WorldKV: Efficient World Memory with World Retrieval and Compression

Researchers introduced WorldKV, a training-free framework for video diffusion models that improves long-horizon memory.

It stores evicted KV-cache chunks, retrieves the most relevant ones, and compresses redundant tokens to preserve scene consistency.

The method enables persistent world generation while delivering about 2x the throughput of full-KV inference.

WorldKV targets the key tradeoff between memory fidelity and speed, without requiring fine-tuning.