confusion about inversion

[123dddwn]

I understand that many papers related to image editing use an inversion operation to recover the denoising trajectory of the reference image, and then inject the style of the reference image based on this trajectory during the actual denoising process.
However, this work does not include such an inversion step. I would like to know how it manages to successfully inject the reference image style without relying on the denoising trajectory.

[dingangui]

The core advantage of inversion-based methods lies in their ability to preserve the structure of the original image, allowing modifications to be localized to specific regions. This paradigm is particularly well-suited for image editing; for example, if the left side of an image contains a cat and the right side a table, an inversion-based method can easily transform the cat into a dog while leaving the table unchanged. However, in the context of image customization, the goal is not to retain the original image’s structure but rather to preserve the identity of the concept. For instance, when the input is a cat, the output should maintain the “cat” concept even if it appears in any new context. In this scenario, it is unnecessary to rely on inversion to maintain structure. Instead, one can start from random noise and guide the denoising process to reference the features of the target concept directly.

[123dddwn]

Thank you for your answer!