---
title: "NVIDIA's MOTIVE Beats Full Training Data Using Just 10% of Video Clips"
subtitle: "NVIDIA's MOTIVE framework scores every training clip by its motion influence, letting you fine-tune video models on just 10% of data for a 74.1% human preference win rate"
url: "https://alphasignal.ai/news/nvidia-s-motive-beats-full-training-data-using-just-10-of-video-clips"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "NVIDIA AI"
company_url: "https://alphasignal.ai/company/nvidiaai"
published: "2026-07-07T19:28:29.000Z"
updated: "2026-07-08T02:23:24.964Z"
type: "News"
topics:
  - "Video"
  - "Data"
subtopics:
  - "Video Generation"
  - "Data Curation"
  - "Synthetic Data"
upvotes: 207
read_time: "7 min"
source_url: "https://research.nvidia.com/labs/sil/projects/MOTIVE/?linkId=100000429733367"
comments: 0
---

# NVIDIA's MOTIVE Beats Full Training Data Using Just 10% of Video Clips

> NVIDIA's MOTIVE framework scores every training clip by its motion influence, letting you fine-tune video models on just 10% of data for a 74.1% human preference win rate

Video generation has made enormous strides in visual quality, but temporal dynamics , the way objects actually move , remain a stubborn weak point. The core problem is surprisingly basic: nobody has had a principled way to tell which training clips are actually teaching a model *how to move*. NVIDIA Research's new framework, [MOTIVE](https://research.nvidia.com/labs/sil/projects/MOTIVE/) (MOTIon attribution for Video gEneration), tackles exactly that. Presented as an Oral and Outstanding Paper Honorable Mention at ICML 2026, it is the first framework to attribute motion , not visual appearance , in video generative models, and to use that signal to curate fine-tuning data.

## The Problem Nobody Solved

Motion is the defining element of videos. Unlike image generation, which produces a single frame, video generative models must capture how objects move, interact, and obey physical constraints. Yet even with rapid progress, a fundamental question remains: which training clips influence the motion in generated videos?

Despite the rapid progress of video generation models, the role of data in influencing motion is poorly understood. Prior work on data attribution , the practice of scoring training examples by how much they influence a model's outputs , was built for images. Naively applying those methods to video collapses motion into appearance, ranking clips high because they share similar backgrounds or object textures, not because they teach similar dynamics. Three specific challenges make this hard:

- **Localizing motion:** attribution must focus on dynamic regions, not static backgrounds
- **Scaling to sequences:** gradients must integrate across time, not just space
- **Capturing temporal structure:** velocity, acceleration, and trajectory coherence cannot be measured one frame at a time

## What MOTIVE Does

MOTIVE is a scalable, motion-centric data attribution framework for video generation that identifies which training clips improve or degrade motion dynamics. The key idea is to compute **influence scores** , a measure borrowed from classical statistics that asks: if I upweight this training example, how much does the model's behavior change? , but to compute them in a way that only counts the motion-related part of the gradient signal.

MOTIVE isolates temporal dynamics from static appearance via motion-weighted loss masks, yielding efficient and scalable motion-specific influence computation. Once every clip in a fine-tuning dataset has a score, you can rank them, take the top 10%, and fine-tune only on that high-influence subset.

## The Pipeline, Step by Step

Motion-gradient computation has three steps: detect motion with AllTracker; compute motion-magnitude patches; and apply loss-space motion masks to focus gradients on dynamic regions. Here is what each step actually does:

1. **Motion detection:** AllTracker (a dense point-tracking model) extracts per-pixel optical flow from every training clip, producing displacement vectors that encode how far each pixel moved between frames.
2. **Motion masking:** Those displacement magnitudes are min-max normalized to [0,1] and downsampled to match the model's latent space. The result is a spatial weight map where moving regions score near 1 and static backgrounds score near 0.
3. **Motion-weighted loss:** The standard diffusion denoising loss is reweighted by these spatial maps before backpropagation. Gradients from static pixels are suppressed; gradients from moving pixels are amplified. The influence score for a training clip is then the cosine similarity between its motion-weighted gradient and the query video's motion-weighted gradient.

Making this tractable at billion-parameter scale required several engineering choices. The method is made scalable via a single-sample variant with common randomness and a projection, computed for each pair of training and query data, aggregated to a final ranking that selects fine-tuning subsets. Concretely, gradients are compressed from roughly 1.4 billion dimensions down to just 512 using a structured random projection (Fastfood), and only a single fixed denoising timestep is used per clip rather than averaging across all timesteps. The paper shows that a projection dimension of 512 achieves 74.7% Spearman correlation with full-gradient rankings, while being orders of magnitude cheaper to store and compute.

A subtle but important fix: raw gradient magnitudes grow with clip length, which would bias rankings toward longer videos regardless of motion quality. MOTIVE normalizes gradients by frame count before projection, which the paper shows reduces spurious length-based correlations by 54%.

