question about training class-conditional LDM

[Yoonho-Na]

Hello

I have trained unconditional LDM and it was successful.

Now, I'm trying to train class-conditional LDMs.

Is it necessary to train autoencoder with conditions as well?

I'm confused because I know that LDM has its own conditioning mechanism via cross attention.

So is it necessary to train class-conditional autoencoder for class-conditional LDM?

[SantiUsma]

It is not necesary. The autoencoder only needs the images because it learns how to represent the images in a latent space so you only have to train it once. LDM later train a Unet with the rest of the conditional mechanism, but the autoencoder could be the same in the in the uncoditional LDM or in the class-conditional.

[ayu1729]

Hello

I have trained unconditional LDM and it was successful.

Now, I'm trying to train class-conditional LDMs.

Is it necessary to train autoencoder with conditions as well?

I'm confused because I know that LDM has its own conditioning mechanism via cross attention.

So is it necessary to train class-conditional autoencoder for class-conditional LDM?

Can you please tell how did you train unconditional LDM? I want to fine tune unconditional LDM on my dataset.

[zy-charon]

Hello, if images are used as condition input for LDM, where should I modify the training settings?

[tangwy98]

你好

我已经训练了无条件的LDM，并且很成功。

现在，我正在尝试训练类条件 LDM。

是否有必要使用条件训练自动编码器？

我很困惑，因为我知道LDM通过交叉注意力有自己的调节机制。

那么是否有必要为类条件LDM训练类条件自动编码器呢？

Can you explain how you trained and the specific details? I want to train class condition ldm on my own dataset