Video-Mirai closes the representation-level planning gap of causal video generators by letting future segments supervise the current causal state, only at training time. At inference the foresight encoder and predictor are discarded; the deployed generator keeps its causal architecture, FLOPs, and KV-cache behavior identical to the baseline.

The trainer writes checkpoints in this layout, so once downloaded the file is loaded directly by torch.load:

See inference.py and scripts/inference.sh for the full inference entry point.

The causal generator rolls out causally under the same mask used at inference. A frozen foresight encoder (the same Wan2.1-14B used by the DMD score teacher) reads the completed rollout including future segments and produces future-aware feature targets. A lightweight predictor maps each causal hidden state to its fused target via a cosine loss. After training, the foresight encoder and predictor are discarded. The deployed generator keeps its causal architecture, FLOPs, and KV-cache behavior unchanged.

See the paper and project page for the full method, ablations, and qualitative results.

Apache 2.0. The Wan2.1 backbone weights this model depends on are released by the Wan-AI team under their own license; please review and comply with their terms when redistributing.