Video generation supports digital content creation, with applications in virtual avatars, AI-driven filmmaking, interactive game engines, and world models. State-of-the-art systems achieve controllability primarily through image-to-video (I2V) generation, in which an injected anchor image determines the identity, composition, and style of the resulting video. In modern pipelines, this anchor is synthesized by a text-to-image (T2I) generator. This raises a largely overlooked question: Are images obtained directly from the T2I model truly suitable for current video models? Our analysis shows that anchor aesthetic and quality scores are weakly correlated with downstream video quality. The structure-and-texture score is negatively correlated with I2V video quality in both the visual and motion dimensions. Fine-tuning the T2I model on an image reward accordingly yields anchors with visibly denser structural and character detail: they score higher on image metrics, yet lower the video reward on every I2V model we test and change downstream motion unpredictably. We therefore introduce Golden Anchor Learning (GAL), to our knowledge the first framework for learning golden anchors by optimizing a T2I generator with downstream video feedback. GAL evaluates candidate anchors through the videos they induce under a frozen I2V model and converts video rewards into image-level preferences for DPO fine-tuning of the T2I generator. On all four I2V models, GAL raises the VBench average by 2.97%--8.46% over the untuned generator. The rendered clips also score higher under VideoReward-Overall, and their extracted frames under HPSv3 and UniPercept. GAL yields consistent motion gains across all four I2V models, as measured by a 5.62%--38.24% increase in VBench dynamic degree. Although GAL is trained only on HunyuanVideo-1.5 renders, these gains extend to the other three I2V models.
Figures & tables
Figure 1: In the shown examples, the GAL anchor induces more in-scene motion in every one-second interval across all four I2V models. Top: within each I2V-model block, rows are clips rendered from Base (the anchor produced by the untuned image model), GAL-Image (trained based on image rewards), and our GAL anchor (trained based video rewards), under the same prompt and renderer; four frames span each clip. Bottom: line plots of per-interval motion for the same clips, measured as the mean residual optical-flow magnitude inside each one-second interval ( Appendix G ).
Figure 2: Mismatch between image scores and video rewards. Pearson correlation between UniPercept scores of the anchor (IAA: aesthetics; IQA: quality; ISTA: structure and texture) and VideoReward scores of its HunyuanVideo-1.5 render (VQ: visual quality; MQ: motion quality; TA: text alignment) over 49,416 (prompt–anchor–video) records. Aesthetics and quality are nearly uncorrelated with every video axis, and structure/texture is anti-correlated with visual and motion quality.
Figure 3: Overview of GAL. Training (top): For each prompt, we sample K=4 anchor images from the T2I generator, render each into a video with a frozen I2V model, and score the videos with a frozen video-reward model. We use these scores to construct image-level preference pairs and fine-tune only the upstream T2I generator with cross-modal DPO. Video rendering, scoring, and pair construction are performed offline. Inference (bottom): The fine-tuned T2I generator produces anchor images that improve video quality, while the downstream I2V model remains unchanged.
Require: Prompt set P ; pretrained T2I policy πψ0 ; frozen reference πref ; frozen I2V renderer G ; frozen video reward Rv ; T2I noising-context distribution qc ; sibling count K ; margin m ; DPO scale β ; step size η
Phase 1: Offline preference construction
1:
Dpref←∅
2:
for p∈P do
3:
Sample K sibling anchor images x1,…,xK∼πψ0(⋅∣p)
via Eq. 10 with independent Brownian paths
4:
for k=1,…,K do
Algorithm 1 GAL: Golden Anchor Learning
Variant
Dynamic Degree
Imaging Quality
Aesthetic Quality
Subject Consistency
Motion Smoothness
Overall Consistency
VBench Average
HunyuanVideo-1.5
Base
0.8241
0.6586
0.6160
0.9577
0.9922
0.2247
0.7122
GAL (ours)
0.8704
0.7111
0.6426
0.9601
0.9909
0.2248
0.7333
(+5.62%)
(+7.97%)
(+4.32%)
(+0.25%)
(−0.13%)
(+0.07%)
(+2.97%)
LTX-Video (transfer)
Base
0.5509
0.6493
0.5909
0.9557
0.9922
0.2537
0.6654
Table 1: Main VBench comparison across four I2V models. GAL (ours), the GAL-VideoReward variant, improves the VBench average, dynamic degree, imaging quality, and aesthetic quality over Base on every model. Parentheses give percentage changes from Base; bold marks the best within each model. Open-source results average three I2V seeds; Kling v3 Pro uses one seed.
