Modern text-to-video (T2V) models generate visually compelling videos, yet meaning-preserving changes in prompt wording can substantially alter semantic realization and overall quality. Existing prompt rewriting modifies the inference input, whereas conventional post-training constructs supervision from one prompt formulation at a time; neither explicitly optimizes reliability across equivalent forms. We introduce VARI (Verified Alignment over Representations of Intent), a group-structured framework that treats automatically verified formulations of the same visual intent as a shared supervision unit. VARI builds matched form-seed rollout pools, cross-scores each video against all forms, and applies a wording-independent structured intent audit to identify valid realizations and hard semantic failures. Group-Consistent Distillation (GCD) aligns every verified form with interchangeable valid realizations, while Form-Contrastive Flow-DPO (FC-DPO) learns group-level preferences from audited positive-negative pairs, prioritizing same-seed, different-form contrasts. On HunyuanVideo-1.5, VARI attains the highest T2V-CompBench, VBench, and StoryEval scores among the evaluated methods. On prompt-form extensions of T2V-CompBench and VBench, it improves weakest-form alignment over the strongest prompt-level baselines by 2.45 and 3.55 points and reduces cross-form drop by 30.5% and 32.2%, respectively, while preserving visual quality. Human evaluation, natural-caption controls, prompt-enhancement experiments, diversity analyses, and transfer to LTX-2.3 further support the robustness and transferability of these gains, while VARI requires no additional inference-time module.
Figures & tables
Figure 1: Overview of VARI. (A) Verified forms and shared seeds define a rollout grid. Cross-scoring provides mean support μ , weakest-view support w , dispersion d , and quality Q ; a structured audit against Tz assigns semantic type, and GroupScore ranks candidates within type. (B) GCD distills retained positives across all forms, while FC-DPO lifts audited intent-level preferences to the full group, prioritizing same-seed, different-form pairs.
Method
VBench Total
VBench Qual.
VBench Sem.
T2V-C Avg.
StoryEval
Base
80.11
82.83
69.23
47.69
32.11
VideoDPO
81.06
83.68
70.59
49.60
34.62
OnlineVPO
81.60
84.19
71.24
50.49
35.31
GCD
81.15
83.46
71.91
50.48
35.73
VARI
82.67
84.91
73.71
52.13
37.58
Table 1: Standard benchmark results. T2V-C averages seven official categories; StoryEval reports strict event completion.
Method
Cons. Attr.
Dyn. Attr.
Spatial Rel.
Motion Bind.
Action Bind.
Interaction
Numeracy
Avg.
Base
79.82
7.47
53.11
40.29
69.31
60.57
23.24
47.69
VideoDPO
81.34
8.25
55.56
41.89
71.24
63.49
25.42
49.60
OnlineVPO
82.31
10.04
56.72
43.91
72.13
64.49
23.81
50.49
GCD
83.26
8.66
56.97
39.88
72.71
66.61
25.29
50.48
VARI
86.64
9.58
58.55
40.42
75.06
67.78
26.87
52.13
Table 2: T2V-CompBench categories. VARI gives the best average; OnlineVPO leads dynamic attribute and motion binding.
Method
T2V-CompBench-R
VBench-R Semantic
VBench-R Quality
Mean ↑
Worst ↑
Drop ↓
FIFR ↓
Mean ↑
Worst ↑
Drop ↓
FIFR ↓
Mean ↑
Base
47.18
44.47
4.92
11.14
68.29
61.27
11.74
11.38
82.78
VideoDPO
49.54
47.08
3.97
8.93
70.16
65.18
9.19
8.72
83.51
OnlineVPO
50.09
46.95
4.12
8.21
70.79
65.07
8.61
8.26
83.94
GCD
50.39
48.81
2.52
5.50
71.36
67.12
6.51
6.08
83.47
VARI
51.25
49.53
2.76
4.43
72.68
68.73
5.84
5.27
84.24
Table 3: Prompt-form group extensions of the official evaluators; VBench quality is reported separately.
Figure 2: Representative mining-view score matrices for one intent group. Rows are generated candidates and columns are verified scoring forms. Cell variation is diagnostic evidence, not form-dependent ground truth for a fixed video.
Figure 3: Group-level VBench-R semantic distributions. VARI raises worst-form scores and reduces cross-form drops.
Figure 4: Source/equivalent-form cases. Baselines violate spatial relation or action binding; VARI preserves both across forms. The panels show temporally ordered frame strips from the evaluated video clips.
Appendix figures & tables29 assets
Supplementary material from the paper’s appendix.
