Iterative text-to-video generation relies on feedback to correct compositional and temporal failures. However, existing methods typically require another full generation run for each correction, repeating computation and potentially disrupting already correct content. We introduce PreCoG, an iterative framework for correcting semantic failures during video generation. PreQA diagnoses failures from low-cost previews and constructs corrective conditions, while PreEdit edits cached generation trajectories under these conditions, attenuating orthogonal changes to limit disruption of already correct content. PreCoG completes generation only for the selected candidate, reusing its cached trajectory to reduce redundant computation and help preserve semantic consistency between the preview and the final video. Experiments show that PreCoG achieves the highest VBench and T2V-CompBench scores and approaches VSR's leading VS2 score at 0.56× its inference time. It also outperforms GENMAC on all three quality and alignment metrics at 0.86× its inference time.
Figures & tables
Figure 1: Previews and full renders with a shared trajectory. Previews use matched prompts and seeds with few-step completion. Percentages denote sampling-step ratios.
Figure 2: Preview reliability. Left: preview/full decision agreement using Gemini 3.1 Pro ( Google DeepMind, 2026 ) in separate requests. Cells show percentages of all preview/full pairs. Right: agreement and disagreement versus the number of initial denoising steps, with two bridge and two clean steps. Disagreement is the fraction of pairs with different preview and full-render decisions.
Method
VBench (% ↑ )
T2V-Comp. (% ↑ )
VS2 ( ↑ )
Time (s ↓ )
Value
Δ
Value
Δ
Value
Δ
Value
Ratio
Direct generation
82.06
0.00
55.27
0.00
3.708
0.000
895.0
1.00 ×
Inference-time search
Full best-of- N ( N=4 )
81.29
-0.77
52.62
-2.65
3.751
+0.043
3580.0
4.00 ×
CachedSearch ( N=8 )
82.11
+0.05
55.02
-0.25
3.803
+0.095
4338.9
4.85 ×
Agentic multi-loop generation
Table 1: Quantitative comparison. We report VBench, T2V-CompBench (T2V-Comp.), VS2, and inference time. Δ : change from direct generation. Bold : best. Underlined : second best.
Figure 3: Qualitative comparison on compositional prompts. Generated videos depict temporal attribute changes (left) and object counting (right), with three uniformly sampled frames per video. Colored labels report VBench, T2V-CompBench (T2V-Comp.) for the corresponding category, and VS2 alignment scores for the complete videos.
Variant
VBench (% ↑ )
T2V-Comp. (% ↑ )
VS2 ( ↑ )
Time (s ↓ )
Value
Δ
Value
Δ
Value
Δ
Value
Ratio
Direct generation
82.06
0.00
55.27
0.00
3.708
0.000
895.0
1.00 ×
PreQA
Evaluation w/ full video
83.56
+1.50
60.66
+5.39
3.861
+0.153
2809.8
3.14 ×
Planning w/o planning
82.79
+0.73
56.50
+1.23
3.644
-0.064
1752.7
1.96 ×
Refinement w/o target refinement
82.48
+0.42
53.76
-1.51
3.486
-0.222
1986.4
2.22 ×
Table 2: Component ablations of PreCoG. Metrics follow Table 1 . VS2 uses soft alignment. Δ denotes change from direct generation. Bold : best. Underlined : second best.
Figure 4: Counterfactual projection ablation. Initial preview, w/o counterfactual projection, and w/ counterfactual projection, from left to right. Rows show the first, middle, and last frames.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Image-conditioned previews and full renders. Paired outputs share the reference image, prompt, and seed. Percentages denote sampling-step ratios, not runtime ratios.
Figure 6: Human preference. PreCoG win/tie/loss percentages for visual quality, prompt alignment, and physical naturalness.
Method
SC
BC
MS
AQ
IQ
DD
Avg.
Direct generation
92.18
94.02
98.21
57.99
68.72
81.25
82.06
Full best-of-N (N=4)
92.32
94.60
98.47
58.44
68.92
75.00
81.29
CachedSearch (N=8)
92.09
94.44
98.27
58.22
68.38
81.25
82.11
VQQA
93.08
94.61
98.45
58.34
67.96
68.75
80.20
VSR
92.69
94.26
98.20
57.27
65.13
75.00
80.43
GENMAC
93.92
94.59
98.51
60.25
71.15
78.12
82.76
Appendix
Table 3: VBench scores for each dimension. All methods use the same prompts as Table 1 . Scores are on a 0–100 scale. Avg. is the mean across six dimensions. Bold and underlined values indicate the best and second-best results, respectively.
Method
CAB
DAB
SR
MB
AB
OI
GN
Direct generation
97.14
3.20
54.23
37.95
69.63
83.33
41.41
Full best-of-N (N=4)
97.62
1.28
56.46
28.48
71.11
90.74
22.66
CachedSearch (N=8)
97.14
16.00
68.82
21.73
69.63
90.74
21.09
VQQA
86.19
1.28
68.17
22.71
65.93
83.33
35.94
VSR
95.24
25.60
61.35
35.85
85.19
99.07
30.47
GENMAC
85.24
1.15
67.29
28.82
63.70
100.00
24.22
Appendix
Table 4: T2V-CompBench category scores. Prompts match Table 1 . Category means use a 0–100 scale. Bold and underlined values mark the best and second-best results.
Figure 7: Quality and alignment versus inference cost. Points show the results in Table 1 , with inference time normalized to direct generation. Dashed lines connect Pareto-optimal methods among the compared results; higher scores and lower costs are preferred.
PreQA model
VBench (% ↑ )
T2V-Comp. (% ↑ )
VS2 ( ↑ )
Value
Δ
Value
Δ
Value
Δ
Direct generation
82.06
0.00
55.27
0.00
3.708
0.000
Qwen VL 3 4B
81.72
-0.34
51.86
-3.41
3.584
-0.124
Qwen VL 3 30B
83.31
+1.25
59.68
+4.41
3.827
+0.119
Qwen3.8-27B
83.49
+1.43
62.28
+7.01
3.854
+0.146
Gemini 3.1 Pro
83.46
+1.40
62.08
+6.81
3.901
+0.193
Appendix
Table 5: PreQA model ablation. Δ denotes change from direct generation (Table 1 ). Bold values indicate the best results.
Backbone
VBench (% ↑ )
T2V-Comp. (% ↑ )
VS2 ( ↑ )
Direct
PreCoG
Δ
Direct
PreCoG
Δ
Direct
PreCoG
Δ
T2V
Wan2.2-T2V-A14B
82.06
83.49
+1.43
55.27
62.28
+7.01
3.708
3.854
+0.146
Wan2.2-TI2V-5B
79.80
81.05
+1.25
48.60
54.20
+5.60
3.420
3.610
+0.190
SkyReels-V2-T2V-14B
81.20
82.35
+1.15
52.80
58.90
+6.10
3.580
3.750
+0.170
Helios-Base
80.70
81.65
+0.95
50.40
55.70
+5.30
3.490
3.650
+0.160
Appendix
Table 6: PreCoG across backbones and tasks. Δ denotes change from direct generation.