Long video generation should keep the subject and background consistent while supporting rich, natural motion. However, autoregressive models are often distilled with distribution matching distillation (DMD) over short windows, which does not train them to keep this consistency over long ranges. As a result, the subject and background can drift over time. Methods that manage the key-value (KV) cache retain or retrieve history to reduce drift. Yet this history may itself have drifted. We propose Reference Forcing, which uses the initial frames as an appearance reference. We first build a teacher that uses attention masks to force each region of a new frame to attend to the same region in the initial frames. The teacher finds these regions with off-the-shelf segmentation models. We then distill this behavior into a separate attention path of the student, which needs no external models at inference. We further introduce value correction, which corrects drifted values in the KV cache toward the matching values of drift-free initial frames. Experiments show that Reference Forcing achieves state-of-the-art subject and background consistency among long-video methods on different base models, and its advantage grows with video length. Ablations show both components are needed to gain consistency without losing motion. More visualizations and code are available at https://zhaosheng-thu.github.io/reference-force/
Figures & tables
Figure 1: Examples of two-hour videos generated by Reference Forcing. Each row shows frames at 0, 30, 60, 90, and 120 minutes and, on the right, the consistency of the subject and the background, measured by their semantic similarity to the first frame.
Figure 2: Overview of Reference Forcing. (a) Each chunk Xn is generated from the first chunk X0 , the recent window, and noisy tokens, with the anchor path of (b) and the value correction of (c). (b) In every layer, the student adds an anchor path (LoRA on the query and keys, and the anchor gate w ) to the frozen native attention and learns to match the prediction of a teacher whose anchor path is masked by the subject mask and the background flow. (c) Value correction moves the values of the drifted chunk Xn−2 toward their key-matched values in X0 , with the anchor gate w learned in (b) as strength.
Length
Method
Consistency
Motion
Frame quality
Overall
Subject ↑
Background ↑
Smoothness ↑
Dynamic ↑
Aesthetic ↑
Imaging ↑
Rank
30 s
LongLive [ 48 ]
97.47
96.56
98.68
42.81
60.98
68.67
2
+ MemRoPE [ 17 ]
97.34
96.54
98.70
44.69
60.90
69.25
2
+ LongLive-RAG [ 10 ]
97.27
96.40
98.65
48.59
60.78
69.27
4
+ ∞ -RoPE [ 50 ]
97.11
96.42
98.68
47.97
60.44
68.55
5
+ Deep Forcing [ 51 ]
97.13
96.33
98.35
41.88
60.92
68.89
6
Table 1: Comparison on MovieGenBench. VBench-Long scores of 30-, 60-, and 120-second videos; all methods use the same KV cache capacity (12 latent frames) and start from the same noise. Rank: average rank over the six columns. Bold : best; underline : second.
Figure 3: Qualitative comparison. All methods share the prompt and first chunk; boxes mark subject changes, magnified in the zoom column. Left: the dark visor of the astronaut turns transparent in baselines but stays dark in Reference Forcing. Right: in baselines, the face of the woman becomes lighter and the strap on her left shoulder wider and brighter, whereas Reference Forcing keeps both.
VLM judge
User study
Reference Forcing vs.
Overall
Consistency
Motion
Quality
Overall
Consistency
Motion
Quality
LongLive
78 [67, 86]
86 [77, 93]
78 [67, 86]
63 [51, 73]
72 [64, 80]
76 [68, 83]
72 [64, 80]
73 [64, 81]
MemRoPE [ 17 ]
72 [61, 82]
76 [65, 85]
69 [57, 79]
60 [48, 71]
75 [67, 83]
79 [70, 85]
73 [64, 81]
72 [64, 80]
LongLive-RAG [ 10 ]
58 [46, 68]
69 [57, 79]
64 [52, 74]
44 [33, 55]
74 [65, 81]
79 [70, 85]
71 [63, 79]
75 [66, 82]
∞ -RoPE [ 50 ]
84 [74, 91]
85 [75, 92]
73 [61, 82]
73 [61, 82]
79 [71, 86]
83 [75, 89]
75 [66, 82]
76 [68, 83]
Deep Forcing [ 51 ]
78 [67, 86]
39 [28, 50]
99 [93, 100]
86 [77, 93]
90 [84, 95]
87 [79, 92]
90 [83, 94]
90 [83, 94]
Table 2: Pairwise evaluation by VLM judge and by user study. Percentage of comparisons in which Reference Forcing is preferred over each method on MovieGen prompts, with 95% confidence intervals. Bold marks values whose interval lies above 50. Both prefer Reference Forcing in consistency and motion against most methods.
