LongLive-RAG: A General Retrieval-Augmented Framework for Long Video Generation

Researchers introduced LongLive-RAG, a retrieval-augmented framework for long video generation.

Instead of only using the latest sliding window, it retrieves relevant historical latents as memory.

A query-based retrieval step and Window Temporal Delta Loss help reduce accumulated errors and identity drift.

Experiments across multiple backbones show improved long-video quality and top average VBench-Long performance.