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

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

Instead of using only the recent window, it searches past generated latents at each step as memory.

They also proposed Window Temporal Delta Loss to make retrieval embeddings more sensitive to meaningful temporal change.

Across multiple backbones and lengths, the method improved long-video quality and achieved the best average VBench-Long rank.

The approach helps reduce error accumulation and identity drift with little extra cost.