How to stop video generation from drifting off-track?

Long video generation with autoregressive diffusion models suffers from error accumulation—mistakes in early frames degrade everything that follows. LongLive-RAG treats previously generated frames as a searchable memory bank: instead of conditioning only on the recent sliding window, a lightweight retrieval step finds relevant historical frames to ground generation. A new training loss encourages the retrieval embeddings to capture meaningful temporal changes rather than local redundancy. Results across multiple models show improved video quality and ranking on VBench-Long. Code released.