Cinecourier
Analysis

Reference Forcing, per its authors, cuts drift in long AI video

Reference Forcing reuses a clip's first frames to limit drift in long video; the authors report top consistency scores on MovieGenBench.

The authors of Reference Forcing, a method for long autoregressive video generation, report the highest subject and background consistency on LongLive, although on the Self-Forcing base at 30 s LongLive-RAG has higher subject consistency (97.13 against 96.99) (arXiv, arXiv). According to the paper, on LongLive at 60 s it scores 97.56 for subject consistency and 96.67 for background consistency. The authors also say its advantage grows with video length.

How the method works

Reference Forcing treats the initial frames of a video as an appearance reference. In training, a teacher with attention masks makes every region of a new frame look at the matching region in those initial frames, and off-the-shelf segmentation models locate the matching regions. The teacher’s behavior is then distilled into a separate attention path of the student, so no external models are needed at inference.

The paper adds a second piece, value correction, a training-free method that pulls stale values in the KV cache back toward the corresponding values of the initial frames, which have not drifted. The implementation builds on LongLive-1.3B with all of its weights frozen, and only the anchor path is trained, for 20,000 steps.

Results on MovieGenBench

Published resultsTable 1 of the paper: Comparison on MovieGenBenchRows of Table 1 of the paper (arXiv HTML version), copied cell for cell by our code; a two-row header is shown as "group: column"; an empty name cell repeats the name above it. These are the authors' own figures, not an independent test. The table has 9 columns; 6 shown, with every column that holds a figure this article gives or a measure it names. Not shown: Frame quality (2 columns), Overall.Data read November 18, 2026Source [2]How this format worksDownload the data behind it
Length Method Consistency: Subject ↑ Consistency: Background ↑ Motion: Smoothness ↑ Motion: Dynamic ↑ Source page
30 s LongLive 97.47 96.56 98.68 42.81 https://arxiv.org/html/2611.21559v1
30 s + MemRoPE 97.34 96.54 98.70 44.69 https://arxiv.org/html/2611.21559v1
30 s + LongLive-RAG 97.27 96.40 98.65 48.59 https://arxiv.org/html/2611.21559v1
30 s + ∞ -RoPE 97.11 96.42 98.68 47.97 https://arxiv.org/html/2611.21559v1
30 s + Deep Forcing 97.13 96.33 98.35 41.88 https://arxiv.org/html/2611.21559v1
30 s + Reference Forcing 97.55 96.70 98.73 41.88 https://arxiv.org/html/2611.21559v1
60 s LongLive 97.33 96.49 98.67 42.19 https://arxiv.org/html/2611.21559v1
60 s + MemRoPE 97.32 96.50 98.59 46.80 https://arxiv.org/html/2611.21559v1
60 s + LongLive-RAG 97.31 96.38 98.61 49.53 https://arxiv.org/html/2611.21559v1
60 s + ∞ -RoPE 97.03 96.36 98.54 51.09 https://arxiv.org/html/2611.21559v1
60 s + Deep Forcing 97.31 96.48 98.24 42.89 https://arxiv.org/html/2611.21559v1
60 s + Reference Forcing 97.56 96.67 98.68 43.13 https://arxiv.org/html/2611.21559v1
120 s LongLive 97.06 96.30 98.59 46.45 https://arxiv.org/html/2611.21559v1
120 s + MemRoPE 96.92 96.34 98.54 48.12 https://arxiv.org/html/2611.21559v1
120 s + LongLive-RAG 97.09 96.25 98.57 49.92 https://arxiv.org/html/2611.21559v1
120 s + ∞ -RoPE 96.72 96.20 98.52 52.73 https://arxiv.org/html/2611.21559v1
120 s + Deep Forcing 96.93 96.18 98.49 50.94 https://arxiv.org/html/2611.21559v1
120 s + Reference Forcing 97.27 96.56 98.61 44.69 https://arxiv.org/html/2611.21559v1
30 s Self-Forcing 96.57 95.89 97.96 40.86 https://arxiv.org/html/2611.21559v1
30 s + MemRoPE 96.94 96.23 97.91 61.95 https://arxiv.org/html/2611.21559v1
30 s + LongLive-RAG 97.13 96.28 98.01 60.16 https://arxiv.org/html/2611.21559v1
30 s + ∞ -RoPE 96.84 96.20 97.94 64.61 https://arxiv.org/html/2611.21559v1
30 s + Deep Forcing 96.96 96.15 97.86 54.77 https://arxiv.org/html/2611.21559v1
30 s + Reference Forcing 96.99 96.30 98.04 65.08 https://arxiv.org/html/2611.21559v1

