Training strong video generation models usually requires massive datasets, large parameter counts, and substantial compute. In this work, we ask whether strong text-to-video quality is possible at a much smaller budget: fewer than 10M clips and less than 100,000 H200 GPU hours. Our core claim is that part of the answer lies in how model capacity is organized, not only in how much of it is used. In video generation, prompt alignment, temporal consistency, and fine-detail recovery can interfere with one another when they are handled through the same pathway. Motif-Video 2B addresses this by separating these roles architecturally, rather than relying on scale alone. The model combines two key ideas. First, Shared Cross-Attention strengthens text control when video token sequences become long. Second, a three-part backbone separates early fusion, joint representation learning, and detail refinement. To make this design effective under a limited compute budget, we pair it with an efficient training recipe based on dynamic token routing and early-phase feature alignment to a frozen pretrained video encoder. Our analysis shows that later blocks develop clearer cross-frame attention structure than standard single-stream baselines. On VBench, Motif-Video~2B reaches 83.76%, surpassing Wan2.1 14B while using 7× fewer parameters and substantially less training data. These results suggest that careful architectural specialization, combined with an efficiency-oriented training recipe, can narrow or exceed the quality gap typically associated with much larger video models.
Figures & tables
Figure 1: Representative generations from Motif-Video 2B. Frames are captured from videos generated by our 2B-parameter text-to-video model across a diverse set of prompts, illustrating the combination of prompt fidelity, temporal coherence, and visual detail that we target throughout this work. The banner is intended as a qualitative teaser; later sections analyze the architectural and training choices that make these generations possible under a micro-budget training regime.
Figure 2: Overview of Motif-Video 2B. Text is encoded by T5Gemma2, while video frames are compressed by the Wan2.1 VAE into spatiotemporal latents and patchified into tokens. The transformer backbone follows a three-stage design that separates early modality fusion, joint text-video representation learning, and final detail reconstruction: 12 dual-stream layers preserve modality-specific processing during early fusion, 16 single-stream layers build a joint text-video representation, and 8 DDT decoder layers serve as a dedicated decoder for high-frequency detail reconstruction. Shared Cross-Attention is attached to the single-stream stage to reinforce text conditioning under long-context token imbalance by using learned query/output projections while reusing the enclosing block’s self-attention key and value projections. The denoised latent is then unpatchified and decoded by the VAE to produce the final video.
Figure 3: Attention structure in dual-stream vs. single-stream vs. DDT decoder layers. Compared with dual and single-stream layers, DDT decoder layers show stronger inter-frame attention structure, where each frame attends more to temporally adjacent frames. The blue box denotes the encoder hidden state: text tokens in the dual-stream and single-stream cases, and the video output tokens from the encoder layers in the decoder case.
Figure 4: Intermediate-layer text-attention drop in single-stream blocks. We compare attention maps from a representative intermediate layer in dual-stream and single-stream stages. Relative to dual-stream, the single-stream intermediate layer allocates substantially less attention mass to text tokens, indicating weaker text conditioning under joint-token competition.
Figure 5: Zero-init alone does not save a cross-attention whose K,V geometry is ungrounded. Both variants are inserted into the same pretrained 360p checkpoint with WOcross=0 , making both forward passes identical to the base model at step 0. After 1,000 steps of continued training under matched optimizer settings, data, and learning rate, the SkyReels-V4–style cross-attention ( top ) collapses: outputs degenerate to near-black frames with fragmented, incoherent structure, while Shared Cross-Attention ( bottom ) continues training without disruption and produces coherent scenes. Each column shows samples from the same prompt under the same seed. The contrast directly supports the manifold argument of Section 3.3 : grounding K,V in self-attention’s existing projections is what makes a new module stable to insert mid-training.
Stage
Task
Resolution
Frames
Text Encoder
REPA
Notes
1
T2I
144p
1
Sent. Emb.
DINOv2
Spatial bootstrap
2
T2IV
144p
33
Sent. Emb.
V-JEPA
3
T2IV
144p
65
Sent. Emb.
V-JEPA
4
T2IV
360p
65
T5Gemma2
–
REPA disabled
5
T2IV
480p
65
T5Gemma2
–
Res. bridge first
6
T2IV
480p
121
T5Gemma2
–
Table 1: Simplified training curriculum for Motif-Video 2B. Joint image–video training is used at all video stages. REPA is disabled from Stage 4 onward following evidence that alignment becomes counterproductive after early convergence [ 40 ] . Shared Cross-Attention is introduced at Stage 9 to address semantic degradation observed at 720p
Figure 6: Dense features from V-JEPA 2.0. The visualization highlights that, while V-JEPA 2.0 captures global motion structure well, its dense features are less spatially coherent than would be ideal for dense REPA supervision in video generation.