I2V model
Variant
VideoReward Overall
HPSv3 (video frames)
UniPercept (video frames)
HunyuanVideo-1.5
Base
−0.014
7.0025
51.4375
GAL (ours)
+0.131
7.6286
55.5926
(+0.145)
(+0.6261)
(+4.1550)
LTX-Video (transfer)
Base
−0.328
6.4788
50.0465
GAL (ours)
−0.224
7.3407
54.1946
(+0.103)
(+0.8619)
(+4.1481)
Table 2: Main reward-model comparison across four I2V models. GAL (ours) improves all three reward scores over Base on every model. VideoReward-Overall scores rendered clips; HPSv3 and UniPercept score their extracted frames. Parentheses give absolute changes from Base.
Variant
Dynamic Degree
Imaging Quality
Aesthetic Quality
Subject Consistency
Motion Smoothness
Overall Consistency
VBench Average
HunyuanVideo-1.5
Base
0.8241
0.6586
0.6160
0.9577
0.9922
0.2247
0.7122
GAL-Image
0.7917
0.7302
0.6453
0.9568
0.9902
0.2230
0.7229
(−3.93%)
(+10.87%)
(+4.76%)
(−0.09%)
(−0.20%)
(−0.76%)
(+1.50%)
GAL-HPSv3
0.8796
0.7035
0.6310
0.9529
0.9882
0.2248
0.7300
(+6.74%)
(+6.82%)
(+2.43%)
(−0.50%)
(−0.41%)
(+0.03%)
(+2.50%)
Table 3: Reward-signal ablation within GAL, evaluated with VBench across four I2V models. Parentheses give percentage changes from Base; bold marks the best within each model. GAL-VideoReward is the default variant.
Variant
VideoReward Overall
HPSv3 (video frames)
UniPercept (video frames)
Base
−0.014
7.0025
51.4375
GAL-Image
−0.531
7.9723
57.2618
(−0.517)
(+0.9698)
(+5.8243)
GAL-HPSv3
−0.493
7.9509
55.7763
(−0.478)
(+0.9484)
(+4.3388)
GAL-Unipercept
−0.171
7.9841
58.6379
Table 4: Reward-signal ablation within GAL, evaluated by video and frame rewards on HunyuanVideo-1.5. Among the compared variants, GAL-VideoReward alone improves all three scores over Base on HunyuanVideo-1.5 and all three transfer models (Table 8 in Appendix H ). We therefore use it as the default variant. Metrics and parentheses follow Table 2 .
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
m
Pairs
Note
0.00
74,124
raw, drop ties only
0.05
52,084
operating point
0.10
33,781
0.15
20,421
0.20
11,672
Appendix
Table 5: DPO pair count as a function of the absolute reward margin threshold m on the GAL-VideoReward pair manifest. All reward variants use the same operating point, m=0.05 , on their own reward-specific pair manifests (GAL-VideoReward: 52,084 pairs), each landing near a ≈5×104 -pair budget. A controlled pair-count sweep that holds data volume fixed across margins is listed as future work.
Split
# prompts
Source
Overlap with other splits
DPO train
12,354
Lexica corpus (cleaned, deduplicated)
0
VBench eval
93 / 72 per dim
Official VBench per-dimension sets
0
Kling eval
50
Held-out Lexica (disjoint from train)
0
Appendix
Table 6: Prompt sources and disjointness audit. All overlap counts are computed after prompt canonicalization. The training and Kling pools are a disjoint train / held-out split of the Lexica corpus; the VBench pool is the official suite.
