Recent advances in video generation models have achieved impressive results. However, these models heavily rely on the use of high-quality data that combines both high visual quality and high motion quality. In this paper, we identify a key challenge in video data curation: the Motion-Vision Quality Dilemma. We discovered that visual quality and motion intensity inherently exhibit a negative correlation, making it hard to obtain golden data that excels in both aspects. To address this challenge, we first examine the hierarchical learning dynamics of video diffusion models and conduct gradient-based analysis on quality-degraded samples. We discover that quality-imbalanced data can produce gradients similar to golden data at appropriate timesteps. Based on this, we introduce the novel concept of Timestep selection in Training Process. We propose Timestep-aware Quality Decoupling (TQD), which modifies the data sampling distribution to better match the model's learning process. For certain types of data, the sampling distribution is skewed toward higher timesteps for motion-rich data, while high visual quality data is more likely to be sampled during lower timesteps. Through extensive experiments, we demonstrate that TQD enables training exclusively on separated imbalanced data to achieve performance surpassing conventional training with better data, challenging the necessity of perfect data in video generation. Moreover, our method also boosts model performance when trained on high-quality data, showcasing its effectiveness across different data scenarios.
Figures & tables
Figure 1: The motion-vision dilemma in video data. (a) Illustrates two cases: a high-MQ, low-VQ action scene (left) and a low-MQ, high-VQ static scene (right), with their corresponding optical flow maps below visualizing the difference in motion intensity. (b) The 2D density plot over Koala36M dataset confirms this trade-off. MQ and VQ exhibit a negative correlation, with the majority of samples ( 56.3% total) falling into the high-VQ/low-MQ or low-VQ/high-MQ quadrants, as defined by the median lines. Data with both high qualities is less common ( 21.9% ).
Figure 2: (a) Video diffusion models exhibit hierarchical denoising: high-noise stages ( t=0.9,0.6 ) establish motion and composition, while low-noise stages ( t=0.3,0.0 ) refine details and textures. (b) Gradient analysis under quality degradations, averaged over 120 samples with three VQ degradation types (blur, compression, noise) and MQ degradation (shuffle). VQ degradation aligns with the original at high timesteps, while MQ degradation aligns at low timesteps, revealing that quality-imbalanced samples can match golden data’s gradients at appropriate timesteps and should be strategically allocated across the denoising process.
Figure 3: Overview of our timestep selective training process. Given bimodal quality distributions in training data (left), TQD employs quality-based sample dropout (middle) and adaptive timestep sampling via Beta distributions (right). Samples with high motion intensity are directed toward large timesteps, those with high visual quality toward small timesteps, achieving specialized learning across denoising process.
Model
Setting
Method
VBench
VideoAlign
IQ
Aesthetic
Dynamic
MotionSmooth
MQ
VQ
Wan 1.3B
-
Basemodel
0.6935
0.5874
0.5938
0.9889
2.1993
3.3236
Set-A
Baseline
0.6916
0.5722
0.5312
0.9935
2.1388
3.2537
TQD
0.7010
0.5757
0.6384
0.9923
2.2557
3.3450
Set-B
Baseline
0.6769
0.5702
0.5224
0.9923
2.0905
3.2338
TQD
0.6921
0.5668
0.5447
0.9935
2.1477
3.2679
Table 1: Quantitative comparison on VBench and VideoAlign metrics across different data settings. Best results in each setting are shown in bold .
Model
SA ↑
PC ↑
Wan 1.3B
3.738
3.690
Baseline
3.786
3.709
Data Filter Only
3.795
3.756
TQD (Ours)
3.820
3.843
Table 2: Quantitative comparison on Subject Alignment (SA) and Prompt Consistency (PC) metrics with VideoPhy2.
Figure 4: Qualitative comparison. TQD demonstrates superior performance across multiple quality dimensions. Top rows: improved visual quality with reduced hand distortion (left) and better motion coherence matching prompt dynamics (right). Bottom rows: enhanced physical plausibility in foam dissipation (left) and liquid dynamics (right), as highlighted in red boxes.
Method
VBench
VideoAlign
IQ
Aesthetic
Dynamic
MotionSmooth
MQ
VQ
Baseline
0.6916
0.5722
0.5312
0.9935
2.1388
3.2537
+ Adaptive timestep
0.6972
0.5735
0.6340
0.9937
2.2193
3.3044
+ Quality dropout
0.6956
0.5701
0.5357
0.9916
2.1909
3.2921
+ Both (TQD)
0.7010
0.5757
0.6384
0.9923
2.2557
3.3450
Table 3: Component ablation of TQD on Wan Set-A. We VBench metrics alongside VideoAlign scores to dissect motion-visual synergies.
