Speculative Forcing for Efficient Autoregressive Video Generation

Hyunwook Choi; Seungryong Kim; Dongyoon Han; Sangdoo Yun

cs.CVOct 15, 2026

Speculative Forcing for Efficient Autoregressive Video Generation

Authors: Hyunwook Choi, Seungryong Kim, Dongyoon Han, Sangdoo Yun

Organizations: KAIST AI · NAVER AI Lab

Abstract

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.

Figures & tables

Figure 1: Compared with full-step sampling (a), Speculative Forcing treats intermediate token estimates as drafts, accepts those that have converged, and refines only the rest (b), achieving comparable quality with fewer denoising steps and higher throughput (c).

Figure 2: Token-wise convergence during denoising. We visualize latent differences between each chunk’s $k$ -step prediction and the baseline four-step output. (a) In video generation, background and object regions that remain similar to history tend to converge early, while other regions continue to change. (b) In world generation, previously observed regions often converge early, whereas newly revealed regions tend to require further denoising.

Figure 3: Speculative Forcing. Baseline runs every token through the full denoising schedule. Speculative Forcing instead treats each token’s intermediate clean estimate as a draft and asks a verifier to predict its remaining gap

\mathrm{rel}_{k}(i)

to the full-step output. Tokens predicted to have converged are accepted and committed, after which they enter the next step only as keys and values, so the following denoising pass is computed for the rejected tokens alone.

Method	Mean Steps $\downarrow$	Through. (FPS)	Evaluation scores $\uparrow$
Method	Mean Steps $\downarrow$	Through. (FPS)	Total	Quality	Semantic
LongLive ( Yang et al., 2025a )
1-step	1.00	67.97	79.06	80.03	75.17
2-step	2.00	44.54	82.46	83.06	80.09
3-step	3.00	33.15	83.08	83.45	81.58
4-step	4.00	26.40	83.08	83.49	81.43

Table 1: Comparison on 5s videos with LongLive. Mean Steps is the measured average number of denoising passes per token. Standard VBench 16-dimension protocol, throughput on a single H100.

Method	Mean Steps $\downarrow$	Through. (FPS)	Drift $\downarrow$	Total Score $\uparrow$	Quality Score $\uparrow$	Semantic Score $\uparrow$	VBench-Long dimension scores $\uparrow$
Method	Mean Steps $\downarrow$	Through. (FPS)	Drift $\downarrow$	Total Score $\uparrow$	Quality Score $\uparrow$	Semantic Score $\uparrow$	Subject Consistency	Background Consistency	Motion Smoothness	Dynamic Degree	Aesthetic Quality	Imaging Quality
LongLive ( Yang et al., 2025a )
1-step	1.00	61.6	2.930	80.82	81.87	76.64	96.94	95.76	98.24	49.63	60.09	63.81
2-step	2.00	40.9	1.992	82.32	82.99	79.61	97.67	96.38	98.60	44.81	62.92	67.66
3-step	3.00	30.6	1.738	82.60	83.09	80.67	98.09	96.75	98.77	38.98	63.71	68.69
4-step	4.00	24.5	1.722	82.69	83.29	80.30	98.22	96.99	98.85	39.17	64.43	68.52

Table 2: Comparison on 30s long videos with LongLive. Evaluated on VBench-Long.

Method	Mean Steps $\downarrow$	Through. (FPS)	RPE		Drift $\downarrow$	Quality Score $\uparrow$	VBench dimension scores $\uparrow$
Method	Mean Steps $\downarrow$	Through. (FPS)	Trans $\downarrow$	Rot $\downarrow$	Drift $\downarrow$	Quality Score $\uparrow$	Subject Consistency	Background Consistency	Motion Smoothness	Dynamic Degree	Aesthetic Quality	Imaging Quality
Lingbot-World 2.0 ( Gao et al., 2026 )
1-step	1.00	46.9	0.552	0.360	5.579	77.19	81.31	87.23	96.98	96.50	58.38	66.63
2-step	2.00	30.7	0.345	0.255	3.999	81.68	87.71	90.28	97.10	98.50	65.98	70.65
3-step	3.00	22.8	0.329	0.251	3.361	82.38	88.42	90.67	97.13	99.00	67.14	71.61
4-step	4.00	18.1	0.328	0.253	3.094	82.62	88.50	90.83	97.23	99.00	67.18	72.24

Table 3: World-model comparison on dynamic videos. Results with the autoregressive Lingbot-World 2.0 ( Gao et al., 2026 ) , evaluated on VBench. RPE is the relative pose error of the camera trajectory. Quality Score aggregates the six dimensions reported here.

