AI video, reported from the sources

Analysis

ScoutCache, a training-free method, reports up to 2.28× speedup on Self-Forcing

Authors say ScoutCache speeds few-step autoregressive video diffusion up to 2.28× on Self-Forcing and 1.86× on LongLive without changing weights.

Cover: ScoutCache speeds up video diffusion without training. 2.28× top speedup on Self-Forcing, authors' own experiments. ScoutCache, Self-Forcing, LongLive, FlowCache, MotionCache (Slow)

The authors of ScoutCache, a paper on arXiv, say their training-free method speeds up few-step autoregressive video diffusion by up to 2.28× on Self-Forcing and 1.86× on LongLive, while improving generation quality (arXiv). According to the paper, this is done without modifying pretrained model parameters, so existing checkpoints can be accelerated as they are. All figures below come from the authors’ own experiments, not from an independent test.

What the method does

ScoutCache is described as a framework that jointly optimizes denoising, historical KV reading, and cache construction. In the paper’s setup, the historical cache holds four frame regions: the first latent frame of the video as a fixed anchor and the three frames from the preceding chunk (arXiv). Among the baselines, FlowCache and Forcing-KV already compress or prune historical KV caches, while ScoutCache combines denoising, historical KV reading and cache construction in one framework.

The authors evaluate it on the publicly released 1.3B checkpoints of Self-Forcing and LongLive. One component, Conditional Denoising Readout (CDR), is measured on its own: the paper says it cuts the normalized update error from 1.000 with no update to 0.596.

Speedups by video length

On Self-Forcing, the paper reports a 1.60× speedup on VBench and 2.19×, 2.25×, and 2.28× on MovieGen at 30, 60, and 120 seconds respectively. The headline 2.28× therefore corresponds to the longest MovieGen setting, at 120 seconds. The paper attributes the speedup’s rise with video duration to the shrinking relative overhead of the initial calibration.

The 1.86× figure belongs to LongLive, a separate backbone. Practitioners should read the two maxima as results for different models, not as one number for the whole method.

Quality scores in Table 6

Published resultsTable 6 of the paper: Detailed video quality scores on Self-ForcingRows of Table 6 of the paper (arXiv HTML version), copied cell for cell by our code. These are the authors' own figures, not an independent test. The first column is the block of the table the row sits in. The table has 11 columns; 5 shown, with every column that holds a figure this article gives or a measure it names. Not shown: Tea Cache, MotionCache (Fast) and 4 more. Rows: 10 of 24, with every row whose value this article gives, or that it names, and rows of MovieGen (120s), the condition of its headline result. Of its 4 blocks, not shown: MovieGen (30s) (7 rows), MovieGen (60s) (7 rows).Data read November 18, 2026Source [2]How this format worksDownload the data behind it
Table block Metric Self- Forcing Flow Cache MotionCache (Slow) ScoutCache Source page
VBench (5s) Quality Score 84.06 84.04 72.96 84.38 https://arxiv.org/html/2611.21038
VBench (5s) Semantic Score 71.36 71.64 52.75 71.28 https://arxiv.org/html/2611.21038
VBench (5s) Total Score 81.52 81.56 68.92 81.76 https://arxiv.org/html/2611.21038
MovieGen (120s) Subject Consistency 96.54 96.35 94.00 97.29 https://arxiv.org/html/2611.21038
MovieGen (120s) Background Consistency 96.66 96.34 95.28 95.43 https://arxiv.org/html/2611.21038
MovieGen (120s) Motion Smoothness 97.99 98.14 95.36 98.09 https://arxiv.org/html/2611.21038
MovieGen (120s) Dynamic Degree 28.24 28.68 17.57 49.14 https://arxiv.org/html/2611.21038
MovieGen (120s) Aesthetic Quality 51.27 51.14 45.66 56.77 https://arxiv.org/html/2611.21038
MovieGen (120s) Imaging Quality 61.85 59.64 51.04 68.82 https://arxiv.org/html/2611.21038
MovieGen (120s) Total Score 75.03 74.62 68.56 79.11 https://arxiv.org/html/2611.21038

Table 6 of the paper gives detailed quality scores on Self-Forcing, with VBench Quality, Semantic and Total scores for the 5-second block and six VBench-Long dimensions plus a Total Score for longer videos. In the VBench (5s) block, the Total Score is 81.52 for Self-Forcing, 81.56 for Flow Cache and 68.92 for MotionCache (Slow). Flow Cache essentially matches the unaccelerated Self-Forcing baseline on this measure, while MotionCache (Slow) loses a large share of quality.

The VBench (5s) Semantic Score is 71.36 for Self-Forcing, 71.64 for Flow Cache and 52.75 for MotionCache (Slow). The Semantic Score follows a similar pattern: Flow Cache is at 71.64 against Self-Forcing’s 71.36, while MotionCache (Slow) is far lower at 52.75.

How to read the claims

The paper states that ScoutCache consistently improves generation quality while accelerating generation. That is a claim by the authors, evaluated on VBench and MovieGen with two 1.3B backbones. Because the method needs no training, a practitioner running either checkpoint could in principle test it without retraining, and the speed figures above give the range worth checking against their own clips. The baselines in Table 6 show why quality must be reported next to speed: the gap between Self-Forcing and MotionCache (Slow) is far larger than the gap between Self-Forcing and Flow Cache.

How we built this

Published results

Table 6 of the paper: Detailed video quality scores on Self-Forcing

Rows of Table 6 of the paper (arXiv HTML version), copied cell for cell by our code. These are the authors' own figures, not an independent test. The first column is the block of the table the row sits in. The table has 11 columns; 5 shown, with every column that holds a figure this article gives or a measure it names. Not shown: Tea Cache, MotionCache (Fast) and 4 more. Rows: 10 of 24, with every row whose value this article gives, or that it names, and rows of MovieGen (120s), the condition of its headline result. Of its 4 blocks, not shown: MovieGen (30s) (7 rows), MovieGen (60s) (7 rows).

arXiv

Sources

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

  1. arXivScoutCache: Efficient Few-Step Autoregressive Video Diffusion via Conditional Readout and Selective Cachingarxiv.orgPrimary source · accessed November 18, 2026
  2. arXivScoutCache: Efficient Few-Step Autoregressive Video Diffusion via Conditional Readout and Selective Cachingarxiv.orgPrimary source · accessed November 18, 2026
  • video diffusion
  • autoregressive
  • kv cache
  • self forcing
  • longlive
  • vbench