The codes are adapted from: 1) ITI-GEN: https://github.com/humansensinglab/ITI-GEN; 2) DAAM: https://github.com/castorini/daam

Please find the additional necessary files ./models , ./iti_gen and ./ckpts files at link

FairQueue

• ./ckpts contain the saved ITI-GEN embeddings listed in the main paper. Note that the "basis_final_embed_29.pt" are the final embeddings utilized in sample generation (this is the same as the orignal base code).

1. Run generation_Mixing.py file with the command:

Python generation_Mixing.py --attr-list <tSA> --prompt-path <Location of Prompt path ending with 'basis_final_embed_29.pt'>

Cross-attention Analysis

• Put the folder './ckpts' under the project folder
• Establish the environment following either requirement.txt or environment.yml (similar to daam 0.1.0 in https://github.com/castorini/daam)
• Replace the original file 'pipeline_stable_diffusion.py' under the environment '/lib/python3.8/site-packages/diffusers/pipelines/stable_diffusion' with the one provided in the folder
• One may run 'cros_attn_analysis.py' with the command:

./run_cros_attn_analysis.sh --tSA "Smiling" --prompts_mode "H2I" --sample_list "[1,2]" --step_start_switch 15

One may find 'cros_attn_analysis.py' for other parameters that can be assigned