本文详细介绍了LoRA的基本原理，它是通过调整Transformer注意力权重的可学习模型。着重讲解了LoRA的训练流程，以及模型文件内容，包括lora_down.weight、lora_up.weight和alpha等权重组成部分。最后，文章展示了LoRAkey的可视化实例，帮助理解每个权重的关键部分。
介绍
相信在SD的生图过程中，我们对LoRA都不陌生，但是它的模型文件是什么样子的，保存的东西是什么，如何起作用的，接下来将详细探讨一下。
一、LoRA基本原理
LoRA(Learnable Re-Weighting),是一种重加权模型。LORA模型将神经网络中的每一层看做是一个可加权的特征提取器，每一层的权重决定了它对模型输出的影响。通过对已有的SD模型的部分权重进行调整，从而实现对生图效果的改善。（大部分LoRA模型是对Transformer中的注意力权重进行了调整）
LoRA从技术角度来讲很简单，基本流程如下图：
蓝色部分表示原来的预训练权重，橙色部分则是lora需要训练的权重A和B。
二、LoRA的训练与推理
训练阶段：上图中的A和B是可训练的权重，在训练阶段，蓝色部分冻结，训练A和B，最后保存权重也仅保存A和B的相关参数。基本步骤如下：
推理阶段：正常使用W = W0+BA来更新模型权重。
二、LoRA文件内容
lora模型中每层的权重包含3个部分，分别为lora_down.weight 、 lora_up.weight 和 alpha。其中down和up分别为lora模型的上下层权重分别对应了B和A权重，alpha也是一个可学习的参数。lora模型每层的权重可表示为：
w
=
a
l
p
h
a
∗
d
o
w
n
∗
u
p
w = alpha * down * up
w=alpha∗down∗up
三、LoRA key可视化
以目前最流行的LCM lora为例进行一下可视化，key如下（部分）：
lora_unet_down_blocks_0_downsamplers_0_conv.alpha
lora_unet_down_blocks_0_downsamplers_0_conv.lora_down.weight
lora_unet_down_blocks_0_downsamplers_0_conv.lora_up.weight
lora_unet_down_blocks_0_resnets_0_conv1.alpha
lora_unet_down_blocks_0_resnets_0_conv1.lora_down.weight
lora_unet_down_blocks_0_resnets_0_conv1.lora_up.weight
lora_unet_down_blocks_0_resnets_0_conv2.alpha
lora_unet_down_blocks_0_resnets_0_conv2.lora_down.weight
lora_unet_down_blocks_0_resnets_0_conv2.lora_up.weight
lora_unet_down_blocks_0_resnets_0_time_emb_proj.alpha
lora_unet_down_blocks_0_resnets_0_time_emb_proj.lora_down.weight
lora_unet_down_blocks_0_resnets_0_time_emb_proj.lora_up.weight
lora_unet_down_blocks_0_resnets_1_conv1.alpha
lora_unet_down_blocks_0_resnets_1_conv1.lora_down.weight
lora_unet_down_blocks_0_resnets_1_conv1.lora_up.weight
lora_unet_down_blocks_0_resnets_1_conv2.alpha
lora_unet_down_blocks_0_resnets_1_conv2.lora_down.weight
lora_unet_down_blocks_0_resnets_1_conv2.lora_up.weight
lora_unet_down_blocks_0_resnets_1_time_emb_proj.alpha
lora_unet_down_blocks_0_resnets_1_time_emb_proj.lora_down.weight
lora_unet_down_blocks_0_resnets_1_time_emb_proj.lora_up.weight
lora_unet_down_blocks_1_attentions_0_proj_in.alpha
lora_unet_down_blocks_1_attentions_0_proj_in.lora_down.weight
lora_unet_down_blocks_1_attentions_0_proj_in.lora_up.weight
lora_unet_down_blocks_1_attentions_0_proj_out.alpha
lora_unet_down_blocks_1_attentions_0_proj_out.lora_down.weight
lora_unet_down_blocks_1_attentions_0_proj_out.lora_up.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_attn1_to_k.alpha
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_attn1_to_k.lora_down.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_attn1_to_k.lora_up.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_attn1_to_out_0.alpha
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_attn1_to_out_0.lora_down.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_attn1_to_out_0.lora_up.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_attn1_to_q.alpha
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_attn1_to_q.lora_down.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_attn1_to_q.lora_up.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_attn1_to_v.alpha
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_attn1_to_v.lora_down.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_attn1_to_v.lora_up.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_attn2_to_k.alpha
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_attn2_to_k.lora_down.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_attn2_to_k.lora_up.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_attn2_to_out_0.alpha
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_attn2_to_out_0.lora_down.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_attn2_to_out_0.lora_up.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_attn2_to_q.alpha
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_attn2_to_q.lora_down.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_attn2_to_q.lora_up.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_attn2_to_v.alpha
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_attn2_to_v.lora_down.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_attn2_to_v.lora_up.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_ff_net_0_proj.alpha
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_ff_net_0_proj.lora_down.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_ff_net_0_proj.lora_up.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_ff_net_2.alpha
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_ff_net_2.lora_down.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_ff_net_2.lora_up.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_attn1_to_k.alpha
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_attn1_to_k.lora_down.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_attn1_to_k.lora_up.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_attn1_to_out_0.alpha
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_attn1_to_out_0.lora_down.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_attn1_to_out_0.lora_up.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_attn1_to_q.alpha
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_attn1_to_q.lora_down.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_attn1_to_q.lora_up.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_attn1_to_v.alpha
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_attn1_to_v.lora_down.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_attn1_to_v.lora_up.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_attn2_to_k.alpha
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_attn2_to_k.lora_down.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_attn2_to_k.lora_up.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_attn2_to_out_0.alpha
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_attn2_to_out_0.lora_down.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_attn2_to_out_0.lora_up.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_attn2_to_q.alpha
