Any Forcing: Any-Step Generality and One-Step Stability for Autoregressive Video Generation
Authors: Jiahua Xu, Yijie Fang, Nannan Wang, Xunhao Lin, Li Cao, Lei Hao, Mingrui Zhu, Fei Gao, Xinbo Gao
Organizations: State Key Laboratory of Integrated Services Networks, Xidian University, Xi’an, China · Technology Planning Dept of the Computing Technology Research Institute, Huawei, Shanghai, China · Huawei Noah’s Ark Institute, Huawei, Shenzhen, China
Autoregressive video generation enables efficient inference through streaming prediction. However, existing methods are typically restricted to fixed sampling steps, limiting the flexibility to balance generation quality and efficiency, while one-step generation suffers from degraded visual quality and error accumulation over long sequences, leading to content drift and generation collapse. To address these issues, we propose Any Forcing, a unified framework for achieving Any-Step Generality and One-Step Stability. Specifically, we introduce a unified adversarial discriminator that jointly performs real-distribution alignment, inter-step consistency alignment, and first-frame quality enhancement, enabling flexible sampling while preserving one-step generation quality. Furthermore, we propose a training-free adaptive long-term memory selection mechanism that selects historical states based on global semantic relevance and local value contribution, while remapping their temporal positions to preserve temporal relationships within the memory cache, thereby enabling stable one-step long-video generation. Experiments demonstrate that Any Forcing achieves high-quality and stable one-step autoregressive video generation, achieving 20.9 FPS and 24.7 FPS for short and long video generation on a single H100 GPU, respectively.
Figures & tables
Figure 1: Visual comparison of short- and long-video generation. Existing methods exhibit degraded visual quality and suffer from long-horizon error accumulation under one-step sampling. Any Forcing achieves stable one-step generation and long-horizon consistency.
Figure 2: Motivation of Any Forcing. Existing methods struggle to balance sampling-step generality and one-step quality, while insufficient global relevance or local contribution undermines one-step stability in long-video generation. Any Forcing addresses these challenges through a unified adversarial discriminator and adaptive long-term memory selection mechanism.
Figure 3: Overview of Any Forcing. Left: A unified discriminator jointly performs real-distribution alignment, inter-step consistency alignment, and first-frame enhancement, complemented by teacher ODE endpoint supervision during early training. Right: Adaptive long-term memory selection combines a global sink, a dynamic memory pool, and recent local context. Global semantic relevance and local value contribution determine memory admission before generating the next frame, followed by position mapping within the trained RoPE horizon. For better visualization, we use video frames as an example to show the mechanism, where Fj represents the j -th frame.
Method
NFE
Total ↑
Quality ↑
Sem. ↑
IQ ↑
AQ ↑
OC ↑
NFE
Total ↑
Quality ↑
Sem. ↑
IQ ↑
AQ ↑
OC ↑
Base model
Wan 2.1-1.3B [ 24 ]
50
83.99
84.82
80.66
67.52
65.70
94.56
20
83.63
84.52
80.06
65.71
65.19
93.81
Few steps
CausVid [ 34 ]
4
83.05
84.14
78.70
68.44
66.04
93.35
2
81.03
82.26
76.11
63.25
62.55
90.22
rCM [ 40 ]
4
83.11
84.48
77.65
66.30
63.33
92.96
2
81.06
83.23
72.34
59.84
62.04
90.35
Self Forcing [ 8 ]
4
84.20
84.88
81.48
69.44
66.68
94.37
2
81.79
83.13
76.40
66.02
63.39
90.40
Table 1: Comparison on VBench for 5-second video generation at 480p resolution. For the few-step setting, the left and right groups use four and two denoising steps, respectively. For the one-step setting, 1∗ and 1† use four and two denoising steps for the first latent frame, respectively, followed by one step for subsequent frames. Best and second-best results within each sampling group are bold and underlined , respectively. Sem.: Semantic. We report three representative quality and semantic metrics here: IQ: Imaging Quality; AQ: Aesthetic Quality; OC: Object Class. Detailed results can be found in Supplementary Material E .
