fan Wu, and Ziyu Liu. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314 , 2025.

Wang et al. [2024a] Jiachen T Wang, Prateek Mittal, Dawn Song, and Ruoxi Jia. Data shapley in one training run. arXiv preprint arXiv:2406.11011 , 2024a.

Wang et al. [2024b] Jiachen T Wang, Tianji Yang, James Zou, Yongchan Kwon, and Ruoxi Jia. Rethinking data shapley for data selection tasks: Misleads and merits. arXiv preprint arXiv:2405.03875 , 2024b.

Wang et al. [2024c] Jiangshan Wang, Yue Ma, Jiayi Guo, Yicheng Xiao, Gao Huang, and Xiu Li. Cove: Unleashing the diffusion feature correspondence for consistent video editing. Advances in Neural Information Processing Systems , 37:96541–96565, 2024c.

Wang et al. [2023a] Sheng-Yu Wang, Alexei A Efros, Jun-Yan Zhu, and Richard Zhang. Evaluating data attribution for text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023a.

Wang et al. [2025] Sheng-Yu Wang, Aaron Hertzmann, Alexei A Efros, Richard Zhang, and Jun-Yan Zhu. Fast data attribution for text-to-image models. arXiv preprint arXiv:2511.10721 , 2025.

Wang et al. [2023b] Xiang Wang, Shiwei Zhang, Han Zhang, Yu Liu, Yingya Zhang, Changxin Gao, and Nong Sang. Videolcm: Video latent consistency model. arXiv preprint arXiv:2312.09109 , 2023b.

Wiedemer et al. [2025] Thaddäus Wiedemer, Yuxuan Li, Paul Vicol, Shixiang Shane Gu, Nick Matarese, Kevin Swersky, Been Kim, Priyank Jaini, and Robert Geirhos. Video models are zero-shot learners and reasoners. arXiv preprint arXiv:2509.20328 , 2025.

Wu et al. [2023] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF international conference on computer vision , 2023.

Wu et al. [2024] Xindi Wu, Mengzhou Xia, Rulin Shao, Zhiwei Deng, Pang Wei Koh, and Olga Russakovsky. Icons: Influence consensus for vision-language data selection. arXiv preprint arXiv:2501.00654 , 2024.

Wu et al. [2025] Xindi Wu, Hee Seung Hwang, Polina Kirichenko, and Olga Russakovsky. Compact: Compositional atomic-to-complex visual capability tuning. arXiv preprint arXiv:2504.21850 , 2025.

Xie et al. [2024] Tong Xie, Haoyu Li, Andrew Bai, and Cho-Jui Hsieh. Data attribution for diffusion models: Timestep-induced bias in influence estimation. arXiv preprint arXiv:2401.09031 , 2024.

Yang et al. [2024] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072 , 2024.

Zhang et al. [2023] Tianjun Zhang, Yi Zhang, Vibhav Vineet, Neel Joshi, and Xin Wang. Controllable text-to-image generation with gpt-4. arXiv preprint arXiv:2305.18583 , 2023.

Zheng et al. [2023] Xiaosen Zheng, Tianyu Pang, Chao Du, Jing Jiang, and Min Lin. Intriguing properties of data attribution on diffusion models. arXiv preprint arXiv:2311.00500 , 2023.

Zhou et al. [2024] Shangchen Zhou, Peiqing Yang, Jianyi Wang, Yihang Luo, and Chen Change Loy. Upscale-a-video: Temporal-consistent diffusion model for real-world video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 2535–2545, 2024.

Zhu et al. [2025] Yixuan Zhu, Jiaqi Feng, Wenzhao Zheng, Yuan Gao, Xin Tao, Pengfei Wan, Jie Zhou, and Jiwen Lu. Astra: General interactive world model with autoregressive denoising. arXiv preprint arXiv:2512.08931 , 2025.

