some questions about the code.

[A1ita99]

1. Does demo.py implement the DiffProtect-fast version?
   Since the reconstructed image during the attack process uses xT = xT_all[start_t * (T_enc // T_atk) - 1] with a step of 5, while the final reconstructed image uses xT = xT_all[-1] adv_render_img = model.render(xT, z_adv, T=T_inf), with a step of 100, why does the final reconstructed image still exhibit the effects of the attack? The noise patterns are different.

2. How can I implement the DiffProttect(non-fast) version？
   x_adv = self.decoder.render(xT, xvar + delta, T=100) During the attack process, when I set T_atk=100, it requires a large amount of GPU memory. My A100 GPU with 80GB of VRAM cannot handle 100 iterations. How can I address this issue？

[Fujiaoji]

I have the same question for 1. Do you figure it out? I find that the adv_render_img is different produced by these two methods. Thanks. @joellliu