Figure 7: Overview of the training-data construction pipeline. The raw pool is split into Image Real, Image Synthetic, Video Real, and Video Synthetic branches. An initial sanitation stage removes broken files, abnormally small files, near-duplicates (SSCD-based), NSFW content, and watermarked content. Surviving clips are progressively filtered by resolution, clip length, motion, and aesthetic signals as they advance through the 144p, 360p, 480p, and 720p training stages, and by stricter aesthetic, domain, and dynamic-motion criteria before the cross-attention refinement and final 720p SFT stage. The Sankey diagram visualizes how flows contract from the raw pool toward the curated training and SFT corpora.
Figure 8: Subject composition of the cross-attention fine-tuning corpus. The corpus was assembled iteratively by curating additional clips from underperforming categories. Left: image distribution. Right: video distribution.
Figure 9: Overview of our offline bucket-balanced sampler for WebDataset-formatted video corpora on W GPUs. An offline planner consumes clip metadata to apply filtering and frame-resolution bucketing, assigns tar shards to ranks, and emits per-rank schedules that preserve sequential archive reads during training.
Method
Steps / epoch
Utilization
DL latency
Baseline (clip shuffle)
N
≈20%
0.05 s/step
Offline sampler + greedy
≈4.6N
≈76%
< 0.001 s/step
Offline bucket sampler + SA (ours)
≈5.4N
≈90%
< 0.001 s/step
Table 2: Relative data utilization per epoch on W GPUs.
Model
Params
Aggregate
Quality Dimensions
Semantic Dimensions
Total ↑
Quality ↑
Semantic ↑
Subj. Cons.
Bg. Cons.
Temp. Flick.
Motion Sm.
Dyn. Degree
Aesth. Qual.
Img. Qual.
Object Class
Multiple Obj.
Human Action
Color
Spatial Rel.
Scene
App. Style
Temp. Style
Ovr. Consist.
Veo 3 [ 42 ] †
–
85.06
85.70
82.49
97.36
96.89
99.30
99.16
72.43
63.81
68.23
93.89
82.20
99.40
82.48
84.26
57.43
23.55
25.97
27.88
Sora [ 4 ] †
–
84.28
85.51
79.35
96.23
96.35
98.87
98.74
79.91
63.46
68.28
93.93
70.85
98.20
80.11
74.29
56.95
24.76
25.01
26.26
Luma †
–
83.61
83.47
84.17
97.33
97.43
98.64
99.35
44.26
65.51
66.55
94.95
82.63
96.40
92.33
83.67
58.98
24.66
26.29
28.13
HunyuanVideo †α [ 21 ]
–
83.43
85.07
76.88
97.22
97.60
99.39
99.05
71.94
60.28
67.24
83.48
66.71
94.40
89.79
72.13
54.46
22.21
24.52
26.95
MiniMax-Video-01 †
–
83.41
84.85
77.65
97.51
97.05
99.10
99.22
64.91
63.03
67.17
87.83
76.04
92.40
90.36
75.50
50.68
20.06
25.63
27.10
Table 3: VBench T2V evaluation across all 16 fine-grained dimensions (scores in %). Bold and underline denote the best and second-best results among open-source models with full dimension scores, respectively. † Closed-source; excluded from open-source rankings. p Evaluated with prompt optimisation (Qwen-rewritten prompts for Wan2.2; SAT-enhanced for CogVideoX1.5-5B). α Updated HunyuanVideo API checkpoint (2025-05-22); open-source weights not released for this version. ⋆ SANA-Video aggregate scores from our unified evaluation protocol [ 8 ] ; individual dimension scores not publicly reported by the authors. § Scores from the original paper [ 53 ] under a T2I2V pipeline (FLUX anchor frame); individual dimension scores not reported. Direct comparison with standard T2V models should be interpreted with caution. All other scores sourced from the VBench Leaderboard [ 18 ] .
Figure 10: Shared Cross-Attention contribution across single-stream encoder blocks and denoising steps ( 1280×736 , 121 frames, 50 steps, σ∈[1.00,0.29] ). Left : Frobenius norm of the cross-attention output WOcrossAttn(Q,K,V) per block (row) and step (column). Right : ratio of the cross-attention residual norm to the self-attention output norm ∥hv∥ . No block falls below 5.2%; the global mean is 7.6% and the maximum 21.7%. Activity peaks at block 0 and in the high-noise regime, stabilising after step 22 ( σ≈0.96 ). The cross-attention residual is nearly orthogonal to the self-attention output (global cosine ≈−0.008 ), confirming that the module injects text-grounded information along directions absent from the self-attention pathway.
