Implementation of Algorithm 2

[backpropagator]

Thanks for the amazing paper, I really loved reading it.

I can see that the code provided in pipeline_ead.py is for Algorithm 3 (which includes using a layout branch). I wanted to know if there are any plans to provide the code for Algorithm 3, or is there some minor changes that one can do to get it.

Thanks.

[FreeButUselessSoul]

I think backpropagator is referring to the implementation of "Algorithm 2 DDCM for inversion-free image editing''? One can find it in pipeline_ead.py, line 50, in the definition of ddcm_sampler(). Hope it can help :)

[backpropagator]

Thanks for the prompt reply. I thought the same, but after carefully going through the code, I think Line 83 and 84 are not how they should be? These lines resemble the reverse sampling Equation (3) of the paper. However, Line 7 and 8 of Algorithm 2 of the paper seem to suggest that the direction term should not be used.

Am I missing something here?

Again, thanks for your response.

[h6kplus]

Hi! Thank you for your interest in our paper.

For Algorithm 3, you can find the implementation in class AttentionControlEdit() in app_infedit.py, where both the self attention control and cross attention control are implemented, if you want to find how these controls are injected, you can refer to register_attention_control() in ptp_utils.

For Algorithm2, the two lines you referred to are the implementation of LIne 7 and 8 of Algorithm 2. The two dir_xt will vanish to 0 since eta = 1 here if you look at line 79.

Hope it can help you!

[backpropagator]

Thanks for the quick reply! Really appreciate it.

Algorithm 3 seems fine. And thanks for the explanation, that clears up Algorithm 2 as well. I was just wondering if there is any reason to keep the dir_xt terms in the code?

I mean for any eta != 1, e_c won't be as desired (as described in Section 3 of paper). So, is there any reason to keep that term there? or can we just remove that term?

[h6kplus]

We haven't investigated the impact of different eta on the performance of the model. Here eta serves as a coefficient to control in injection of the noise, and it follows the setting of $\sigmat = \eta\sqrt{1-\alpha{t-1}}$, where we use $\eta=1$