question on reproducing i2p experiment on ESD

[enkeejunior1]

Thank you for your awesome work.

What should the placeholder_token be for the i2p experiment? Currently, it's --placeholder_token="<art-style>" --initializer_token="art", but I'm asking if this is correct.

Thank you in advance.

[damon-demon]

I am also trying to reproduce it.

Based on the code in the esd folder , the following are my setting used for I2P Exps:

• --learnable_property 'person' <-- Check Line 518 in concept_inversion.py

• --placeholder_token '<nude-person>' <-- Textual inversion is to learn a word embedding, so here you can name whatever you want. Check Line 631 in concept_inversion.py

• --initializer_token 'nude' <-- Here define the initialization of word embedding to be optimized.

If you are working on adversarial prompt attacks for unlearned diffusion model, you can also check my recent work "To Generate or Not? Safety-Driven Unlearned Diffusion Models Are Still Easy To Generate Unsafe Images ... For Now"

☀️☀️☀️