Reproduce editing result from paper

[nttung1110]

Hi,

First of all, thanks for such an amazing work. The results in the paper looks pretty fascinating. I tried to use exactly the same code without changing anything (even installing the correct package, i.e. diffuser 0.15.0) but got unexpected result. Here are the two results for test case: person, and corgi.

As u can see, this code might work for corgi test case but not for person. Could I ask if u also got the same result from your side? Many thanks

[ljzycmd]

Hi @nttung1110,

Thanks for your interest. I have checked my original implementation to edit the person. You may pass the intermediate inverted features latents_list to the editing process to reproduce the results presented in the paper.

# File: playground_real.ipynb
image_masactrl = model(prompts,
                       latents=start_code,
                       guidance_scale=7.5,
                       ref_intermediate_latents=latents_list)

The results:

[nttung1110]

Hi,

Thanks for your quick response. Here is what I have after following your instruction:

The result seems to improve a bit. Does this suggest that I have to manually observe the result to decide whether to pass the intermediate inverted features for every single case? Many thanks!