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_attn2_to_q.lora_down.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_attn2_to_q.lora_up.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_attn2_to_v.alpha
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_attn2_to_v.lora_down.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_attn2_to_v.lora_up.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_ff_net_0_proj.alpha
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_ff_net_0_proj.lora_down.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_ff_net_0_proj.lora_up.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_ff_net_2.alpha
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_ff_net_2.lora_down.weight
lora_unet_down_blocks_1_attentions_0_transformer_blocks_1_ff_net_2.lora_up.weight
lora_unet_down_blocks_1_attentions_1_proj_in.alpha
lora_unet_down_blocks_1_attentions_1_proj_in.lora_down.weight
lora_unet_down_blocks_1_attentions_1_proj_in.lora_up.weight
lora_unet_down_blocks_1_attentions_1_proj_out.alpha
lora_unet_down_blocks_1_attentions_1_proj_out.lora_down.weight
lora_unet_down_blocks_1_attentions_1_proj_out.lora_up.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_attn1_to_k.alpha
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_attn1_to_k.lora_down.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_attn1_to_k.lora_up.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_attn1_to_out_0.alpha
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_attn1_to_out_0.lora_down.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_attn1_to_out_0.lora_up.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_attn1_to_q.alpha
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_attn1_to_q.lora_down.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_attn1_to_q.lora_up.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_attn1_to_v.alpha
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_attn1_to_v.lora_down.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_attn1_to_v.lora_up.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_attn2_to_k.alpha
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_attn2_to_k.lora_down.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_attn2_to_k.lora_up.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_attn2_to_out_0.alpha
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_attn2_to_out_0.lora_down.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_attn2_to_out_0.lora_up.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_attn2_to_q.alpha
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_attn2_to_q.lora_down.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_attn2_to_q.lora_up.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_attn2_to_v.alpha
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_attn2_to_v.lora_down.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_attn2_to_v.lora_up.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_ff_net_0_proj.alpha
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_ff_net_0_proj.lora_down.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_ff_net_0_proj.lora_up.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_ff_net_2.alpha
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_ff_net_2.lora_down.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_0_ff_net_2.lora_up.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_attn1_to_k.alpha
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_attn1_to_k.lora_down.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_attn1_to_k.lora_up.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_attn1_to_out_0.alpha
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_attn1_to_out_0.lora_down.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_attn1_to_out_0.lora_up.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_attn1_to_q.alpha
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_attn1_to_q.lora_down.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_attn1_to_q.lora_up.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_attn1_to_v.alpha
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_attn1_to_v.lora_down.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_attn1_to_v.lora_up.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_attn2_to_k.alpha
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_attn2_to_k.lora_down.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_attn2_to_k.lora_up.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_attn2_to_out_0.alpha
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_attn2_to_out_0.lora_down.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_attn2_to_out_0.lora_up.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_attn2_to_q.alpha
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_attn2_to_q.lora_down.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_attn2_to_q.lora_up.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_attn2_to_v.alpha
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_attn2_to_v.lora_down.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_attn2_to_v.lora_up.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_ff_net_0_proj.alpha
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_ff_net_0_proj.lora_down.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_ff_net_0_proj.lora_up.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_ff_net_2.alpha
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_ff_net_2.lora_down.weight
lora_unet_down_blocks_1_attentions_1_transformer_blocks_1_ff_net_2.lora_up.weight
lora_unet_down_blocks_1_downsamplers_0_conv.alpha
lora_unet_down_blocks_1_downsamplers_0_conv.lora_down.weight
lora_unet_down_blocks_1_downsamplers_0_conv.lora_up.weight
扫一扫
3410
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
