Few-step autoregressive video generation enables low-latency streaming, yet prevailing training pipelines separate causal adaptation from distribution matching. Recent representation-based text-to-video methods retain this separation by starting with an ODE initialization stage, even when the target distribution is defined directly by reference videos. We introduce \method, the first single-stage framework that uses reference videos both as paired examples of how a history continues and as a target population for self-generated rollouts. Specifically, Direct Forcing is trained with two complementary signals: (1) Causal Continuation Alignment learns paired spatiotemporal correspondence under reference histories; and (2) Rollout Distribution Matching employs maximum mean discrepancy to align the representation distributions of self-generated and reference videos. To coordinate these signals, we introduce Dynamic Gradient Balancing, which adapts their relative gradient contributions during training. Experiments with Wan2.1-1.3B show that Direct Forcing learns a four-step autoregressive generator in 250 updates without a separate ODE initialization stage or online diffusion-score supervision, achieving 84.63 VBench Total and 86.01 Quality.
Figures & tables
Figure 1: Single-stage training for few-step AR generation. Direct Forcing jointly learns causal alignment and distribution matching, achieving 84.63 VBench Total in 250 updates. Baseline counts include inherited initialization (Table 1 ).
Figure 2: Temporal discontinuity under MMD-only training. Consecutive frames 8–10 show an abrupt change at the first chunk boundary.
Method
Training path
AR teacher training
ODE init.
Distribution matching
Total
Self Forcing [ 16 , 41 ]
ODE init. (BID teacher) → DMD
–
3,000
750
3,750
Causal Forcing [ 41 ]
ODE init. (AR teacher) → DMD
2,000
1,000
750
3,750
ViRDM [ 24 ]
ODE init. (AR teacher) → MMD
2,000
1,000
20
3,020
Elastic Forcing [ 38 ]
ODE init. (BID teacher) → MMD
–
3,000
150
3,150
Direct Forcing (ours)
Joint CCA + RDM
–
–
250
250
Table 1: Training paths from bidirectional Wan2.1-1.3B to four-step AR generation. BID denotes bidirectional. Numeric entries report optimization updates for causal-teacher training, ODE initialization, and distribution matching; Total includes inherited stages. Our 250 updates jointly optimize Causal Continuation Alignment (CCA) and Rollout Distribution Matching (RDM). Backbone pretraining and offline data/feature preparation are excluded.
Figure 3: Direct Forcing overview. A shared few-step causal generator learns from reference histories with Causal Continuation Alignment (CCA; Sec. 3.2 ) and self-generated histories with Rollout Distribution Matching (RDM; Sec. 3.3 ). Frozen encoders provide paired cosine alignment and video–text MMD, respectively. Dynamic Gradient Balancing (Sec. 3.4 ) first matches gradient norms, then weights the CCA gradient by its current fitting residual before combining both gradients to update the shared generator.
Model
# Params
Resolution
Throughput ↑
Latency (s) ↓
Total ↑
Quality ↑
Semantic ↑
Bidirectional Video Diffusion Models
LTX-Video-1.9B [ 14 ]
1.9B
768×512
8.98
13.5
79.83
81.88
71.62
Wan2.1-1.3B [ 30 ]
1.3B
832×480
0.78
103
83.37
84.30
79.65
Autoregressive Video Diffusion Models
NOVA [ 7 ]
0.6B
768×480
0.88
4.1
80.31
80.66
78.92
Pyramid Flow [ 19 ]
2.0B
640×384
6.7
2.5
80.75
83.41
70.11
Table 2: Text-to-video generation on VBench. Models are grouped by generation scheme. Scores are percentages. The highlighted row is our 250-update model. ∗ ViRDM denotes its reported variant without optical-flow regularization. Our main model also excludes this auxiliary objective; Appendix C.4 presents our optical-flow regularization ablation.
Figure 4: Effect of training duration. VBench scores of our main configuration (RDM batch size 128) from 50 to 250 updates. Each marker denotes an evaluated checkpoint.
Configuration
Subj. ↑
Smooth. ↑
Aes. ↑
Dyn. ↑
Sem. ↑
CCA only
97.08
99.50
59.13
13.89
72.21
RDM only
84.17
96.82
63.95
11.11
71.93
CCA init → RDM-only
94.08
98.56
64.35
74.44
77.88
CCA + RDM jointly
95.69
98.77
65.07
54.72
77.99
Table 3: CCA and RDM training. Subj./Smooth./Aes./Dyn.: Subject Consistency/Motion Smoothness/Aesthetic Quality/Dynamic Degree (VBench quality); Sem.: aggregate Semantic. Sequential uses 150+150 updates; others use 150.
