No fine-tuning or extra module needed, why losses and learning rate?

[DarrenIm]

Hi, interesting work. As mentioned in the paper:

All content editing and preservation signals in our proposed method come from the image itself. It allows for a direct translation of T2I generation ability in diffusion models to image editing tasks without the need for any model fine-tuning or training.

I'm confused by few losses and even learning rate defined in Eq. 5 and Eq. 6, what are they used for? Thanks in advance.

[MC-E]

Our method is designed based on the score function (https://arxiv.org/abs/1907.05600). This approach deals with the diffusion process from a continuous perspective, where a gradient needs to be calculated for each diffusion step. The loss function here serves as additional guidance for the gradient.