## Two Selection Modes

MOTIVE supports two practical workflows depending on your goal:

- **Single-query selection:** You provide one target motion (e.g., a video of a coin spinning), score every training clip against it, and take the top-K. Best for specialist fine-tuning on a specific motion type.
- **Multi-query aggregation (majority voting):** You provide multiple query videos across different motion categories. A training clip earns a vote for each query where its score exceeds a percentile threshold. Clips that rank high across many motion types get selected. Best for generalist fine-tuning where you want broad motion improvement.

## Benchmarks and Results

The team evaluated on two large-scale video datasets , VIDGEN-1M and 4DNeX-10M , using 10,000 clips each. The primary model was Wan2.1-T2V-1.3B, a widely used open-source text-to-video backbone, with additional experiments on the larger Wan2.2-TI2V-5B. Evaluation used [VBench](https://arxiv.org/html/2601.08828), a standard benchmark covering six dimensions: subject consistency, background consistency, motion smoothness, dynamic degree, aesthetic quality, and imaging quality. Motion smoothness and dynamic degree were the primary targets.

| Method | Dynamic Degree (%) | Motion Smooth. (%) | Subject Consist. (%) |
| --- | --- | --- | --- |
| Base model | 39.6 | 96.3 | 95.3 |
| Full fine-tuning (100% data) | 42.0 | 96.3 | 95.9 |
| Random selection (10% data) | 41.3 | 96.3 | 95.3 |
| MOTIVE w/o motion masking (10%) | 43.8 | 96.3 | 95.4 |
| **MOTIVE (10% data)** | **47.6** | **96.3** | **96.3** |

With MOTIVE-selected high-influence data, the method improves both motion smoothness and dynamic degree on VBench, achieving a 74.1% human preference win rate compared with the pretrained base model. Critically, MOTIVE identifies clips that strongly affect motion and guides data curation that improves temporal consistency and physical plausibility. The 47.6% dynamic degree score using only 10% of training data beats the full fine-tuning score of 42.0% , meaning careful data selection outperforms using everything.

Human evaluation involved 17 annotators comparing 50 video pairs across 10 motion categories (bounce, compress, explode, float, free fall, roll, slide, spin, stretch, swing). MOTIVE won 74.1% of comparisons against the base model, 58.9% against random selection, and 53.1% against full fine-tuning.

## The Real Breakthrough

To the authors' 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. That distinction matters more than it might seem. The ablation comparing MOTIVE with and without motion masking is telling: without it, dynamic degree reaches 43.8%; with it, 47.6%. The motion mask is doing real work , it is not just selecting clips that happen to have a lot of movement. The paper shows that the top 10% selected clips have a mean motion magnitude only 4.3% higher than the bottom 10%, confirming that influence scores capture something fundamentally different from raw motion quantity.

The compute story is also practical. On a single A100 GPU, computing influence scores for 10,000 clips takes about 150 hours , but that cost is a one-time investment. Once gradients are stored as 512-dimensional vectors, adding a new query takes about 54 seconds. The process is also embarrassingly parallel: NVIDIA said 74 of its papers were accepted at ICML 2026, and MOTIVE is among the most practically deployable of the lot, since the gradient computation scales linearly across GPUs and can be reduced to roughly 2.3 hours on 64 A100s.

## Limitations and What's Next

The framework has honest limitations. Motion saliency depends on the quality of the optical flow tracker , severe occlusions or transparent objects can degrade the motion masks. Camera-only motion (where the scene moves but objects don't) is hard to separate from genuine object dynamics without additional signals like camera pose. And the per-clip attribution misses the possibility that only certain segments of a clip carry the useful motion signal.

The [paper](https://arxiv.org/html/2601.08828) flags several future directions: finer-grained attribution at the motion-segment level, closed-loop curation that iterates between attribution and fine-tuning, and extensions to world models and audio. Code is listed as coming soon on the [project page](https://research.nvidia.com/labs/sil/projects/MOTIVE/). For anyone building or fine-tuning video generation models today, MOTIVE offers a concrete answer to a question the field has been sidestepping: not all training data is equal, and now there is a way to tell which clips actually matter for motion.

## Sources

- [X post by @nvidiaai](https://x.com/nvidiaai/status/2074576305790546139)
- [MOTIVE: Motion Attribution for Video Generation — NVIDIA Research Project Page](https://research.nvidia.com/labs/sil/projects/MOTIVE/)
- [Motion Attribution for Video Generation — arXiv paper](https://arxiv.org/html/2601.08828)
- [ICML 2026 Orals — MOTIVE listing](https://icml.cc/virtual/2026/events/oral)
- [Nvidia says its models underpin ICML 2026 research — IT Brief](https://itbrief.news/story/nvidia-says-its-models-underpin-icml-2026-research)