Method
30-second video generation
60-second video generation
NFE
Total ↑
Quality ↑
Sem. ↑
IQ ↑
AQ ↑
OC ↑
NFE
Total ↑
Quality ↑
Sem. ↑
IQ ↑
AQ ↑
OC ↑
Few steps
CausVid [ 34 ]
4
78.54
81.08
68.38
60.30
55.42
81.05
4
74.92
79.29
57.42
56.80
50.78
65.50
LongLive [ 28 ]
4
82.81
83.34
80.69
68.73
63.90
95.12
4
82.73
83.24
80.69
68.57
64.16
94.43
One step
Infinity RoPE [ 30 ]
1∗
84.51
85.83
79.22
70.52
62.77
94.88
1∗
84.12
85.57
78.31
70.80
61.46
93.41
Table 2: Comparison on VBench for 30- and 60-second video generation at 480p resolution. 1∗ and 1† use four and two sampling steps for the first latent frame, respectively, followed by one step for subsequent frames. Best and second-best results within each one-step sampling group and video length are bold and underlined , respectively.
Figure 4: Visual comparisons of one-step short-video (top) and long-video (bottom) generation. Any Forcing achieves high-quality video generation while preserving stable subject appearance and scene structure over time. All methods adopt the 1∗ sampling configuration. More visualization results are available in Supplementary Material E .
Figure 5: Human preference results. The corresponding sampling settings are indicated to the right of the bars. We report the win rate of Any Forcing against others in blind pairwise comparisons.
Figure 6: Performance analysis of short (left) and long (right) video generation on a single H100 GPU. 1∗ and 1† use four and two first-frame steps, respectively, followed by one step per frame. The bubble size indicates the number of sampling steps (left) and the historical-frame budget H , excluding the current frame (right).
Method
Total
Quality
Semantic
Total
Quality
Semantic
Any Forcing
85.06
86.19
80.54
84.79
85.83
80.64
w/o CA
84.66
85.67
80.62
84.14
85.11
80.25
w/o RA
83.78
85.19
78.15
83.86
85.12
78.85
w/o FE
83.52
84.86
78.15
81.86
83.22
76.45
w/o ODE
84.67
85.83
80.00
83.10
84.17
78.84
Table 3: Ablation study of the proposed components on VBench. The left and right groups correspond to 1∗ and 1† sampling, respectively. Best and second-best results within each sampling group are bold and underlined , respectively.
Method
IQ
AQ
OC
MO
SR
TF
Scene
TS
Any Forcing
71.71
59.38
93.50
82.08
89.14
99.34
59.70
23.28
w/o GSR
71.68
59.33
92.67
81.55
89.55
99.37
59.36
23.20
w/o LVC
71.69
59.35
93.27
81.72
89.73
99.33
59.15
23.09
w/o PM
70.90
58.99
92.45
81.44
89.51
99.27
59.05
22.95
Table 4: Ablation study of adaptive long-term memory selection on VBench. GSR: Global Semantic Relevance; LVC: Local Value Contribution; PM: Position Mapping. Best and second-best results are bold and underlined , respectively.
λFE
Total
Quality
Semantic
Total
Quality
Semantic
0.01
83.46
84.95
77.49
81.59
82.94
76.23
0.05
85.06
86.19
80.54
84.79
85.83
80.64
0.10
83.54
84.88
78.16
81.64
82.78
77.08
Table 5: Sensitivity analysis of the first-frame enhancement weight λFE on VBench. The left and right groups correspond to 1∗ and 1† sampling, respectively. Best and second-best results within each sampling group are bold and underlined , respectively.
Figure 7: Sensitivity analysis of recent-context length (left) and long-term memory capacity (right). Recent frames denote the number of retained local-context frames after candidate removal.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Hyperparameter
Setting
Generator initialization
ODE-initialization causal model from Causal Forcing
Flow denoising for fake score; DMD + multi-task GAN + ODE endpoint supervision for generator
Training frames
21 latent frames; 16×60×104 per frame
Frames per autoregressive block
1
Training sampling steps
Randomly selected from {1,2,3,4}
Appendix
Table 6: Implementation settings of Any Forcing.
Figure 14
Figure 8: Interface for the blind pairwise user study. Participants compare videos generated from the same prompt and select their preferred result based on visual quality and semantic alignment.
Figure 9: Adversarial training dynamics with and without ODE endpoint warm-up over the first 300 training steps. Columns show real-distribution alignment, inter-step consistency, and first-frame enhancement; rows show generator and discriminator losses, respectively. The shaded region indicates the 200-step warm-up stage.
