Prompt-Adapter Context Routing for Parameter-Efficient Multi-Shot Long Video Extrapolation
Anna Córdoba, Adam Puente Tercero, Nerea Angulo Hijo, Mar Linares Tercero, Julia Barrientos, Ainhoa Miranda, Jesús Olivera
Read on arXiv →
Key claim
PACR-Video enables stable long video extrapolation efficiently.
In plain English
Long video extrapolation often struggles with maintaining coherence and quality over extended durations. Current methods typically require extensive fine-tuning of generators, which can be resource-intensive and inefficient. PACR-Video changes this by using a frozen text-to-video diffusion transformer augmented with low-rank adapters and a prompt bank to ensure consistency and quality across shots. Builders might care because this approach allows for efficient video generation while preserving important visual and narrative elements.
Introduces a novel parameter-efficient framework for long video extrapolation.
Demonstrates solid performance across multiple benchmarks with relevant comparisons.
Deep reliability assessment
The methodology supports parameter-efficient long video extrapolation by using lightweight temporal adapters and a recursive prompt bank, but the claims of outperforming all baselines on multiple metrics may be overclaimed without detailed comparative analysis.
Reproducibility
No open source code or dataset is mentioned in the provided text.
Key figure
Figure 1 illustrates the PACR-Video method, which extrapolates long multi-shot videos by summarizing prior shots into compact prompts and steering a frozen generator through lightweight adapters.