EverAnimate: Minute-Scale Human Animation via Latent Flow Restoration

Researchers introduced EverAnimate, a post-training method for long-horizon human video generation.

It keeps a persistent latent context across chunks and uses a restorative sampling objective to reduce drift and preserve fidelity.

The approach improves character identity and video quality over prior methods on both short and long sequences, including clips up to 90 seconds.

EverAnimate achieves these gains with lightweight fine-tuning via LoRA, tackling a major failure mode in long-form AI video.