Appendix A Notation

附录 A 符号

Symbol	Description
Acronyms and Basic Notation
VAE	Variational Autoencoder
DiT	Diffusion Transformer backbone
$\mathbf{I}$	Identity matrix
Video Generation
$p_{\bm{\theta}}(\mathbf{v}\mid\mathbf{c})$	Conditional video generator with parameters $\bm{\theta}$
$\mathbf{v}\in\mathbb{R}^{F\times H\times W\times 3}$	Video clip with frames $F$, height $H$, width $W$
$\mathbf{c}$	Conditioning signal such as text or multimodal metadata
$\bm{\theta}$	Trainable model parameters
$f\in\{1,\ldots,F\}$	Frame index
$h\in\{1,\ldots,H\}$, $w\in\{1,\ldots,W\}$	Spatial indices for height and width respectively
$\tilde{h}$, $\tilde{w}$	Latent grid indices
Datasets
$\mathcal{D}=\{(\mathbf{v}_{n},\mathbf{c}_{n})\}_{n=1}^{N}$	Training corpus with size $N$ and index $n$
$\mathcal{D}_{\textnormal{ft}}\subseteq\mathcal{D}$	Fine-tuning dataset
$\mathcal{S}\subseteq\mathcal{D}$	Selected influential subset
$k\in\{1,\ldots,K\}$	The selected subset size
$Q$	Number of query data
$q\in\{1,\ldots,Q\}$	Query index
$\mathbf{x}$	Generic input data pair
$\mathbf{x}_{\textnormal{test}}$, $\mathbf{x}_{n}$	Test/query pair and training pair
$\hat{\mathbf{v}}$, $\hat{\mathbf{c}}$	Query video and its conditioning
Latent Space and Diffusion Components
$E$	VAE encoder
$\mathbf{h}=E(\mathbf{v})\in\mathbb{R}^{F\times(H/s)\times(W/s)\times C}$	Latent video with spatial factor $s$ and channels $C$
$\mathbf{z}$	Noisy latent variable used in diffusion or flow matching
$\bm{\epsilon}\sim\mathcal{N}(0,\mathbf{I})$	Gaussian noise
$\bm{\epsilon}_{\bm{\theta}}(\mathbf{z},\mathbf{c},t)$	Predicted noise network in diffusion training
$\mathbf{f}_{\bm{\theta}}(\mathbf{z},\mathbf{c},t)$	Time-indexed vector field in flow matching
$\dot{\mathbf{z}}$	Time derivative of the latent trajectory
$\alpha_{t},\sigma_{t}$	Scheduler signal and noise scales at timestep $t$
$\bm{\epsilon}_{\textnormal{target}}$	Target noise or velocity used for supervision
$t\in\{1,\ldots,T\}$	Diffusion or flow-matching timestep, with total timesteps $T$
$t_{\textnormal{fix}}$, $\bm{\epsilon}_{\textnormal{fix}}$	Fixed timestep and shared noise draw used for low-variance gradients

Table 3: Glossary and notation.

符号	说明
缩写与基本记号
VAE	变分自编码器（Variational Autoencoder）
DiT	Diffusion Transformer 主干
$\mathbf{I}$	单位矩阵
视频生成
$p_{\bm{\theta}}(\mathbf{v}\mid\mathbf{c})$	参数为 $\bm{\theta}$ 的条件视频生成器
$\mathbf{v}\in\mathbb{R}^{F\times H\times W\times 3}$	帧数为 $F$、高为 $H$、宽为 $W$ 的视频片段
$\mathbf{c}$	条件信号，例如文本或多模态元数据
$\bm{\theta}$	可训练的模型参数
$f\in\{1,\ldots,F\}$	帧索引
$h\in\{1,\ldots,H\}$，$w\in\{1,\ldots,W\}$	高、宽方向的空间索引
$\tilde{h}$，$\tilde{w}$	潜空间网格索引
数据集
$\mathcal{D}=\{(\mathbf{v}_{n},\mathbf{c}_{n})\}_{n=1}^{N}$	规模为 $N$、索引为 $n$ 的训练语料
$\mathcal{D}_{\textnormal{ft}}\subseteq\mathcal{D}$	微调数据集
$\mathcal{S}\subseteq\mathcal{D}$	选出的有影响子集
$k\in\{1,\ldots,K\}$	所选子集的规模
$Q$	查询数据的数目
$q\in\{1,\ldots,Q\}$	查询索引
$\mathbf{x}$	一般的输入数据对
$\mathbf{x}_{\textnormal{test}}$，$\mathbf{x}_{n}$	测试/查询对与训练对
$\hat{\mathbf{v}}$，$\hat{\mathbf{c}}$	查询视频及其条件
潜空间与扩散组件
$E$	VAE 编码器
$\mathbf{h}=E(\mathbf{v})\in\mathbb{R}^{F\times(H/s)\times(W/s)\times C}$	空间因子为 $s$、通道数为 $C$ 的潜空间视频
$\mathbf{z}$	扩散或流匹配中使用的加噪潜变量
$\bm{\epsilon}\sim\mathcal{N}(0,\mathbf{I})$	高斯噪声
$\bm{\epsilon}_{\bm{\theta}}(\mathbf{z},\mathbf{c},t)$	扩散训练中的预测噪声网络
$\mathbf{f}_{\bm{\theta}}(\mathbf{z},\mathbf{c},t)$	流匹配中随时间索引的向量场
$\dot{\mathbf{z}}$	潜空间轨迹对时间的导数
$\alpha_{t},\sigma_{t}$	时间步 $t$ 上调度器的信号尺度与噪声尺度
$\bm{\epsilon}_{\textnormal{target}}$	用于监督的目标噪声或速度
$t\in\{1,\ldots,T\}$	扩散或流匹配的时间步，总步数为 $T$
$t_{\textnormal{fix}}$，$\bm{\epsilon}_{\textnormal{fix}}$	用于低方差梯度的固定时间步与共享噪声抽取

表 3：术语与符号表。