Appendix
Family
Allowed surface variation
Lexical
Synonyms and local wording changes without semantic strengthening
Syntactic
Clause reordering, attachment changes, or role-preserving voice conversion
Caption-like
Natural descriptive-caption syntax with the same commitments
Concise
Removal of linguistic redundancy while retaining every visual constraint
Relation-focused
Reordering that foregrounds an existing spatial or interaction relation
Action-focused
Reordering that foregrounds an existing action or event sequence
Appendix
Table 6: Prompt-form transformation families. A candidate is rejected whenever its visual content changes, regardless of its nominal family.
Compiler stage
Count
Rate (%)
Generated candidates
53,487
100.00
Malformed / near-duplicate removed
2,713
5.07
Field-wise verifier inputs
50,774
94.93
Verifier-accepted candidates
34,318
64.16
Retained rewrites (four per source)
20,000
37.39
Complete five-form groups
5,000
100.00 intents
Appendix
Table 7: Prompt-form generation and verification funnel. Candidate-stage rates use 53,487 as the denominator; complete-group coverage is reported over 5,000 source intents.
Critical field
Equivalent / total
Precision (%)
Entity / count
59/60
98.33
Spatial relation
57/60
95.00
Action–role binding
58/60
96.67
Temporal order
56/60
93.33
Attribute / state / other
58/60
96.67
Overall
288/300
96.00
Appendix
Table 8: Human semantic-equivalence audit of accepted prompt forms.
Audit subset
Equivalent / total
Precision (%)
Training accepted pairs
190/200
95.00
Benchmark accepted pairs
98/100
98.00
All accepted pairs
288/300
96.00
Complete benchmark groups (all four rewrites valid)
97/100
97.00
Appendix
Table 9: Source-specific pair audit and complete-group benchmark audit. Group success requires all four rewrites to preserve the shared target.
Compiler outcome
Count
Rate (%)
At least one audited positive
4,371
87.42
At least one audited negative
3,228
64.56
At least one preference pair
3,069
61.38
GCD-only intent
1,302
26.04
No retained supervision
629
12.58
Matched-seed retained pairs
4,236
71.71
Appendix
Table 10: Supervision yield. Intent rates use 5,000 source intents; pair-provenance rates use 5,907 retained preference pairs.
Commitments
N
Positive
Negative
Pair
Jz=3 – 4
1,617
1,492 (92.3)
913 (56.5)
876 (54.2)
Jz=5 – 6
2,356
2,068 (87.8)
1,539 (65.3)
1,474 (62.6)
Jz≥7
1,027
811 (79.0)
776 (75.6)
719 (70.0)
All intents
5,000
4,371 (87.4)
3,228 (64.6)
3,069 (61.4)
Appendix
Table 11: Supervision yield by the number of atomic commitments. Cells report count (rate %). Min/max audit aggregation is deliberately more conservative for longer structured targets.
Item
Data and generation configuration
Main / transfer backbone
HunyuanVideo-1.5 / LTX-2.3 (22B dev)
Trainable component
Video DiT only; text encoders and VAE frozen
Training intents
5,000
Prompt-form groups
One source form and four verified semantic-preserving rewrites per intent
Generation seeds
Four intent-specific seeds reused across all verified forms
Table 12: Core implementation configuration. Compared methods use matched generation settings and total sampling budgets whenever permitted by the method definition.
Method
Rollouts
Scoring / supervision
Optimization pattern
Inference overhead
VideoDPO
100k
Single-form reward and pair construction
Fixed-pool preference training
None
OnlineVPO
100k
Online reward and pair refresh
Interleaved generation and updates
None
VARI
100k
Cross-form scores; shortlist-only audit
Parallel offline GCD + FC-DPO
None
Appendix
Table 13: Per-run protocol comparison under a matched video-rollout budget. API calls for the shortlist audit are logged separately by the compiler.
T2V-C category
Consist.
Dynamic
Spatial
Motion
Action
Interact.
Numeracy
τb (%)
78.07
7.18
50.22
36.31
65.09
56.37
20.19
Appendix
Table 14: Validation-fixed success thresholds used by T2V-CompBench-R FIFR. VBench-R analogously uses the base-model validation median of each matched semantic dimension.
Method
T2V-C Avg.
T2V-C-R Worst
VBench-R Worst
OnlineVPO
50.49±0.23
46.95±0.27
65.07±0.29
GCD
50.48±0.18
48.81±0.21
67.12±0.25
VARI
52.13±0.19
49.53±0.22
68.73±0.26
Appendix
Table 15: Mean ± standard deviation over three independent training runs.
