How do I use lucid to generate a picture that corresponds to the activation vector at a certain level as much as I specify？

[SHIER12138]

Hello: The idea is that it looks something like this, and I need to find an image X, where the output at a certain layer is as close as possible to the activation vector that I've been recording。

Where g is the neural network, X is the picture, and t is the activation record I saved before

Can L2 in objectvies be used to achieve this?

ps:I found this formula in this article https://arxiv.org/pdf/1905.07072v1.pdf

My English is not very good, I hope I made myself clear