Unverified50% confidenceFactExact time

光流计算连续两帧之间每个像素的位移向量形成稠密运动场，经典算法包括Lucas-Kanade和Horn-Schunck，RAFT等深度学习方法提升了精度

1

Sources

50%

Confidence

Long-term

Relevance

8/27/2026

First Seen

Sources

AI Video Generation Explained: Diffusion Models vs. Motion Transfer

Related Entities

光流concept Lucas-Kanadeconcept Horn-Schunckconcept RAFTconcept

Related Claims

Unverified高清放大模块的二次推理过程会额外消耗约1-2G显存68% similar Unverified矮星采用不对称量化方案：专家层压至2比特，注意力层保留8比特，路由层保留原始精度68% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/810069

API

curl https://kongchang.com/api/v1/knowledge/claims/810069

MCP

get_claim(id=810069)