Figure 5: Gradient scales and fitting progress. (a) Fixed-weight CCA/RDM gradient norm ratio. (b) GNM-only losses relative to update 1; RDM is measured before self-normalization.
Batch size
32
64
128
256
512
1024
Total ↑
81.27
82.40
82.36
82.96
82.89
82.81
Quality ↑
82.10
83.51
83.46
83.95
84.06
83.89
Semantic ↑
77.95
77.99
77.99
78.97
78.21
78.47
Dynamic ↑
33.33
54.72
47.22
47.78
54.17
51.67
Table 4: RDM batch size with VideoMAE v2 L16 for both components. CCA batch size is 32; all runs use 150 updates.
Representation
Total ↑
Quality ↑
Semantic ↑
Dynamic ↑
Latent FM
81.76
82.77
77.75
51.39
V-JEPA 2.1 L16
81.77
83.12
76.35
40.56
V-JEPA 2.1 L24
82.05
83.30
77.08
58.06
DiT L9
82.56
83.53
78.68
58.89
VideoMAE v2 L16
82.40
83.51
77.99
54.72
VideoMAE v2 L24
82.80
84.00
78.01
58.06
Table 6: Representation for causal continuation alignment, with rollout distribution matching fixed to VideoMAE v2 L16. Latent FM denotes velocity regression in latent space.
Continuation
Distribution
Total ↑
Qual. ↑
Sem. ↑
Dyn. ↑
V24
V16
82.80
84.00
78.01
58.06
V24
D9
83.49
84.96
77.58
68.61
V24
V16+D9
84.11
85.47
78.67
74.72
V24+D9
V16+D9
83.68
84.99
78.43
71.94
V24+D9
V16+D9+V24
83.46
84.71
78.45
75.00
Table 8: Representation combinations after 150 updates. Vk : VideoMAE v2 L k ; D9 : DiT L9. Gray marks the selected combination.
Figure 7: Qualitative comparison on four text-to-video prompts. For each case, rows show ViRDM, Elastic Forcing, and Direct Forcing from top to bottom; columns show frames 0, 40, and 80. These lengthy prompts are abbreviated above the panels for readability.
Metric
Self-Forcing++
Direct Forcing
Subject consistency
97.00
97.07
Background consistency
95.55
96.11
Motion smoothness
98.39
98.61
Dynamic degree
55.36
53.13
Aesthetic quality
53.76
56.14
Imaging quality
67.87
67.15
Table 9: Long-video generation on MovieGen-128 at 50 seconds. Self-Forcing++ scores follow its reported results [ 6 ] . All scores are percentages; higher is better.
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
Shared space
Total ↑
Qual. ↑
Sem. ↑
Dyn. ↑
VideoMAE v2 L16
82.40
83.51
77.99
54.72
VideoMAE v2 L24
82.69
84.06
77.24
55.83
V-JEPA 2.1 L16
79.69
81.06
74.21
25.28
V-JEPA 2.1 L24
80.51
81.79
75.39
25.56
DiT L9
82.32
83.76
76.55
66.94
Appendix
Table 10: Shared representations for continuation alignment and rollout distribution matching. VideoMAE v2 L24 yields the highest Total and Quality, while DiT L9 gives the highest Dynamic.
Training decoder
Total ↑
Qual. ↑
Sem. ↑
Dyn. ↑
TAEW2.1
82.40
83.51
77.99
54.72
Wan VAE (truncated)
81.70
82.62
78.03
51.11
Appendix
Table 11: Training decoders with both components using VideoMAE v2 L16. Evaluation uses the original Wan VAE.
Steps
Total ↑
Quality ↑
Semantic ↑
Dynamic ↑
4
84.11
85.47
78.67
74.72
2
83.77
85.20
78.03
70.00
1
81.58
82.75
76.94
69.17
Appendix
Table 12: Denoising-step ablation of Direct Forcing after 150 updates. Steps refer to later chunks in RDM rollouts and inference; the first chunk uses four steps. CCA retains the four-level noise grid for all chunks.