Attribution-critical variant
Mean
Worst
Drop
FIFR
GCD
50.46±0.18
48.88±0.21
2.57±0.13
5.41±0.24
Intent-audit-only FC-DPO
50.94±0.15
49.09±0.19
2.61±0.12
5.02±0.22
VARI w/o matched-seed pairs
51.06±0.19
49.28±0.18
2.46±0.16
4.87±0.20
VARI (matched-seed prioritized)
51.17±0.13
49.51±0.16
2.29±0.11
4.44±0.18
Appendix
Table 16: Mean ± standard deviation over three paired training branches on the fixed development split.
Mined type
Human precision
Visual usability
Indep. agreement
Group-consistent positive
96/100
91/100
88/100
Low-score negative
76/100
47/100
68/100
Dispersion-only negative
58/100
81/100
50/100
Intent-audit negative
85/100
76/100
78/100
Audited FC negative
91/100
79/100
84/100
Appendix
Table 17: Blind supervision audit. Bold marks the best result among negative-mining variants; each cell reports valid counts out of N=100 candidates per type.
Evaluation source
Retrieved candidates
Confirmed shared intents
VBench
13
0
T2V-CompBench
23
0
StoryEval
8
0
VATEX
17
0
MSR-VTT
26
0
Appendix
Table 18: Training–evaluation overlap audit. Exact normalized overlap is zero for every source.
Method
Appearance div.
Motion div.
CF margin
Base
100.00
100.00
11.36
OnlineVPO
98.20
97.40
12.01
GCD
95.70
95.30
12.71
VARI
96.50
95.80
13.90
Appendix
Table 19: Diversity retention relative to Base (%) and counterfactual semantic margin.
Change type
Base
OnlineVPO
GCD
VARI
Count
10.77
11.41
11.68
13.12
Spatial relation
13.26
14.01
14.82
15.97
Action-role swap
11.88
12.59
13.37
14.53
Temporal order
9.52
10.04
10.96
11.98
Appendix
Table 20: Counterfactual margin by minimally changed semantic commitment.
Category
Base
VideoDPO
OnlineVPO
GCD
VARI
Consistency
79.82
81.34
82.31
83.26
86.64
Dynamic
7.47
8.25
10.04
8.66
9.58
Spatial
53.11
55.56
56.72
56.97
58.55
Motion
40.29
41.89
43.91
39.88
40.42
Action
69.31
71.24
72.13
72.71
75.06
Interaction
60.57
63.49
64.49
66.61
67.78
Appendix
Table 21: T2V-CompBench category scores under the unified evaluation protocol. Values are percentages; the table is transposed to preserve a stable single-column layout.
Variant
Mean
Worst
Drop
FIFR
Reward-SFT
48.58
43.82
5.97
13.11
Prompt-DPO
49.91
46.77
4.49
9.24
FC-DPO
50.31
47.68
4.01
7.76
GCD
50.46
48.88
2.57
5.41
GCD + prompt-DPO
50.82
48.64
3.28
6.19
VARI
51.17
49.51
2.29
4.44
Appendix
Table 22: Training-objective ablation on T2V-CompBench-R.
Variant
Mean
Worst
Drop
FIFR
Reward-SFT
48.58
43.82
5.97
13.11
Prompt-level DPO
49.91
46.77
4.49
9.24
Multi-form prompt-DPO
50.08
47.03
4.34
8.61
Direct worst-form DPO
50.39
48.17
3.12
6.38
Form-specific distillation
49.93
47.51
3.69
7.19
Single-positive GCD
50.18
48.56
2.89
5.93
Appendix
Table 23: Full causal ablation on the fixed T2V-CompBench-R development split.
Mining score
μ
w
−d
Worst ↑
Drop ↓
FIFR ↓
Single-form
✗
✗
✗
41.08
16.89
14.94
Mean only
✓
✗
✗
44.17
13.43
11.18
Mean + worst
✓
✓
✗
46.12
10.21
8.24
Mean + worst - drop
✓
✓
✓
46.95
7.48
5.96
Full + Q
✓
✓
✓
46.87
7.64
6.08
Appendix
Table 24: GroupScore-component ablation with T2V-CompBench-R evaluation.
Table 30: Verified VATEX multi-caption diagnostic scored with held-out VisionReward.
Method
Mean
Worst
Drop
FIFR
Base
61.63
57.17
6.67
10.22
VideoDPO
63.61
59.33
6.07
8.76
OnlineVPO
64.05
59.24
6.37
9.00
GCD
65.21
61.69
4.96
6.33
VARI
65.88
62.34
4.69
5.84
Appendix
Table 31: Disjoint MSR-VTT control under the same held-out reward protocol.
