Prompt-Adapter Context Routing for Parameter-Efficient Multi-Shot Long Video Extrapolation

2026-07-08 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, medium

Summary

PACR-Video is a parameter-efficient framework for multi-shot long video extrapolation that maintains coherence across extended sequences. It operates by keeping a text-to-video diffusion transformer frozen, augmenting it with low-rank temporal adapters steered by learned shot-role prompt tokens. Instead of fine-tuning the entire generator or using large external memories, PACR-Video employs a recursive prompt bank to store compact entity, location, action, and style prompts from prior shots. These prompts are then routed through adapter gates based on predicted narrative dependencies. A Shot-Local/Story-Global tuning objective, combining next-shot reconstruction, cross-shot identity contrast, and prompt sparsity regularization, helps reduce long-horizon drift. The framework also uses an adapter composition schedule to balance early-shot visual consistency with later-shot event progression. Evaluated on benchmarks like FlintstonesSV, Pororo-SV, and MovieNet, PACR-Video significantly outperforms various baselines, including VideoCrafter2 and ReCA, in quality, alignment, and consistency.

Key takeaway

For Machine Learning Engineers developing long-form video generation systems, PACR-Video offers a compelling alternative to full model fine-tuning. You should consider its parameter-efficient approach. It uses lightweight temporal adapters and a recursive prompt bank, maintaining narrative coherence and identity consistency across multiple shots. This method reduces computational overhead and improves video quality and semantic alignment, enabling more effective scaling of long video projects.

Key insights

PACR-Video achieves coherent long video extrapolation by routing compact prompts through lightweight adapters in a frozen diffusion model.

Principles

• Freezing large models with small adapters is efficient.
• Recursive prompt banks maintain long-horizon coherence.
• Gated routing enables selective context reuse.

Method

PACR-Video freezes a text-to-video diffusion transformer, adding low-rank temporal adapters. It uses a recursive prompt bank for entity, location, action, and style prompts, routing them via adapter gates based on narrative dependencies.

In practice

• Use low-rank adapters for parameter efficiency.
• Implement prompt banks for long-term memory.
• Apply gated routing for context selection.

Topics

• Video Extrapolation
• Multi-Shot Video Generation
• Parameter-Efficient Learning
• Diffusion Models
• Prompt Engineering
• Temporal Adapters
• Context Routing

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Prompt-Adapter Context Routing for Parameter-Efficient Multi-Shot Long Video Extrapolation — Compact prompt routing and lightweight temporal adaptation enable stable long video extrapolation.
• Veda: Scalable Video Diffusion via Distilled Sparse Attention — Generation quality in sparse attention depends on mask alignment with full attention's tile-wise geometry, not just sparsity ratio.
• Mode Seeking meets Mean Seeking for Fast Long Video Generation — Decoupling local fidelity and long-term coherence enables minute-scale video generation from limited long-form data.
• TetherCache: Stabilizing Autoregressive Long-Form Video Generation with Gated Recall and Trusted Alignment — Stabilizing long-form autoregressive video generation requires selective memory recall and statistical alignment of cached features.
• MilliVid: Hierarchical Latents for Long-Range Consistency in Video Generation — Hierarchical latent token compression and coarse-to-fine rollout enable long-range consistency in video generation by managing detail levels efficiently.
• MobileWan: Closing the Quality Gap for Mobile Video Diffusion — Server-scale 5B-parameter video diffusion models can run on mobile via recurrent reformulation and structured compression.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.