Game world models have made rapid progress toward real-time interactive generation. Yet action control is still predominantly implemented as feature conditioning, leaving the model to learn from training data how each action should move the visual world. This formulation is poorly matched to a broad class of planar-camera worlds, such as top-down and isometric games. In these worlds, player motion induces a geometrically well-defined planar camera displacement. We observe that the key to spatially consistent control in such worlds is to represent the effect of each action directly in world coordinates. To this end, we introduce ActionRoPE, a parameter-free action interface that converts action-driven camera displacement into world-coordinate rotary positions. As a result, tokens corresponding to the same world location retain the same positional address across time. These world coordinates further expose a spatial asymmetry in generation: regions covered by the initial observation are geometrically known, while newly revealed regions remain to be synthesized. We therefore propose Coordinate-Conditioned Noise Scheduling (CNS), which assigns token-wise denoising schedules according to this known-new structure. It allows observed regions to stabilize earlier, while new content is generated with reference to them. Compared with four representative action-conditioning mechanisms, our approach follows the commanded motion more faithfully and reduces the revisit drift by 66% over the strongest baseline. The same interface transfers across video backbones from 1.3B to 33B parameters and to out-of-distribution scenes. Project Page: https://teawhite.cn/ActionRoPE/.
Figures & tables
Figure 1: Action as world coordinates. (a) In a planar-camera world, each action corresponds to a known translation of the scene. (b) Standard RoPE re-addresses the same scene token after the camera moves, while ActionRoPE adds the camera displacement to each token’s position, keeping its index across frames. (c) ActionRoPE enables precise motion and revisits, control in unseen scenes, and variable camera speed, without adding parameters to the control path.
Figure 2: Overview of ActionRoPE. The left panel shows the video DiT, where ActionRoPE replaces the screen-relative RoPE before self-attention and the known/new mask κ is added to the patch embedding through a zero-initialized convolution. By binding camera motion to world coordinates, ActionRoPE keeps the same positional index for the same scene location across views, as illustrated in the upper-right panel. The same coordinates reveal which regions are known or newly revealed, and CNS assigns them different denoising schedules, as shown in the lower-right panel.
Figure 3: Qualitative comparison of revisit consistency. For each scene, the leftmost column shows the first frame. For each method, the top row shows the final frame after a full round trip, and the bottom row shows its absolute error against the first frame (blue indicates low and red indicates high error). Our method keeps the error low across the whole scene, while the four baselines show clearly larger errors that come from shifted or regenerated content.
Figure 4: Control on out-of-distribution first frames. Each row starts from a first frame in a visual style that is absent from the training game (watercolor and bamboo forest, respectively), and the columns show the generated frames at increasing frame index f .
Figure 5: Control on different backbones. The same first frame and command are given to ActionRoPE on SkyReels-V2, LTX-2.5 and MiniMax-H3. The camera moves right and turns back at f=40 , and the columns show frames at increasing time.
Motion fidelity
Direction
Revisit consistency
Method
Control Params
slope
R2
angle ↓
PSNR ↑
SSIM ↑
NCC ↑
DINO ↑
drift ↓
Wan + Prompt
0
0.911
− 8.46
2.19
20.98
0.613
0.684
0.852
26.6
Wan + Linear
+0.18 M
0.857
0.57
3.12
20.40
0.585
0.625
0.834
35.3
Wan + AdaLN
+66.5 M
0.784
− 0.52
1.92
21.11
0.604
0.685
0.862
22.6
Wan + Cross-Attn
+98.6 M
0.798
− 7.18
3.05
19.48
0.529
0.515
0.817
40.2
Ours
0
0.981
0.952
0.68
23.94
0.691
0.831
0.892
7.8
Table 1: Quantitative comparison results with baselines. On the same Wan backbone, our method achieves the best motion fidelity, directional compliance and revisit consistency. It adds no parameters to the control path and only 0.015 M parameters for the known/new mask channel.
Backbone
slope
R2
angle ↓
PSNR ↑
SSIM ↑
NCC ↑
DINO ↑
drift ↓
SkyReels-V2-DF 1.3B
1.047
0.919
1.46
22.40
0.648
0.757
0.858
15.3
Wan2.2-TI2V 5B
0.981
0.952
0.68
23.94
0.691
0.831
0.892
7.8
LTX-2.5 22B
1.014
0.845
0.59
23.55
0.717
0.810
0.800
12.8
MiniMax-H3 33B
1.082
0.861
1.90
21.97
0.647
0.726
0.831
22.0
Table 2: Quantitative results across backbones. With the same action interface, our method controls the camera on backbones ranging from 1.3B to 33B parameters. MiniMax-H3 generates 73 frames per clip, which are temporally resampled to 81 frames for evaluation.