Dimension
Counts
Raw shares (%)
W–L
95% CI
α
W
T
L
W
T
L
(pt.)
Intent faithfulness
352
96
152
58.67
16.00
25.33
33.33
[25.9, 40.6]
0.59
Cross-form consistency
378
83
139
63.00
13.83
23.17
39.83
[32.4, 47.1]
0.63
Visual quality
319
101
180
53.17
16.83
30.00
23.17
[14.9, 31.3]
0.57
Overall preference
347
106
147
57.83
17.67
24.50
33.33
[26.3, 40.4]
0.60
Appendix
Table 32: Detailed blinded human evaluation: VARI versus OnlineVPO. Each dimension contains 600 judgments.
Figure 5: Four additional held-out cases with ordered frames. (a) VARI preserves five salient robots and the group dance across both forms. (b) It maintains one coherent game layout and stable participants. (c) It moves the car past a stable church. (d) It realizes progressive snowman collapse and volume loss.
Comparison
Forms
Material shown
Main-paper spatial relation
2×3
Displayed strips (6 evaluated clips)
Main-paper action binding
2×3
Displayed strips (6 evaluated clips)
Appendix paired-form cases
2×2
Static temporally ordered panels
Appendix dynamics cases
2×1
Static temporally ordered panels
Appendix
Table 33: Qualitative evidence shown in this preprint. Each count is prompt forms × compared methods or intents, as applicable.
Text-to-video (T2V) models are highly sensitive to prompt specificity, while iterative prompt optimization typically improves generation quality only through repeated video generation and vision-language model evaluation at inference time. We propose TIDE, a framework that distills this costly search process into an interpretable and reusable prompt-rewriting Skill. TIDE formulates prompt rewriting as a set of source-conditioned rule modules, each targeting a distinct semantic dimension. Once distilled, the frozen Skill activates the relevant rules and generates both positive and negative prompts in a single LLM call, followed by only one T2V generation, without test-time visual search or backbone fine-tuning. Evolved using only 100 prompts over 14 hours, TIDE achieves prompt-optimization performance comparable to computationally expensive search-based methods. Moreover, when transferred to previously unseen video-generation models, TIDE outperforms existing approaches specifically designed or trained for those models. Code is available at https://github.com/xiningin/TIDE_T2V.
Chengxing Zhou, Yuliang Huang, Jingwei Song
Sun Yat-sen University, China · The University of Hong Kong, China
Reward models for text-to-video (T2V) generation guide post-training but often fail at fine-grained semantic alignment. We trace this to two structural weaknesses in existing reasoning-based reward models: they do not systematically verify every condition described in the prompt, and the visual evidence supporting each judgment remains implicit in their free-form reasoning. We propose SG-PVR, a video reward model that addresses these limitations through plan-and-verify reasoning grounded in spatio-temporal scene graphs. The verification plan decomposes the prompt into atomic claims, making the set of requirements to be checked explicit. The spatio-temporal scene graph, encoding entities, attributes, and temporally grounded relations, is extracted from the video and maintained as a persistent structured visual reference throughout reasoning. Each claim is verified against both the video and the scene graph, anchoring judgments in explicit visual evidence. SG-PVR achieves strong performance on semantic alignment, including fine-grained temporal semantics. As a test-time reranker, it further enhances compositional alignment in T2V generation.
Hyomin Kim, Junghye Kim, Joanie Hayoun Chung +4
Department of Artificial Intelligence, Korea University · Department of Statistics, Korea University
While large-scale datasets have driven significant progress in Text-to-Video (T2V) generative models, these models remain highly sensitive to input prompts, demonstrating that prompt design is critical to generation quality. Current methods for improving video output often fall short: they either depend on complex, post-editing models, risking the introduction of artifacts, or require expensive fine-tuning of the core generator, which severely limits both scalability and accessibility. In this work, we introduce 3R, a novel RAG based prompt optimization framework. 3R utilizes the power of current state-of-the-art T2V diffusion model and vision language model. It can be used with any T2V model without any kind of model training. The framework leverages three key strategies: RAG-based modifiers extraction for enriched contextual grounding, diffusion-based Preference Optimization for aligning outputs with human preferences, and temporal frame interpolation for producing temporally consistent visual contents. Together, these components enable more accurate, efficient, and contextually aligned text-to-video generation. Experimental results demonstrate the efficacy of 3R in enhancing the static fidelity and dynamic coherence of generated videos, underscoring the importance of optimizing user prompts.