Figure 4: Qualitative comparisons on video generation and world modeling. Both our head and training-free verifiers preserve visual quality comparable to the four-step baseline without visibly increasing quality drift over the illustrated rollouts on our benchmark.

Figure 5: Qualitative ablations of verification and commitment granularity. Our learned verifier handles distortions that challenge training-free verification. Premature chunk-wise acceptance propagates errors to later chunks, while motion-based verification struggles to distinguish camera-induced changes from distortions.

Table 9

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Verifier signal	$\epsilon=0.16$			$\epsilon=0.24$			$\epsilon=0.32$
Verifier signal	$k=1$	$k=2$	$k=3$	$k=1$	$k=2$	$k=3$	$k=1$	$k=2$	$k=3$
Learned verifier (head) design
On-policy training	0.932	0.909	0.905	0.935	0.889	0.902	0.941	0.871	0.887
Per-token MLP	0.936	0.910	0.913	0.940	0.894	0.911	0.946	0.876	0.888
Ours H	0.933	0.911	0.910	0.934	0.890	0.906	0.944	0.871	0.887
Latent-based verification signals

Appendix

Table A: Verification accuracy on LongLive. We report AUROC

\uparrow

for identifying tokens requiring further refinement, defined by

\mathrm{rel}_{k}(i)>\epsilon

Method	Mean Steps $\downarrow$	Through. (FPS)	Drift $\downarrow$	Quality Score $\uparrow$	Imaging Quality $\uparrow$
On-policy training	2.57	24.5	2.940	82.67	72.40
Per-token MLP	2.52	24.8	2.899	82.67	72.48
w/o spatiotemporal smoothing	2.55	24.7	2.882	82.58	72.54
Ours H	2.55	24.5	2.839	82.65	72.65

Appendix

Table B: Verifier ablations on Lingbot-World 2.0. Learned verifier variants are calibrated to our operating point (

\epsilon=0.24

Method	Mean Steps	Through. (FPS)	Total	Quality	Subject Consist.	Backgr. Consist.	Temporal Flicker	Motion Smooth.	Dynamic Degree	Aesthetic Quality	Imaging Quality
1-step	1.00	67.97	79.06	80.03	90.43	93.51	99.20	98.24	45.00	61.06	63.97
2-step	2.00	44.54	82.46	83.06	95.88	95.84	99.35	98.59	46.67	65.05	67.65
3-step	3.00	33.15	83.08	83.45	96.91	96.75	99.40	98.76	41.11	66.55	68.39
4-step	4.00	26.40	83.08	83.49	97.13	96.98	99.38	98.84	37.78	67.21	68.82
Surprise Forcing ( Shi et al., 2026 )	3.59	27.38	83.11	83.53	97.13	96.99	99.39	98.83	38.33	67.23	68.80
Ours TF (matched to $\varepsilon{=}0.24$ )	1.97	39.77	83.16	83.55	96.97	96.51	99.43	98.76	41.67	66.64	68.83

Appendix

Table C: Quality evaluation on LongLive. Evaluated on VBench.

Method	Mean Steps	Through. (FPS)	Total	Semantic	Object Class	Multiple Objects	Human Action	Colour	Spatial Relation	Scene	Appear. Style	Temporal Style	Overall Consist.
1-step	1.00	67.97	79.06	75.17	90.27	64.57	94.80	91.03	68.39	46.45	21.11	23.28	26.62
2-step	2.00	44.54	82.46	80.09	95.52	79.88	97.20	90.92	77.32	55.77	20.62	24.06	26.92
3-step	3.00	33.15	83.08	81.58	96.80	87.20	96.60	91.25	78.87	58.75	20.58	24.20	26.78
4-step	4.00	26.40	83.08	81.43	96.27	87.26	96.20	90.78	78.21	59.65	20.55	24.21	26.65
Surprise Forcing ( Shi et al., 2026 )	3.59	27.38	83.11	81.42	96.28	87.18	96.20	90.87	78.22	59.52	20.55	24.21	26.66
Ours TF (matched to $\varepsilon{=}0.24$ )	1.97	39.77	83.16	81.62	96.49	86.23	97.40	91.04	79.47	59.46	20.47	24.23	26.72

Appendix

Table D: Semantic evaluation on LongLive. Evaluated on VBench.

