High-resolution generation enables video world models to render fine-grained scene structures and textures, but directly generating high-resolution video incurs substantial computational and memory costs. We present ScaleWorld, a causal video model that unifies low-resolution world prediction and high-resolution refinement within a single shared backbone. By refining its own predictions, ScaleWorld produces high-resolution world videos without requiring a separate refinement model. For each video chunk, the shared backbone first completes a low-resolution prediction and then performs one-step high-resolution refinement conditioned on that prediction. To support efficient streaming generation, we maintain separate causal states with a long low-resolution history for prediction and a short high-resolution history for refinement. To address task interference during distillation, we separate specialist optimization from capability consolidation and introduce task-asymmetric on-policy distillation (TA-OPD). This procedure transfers prediction and refinement behaviors from specialized teachers into a shared student through local velocity matching under their respective sampling and history policies. {Extensive experiments demonstrate that ScaleWorld achieves high-quality world generation with improved computational and memory efficiency.
Figures & tables
Figure 2: Overview of ScaleWorld . Top: Joint prediction and refinement training precedes causal adaptation, few-step specialist training, and TA-OPD. Center: One shared backbone completes each camera-controlled LR prediction and then refines it using the upsampled latent as a condition. WM maintains longer clean history; refinement retains shorter noisy-forward history. HR outputs do not enter the WM history. Bottom: During consolidation, frozen specialists supervise the shared student through local velocity matching, with matched current queries and network-specific KV histories.
Property
WM
Ref
Output resolution
low
high
Conditions
image, camera, text
LR latent, text
Sampling
multi-step
single-step
History span
long
short
History construction
clean recommit
noisy-forward
Table 1: Prediction and refinement share parameters while using distinct conditions, sampling procedures, and causal memories.
Figure 3: Native world prediction. Initial observations, camera trajectories, and generated frames from HY-World 1.5, LingBot-World 2.0, Matrix-Game 3.5, and our WM role.
Figure 4: High-resolution refinement. Ultra Flash, bicubic 2× , FlashVSR, and our model on two examples; enlarged regions expose differences in structure and texture.
Method
Average ↑
Video Quality ↑
Setting Adherence ↑
Interaction Adherence ↑
Consistency ↑
Physical ↑
(a) Complete HR generation
Separate cascade ( W→R )
81.58
83.72
78.21
86.80
87.04
72.13
Shared U→U (clean-recommit KV)
81.53
83.83
77.81
86.98
87.01
72.03
Shared U→U (noisy-forward KV)
81.60
83.43
78.73
87.08
86.92
71.82
(b) Native world prediction without refinement
Matrix-Game 3.5
74.760
78.46
72.10
83.12
79.63
60.51
Table 2: WBench generation quality and capability retention. (a) Complete HR generation with 480×832 prediction and 960×1664 refinement. (b) Native world prediction without refinement. W , R , and U denote the WM specialist, refinement specialist, and unified model. Bold marks the highest score per metric within each group.
World model
Refiner
Avg. ↑
Qual. ↑
Set. ↑
Inter. ↑
Cons. ↑
Phys. ↑
WorldPlay
Ultra Flash
78.630
79.69
72.27
85.62
87.43
68.13
WorldPlay
FlashVSR
79.160
80.08
73.09
86.52
87.09
69.01
LingBot-World 2.0
Ultra Flash
81.070
81.94
79.29
84.73
84.36
75.03
LingBot-World 2.0
FlashVSR
81.030
81.98
79.86
84.93
84.44
73.96
Matrix-Game 3.5
Ultra Flash
73.990
78.56
70.81
83.43
79.48
57.69
Matrix-Game 3.5
FlashVSR
75.060
78.72
72.53
83.14
79.51
61.41
Table 3: WBench scores on final HR videos. Each external WM supplies the same frozen predictions to its two refiners. Bold marks the highest score per column.
Refiner
MUSIQ ↑
CLIP- IQA+ ↑
LR PSNR ↑
MC L1 ↓
MC HF L1 ↓
FlashVSR center-crop protocol: 896×1664 output
FlashVSR Full
61.9861
0.5167
23.4840
0.0327
0.0140
Full frame: 480×832→960×1664
Ultra Flash v1.1
53.7366
0.4417
22.2844
0.0330
0.0090
Ours (clean)
56.7657
0.5156
25.3340
0.0208
0.0081
Ours (noisy)
58.9656
0.5185
26.2409
0.0212
0.0086
Table 4: Refinement of identical frozen LR videos from our WM. Compare rows only within each spatial protocol. Bold marks the best full-frame scores. PSNR measures LR consistency, not HR reconstruction; MC residuals use grayscale in [0,1] . Clean/noisy denote clean-recommit/noisy-forward KV histories.
