MooreThreads / Moore-AnimateAnyone

Character Animation (AnimateAnyone, Face Reenactment)
Apache License 2.0
3.21k stars 254 forks source link

Different base model in stage 1 and stage 2 #155

Open Naive-Bayes opened 4 months ago

Naive-Bayes commented 4 months ago

Any difference between stable-diffusion-v1-5 and sd-image-variations-diffusers? The author set "base_model_path: './pretrained_weights/sd-image-variations-diffusers'" in stage 1 and "'./pretrained_weights/stable-diffusion-v1-5'" in stage 2.

In Chinese: sd-image-variations-diffusers和stable-diffusion-v1-5这两个库有啥区别吗,从目录结构上看就是在sd-image-variations-diffusers中多了image encoder。 为啥第一个阶段用sd-image-variations-diffusers,第二个阶段用stable-diffusion-v1-5?

xiecon commented 4 months ago

我看代码中,第二阶段的sd1.5会被第一阶段训练的unet权重覆盖,所以使用哪个都不重要

aleeyang commented 3 months ago

我看代码中,第二阶段的sd1.5会被第一阶段训练的unet权重覆盖,所以使用哪个都不重要

那第一阶段使用stable-diffusion-v1-5 或者sd-image-variations-diffusers有什么区别吗

Naive-Bayes commented 3 months ago

我看代码中,第二阶段的sd1.5会被第一阶段训练的unet权重覆盖,所以使用哪个都不重要

是的,读代码的时候发现那个就是初始化一下结构,后面会用第一阶段训好的权重做reload