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.
Figures & tables
Figure 1: Clean Forcing reduces drift during AR rollout. The first 20 s of two 50 s generations with the same prompt and seed are shown. The adapted AR base (top) becomes oversaturated and loses the glass, while the same model with the merged Clean Forcing corrector (bottom) preserves the scene and pouring motion.
Figure 2: Overview. (1) At the shared noisy state zt , Δv:=vθ(zt,hclean,t)−vθ(zt,hgen,t) is the history-induced velocity gap. The corresponding one-step clean-latent predictions differ by σt∥Δv∥ , contributing to drift over successive chunks. (2) The LoRA student conditions on hgen and matches the frozen teacher’s prediction under hclean ( LCF ) ( ϕ denotes the weights of the LoRA corrector rϕ ). (3a) We generate new training pairs Dϕ with the corrector active and continue training on D0∪Dϕ , where D0 contains pairs from uncorrected rollouts. (3b) Lcon backpropagates the next chunk’s error through the undetached prediction x^0 .
Official VBench ↑
Motion
Long-Horizon Quality
Method
Subj.
Backg.
Aesth.
Imag.
Smooth
Dynamic
MUSIQ ↑
Δ -drift ↓
Matched base
Adapted base
65.7
76.7
44.9
47.9
91.1
29.7
44.8
+11.27±0.81
Context noise ( σ=0.2 )
67.2
77.8
45.2
51.4
85.8
32.0
49.1
+12.05±0.76
History guidance ( w=1.2 )
64.6
76.1
42.4
47.7
94.5
0.8
42.3
+20.55±0.65
Pathwise TTC ( {500,250} )
59.7
74.0
40.2
37.4
91.0
64.1
35.5
+13.96±0.91
Table 1: Quantitative comparison on 128 MovieGen prompts at 50 s, with ± denoting SEM. Δ -drift is first-20% minus full-video MUSIQ ( Po et al., 2025 ) . Across its variants, Clean Forcing achieves the lowest drift and highest aesthetics. It ranks second in subject and background consistency, imaging quality, and MUSIQ. Among shared-prompt rows, bold with light red marks the best value and underline with light green the second best. Published † values use different prompts, and dynamic degree only checks for motion collapse; neither is ranked (Appendix C ).
Figure 3: Per-frame quality over 50 s. Mean MUSIQ ± SEM across 128 prompts. Both Clean Forcing variants remain stable as the base and training-free baselines decline. Ripples mark chunk boundaries.
Figure 4: Qualitative comparison over 50 s. Frames are sampled at equal intervals for the spiral staircase (top, p098) and mechanical clock (bottom, p101) prompts. The base and training-free baselines develop saturation or structural artifacts, while Self Forcing shifts color. Both Clean Forcing variants retain the scene throughout the rollout.
Table 2: Ablation studies. We extend held-out clips by 48 latent frames ( ∼12 s), using paired clips and seeds across comparable rows. Bold and underlining indicate the best and second-best reported values. (Drift removed is measured relative to the unadapted base; only this metric was recorded for the DAgger-only checkpoint. Factual targets a collapse motion, making drift and MUSIQ uninformative (Appendix C ). GT-anchored targets b were evaluated on a separate dataset and increased drift, so their values are not directly comparable.)
Video diffusion models generate or edit temporally coherent video by iteratively denoising latent sequences conditioned on text, images, motion, or other signals.
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
Autoregressive video diffusion models enable real-time streaming video generation. However, errors introduced during self-rollout accumulate over long horizons, manifesting as color drift, motion stagnation, and eventual visual collapse. In this paper, we characterize this phenomenon from a frequency-domain perspective: error accumulation appears as a pronounced energy drift in the low-frequency bands. We further investigate the effectiveness of attention sink in the frequency domain, and find that it improves the video quality by alleviating the spectral energy drift to some extent, but cannot fully resolve it. Motivated by the above analysis, we propose FreqForcing, a training-free framework that addresses error accumulation in long-video generation via Spectral Self-Anchoring (SSA). The proposed SSA leverages the low-frequency components of anchor attention to maintain long-horizon visual stability, while preserving dynamic motion through the high-frequency components of local attention. Our FreqForcing extends Self-Forcing pretrained on 5s clips to two-minute generation, achieving 24x extrapolation. Extensive experiments show that FreqForcing outperforms existing training-free methods quantitatively and qualitatively while remaining competitive with representative training-based approaches.
Modern autoregressive (AR) video diffusion models excel at short-horizon video generation, yet generating long videos remains challenging due to drifting, where colors and textures shift, and motion dynamics decay. Existing works primarily rely on KV conditioning, which selects or modifies cached key-value (KV) entries to mitigate drifting. However, we observe that KV conditioning alone is insufficient as it assumes cached KV entries remain in-distribution. This assumption fails beyond the training horizon: nothing constrains the construction of KV entries during rollout, giving rise to the KV-provenance problem where cached entries themselves become out-of-distribution (OOD). To address this, we propose In-Distribution Forcing (ID-Forcing), a test-time framework that aligns both KV caching and KV conditioning with training configurations. Its key mechanism, self-caching, prevents OOD KV entries at their source. Each chunk is cached without attending to prior KV entry, keeping the rolling window exactly in-distribution. Consequently, ID-Forcing seamlessly extends short-horizon models to minute-scale video generation. Extensive evaluations show that our method remains competitive on standard video generation benchmark while substantially outperforming prior work in mitigating drifting, as validated by both our drift metrics and a user study.
Jeongwoo Shin, Youngyoon Choi, Sangwoo Jo +5
Seoul National University · Korea University · Sungkyunkwan University +1