Figure 11: Selected single-frame samples from Motif-Video 2B across a range of subjects and visual styles. Each tile is a frame drawn from an independently generated text-to-video clip. The grid is intended to convey the breadth of domains the model handles, including photographic scenes, stylized and fantastical content, close-up subjects, and wide landscapes, rather than to claim uniform quality across all prompts.
Figure 12: Image-to-video generation results. The leftmost panel is the input image, and the model preserves its original appearance while generating temporally coherent video content from it.
Model
Params
Total ↑
Prompt-following ↑
Video-fidelity ↑
Wan2.1
14B
1114.8
1054.4
1101.6
HunyuanVideo-1.5
6.5B
1083.0
1045.2
1043.8
Wan2.2
5B
1062.4
1039.1
1049.9
LTX-Video 2
14B
1041.5
1005.0
1043.7
Motif-Video (ours)
2B
1026.2
1015.1
1047.3
Wan2.1
1.3B
1015.6
1010.1
1006.5
Table 4: Human evaluation results. Pairwise preferences are converted to ELO ratings. Total aggregates all pairwise judgments across both axes; prompt-following measures whether the generated video matches the input text; video-fidelity measures visual coherence and plausibility, independent of the prompt. Models are sorted by Total ELO.
Figure 13: Example of generated results from the arena. Prompt: ”A guitarist sits on a fire escape playing at twilight, fingers moving in relaxed patterns along the neck of a scratched acoustic guitar. Shot on a 40mm lens with a slow crane-up from the street below, the brick wall beside him glows deep orange as the last sun hits it and the sky above shifts toward indigo. He wears a loose denim shirt rolled to the elbows, a leather bracelet knocking softly against the guitar body. A potted plant beside him sways in the warm updraft. The camera rises past him to reveal the skyline beginning to glitter with early window lights.” Note that even high-fidelity models sometimes fail on certain examples, but we have observed such models, for example Wan2.1 14B, consistently show better results, unlike the example shown here.
Figure 14: Micro-scale semantic distortion. Three characteristic failures at the sub-object level: distorted hand anatomy on a close-up instrument subject (left), broken body structure under a high-motion skydiving prompt (middle), and attribute leakage between co-present animals in a multi-subject scene (right). The generations may remain category-correct (guitar, skydiver, cat and dog), leading VBench’s semantic dimensions award credit, but a human viewer flags the artifact on first inspection.
Figure 15: Temporal failure modes. Top: physically implausible liquid dynamics in a wine-splash prompt: the motion is locally smooth but violates gravity and surface tension. Middle: loss of temporal coherence under high scene complexity in a cavalry-charge prompt, where subject identities blur across frames and multi-agent spatial relationships fail to persist. Bottom: unintended mid-clip scene transition, where the prompted setting drifts into an unrelated composition partway through the sequence.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 16: Additional qualitative human-centered generations. Representative frames from videos involving human subjects, included as supplementary qualitative results.
Figure 17: Additional image-to-video results. The leftmost panel is the input image, and the remaining panels show representative generated video frames.
Figure 18: Qualitative effect of Shared Cross-Attention. For each prompt, the top row shows generation with Shared Cross-Attention enabled; the bottom row shows the same prompt and seed with cross-attention disabled on all 16 single-stream encoder blocks (360p, 50 steps, 121 frames).
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
We present Paris 2.0, the first video generation model pre-trained through decentralized computation. Its training recipe builds upon Paris 1.0 (arXiv:2510.03434), the first ever open-weight Decentralized Diffusion Model (DDM), which showed that image generation can be trained without a monolithic GPU cluster. However, temporally coherent video generation had remained an open problem under decentralized training, and Paris 2.0 closes it. In low-resolution text-to-video training, against a monolithic model trained on the same data under a matched total compute budget, Paris 2.0 cuts Frechet Video Distance (FVD) from 561.04 to 279.01, a ~2.0x improvement, and lifts CLIP text-video similarity and aesthetic score.
Rapid progress in video generation has led to a plethora of models that differ substantially in capability and generation cost. This raises a natural question: can each request be efficiently routed to an appropriate model? We find that even when the consensus of the other annotators is used as an oracle, it agrees with each annotator's own favorite only 34-55% of the time. Motivated by this observation, we introduce TasteRoute, a personalized video-generation router that selects a generator jointly based on the input request, user preferences, and available generation budget. Across text-to-video and image-to-video settings, TasteRoute is competitive with strong simple baselines on preference routing while reducing average generation cost. The cost saving increases under higher budget caps. Finally, we release TasteRoute-3k, a human-annotated dataset containing multi-model video comparisons, quality judgments, preference rankings, and user-profile signals to facilitate future research on personalized and cost-aware video routing.
Zhi Rui Tam, Chao-Chung Wu, Sin-Han Yang +8
National Taiwan University · Appier Inc. · Cornell University