Camera-conditioned video diffusion models often suffer from inaccurate camera control and geometric inconsisten- cies because camera constraints and estimated geometry are only weakly enforced during denoising. Existing reme- dies often require costly geometry-aware fine-tuning or ad- ditional 3D supervision. We introduce GeoART, a training- free self-refinement framework that improves camera con- trollability through geometry-aware test-time scaling. At selected denoising steps, GeoART generates multiple can- didates within the Predict-and-Perturb cycle and eval- uates them using a geometry prior, without directly modify- ing the denoising trajectory. A confidence gate suppresses unreliable geometric evidence, while rank-calibrated prob- abilistic selection combines score magnitude and relative rank to choose the next refinement state. Experiments demonstrate consistent improvements in camera controlla- bility, multi-view consistency, and visual quality over exist- ing sampling and self-refinement methods, without training or modifying the pretrained video diffusion model
Figures & tables
Figure 2: GeoART framework. At selected denoising intervals, GeoART generates multiple self-refinement candidates, evaluates them using confidence-gated geometric scores, and probabilistically selects promising candidates. After ( Kt ) refinement iterations, the refined state is returned to the original denoising trajectory without model updates or geometry-based gradient guidance.
Figure 3: Comparison with baseline methods. Our method achieves higher visual quality, fewer artifacts, and more accurate camera control. Red lines mark the ground-truth positions of selected vertical structures (e.g., columns and poles) and remain fixed across methods to visualize positional drift.
Model
Views
Method
NFE ↓
PSNR ↑
SSIM ↑
LPIPS ↓
Trans. ↓
Rot. ↓
FID ↓
FVD ↓
BG Cons. ↑
Subj. Cons. ↑
Wan
One
FM
50
15.23
0.409
0.415
0.332
0.247
109.43
108.47
0.908
0.805
FM ( 2× NFE)
100
15.12
0.404
0.403
0.297
0.225
105.83
102.16
0.917
0.807
Self-Refine [ 27 ]
78
15.26
0.425
0.399
0.285
0.210
109.28
88.53
0.919
0.822
CFG-Ctrl [ 58 ]
50
15.19
0.406
0.408
0.335
0.266
127.89
140.85
0.911
0.803
Z-Sample [ 5 ]
130
14.96
0.392
0.425
0.351
0.263
110.86
98.87
0.914
0.821
Video-T1 [ 41 ]
500
15.40
0.411
0.405
0.310
0.225
105.65
104.33
0.918
0.826
Table 1: Quantitative comparison using Wan and Uni3C under one-view and two-view settings. Best, second-best, and third-best values within each model/view setting are highlighted in green , blue and yellow , respectively.
Figure 4: Hyperparameter analysis. Qualitative and quantitative effects of the candidate number N (a,c) and selection interval Δ (b,d). The red line in (a) marks the ground-truth object position and remains fixed across settings.
Configuration
PSNR ↑
SSIM ↑
LPIPS ↓
Trans. ↓
Rot. ↓
Component and selection ablations
Single candidate ( N=1 )
17.008
0.429
0.393
0.264
0.192
Uniform selection
17.424
0.435
0.396
0.236
0.173
Greedy selection
17.594
0.433
0.384
0.219
0.194
w/o rank calibration
17.602
0.461
0.357
0.228
0.172
w/o confidence gate
17.370
0.446
0.363
0.237
0.189
Table 2: Ablation study. Component, selection-strategy, and geometry-evaluator ablations using Wan on a fixed subset of 50 held-out scenes drawn from the evaluation set used for baseline comparison with resolution 256×256 .
Figure 5: Qualitative ablation study.
Figure 6: Effect of geometry evaluators. GeoART produces consistent results across different off-the-shelf geometry estimators, demonstrating robustness to the choice of geometry prior.
Hyperparameter
Value
Spatial resolution
256×256
Number of frames
33
Number of denoising steps
50
Denoising strength
1.0
Classifier-free guidance scale
3.0
Flow-scheduler shift
5.0
Table 3: Hyperparameters of GeoART.
τc
PSNR ↑
SSIM ↑
LPIPS ↓
Trans. ↓
Rot. ↓
FVD ↓
0.1
17.17
0.400
0.316
0.216
0.187
103.14
0.2
17.64
0.493
0.296
0.159
0.118
76.35
0.3
17.76
0.498
0.275
0.135
0.096
75.99
0.4
17.67
0.490
0.293
0.149
0.105
77.79
0.5
17.53
0.475
0.314
0.172
0.144
85.22
0.6
17.41
0.430
0.306
0.206
0.185
94.36
Table 4: Confidence-threshold sensitivity.
Refinement schedule R
PSNR ↑
SSIM ↑
LPIPS ↓
Trans. ↓
Rot. ↓
FVD ↓
{(5,10,2)}
17.26
0.481
0.298
0.201
0.143
94.73
{(5,15,2)}
17.64
0.496
0.281
0.162
0.115
79.74
{(5,20,2)}
17.70
0.496
0.283
0.175
0.124
80.87
{(5,30,2)}
17.64
0.494
0.286
0.171
0.122
81.12
{(40,45,1)}
17.48
0.485
0.289
0.168
0.119
92.84
{(5,15,2),(40,45,1)}
17.76
0.498
0.275
0.135
0.096
75.99
Table 5: Sensitivity to the stochastic-refinement schedule.
Method
PSNR ↑
SSIM ↑
LPIPS ↓
Trans. ↓
Rot. ↓
FM
16.69
0.386
0.434
0.289
0.243
Self-Refine [ 27 ]
17.24
0.394
0.426
0.256
0.204
CFG-Ctrl [ 58 ]
16.88
0.388
0.443
0.284
0.253
Video-T1 [ 41 ]
17.13
0.392
0.412
0.221
0.183
Ours
17.83
0.439
0.387
0.185
0.144
Table 6: Higher-resolution evaluation. Comparison at 512×512 resolution on held-out sequences.
Figure 7: Failure Case.
Figure 8: Comparison with baseline methods.
Figure 9: Comparison with baseline methods.
Figure 10: Qualitative comparison at 512×512 resolution.
Figure 11: Qualitative comparison at 512×512 resolution.
Camera-conditioned video generation uses camera poses and scene geometry to create temporally coherent videos from specified viewpoints and trajectories.
VCIP & TMCC & DISSec, College of Computer Science, Nankai University · Beijing University of Posts and Telecommunications · LIX, École Polytechnique, IP Paris