Autoregressive (AR) modeling is arguably the only way to extend a fixed-length video diffusion model to longer or even streaming videos. However, existing AR diffusion methods use only the previously predicted chunk as context for the next, discarding the intermediate predictions that constitute the chunk's generation history. In contrast, conventional AR models such as LLMs condition on the entire history of preceding tokens. To this end, we propose Trajectory Forcing (TraF), an AR diffusion method that exploits the denoising trajectories of previous video chunks as context. TraF operates under two training regimes: (1) with ground-truth videos, where pseudo-trajectories are constructed from Gaussian interpolations and re-predicted by the model, and (2) with a teacher model, where real trajectories are obtained from the student's own denoising rollout. The first regime naturally provides a strong initialization for the second. On the 5-second VBench benchmark, TraF achieves a Total score of 85.15, outperforming Self Forcing (+0.84) and Causal Forcing (+1.11), with Quality and Semantic improving simultaneously. On 30-second generation---6× the training horizon---TraF achieves the highest overall quality among all compared methods.
Figures & tables
Figure 1: Standard AR vs. AR diffusion. (a) In standard AR (Eq. 1 ), each token is atomic and requires no further generation. The next token is naturally conditioned on all preceding tokens. (b) In AR diffusion, each chunk is generated through T denoising steps that form a trajectory. Following Eq. ( 2 ), each denoising step should be conditioned on the full trajectories of all preceding chunks.
Method
SL
RL
Fixed-point
Teacher Forcing Williams and Zipser [1989] , Jin et al. [2024] , Teng et al. [2025] , Yuan et al. [2026]
GT x(n,T)
✗
Diffusion Forcing Chen et al. [2024] , Song et al. [2025] , Yin et al. [2025] , Wu et al. [2025] , Guo et al. [2025] , Po et al. [2025]
Corrupted x(n,T)
✗
Self Forcing Huang et al. [2025] , Zhu et al. [2026] , Liu et al. [2025] , Yang et al. [2025] , Cui et al. [2025]
✗
Endpoint x(n,T)
Trajectory
Trajectory Forcing (Ours)
Trajectory x(n,⋅)
Trajectory x(n,⋅)
Table 1: AR diffusion methods categorized by their use of the denoising trajectory. Each column reports the conditioning used under the corresponding training stage. ✗ indicates that the method does not use that stage. Prior methods condition on a fixed point regardless of the denoising step, whereas Trajectory Forcing conditions on the denoising trajectory x(n,⋅) .
Figure 3
Figure 2: Trajectory conditioning and γ ablation. Left: Each denoising step of the current chunk conditions on the preceding chunk’s trajectory. γ determines the selected position ( γ→∞ : endpoint-only conditioning, as in existing methods). Middle: Unregularized continuous targets Sγ ; finite-grid selection is detailed in Appendix A . Right: The tested finite values of γ outperform endpoint-only conditioning in Total. γ=2 – 3 forms a broad optimum.
Method
#Params
Teacher
Throughput
Latency
Total ↑
Quality ↑
Semantic ↑
(FPS) ↑
(s) ↓
Bidirectional Video Diffusion
LTX-Video HaCohen et al. [2024]
1.9B
✗
8.98
13.5
79.83
81.88
71.62
Wan2.1-1.3B Team Wan et al. [2025]
1.3B
✗
0.78
103
83.37
84.30
79.65
Autoregressive Video Diffusion
MAGI-1 Teng et al. [2025]
4.5B
✗
0.19
282
78.88
81.67
67.72
Table 2: Short-video evaluation on VBench (946 prompts, all 16 dimensions). Result sources are detailed in Appendix C . Throughput and latency are reported in FPS and seconds, respectively. —: not reported.
Method
Temporal Quality ↑
Frame Quality ↑
Text Align. ↑
Total ↑
Video- Score2 ↑
Quality Drift ↓
CausVid Yin et al. [2025]
83.97
62.66
58.88
68.50
3.549
2.72
HiAR Zou et al. [2026]
82.82
65.32
62.42
70.19
3.758
1.60
Causal Forcing Zhu et al. [2026]
89.26
59.58
62.61
70.48
3.589
4.36
Self Forcing Huang et al. [2025]
85.35
61.49
66.33
71.06
3.719
2.21
Trajectory Forcing (Ours)
90.39
65.46
66.08
73.98
3.771
1.38
Rolling Forcing Liu et al. [2025]
82.04
65.36
66.77
71.39
3.804
0.56
Table 3: Long-video evaluation (128 prompts, 30 s). Temporal Quality, Frame Quality, and Text Alignment are axis aggregates following CausVid Yin et al. [2025] . Total is their equal-weighted average. The VideoScore2 He et al. [2025] column reports the text-alignment sub-score. Gray : long-video-specific training.
Figure 3: Qualitative comparison on 30-second videos. Three frames are sampled at t=0 s, t=15 s, and t=30 s for four representative MovieGen prompts: stylish woman (top-left), paper airplanes (top-right), art gallery tour (bottom-left), and rabbit reading newspaper (bottom-right). Trajectory Forcing (Ours) maintains subject coherence and color stability across the full 30 seconds, whereas Causal Forcing exhibits noticeable color drift after t=15 s, HiAR produces visually stable but near-static outputs, Self Forcing shows gradual detail loss in later segments, and CausVid degrades earliest, with pronounced color drift and texture artifacts by t=15 s.