Setting
IQ
AQ
OC
MO
SR
TF
Scene
TS
Any Forcing
71.71
59.38
93.50
82.08
89.14
99.34
59.70
23.28
w/ noisy query
71.74
59.36
93.31
81.84
89.77
99.34
59.88
23.10
Appendix
Table 7: Effect of including the current noisy query in local value evaluation.
Figure 10: Qualitative examples of generation with the current noisy query included in local value evaluation. Red arrows indicate regions of content degradation, including duplicated sheep heads.
Setting
IQ
AQ
OC
MO
SR
TF
Scene
TS
Any Forcing
71.71
59.38
93.50
82.08
89.14
99.34
59.70
23.28
w/o collisions
71.65
59.34
92.90
82.09
89.03
99.31
58.95
23.13
Appendix
Table 8: Effect of preventing positional collisions.
Figure 11: Comparison across all 16 VBench dimensions. Top: 5s generation with uniform four-step and two-step sampling, followed by one-step generation with four and two first-frame steps. Bottom: corresponding one-step comparisons for 30s (left two panels) and 60s (right two panels). Any Forcing is highlighted in red. The radar plots use axis-specific display ranges shared by all methods within each panel; outermost labels denote axis limits rather than method scores. SC: Subject Consistency; BC: Background Consistency; TF: Temporal Flickering; MS: Motion Smoothness; DD: Dynamic Degree; AQ: Aesthetic Quality; IQ: Imaging Quality; OC: Object Class; MO: Multiple Objects; HA: Human Action; SR: Spatial Relationship; AS: Appearance Style; TS: Temporal Style; Overall: Overall Consistency.
Figure 12: Comparisons of 60-second video generation on a Tokyo street scene (top) and a coastal landscape (bottom). All methods adopt the 1∗ sampling configuration.
Figure 13: Comparisons of Any Forcing under different sampling schedules. The top, middle, and bottom groups show 5-, 30-, and 60-second generation, respectively.
Autoregressive video generation creates videos frame by frame or in chunks, conditioning each prediction on prior context for streaming and interactive use.
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.
We present AAD-1, an Asymmetric Adversarial Distillation framework for One-step autoregressive image-to-video generation. State-of-the-art methods adopt adversarial distillation but suffer from motion collapse and training instability, resulting in static videos. AAD-1 addresses these challenges through two key designs in architecture and training strategy. Our key architectural insight is to break the symmetry between generator and discriminator. While the generator remains causal to preserve autoregressive sampling capability, the discriminator attends bidirectionally over the full spatiotemporal context and produces a single holistic realism score for the entire video sequence. This asymmetric design enables the discriminator to effectively detect global temporal failures and long-range drift that cause motion collapse in autoregressive generation. To stabilize training, we introduce a phased strategy that first uses distribution matching to bootstrap a stable one-step generator, providing a warm-up phase that brings the student distribution closer to the teacher before adversarial distillation begins. Extensive experiments on VBench demonstrate that AAD-1 achieves state-of-the-art performance in one-step autoregressive video generation.
Haobo Li, Yanhong Zeng, Yunhong Lu +6
AutoLab, SAI, SJTU · Ant Group · Department of Automation, Tsinghua University +2
Few-step autoregressive video generators enable streaming generation, yet they run every token through the final denoising step, regardless of how much refinement it needs. We find this uniform computation is often unnecessary: many tokens already closely match their eventual full-step outputs in the initial steps, while others require refinement until the end. We introduce Speculative Forcing, a framework that treats intermediate clean estimates as drafts and predicts when individual tokens can be safely accepted early. Once accepted, a token is frozen as its final output and undergoes no further denoising, while its cached keys and values remain available as context for the remaining tokens. We explore two alternative verifier designs: a training-free one based on similarity to clean content and stability across steps, and a learned one that reads the frozen generator's intermediate features to estimate each token's remaining gap to its full-step output. Across short-video, long-video, and world generation, Speculative Forcing preserves the quality of full-step sampling while substantially reducing computation. On LongLive, it matches the four-step generator with ~1.7 denoising steps per token, running over 1.5x faster at 41.7 and 37.8 fps. On Lingbot-World 2.0, where camera motion continually changes the visible content, it maintains quality with ~2.5 denoising steps per token and runs 1.35x faster at 24.5 fps.