组合效果预览链接 Combination effect preview link:https://civarchive.com/models/579793
建议资源添加失败，请点击查看其余的素材
Failed to add suggested resources,Please click to view the remaining materials.
使用方式
插件：sd-webui-supermerger
模型A:风格模型1，模型B：基础模型，模型C：风格模型1
融合算法：A+(B-C)*alpha
Calculation Mode：trainDifference
α：0.6
Usage: .
Plugin: sd - webui - supermerger.
Model a: Style Model 1, Model b: Basic Model, Model c: Style Model 1.
Fusion Algorithm: a + (b - c) * alpha.
Calculation mode: traindifference.
α: 0.6.
下载量够多时我会上传用于拆解的模型
When the download volume is sufficient, I will upload the models for disassembly
Description
我想没人想知道提取方式
I think no one wants to know the extraction method.
FAQ
Comments (2)
I think it would be easier to understand intuitively if you wrote it as
BASE+(Style-BASE)*0.4
rather than
Style+(BASE-Style)*0.6
(although 0.4 would not be enough).
I think this is probably the result of repeatedly synthesizing v1-5-pruned-emaonly.safetensors with -1 from a specific model.
I have never tried it, so I do not know if it makes sense to synthesize such a dense style element. Is it effective?
I have previously tried to create a model with fewer v1-5-pruned-emaonly.safetensors, but it did not work very well.
I separated the model by adjusting α to a negative value, but the key to separation is not only α, but also two condensed basic elemental models.
I initially used the formula: base + (style - base) * 0.4, but this mixed method dilutes the basic elements and leads to a decrease in model performance. You can understand it as the basic model obtaining some artistic elements but reducing the basic elements. The lack of basic elements will result in a decrease in the quality of the generated images.
I recommend using style + (base - style) * 0.6, and the value can be set higher. The basic model can actually easily obtain the style of other models.