Figure 4: Similarity to the first frame. DINOv2 similarity of the subject and the background to the first frame, averaged over MovieGen prompts, on LongLive (120 seconds) and on Self-Forcing (30 seconds). Base: LongLive and Self-Forcing without a long-video method. Arrows: the gap between Reference Forcing and Base at the last time.
Figure 7
Figure 7: Anchor gate w on generated videos. Left: frames and w (orange) at 20, 40, and 60 seconds for a changing background (BG), appearing new subjects (New), and a changing pose (Pose). Right: distribution of w for subject tokens (Subj.), background tokens also visible in the first chunk (Bg.), and new content (New). w stays high on the subject and low on new content.
Figure 8: Gain of the anchor path. Top left: change of each metric relative to the default gain of 1. Top right: consistency (mean of subject, background, and L-Consistency) against motion (mean of dynamic degree and tracking strength), each normalized over the seven gains; the upper right is better. Bottom: the same prompt and initial noise with different gains. A higher gain gives more consistency and less motion: a low gain loses the church and the octopus, a high gain nearly freezes the scene, and the default keeps both moving.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Features
Precision (%)
Coverage (%)
External model
Random
0.6
–
no
DINOv2 [ 30 ]
86.5
82.5
yes
Model keys (ours)
86.2
78.9
no
Appendix
Table 3: Matching accuracy of value correction. Matches between recent chunks and the first chunk in 64 videos generated by Reference Forcing, compared with optical flow from RAFT [ 41 ] with a forward-backward check. Precision: share of matches within one token of the token that the flow points to. Coverage: share of the tokens visible in the first chunk under the flow that are matched.
Figure 9: More qualitative comparisons (1/3). Layout as in Figure 3 . The massive cloud figure turns into a thin figure in MemRoPE, LongLive-RAG, and ∞ -RoPE and is blurred in Deep Forcing. Reference Forcing keeps its shape, color, and details from the first frame while it still moves.
Figure 10: More qualitative comparisons (2/3). Layout as in Figure 3 . The colorful street is replaced by a yellow wall in MemRoPE, LongLive-RAG, and ∞ -RoPE. Reference Forcing keeps the street of the first frame while the camera still moves.
Figure 11: More qualitative comparisons (3/3). Layout as in Figure 3 . The small tail of the creature grows several times larger with dark stripes in MemRoPE, LongLive-RAG, and ∞ -RoPE, and a white belly appears. Reference Forcing keeps the shape, color, and details of the first frame while the creature still moves.
Figure 12: Videos of different lengths generated by Reference Forcing. Each row shows seven frames evenly spaced over the video. From top to bottom: a 12-hour video and five 2-hour videos, each generated in one pass without restarts or stitching.
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
Long video generation is a critical step toward building realistic world models, requiring both high visual fidelity and long-range interaction consistency. Recent autoregressive diffusion models enable long-horizon generation through KV cache reuse, yet suffer from two fundamental challenges: failure to preserve long-range interactions due to sliding-window KV cache and error accumulation that progressively degrades generation quality over time. To address these issues, we propose BIFE, a framework that introduces a semantic sparse KV cache for retrieval-based long-range conditioning and a Block Forcing training strategy to enforce cross-block consistency. Together, these designs preserve historical interactions while mitigating drift, enabling stable and coherent minute-long video generation. We also introduce InterVBench, a minute-long video benchmark with fine-grained block-level annotations and Video Drift Error metrics. Extensive experiments on InterVBench and VBench-Long demonstrate that BIFE achieves state-of-the-art performance, including a 22.2% improvement on VDE-Subject and a 19.4% improvement on VDE-Clarity over baselines. Website: https://alibaba-damo-academy.github.io/BIFE. Code: https://github.com/alibaba-damo-academy/BIFE.
Zeyu Zhang, Jinyuan Mao, Shuning Chang +5
DAMO Academy, Alibaba Group · Zhejiang University · Hupan Lab
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