I2V model
Dynamic fraction ↑
Within-prompt var. (ddof=0) ↓
Within-prompt var. (ddof=1) ↓
LTX-Video
0.5556
0.10204
0.11132
CogVideoX-1.5
0.6481
0.07735
0.08439
HunyuanVideo-1.5
0.7963
0.03318
0.03620
WAN
0.8322
0.02865
0.03125
Appendix
Table 7: Anchor-seed sensitivity of the binary VBench dynamic-degree readout. For each of the 72 official VBench dynamic-degree prompts, we sample 12 seed-conditioned sibling anchors from the same T2I generator and render them with each frozen I2V model under a fixed downstream setting. VBench returns a binary label d∈{0,1} for each rendered clip. We report the overall dynamic fraction and the prompt-averaged within-prompt variance across the 12 seeds. Low variance means that changing the seed-conditioned anchor rarely changes the binary dynamic/static decision for the same prompt. WAN has the highest dynamic fraction and the lowest within-prompt variance, indicating a near-saturated and low-sensitivity dynamic-degree readout.
Variant
VideoReward Overall
HPSv3 (video frames)
UniPercept (video frames)
LTX-Video (transfer)
Base
−0.328
6.4788
50.0465
GAL-Image
−0.763
7.7409
55.7652
(−0.435)
(+1.2621)
(+5.7187)
GAL-HPSv3
−0.672
7.7935
55.0136
(−0.344)
(+1.3147)
(+4.9671)
Appendix
Table 8: Reward-model scores on the three transfer I2V models across GAL variants. VideoReward-Overall scores rendered clips; HPSv3 and UniPercept score their extracted frames. Parentheses give the absolute change from Base; bold marks the best value within each model.
Video-to-video (V2V) generation is difficult to evaluate because outputs must both follow editing instructions and preserve frame-level correspondence with the source video, which existing T2V and I2V metrics do not capture. We introduce V2V-Bench, a 11-dimension benchmark organized into five categories: temporal alignment, structural fidelity, transformation quality, video quality, and semantic alignment. V2V-Bench pairs diverse source videos with challenging editing tasks and evaluates two commercial models, Grok Imagine and Gemini Veo3, and one open-source model, Open Sora 2. Results show complementary model strengths: Grok performs better on editing fidelity, while Veo3 achieves stronger visual quality. On six V2V-specific dimensions, V2V-Bench reaches a Spearman correlation of 0.905 with human judgments.
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.
Video reward models are trained and evaluated on human preferences between videos that often differ simultaneously on properties such as prompt alignment, visual fidelity, motion, and generator identity. As a result, aggregate accuracy does not reveal which video properties a reward model can reliably distinguish. We construct preference pairs by controlled interventions that change one property of a generated video while holding the rest fixed. The resulting pairs, by construction, provide preferences for how the videos differ along a specific dimension, without requiring any human annotation. We propose IsoVRBench, a video generation reward modeling benchmark of 8,000 such pairs across eight interventions, and show that state-of-the-art models separate broad quality differences between generators but miss important features such as temporal order. The best reward model reaches only 60.8% on reordered frames against 100% for humans, and rankings on existing benchmarks do not predict this failure. Using the synthetic isolated intervation pairs as training data leads to 1) IsoVRM, a reward model outperforming its counterpart trained with a larger human labeled preference dataset by 3.8 points on VBench-2.0, 2) IsoWan, the video generator post-trained with IsoVRM, outperforming its human-data counterpart in human preferences, and 3) gains that stack when combined with human data. Because intervention pairs are automatically generated rather than collected, the benchmark and training data can be easily extended as generators evolve.
Shuyue Stella Li, Jacqueline He, Jiahui Chen +5
University of Washington · 2Meta Superintelligence Labs