Method
PSNR ↑
SSIM ↑
NCC ↑
DINO ↑
drift ↓
Wan + ActionRoPE
22.81
0.652
0.788
0.885
11.7
Wan + ActionRoPE + CNS
23.94
0.691
0.831
0.892
7.8
Table 3: Ablation study. We remove CNS from our full model on the Wan backbone, while keeping ActionRoPE in both compared models. Without CNS, all five revisit consistency metrics degrade, and the residual drift increases noticeably from 7.8 to 11.7 px.
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: More revisit comparisons (I). For each scene, we show the final frame of each method after the round trip and its error map against the first frame.
Figure 7: More revisit comparisons (II). Same layout as Figure 6 , on three other evaluation scenes.
Figure 8: Measured versus commanded displacement. Each point is one clip, the orange line connects the medians at each speed, and the dashed line marks exact tracking. The slope and R2 are those of Table 1 .
Figure 9: Revisit comparison on out-of-distribution scenes. The three first frames show an arctic tundra, a medieval castle and an underwater ruin. Same layout as Figure 6 .
Figure 10: More out-of-distribution styles. Our method on four first frames with styles that never appear in the training game, under a command that moves right and turns back at f=40 .
Figure 11: Cross-backbone example (I). The same first frame and command on the four backbones.
Figure 12: Cross-backbone example (II). Same setting as Figure 11 , on another evaluation scene.
Figure 13: Sampling with and without CNS. Clean-frame estimates of the last frame during sampling, under a straight walk to the right. The region to the left of the orange line is covered by the first frame.
Figure 14: Revisit with and without CNS. Final frames after the round trip and their error maps, for our full model and the model trained without CNS.
Motion fidelity
Direction
Revisit consistency
Method
slope
R2
angle ↓
PSNR ↑
SSIM ↑
NCC ↑
DINO ↑
drift ↓
Wan + Prompt
0.831
− 2.39
4.66
17.64
0.360
0.664
0.912
30.3
Wan + Linear
0.814
0.470
2.60
16.64
0.359
0.605
0.883
47.7
Wan + AdaLN
0.762
− 0.49
1.71
18.44
0.376
0.721
0.921
16.5
Wan + Cross-Attn
0.762
− 2.90
5.41
17.66
0.363
0.635
0.902
36.0
Ours
0.922
0.926
0.94
19.67
0.417
0.809
0.928
7.6
Appendix
Table 4: Comparison on out-of-distribution scenes. Same metrics as Table 1 , on 50 first frames whose scenes and styles do not appear in the training game.
Method
0.5×
0.75×
1.0×
1.25×
1.5×
Wan + Prompt
2.05
1.36
1.02
0.82
0.68
Wan + Linear
1.14
1.00
0.90
0.84
0.76
Wan + AdaLN
1.16
1.07
0.95
0.75
0.60
Wan + Cross-Attn
1.80
1.20
0.90
0.72
0.60
Ours
0.98
0.99
0.99
0.99
0.96
Appendix
Table 5: Measured-to-commanded displacement ratio per speed. Median over 100 first frames of the evaluation set. The ideal value is 1 , and the closest value in each column is in bold.
Figure 15: Noise level during sampling. A newly revealed token follows the original schedule of Wan2.2 ( S=5 ), while a known token ( κ=1 ) follows a lower curve under each CNS variant.
Schedule
PSNR ↑
drift ↓
angle ↓
Uniform (without CNS)
22.81
11.7
0.67
Power, α=1
23.65
9.1
0.71
Power, α=2
23.57
9.5
0.77
Shift, Sknown=1
23.20
10.7
0.87
Shift, Sknown=2 (ours)
23.94
7.8
0.68
Appendix
Table 6: Noise schedule variants on Wan2.2. Revisit consistency and direction error on the evaluation set.
Backbone
s
Mask
S
Sknown
Fine-tuning
LR
Batch
Steps / CFG
SkyReels-V2-DF 1.3B
16
7,680
5
2
full
10−5
16
50 / 1
Wan2.2-TI2V 5B
32
15,360
5
2
full
10−5
16
50 / 1
LTX-2.5 22B
32
8,192
7.77
3.1
LoRA, rank 128
10−4
16
30 / 3
MiniMax-H3 33B
32
10,752
12
4.8
LoRA, rank 128
10−4
8
50 / 1
Appendix
Table 7: Per-backbone settings. s : pixels per spatial token. Mask: parameters of the known/new mask channel. S , Sknown : CNS shifts. CFG: classifier-free guidance scale at sampling.