Conditioning
Total ↑
Quality ↑
Semantic ↑
Gaussian ( σ=0.22 )
83.87
84.76
80.29
Gaussian ( σ=0.11 )
83.95
84.92
80.08
Trajectory
85.15
85.89
82.20
Table 4: Ablation studies (5-second VBench). Left: noise vs. trajectory comparison, where endpoint predictions with added Gaussian noise of varying magnitude replace trajectory predictions. Right: SL-stage comparison across three conditioning strategies under identical conditions (architecture, data, 20K steps), evaluated with 50-step denoising before distillation.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Component
Setting
Base model
Wan2.1-T2V-1.3B (causal attention)
Video resolution
832×480 , 16 fps, 81 frames (5 s)
RL prompts
VidProM Wang and Yang [2024] ( ∼ 250K after filtering)
RL prompt rewriting
Qwen2.5-7B-Instruct Qwen Team [2024]
Optimizer
Adam ( lr=2×10−6 , β1=0 , β2=0.999 )
Batch size
64
Appendix
Table 5: Training configuration. The settings above the first divider refer to the RL stage.
Metric
TraF − Self Forcing
TraF − Causal Forcing
Total
+2.92[+2.06,+3.78]
+3.50[+2.76,+4.26]
Temporal Quality
+5.04[+3.62,+6.47]
+1.13[+0.23,+2.01]
Frame Quality
+3.97[+3.07,+4.89]
+5.88[+5.01,+6.75]
Text Alignment
−0.25[−2.03,+1.59]
+3.47[+1.77,+5.26]
Appendix
Table 6: Score differences and 95% paired prompt-bootstrap confidence intervals for 30-second evaluation.
Video diffusion models generate or edit temporally coherent video by iteratively denoising latent sequences conditioned on text, images, motion, or other signals.
Autoregressive (AR) diffusion offers a promising framework for generating videos of theoretically infinite length. However, a major challenge is maintaining temporal continuity while preventing the progressive quality degradation caused by error accumulation. To ensure continuity, existing methods typically condition on highly denoised contexts; yet, this practice propagates prediction errors with high certainty, thereby exacerbating degradation. In this paper, we argue that a highly clean context is unnecessary. Drawing inspiration from bidirectional diffusion models, which denoise frames at a shared noise level while maintaining coherence, we propose that conditioning on context at the same noise level as the current block provides sufficient signal for temporal consistency while effectively mitigating error propagation. Building on this insight, we propose HiAR, a hierarchical denoising framework that reverses the conventional generation order: instead of completing each block sequentially, it performs causal generation across all blocks at every denoising step, so that each block is always conditioned on context at the same noise level. This hierarchy naturally admits pipelined parallel inference, yielding a 1.8 wall-clock speedup in our 4-step setting. We further observe that self-rollout distillation under this paradigm amplifies a low-motion shortcut inherent to the mode-seeking reverse-KL objective. To counteract this, we introduce a forward-KL regulariser in bidirectional-attention mode, which preserves motion diversity for causal inference without interfering with the distillation loss. On VBench (20s generation), HiAR achieves the best overall score and the lowest temporal drift among all compared methods.
Kai Zou, Dian Zheng, Hongbo Liu +3
University of Science and Technology of China · Anhui Province Key Laboratory of Digital Security, USTC · The Chinese University of Hong Kong +2
Autoregressive (AR) video diffusion models generate streaming video of arbitrary length by conditioning each new chunk of frames on those already generated. However, they suffer from exposure bias, where a model trained on clean context must continue from its own imperfect outputs, and errors compound until quality collapses within seconds (drift). Existing methods either require cluster-scale base retraining, apply training-free inference corrections with inconsistent results, or train frozen-base correctors for sampling and cache errors. We instead find that drift is largely deterministic and therefore learnable, as holding the noisy state fixed while varying only the history reveals a counterfactual velocity gap that is ~95% systematic across noise seeds. Based on this observation, we introduce Clean Forcing, which trains a 5.9M-parameter LoRA corrector on the frozen base model to regress this gap, with a closed-loop stage that adds DAgger-style aggregation and a drift-contraction objective to expose the corrector to its own rollouts. Because the clean histories can be the base model's own single-shot generations, training needs no real videos, and the merged corrector adds no inference cost. In 50 s text-to-video generation, Clean Forcing reduces Δ-drift from +11.3 to +1.65 without real video and to -0.16 with 40 real clips, improving on the best published result of +3.57 with over 150x less data. Extensive experiments demonstrate that Clean Forcing outperforms prior methods in aesthetic quality and data efficiency with reduced drift for long-horizon video generation.
Wenqing Wang, Joonghyuk Shin, Jonathan Tremblay +2
Northeastern University · Seoul National University · NVIDIA
Causal video generators must predict from the past, but they need not learn only from it. In streaming autoregressive video diffusion, each emitted segment becomes a commitment that future segments must preserve. Standard training, however, does not explicitly supervise causal states with future-informed representation targets. This creates what we call a representation-level planning gap: states that fit the current segment may discard identity, layout, and motion information needed for a consistent future. We introduce Video-Mirai, a training-only method that closes this gap without changing causal inference: the generator rolls out causally, a frozen foresight encoder reads the completed rollout non-causally, and a lightweight predictor distills the resulting stopped-gradient targets into causal states. Future frames supervise representations, never generator inputs. At inference, the encoder and predictor are discarded, leaving the original architecture, per-step FLOPs, and KV-cache behavior unchanged. Video-Mirai improves a strong Causal-Forcing baseline on 5-second VBench from 83.8 to 84.6 in terms of Total Score. On 30-second rollouts beyond the training horizon, subject consistency improves from 84.9 to 88.5 and background consistency from 90.2 to 91.9. Ablations identify future-conditioned targets as a key ingredient, and probes show that future frames become more decodable from current features. Causality should constrain inference, not representation supervision. Our study highlights that visual autoregressive models need foresight. Project Page: https://y0uroy.github.io/Video-Mirai.
Yonghao Yu, Lang Huang, Runyi Li +2
1The University of Tokyo · 2National Institute of Informatics · 3Peking University