Metric
Mixed DMD
CD + mixed DMD
TA-OPD
Avg. ↑
72.000
75.878
80.308
Qual. ↑
79.82
75.72
83.04
Set. ↑
63.94
64.53
77.28
Inter. ↑
76.79
87.98
86.05
Cons. ↑
79.96
82.44
85.64
Phys. ↑
59.49
68.72
69.53
Table 5: Ablation of distillation strategies for native WM prediction. All variants use identical initialization.
Figure 5: Qualitative self-cascade results. Paired LR/HR outputs with camera controls; orange insets show matched regions.
Training step
Global cosine
Negative blocks
400
+0.0969
2/30
600
−0.1748
29/30
700
−0.1775
25/30
800
−0.3421
30/30
Table 6: WM/Ref DMD gradient alignment during mixed training. Negative blocks have cosine <0 ; probe settings: Appendix C.1 .
Metric
Clean- recommit
Noisy-forward (reference)
HR forwards
2
1
HR Q ↑
56.7657
58.9656
PSNR ↑
25.3340
26.2409
TE ↓
0.0208
0.0212
Generator time (s) ↓
0.60
0.30
Generator FPS ↑
26.7
53.3
Table 7: Controlled ablation of refinement-history construction on fixed LR inputs. HR forwards include history construction; generator time is measured per 16-frame chunk before VAE decoding; generator FPS is 16 divided by this time. PSNR measures LR consistency; HR Q is MUSIQ and TE is MC L1. All other training and evaluation settings are identical.
Metric
W→R Resident
U→U Two copies
U→U Shared copy
Copies
2
2
1
Params (B)
10.28
10.28
5.14
Weight GiB
19.16
19.16
9.58
KV GiB
1.77
1.77
1.77
Peak GiB ↓
45.42
45.42
35.84
FPS ↑
10.77
11.94
11.94
Table 8: Measured deployment costs on four GPUs (BF16, batch size one). Weight/KV are per-GPU allocations in GiB; Peak is the maximum allocated memory across GPUs. FPS covers end-to-end generation, including decoding.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Role
Pixels
VAE latent
Patch grid
WM
480×832
30×52
15×26
Ref
960×1664
60×104
30×52
Appendix
Table 9: Reference spatial representations.
Property
WM
Ref
Condition
image, camera, text
LR latent, text
Training layout
[1,4,4,…]
[4,4,…]
Nominal indices
[1000,750,500,250]
[1000]
Temporal span
25 latent frames
8 latent frames
Rollout
25 latent frames
8–16 latent frames
Supervised blocks
6 content blocks
final block
Appendix
Table 10: Reference consolidation settings. Sampling and cache settings are listed separately for each role.
Figure 6: Additional qualitative self-cascade results. Each example shows an initial observation followed by successive camera-controlled rollout frames. The LR rows directly decode WM-predicted latents with the VAE; the HR rows show refinement outputs conditioned on those latents. Key labels indicate camera controls, and orange boxes and insets highlight matched regions.
WM teacher
Ref teacher
WM Inter. ↑
WM Cons. ↑
HR Q ↑
TE ↓
TWM
TRef
86.05
85.64
58.9656
0.0212
Appendix
Table 11: Reference student consolidated from two specialist teachers. WM Inter./Cons. are native WBench Interaction/Consistency; HR Q and TE are fixed-input MUSIQ and MC L1 for noisy-forward refinement.
System
Avg. ↑
Qual. ↑
Set. ↑
Inter. ↑
Cons. ↑
Phys. ↑
MUSIQ ↑
FPS ↑
Cascade ( W→R )
81.58
83.72
78.21
86.80
87.04
72.13
59.2371
10.77
U→U (clean-recommit KV)
81.53
83.83
77.81
86.98
87.01
72.03
56.7657
9.41
U→U (noisy-forward KV)
81.60
83.43
78.73
87.08
86.92
71.82
58.9656
11.94
Appendix
Table 12: Separate specialist cascade versus unified generation at 480×832 LR and 960×1664 HR. WBench Avg. is the mean of its five official dimensions. FPS covers end-to-end generation, including WM prediction, refinement, history construction, and decoding. WBench scores are measured for all three systems. MUSIQ reports the corresponding refiners on the fixed decoded LR inputs of Table 4 , rather than native latent-handoff outputs. Bold marks the highest reported value in each column.