Flow weight w
Total ↑
Qual. ↑
Sem. ↑
Dyn. ↑
0
82.40
83.51
77.99
54.72
10−4
82.93
84.12
78.19
77.50
5×10−4
83.46
84.87
77.84
90.56
Appendix
Table 13: Optical-flow regularization with both components using VideoMAE v2 L16. Increasing the weight improves Total and Dynamic.
Metric
0
10−4
5×10−4
Subject Consistency
95.69
95.51
95.47
Background Consistency
95.62
94.92
94.87
Temporal Flickering
99.15
97.72
97.73
Motion Smoothness
98.77
98.44
98.32
Dynamic Degree
54.72
77.50
90.56
Aesthetic Quality
65.07
64.85
64.72
Appendix
Table 14: Full VBench dimension scores for optical-flow regularization. Columns give flow weight w . Scores are multiplied by 100; higher is better for every metric.
Representation
Total ↑
Quality ↑
Semantic ↑
Dynamic ↑
Global pooling
80.75
81.84
76.39
55.56
Tokens
82.40
83.51
77.99
54.72
Appendix
Table 15: Pooling ablation for rollout distribution matching, with VideoMAE v2 L16 for both training components.
CCA representation
Total ↑
Qual. ↑
Sem. ↑
Dyn. ↑
V-JEPA 2.1 L16
81.14
82.01
77.65
28.61
V-JEPA 2.1 L24
81.62
82.47
78.24
35.56
Appendix
Table 16: V-JEPA 2.1 features for continuation alignment, with rollout distribution matching fixed to VideoMAE v2 L24.
Figure 8: Qualitative effects of the training schedule. Rows show CCA-only, sequential CCA initialization followed by RDM-only, and joint training. CCA-only and joint training use 150 updates; sequential training uses 150+150. All rows use frames 0/16/32 for the school case and 0/40/80 for the airliner case. Prompts are abbreviated.
CCA batch size
Total ↑
Quality ↑
Semantic ↑
Dynamic ↑
16
81.84
82.64
78.61
38.06
32
82.40
83.51
77.99
54.72
64
82.03
83.21
77.32
51.67
Appendix
Table 17: Sensitivity to the CCA batch size, using VideoMAE v2 L16 for both CCA and RDM. The default batch size is 32.
Continuation
Distribution
Total ↑
Qual. ↑
Sem. ↑
Dyn. ↑
V24
V16
82.80
84.00
78.01
58.06
V24
D9
83.49
84.96
77.58
68.61
V24
V16+D9
84.11
85.47
78.67
74.72
V24
V16+D9+V24
83.86
85.31
78.05
77.78
V24+D9
V16
83.35
84.36
79.32
56.39
V24+D9
D9
83.04
84.37
77.72
65.00
Appendix
Table 18: Additional representation combinations after 150 updates. Vk : VideoMAE v2 L k ; D9 : DiT L9. Gray marks the representation configuration used by our main model.
Perturbation
Clean
Noised
Reduction
Cross-video splicing
24.75
17.99
27.29%
Temporal freezing: 50% retained
4.31
3.33
22.68%
Appendix
Table 19: Response of clean and noised DiT L9 features at τ=300 . Distances are multiplied by 100; Reduction gives the relative decrease after noising.
Layer
τ
Total ↑
Quality ↑
Semantic ↑
Dynamic ↑
CCA: RDM uses VideoMAE v2 L16
L9
0
81.88
82.54
79.25
44.72
L9
300
82.56
83.53
78.68
58.89
L9
600
82.10
82.88
78.99
48.33
L9
900
82.02
83.02
78.05
65.28
L25
300
82.90
83.78
79.36
56.39
Appendix
Table 20: DiT layer and encoder time condition τ for CCA and RDM. All DiT inputs are clean latents; the other component uses VideoMAE v2 L16. Gray marks L9 at τ=300 ; best scores are bold within each component.
ViRDM
Elastic Forcing
Direct Forcing
Dynamic ↑
48.61
63.33
75.56
Appendix
Table 21: Dynamic Degree in the main comparison.
Batch size
Update
Total ↑
Quality ↑
Semantic ↑
Dynamic ↑
64
150
84.11
85.47
78.67
74.72
128
250
84.63
86.01
79.08
75.56
256
175
83.87
85.24
78.36
71.94
512
75
83.63
85.00
78.15
65.28
Appendix
Table 22: RDM batch size with VideoMAE v2 L24 for CCA and VideoMAE v2 L16 plus DiT L9 for RDM. Each row reports the highest-Total checkpoint within the first 250 updates; CCA batch size is 32. Gray marks the main setting.