Method	Mean Steps	Through. (FPS)	Drift	Total	Quality	Subject Consist.	Backgr. Consist.	Temporal Flicker	Motion Smooth.	Dynamic Degree	Aesthetic Quality	Imaging Quality
1-step	1.00	61.59	2.930	80.82	81.87	96.94	95.76	99.24	98.24	49.63	60.09	63.81
2-step	2.00	40.91	1.992	82.32	82.99	97.67	96.38	99.27	98.60	44.81	62.92	67.66
3-step	3.00	30.62	1.738	82.60	83.09	98.09	96.75	99.31	98.77	38.98	63.71	68.69
4-step	4.00	24.45	1.722	82.69	83.29	98.22	96.99	99.28	98.85	39.17	64.43	68.52
Surprise Forcing ( Shi et al., 2026 )	3.22	27.61	1.674	82.89	83.52	98.19	97.04	99.35	98.80	41.57	64.80	68.44
Ours TF	1.71	38.78	1.685	82.75	83.28	98.11	96.77	99.44	98.74	41.57	63.83	68.27

Appendix

Table E: Quality evaluation on LongLive. Evaluated on VBench-Long.

Method	Mean Steps	Through. (FPS)	Drift	Total	Semantic	Object Class	Multiple Objects	Human Action	Colour	Spatial Relation	Scene	Appear. Style	Temporal Style	Overall Consist.
1-step	1.00	61.59	2.930	80.82	76.64	90.99	71.92	95.67	92.21	67.74	49.88	21.08	23.39	26.40
2-step	2.00	40.91	1.992	82.32	79.61	93.58	81.67	98.00	91.87	75.71	52.77	20.74	23.92	26.65
3-step	3.00	30.62	1.738	82.60	80.67	92.24	86.96	97.33	91.28	78.54	56.14	20.60	24.11	26.61
4-step	4.00	24.45	1.722	82.69	80.30	92.47	84.48	95.67	89.56	79.43	57.83	20.58	24.09	26.43
Surprise Forcing ( Shi et al., 2026 )	3.22	27.61	1.674	82.89	80.39	92.85	84.55	96.33	89.99	79.43	56.77	20.58	24.14	26.58
Ours TF	1.71	38.78	1.685	82.75	80.65	92.75	84.93	97.33	90.09	79.74	57.76	20.55	23.91	26.63

Appendix

Table F: Semantic evaluation on LongLive. Evaluated on VBench-Long.

Efficient Video Generation

See all 25 papers

One-Forcing: Towards Stable One-Step Autoregressive Video Generation
May 22, 2026Jiaqi Feng, Justin Cui, Yuanhao Ban +1Autoregressive Video Generation One-Step Generative Modeling
Tsinghua University · UCLA
Speculative Decoding for Autoregressive Video Generation
Apr 19, 2026Yuezhou Hu, Jintao ZhangVideo Diffusion Models Autoregressive Diffusion
University of California, Berkeley
Prospective Forcing: Learning to Look Ahead for Autoregressive Long Video Diffusion
Oct 10, 2026Yifan Shen, Yewei Xia, Gongxu Luo +6Autoregressive Video Generation Video Diffusion Models
Mohamed bin Zayed University of AI · Institute of Foundation Models · Carnegie Mellon University

May 22, 2026J/K move · Enter open · S save

One-Forcing: Towards Stable One-Step Autoregressive Video Generation

Jiaqi Feng, Justin Cui, Yuanhao Ban, Cho-Jui Hsieh

Tsinghua University · UCLA

Autoregressive Video Generation One-Step Generative Modeling Knowledge Distillation Adversarial Knowledge Distillation Video Diffusion Models Single-Step Video Generation+1

Open paper AskPDF arXiv

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.

Abstract

Efficient Video Generation

Explore similar work

One-Forcing: Towards Stable One-Step Autoregressive Video Generation

Speculative Decoding for Autoregressive Video Generation

Prospective Forcing: Learning to Look Ahead for Autoregressive Long Video Diffusion