Unverified50% confidenceFactExact time
LoRA 由微软提出，其核心思想是将权重更新量 ΔW 分解为两个低秩矩阵的乘积 BA
1
Sources
50%
Confidence
Long-term
Relevance
7/20/2026
First Seen
Sources
Related Claims
UnverifiedLoRA的数学原理是将权重更新矩阵分解为两个低秩矩阵的乘积，文件体积通常仅几十至几百MB71% similarUnverifiedLoRA在前向传播时原始权重W₀保持冻结，输出变为h = W₀x + (BA)x/√r，对于SD 1.5模型在rank=4时整个LoRA文件约只有3-6MB66% similarUnverifiedLoRA（Low-Rank Adaptation）冻结预训练模型全量权重，仅向特定层注入两个低秩矩阵，文件体积通常只有几 MB 到数十 MB59% similarVerifiedLoRA（Low-Rank Adaptation）文件体积通常只有几十MB，远小于完整模型的数GB大小55% similarUnverifiedLoRA等参数高效微调方法将显存需求压缩至普通消费级GPU可承受的范围55% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/566555API
curl https://kongchang.com/api/v1/knowledge/claims/566555MCP
get_claim(id=566555)