question about the role of prompt depth features

[questioner879]

I’m a beginner in diffusion models, and in your code, directly setting the encoded prompt depth to 0, detaching it, and then adding it to the image features essentially means that the prompt depth has no effect. Wouldn’t that just be using someone else’s model? If I’m mistaken, please forgive me.

我是扩散模型的初学者，在你的代码中直接将prompt depth编码后全部置为0并且detach，再与图像特征相加，那相当于prompt depth没有起到任何作用，这岂不是在用别人的模型？如果我说的不对请见谅

[haotongl]

" directly setting the encoded prompt depth to 0, detaching it"
This operation does not exist in our implementation. You may double-check it. Any issues or concerns are welcome.

[questioner879]

Thank you for your reply, but I think the function of your zero_module for prompt depth encoding is directly setting the encoded prompt depth to 0 and detach it. Specifically, the function of this module is to set all weights and biases of this layer to 0, which also means that all outputs will be set to 0