Method
Chunk boundary ratio ( ≈1 )
Motion Smoothness ↑
ViRDM [ 24 ]
1.393 [1.332, 1.461]
98.12
Direct Forcing
1.040 [0.973, 1.129]
98.52
Appendix
Table 23: Temporal continuity of the ODE-initialized ViRDM pipeline and our single-stage pipeline, both without optical-flow regularization. Brackets denote 95% prompt-bootstrap confidence intervals. Boundary ratios closer to one are better.
Metric
Original data
Seedance 2.5
Dynamic Degree ↑
54.72
62.50
Motion Smoothness ↑
98.77
98.90
Subject Consistency ↑
95.69
93.81
Background Consistency ↑
95.62
93.80
Appendix
Table 24: Full VBench scores of Direct Forcing trained on the indicated reference data. Scores are percentages.
Few-step autoregressive video generation commonly relies on Distribution Matching Distillation (DMD), requiring a bidirectional diffusion teacher and an online fake-score model. We instead learn the rollout distribution directly from reference videos, eliminating both score models during post-training. Our framework minimizes maximum mean discrepancy (MMD) in frozen self-supervised video representation spaces, using a hybrid Nyström--Monte Carlo estimator to balance approximation bias and sampling variance. Memory-efficient replay and gradient subsampling make this objective practical. Using the same architecture and initialization as Self-Forcing, our 1.3B model improves the VBench Total score from 83.80 to 84.64 while retaining 17 FPS. Removing auxiliary score models also enables 14B post-training on eight H200 GPUs. Beyond distillation, learning from reference videos enables the acquisition of new visual styles, semantic concepts, and spatial priors without a target-specific diffusion teacher.
Chi Zhang, Yueyi Liu, Shi Haoyang +5
College of AI, Tsinghua University · IAIR, Xi’an Jiaotong University · Xianghui Academy, Fudan University +1
Recent advances in autoregressive diffusion-based video generation have substantially improved the quality of real-time video synthesis. However, most existing methods still require multiple denoising steps, while reducing sampling to a single step often leads to severe quality degradation: trajectory-based consistency distillation methods often produce videos with weak dynamics, whereas DMD-based methods, such as Self-Forcing, tend to generate blurry frames. We attribute this limitation to the teacher trajectories exhibiting highly concentrated curvature near the high-noise endpoint, which poses a fundamental geometric challenge for one-step distillation under the consistency distillation framework. To address these limitations, we propose One-Forcing, a simple yet effective approach that augments the DMD objective with an auxiliary GAN loss for high-quality and efficient one-step video generation. We further find that framewise autoregression stabilizes adversarial training, enabling higher-quality generation with substantially fewer training iterations than chunkwise autoregression. Experiments on VBench show that One-Forcing achieves a total score of 83.76, establishing state-of-the-art performance among one-step causal video generation methods while remaining competitive with strong multi-step approaches.
Autoregressive video models can generate minute-long videos in real time, but they produce generic subjects from text rather than specific subjects from user-provided images. Existing customization methods either require costly per-subject optimization or use pretrained conditioning networks that jointly process all video frames with bidirectional attention. Neither approach is designed for causal streaming. We present Custom Forcing, a training-free method that stores reference-based anchor frames in the persistent KV cache of a frozen autoregressive video model. However, fixed anchors face two limitations: simple conditioning allows identity to drift, and the text prompt continues to favor a generic subject. To address these problems, drift-adaptive value amplification (DVA) scales reference influence with the degree of identity drift, while anchor contrast guidance (ACG) steers generation away from the generic class prior. Over two-minute rollouts, fixed anchors fall from 0.58 to 0.42 in DINO-I, while Custom Forcing keeps it between 0.58 and 0.62 without reducing motion. Custom Forcing also achieves higher subject similarity than bidirectional customization methods and better preserves identity over 30s than causal image-to-video and reference-to-video models, while generating each frame 9.5-28.5 times faster than these long-video baselines.
Yunseung Ok, Hyunsoo Kim, Minseo Kim +1
Kyung Hee University · The University of Texas at Austin