Cinecourier
News

MemCrafter paper reports 0.908 memorability score and 85.00% human hit ratio

An arXiv paper presents MemCrafter, an agentic prompt-refinement method for memorable text-to-video, and reports gains over Seedance-2.5 and other baselines.

The authors of an arXiv paper present MemCrafter, a method that refines text-to-video prompts so the resulting clips are easier to remember, and report a memorability score of 0.908 (arXiv, arXiv). In their user study, the authors report that MemCrafter reached an average hit ratio of 85.00%, compared with 77.14% for Seedance-2.5. All figures come from the authors’ own experiments, so they show what the method does under the paper’s setup rather than an independent test.

What MemCrafter does

According to the paper, MemCrafter is a brain-informed approach that raises video memorability through iterative, agentic refinement of the prompt. The method works on the text side: it changes the prompt rather than the video model itself. The authors say the signal guiding the refinement achieves a Spearman correlation of ρ=0.637 on the test set.

Baseline comparison in Table 1

Published resultsTable 1 of the paper: Baseline comparisonRows 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.Data read November 18, 2026Source [2]How this format worksDownload the data behind it
Method MS ↑ SMN ↑ VQ ↑ SA ↑ Source page
LTX-2.5 0.759±0.011 0.847±0.005 0.808±0.008 0.203±0.002 https://arxiv.org/html/2611.15107v1
Rewriter 0.794±0.004 0.849±0.002 0.791±0.007 0.259±0.001 https://arxiv.org/html/2611.15107v1
SCMAPR 0.773±0.003 0.853±0.005 0.792±0.007 0.244±0.004 https://arxiv.org/html/2611.15107v1
RAPO 0.773±0.005 0.857±0.002 0.809±0.005 0.257±0.002 https://arxiv.org/html/2611.15107v1
VPO 0.790±0.001 0.851±0.001 0.812±0.006 0.244±0.002 https://arxiv.org/html/2611.15107v1
Seedance-2.5 0.848±0.003 0.859±0.005 0.815±0.005 0.252±0.004 https://arxiv.org/html/2611.15107v1
MemCrafter 0.908±0.003 0.871±0.004 0.829±0.005 0.261±0.006 https://arxiv.org/html/2611.15107v1

Table 1 reports results as mean ± standard deviation, and the authors say MemCrafter achieves the best performance on all four evaluation metrics. On memorability (MS), which comes from the authors’ own neuro-memorability signal m(V) that the method optimizes, MemCrafter scores 0.908±0.003, while Seedance-2.5 scores 0.848±0.003 and LTX-2.5 scores 0.759±0.011. On the independent SMN score, MemCrafter scores 0.871±0.004 against 0.859±0.005 for Seedance-2.5. For video quality (VQ), MemCrafter scores 0.829±0.005 and Seedance-2.5 scores 0.815±0.005. For semantic alignment (SA), MemCrafter scores 0.261±0.006 and Rewriter scores 0.259±0.001. The memorability gap over Seedance-2.5 is large next to the spread in the table, while the SA lead over Rewriter is smaller than the standard deviation reported for MemCrafter. The SMN and VQ margins over Seedance-2.5 are modest, which fits the authors’ reading that memorability gains are not confined to the optimization objective, though it does not prove it.

Quality did not drop

The authors note that video quality is not an explicit part of their optimization objective, yet it improves during optimization. They suggest that memorability-oriented refinement can also benefit perceptual quality. In the table, the VQ of the VPO baseline is 0.812±0.006, below MemCrafter’s 0.829±0.005, so the quality claim holds against that row at least.

Human study and code

For the human check, the authors ran a video recognition task, and MemCrafter reached an average hit ratio of 85.00% against 77.14% for Seedance-2.5. That result matters because the memorability score is a model-based estimate, and people recognising the clips gives a separate reading. The authors say they will release their code to support reproducibility and future research.

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 a price or hardware requirements.

How we built this

Published results

Table 1 of the paper: Baseline comparison

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.

arXiv

Sources

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

  1. arXivFrom Seeing to Remembering: Brain-Informed Agentic Text-to-Video Generation for Enhancing Memorabilityarxiv.orgPrimary source · accessed November 18, 2026
  2. arXivFrom Seeing to Remembering: Brain-Informed Agentic Text-to-Video Generation for Enhancing Memorabilityarxiv.orgPrimary source · accessed November 18, 2026
  • text to video
  • memorability
  • prompt refinement
  • arxiv
  • benchmarks