MOTIVE: Motion Attribution for Video Generation

TL;DR

MOTIVE (Motion Attribution for Video Generation, ICML 2026 Spotlight) is a gradient-based data-attribution method that ranks training videos by their influence on a motion-specific loss, then uses the top-ranked subset as finetuning data for a video generation model. The key trick is a motion-weighted gradient: optical flow from AllTracker is converted into a patch-level motion mask that downweights static-appearance regions before computing per-sample gradients, so attribution scores reflect temporal dynamics rather than texture similarity. Finetuning on the top-10% subset matches or beats training on the full corpus and wins a reported 74.1% human-preference rate against the baseline. Two selection modes are supported: optimize for a single target motion, or aggregate across multiple motion queries by majority vote for diverse coverage.

Key claims

• Gradient-based influence attribution can be made tractable for billion-parameter video models via a single-sample variant with common randomness and a projection, computed per (training, query) pair and aggregated into a final ranking [§1 Scalable Gradient-Based Attribution].
• Motion masking — optical-flow-derived patch weights applied in loss space — isolates temporal dynamics from static appearance during attribution, changing which training clips are deemed influential [§2 Motion Attribution via Motion Masking, qualitative pairs].
• A 10% Motive-selected finetuning subset achieves comparable or better VBench scores (subject consistency, background consistency, motion smoothness, dynamic degree, aesthetic quality, imaging quality) than random 10% selection [Table 1].
• Human evaluators prefer Motive-finetuned outputs over the baseline at a 74.1% win rate across 10 motion categories with 17 participants over 850 pairwise comparisons [Table 2, Results].
• The same attribution framework supports two operating modes: optimize for a single specific motion query, or use majority-vote aggregation across multiple motion queries with a score threshold for diverse motion coverage [§3 Influential Subset Selection].
• Without motion masking, attribution collapses onto appearance similarity; with motion masking, the same query retrieves clips with matching motion patterns even when appearance differs [Qualitative comparison, §2].

Method

Motive has three stages. (1) Motion detection — AllTracker produces dense optical flow over each training clip, which is converted to a patch-level motion-magnitude map. (2) Motion-weighted gradients — the per-sample training loss is masked by the motion map before backprop, so gradient contributions from static regions are suppressed and gradients from dynamic regions dominate. (3) Influence attribution — for each (training, query) pair, a gradient-similarity score is computed under a scalable single-sample-with-common-randomness variant plus a projection, making the per-pair cost low enough to run over a full corpus against target motion queries. Scores are aggregated either by selecting the top-K against a single target motion or by majority-vote thresholding across multiple queries. The resulting subset (typically 10% of the corpus) is then used as the finetuning set for a video generation backbone.

Results

On VBench, Motive’s 10% subset is reported as competitive with or outperforming random 10% selection across the six standard axes (subject/background consistency, motion smoothness, dynamic degree, aesthetic quality, imaging quality) [Table 1]. The human study (50 videos × 17 participants × pairwise = 850 judgments across 10 motion categories) gives Motive a 74.1% win rate over the baseline [Table 2, Results]. The qualitative pairs side-by-side show that with motion masking, retrieved clips share motion patterns rather than appearance — the ablation case for the masking step being load-bearing rather than a cosmetic addition [§2].

Why it’s interesting

Motive is the first filed example of attribution-driven finetuning data selection for a video generation model, and it lands directly on the team’s data-efficiency thesis: at fixed quality, do less work by picking the right 10%. Mechanically it sits next to Nemotron-CLIMB: CLustering-based Iterative Data Mixture Bootstrapping for Language Model Pre-training (CLIMB) — both papers argue that a learned, automated curation gate dominates random sampling — but Motive’s gate is per-sample influence-on-motion rather than per-cluster mixture weights, and operates at finetuning scale rather than pretraining scale. The motion-masking step is specifically what stops attribution from collapsing onto appearance, which is the failure mode any naive influence-function setup on video would hit; this design choice is the most transferable lesson for any Luma project doing targeted finetuning over a motion or dynamics axis.

See also

• Synthetic Training Data — Motive is a curation-gate datapoint over real data, parallel to CLIMB; the broader thesis is that automated selection beats scale at finetuning time.
• Nemotron-CLIMB: CLustering-based Iterative Data Mixture Bootstrapping for Language Model Pre-training — CLIMB curates pretraining-scale mixtures via clustering + benchmark-validated proxy search; Motive does per-sample influence attribution at finetuning scale — same thesis, different mechanism and scale.
• Video-Thinker: Sparking "Thinking with Videos" via Reinforcement Learning — also makes the “10% of the data with the right curation gate matches full-corpus training” claim, but on a video-reasoning MLLM rather than a generative video model.