κmax
CogVideoX
Wan
MQ ↑
VQ ↑
MQ ↑
VQ ↑
10
2.2018
3.2596
2.2282
3.2808
20
2.2345
3.2570
2.2338
3.3020
30
2.2249
3.2543
2.2557
3.3450
40
2.2192
3.2535
2.2503
3.3152
50
2.2199
3.2504
2.2378
3.3089
Table 4: Ablation study on concentration parameter κmax across CogVideoX and Wan architectures.
Figure 5: Timestep sampling distributions under varying quality scores. We visualize the Beta distributions with κbase=4 (top) and κbase=2 (bottom) for samples with different MQ and VQ scores. Left: High MQ, low VQ shifts toward large timesteps. Middle: Equal MQ and VQ yields centered distributions. Right: Low MQ, high VQ concentrates on small timesteps. Note that κbase=2 degenerates to uniform when MQ = VQ (middle-bottom).
Figure 6: Data distribution of our collected training set.
Figure 7: Dense video captioning example. We leverage Gemini 2.5 Pro with structured prompts (top) to generate comprehensive video descriptions. The bottom panel shows an example caption (300+ words) for a wedding scene, demonstrating fine-grained descriptions of subjects’ clothing, spatial dynamics, lighting, and camera techniques. This structured approach ensures annotation quality and consistency across our 280K training dataset.
Method
IQ
Aesthetic
Dynamic
Smooth
MQ
VQ
Uniform (no TQD)
0.6916
0.5722
0.5312
0.9935
2.1388
3.2537
TQD (0% noise)
0.7010
0.5757
0.6384
0.9923
2.2557
3.3450
TQD (10% noise)
0.6985
0.5743
0.6251
0.9931
2.2431
3.3285
TQD (20% noise)
0.6953
0.5731
0.6012
0.9928
2.2114
3.3085
Table 5: Robustness to scorer noise on Wan Set-A. TQD maintains substantial gains over the baseline even with 20% noise.
Video diffusion models generate or edit temporally coherent video by iteratively denoising latent sequences conditioned on text, images, motion, or other signals.
Learning-based video quality assessment (VQA) has advanced rapidly, yet progress is increasingly constrained by a disconnect between model design and dataset curation. Model-centric approaches often iterate on fixed benchmarks, while data-centric efforts collect new human labels without systematically targeting the weaknesses of existing VQA models. Here, we describe MDS-VQA, a model-informed data selection mechanism for curating unlabeled videos that are both difficult for the base VQA model and diverse in content. Difficulty is estimated by a failure predictor trained with a ranking objective, and diversity is measured using deep semantic video features, with a greedy procedure balancing the two under a constrained labeling budget. Experiments across multiple VQA datasets and models demonstrate that MDS-VQA identifies diverse, challenging samples that are particularly informative for active fine-tuning. With only a 5% selected subset per target domain, the fine-tuned model improves mean SRCC from 0.651 to 0.722 and achieves the top gMAD rank, indicating strong adaptation and generalization.
Reliable video generation requires more than high-quality frames to form a coherent story: a model must maintain a persistent state, transporting visual attributes such as identity, scene layout, motion, and fine details across time. Existing training-free methods mainly strengthen cross-frame attention or analyze local attention entropy, but these views do not reveal whether temporal interactions stay in a healthy transport regime. In this work, we study video generation through the perspective of Temporal State Transport. We introduce Spectral Tension, a signed diagnostic that compares local attention diffuseness with global spectral diversity, and use it to identify two opposite temporal failures: fragmented transport and over-mixing hotspots. Based on this diagnosis, we propose Spectral Transport Homeostasis, a training-free regulator that softly corrects pathological temporal states while largely preserving balanced ones. Experiments on pretrained video generation models show that the original model often occupies imbalanced temporal regimes, whereas our method selectively applies larger corrections to the worst temporal hotspots and improves temporal consistency and visual quality without finetuning. Code: https://github.com/lytang63/temporal-state-transport
Luyao Tang, Bingjun Luo, Dong Yi +5
The University of Hong Kong, Hong Kong SAR, China · Tsinghua University, Beijing, China · Harbin Engineering University, Harbin, China +2
Despite the rapid progress of video generation models, the role of data in influencing motion is poorly understood. We present Motive (MOTIon 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, our method improves 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.
Xindi Wu, Despoina Paschalidou, Jun Gao +5
NVIDIA · Princeton University · University of Michigan +3