The paper’s main comparison table scores methods on MovieGenBench using VBench-Long at 30 s, 60 s and 120 s. All methods share the same KV cache capacity of 12 latent frames and start from the same noise. The authors say that at 30 s and 60 s their method has the highest subject consistency, background consistency and motion smoothness, while its dynamic degree stays on par with LongLive.

The 30 s block built on LongLive shows how narrow the margins are: Reference Forcing reaches 97.55 subject consistency against 97.47 for LongLive, and 96.70 background consistency against 96.56. On motion smoothness its 98.73 edges out MemRoPE at 98.70, a gap in the second decimal place. The dynamic column tells a different story: Reference Forcing scores 41.88, the same as Deep Forcing and below LongLive at 42.81 and LongLive-RAG at 48.59. Higher consistency can come with less motion, and the table suggests that is partly the case here, though the authors’ own text calls the difference in dynamic degree on par with LongLive.

These are the authors’ own figures, not an independent test. In the paper’s qualitative comparison, LongLive, MemRoPE, LongLive-RAG and ∞-RoPE make an astronaut’s dark visor transparent and lighten a woman’s face, while Deep Forcing shifts the entire frame to a blurry purple. Reference Forcing keeps the visor and the skin tone as they look in the first frame, according to the authors.

Ablation and longer runs

In an ablation on LongLive, the authors add the anchor path and value correction one at a time. Together they give an L-Consistency gain of +0.097, against +0.044 and +0.034 for each alone, while keeping the motion of LongLive. The two components therefore appear to be complementary rather than redundant, as the paper reads it.

At 120 s, Reference Forcing on LongLive records 97.27 subject consistency and 96.56 background consistency. The project page also shows videos that, according to it, are generated in one pass without restarts or stitching (zhaosheng-thu.github.io).

Where it falls short

The authors list several limits, including that only the first chunk serves as reference, so content that first appears later, such as an object entering the scene, has no reference, and updating the reference during generation is left for future work. The anchor path averages over the matched region of the first chunk, so fine textures can come out a little smoother. Scenes with multiple similar subjects stay hard, since the teacher’s segmentation tracks only one subject.

The teacher also depends on a detector to locate the subject in the first chunk. When the detector finds nothing, the teacher treats the full frame as background and matches it with optical flow. That mostly occurs in scenes lacking a main subject, such as landscapes or crowds. The paper notes that when the detector finds no subject, which mostly happens in scenes such as landscapes or crowds, the teacher treats the whole frame as background and matches it by optical flow.

What the sources we read do not state

We check six facts on every story in the full text of each source. The pages listed under Sources do not state hardware requirements.

How we built this

Published results

Table 1 of the paper: Comparison on MovieGenBench

Rows of Table 1 of the paper (arXiv HTML version), copied cell for cell by our code; a two-row header is shown as "group: column"; an empty name cell repeats the name above it. These are the authors' own figures, not an independent test. The table has 9 columns; 6 shown, with every column that holds a figure this article gives or a measure it names. Not shown: Frame quality (2 columns), Overall.

arXiv

Sources

The sources this article was written from. Links open the original page.

  1. arXivReference Forcing: Long Video Generation with Region-Guided Drift Correctionarxiv.orgPrimary source · accessed November 18, 2026
  2. arXivReference Forcing: Long Video Generation with Region-Guided Drift Correctionarxiv.orgPrimary source · accessed November 18, 2026
  3. zhaosheng-thu.github.iohttps://zhaosheng-thu.github.io/reference-force/: page linked by the paper as its resourcezhaosheng-thu.github.ioPrimary source · accessed November 18, 2026
  • long video
  • autoregressive video
  • longlive
  • vbench
  • research