Repository to the Paper Fair Diffusion: Instructing Text-to-Image Generation Models on Fairness

A demo is available at Huggingface and you can find a recent article about our work at MIT Technology Review

A model editing technique has to be installed first. This repositories builds on Sega, but can be adjusted to other methods, too. Further information about Sega.

It can be installed, e.g., directly via git

pip install git+https://github.com/ml-research/semantic-image-editing.git

It applies semantic guidance to any concept, here e.g. to mitigate gender occupation biases.

from semdiffusers import SemanticEditPipeline
device='cuda'

pipe = SemanticEditPipeline.from_pretrained(
    "runwayml/stable-diffusion-v1-5",
).to(device)

An exemplary usage of our pipeline could look like this:

import torch
gen = torch.Generator(device=device)

gen.manual_seed(21)
out = pipe(prompt='a photo of the face of a firefighter', generator=gen, num_images_per_prompt=1, guidance_scale=7,
           editing_prompt=['male person',       # Concepts to apply 
                           'female person'],
           reverse_editing_direction=[True, False], # Direction of guidance i.e. decrease the first and increase the second concept
           edit_warmup_steps=[10, 10], # Warmup period for each concept
           edit_guidance_scale=[4, 4], # Guidance scale for each concept
           edit_threshold=[0.95, 0.95], # Threshold for each concept. Threshold equals the percentile of the latent space that will be discarded. I.e. threshold=0.99 uses 1% of the latent dimensions
           edit_momentum_scale=0.3, # Momentum scale that will be added to the latent guidance
           edit_mom_beta=0.6, # Momentum beta
           edit_weights=[1,1] # Weights of the individual concepts against each other
          )
images = out.images

We provide test_notebook.ipynb for further experiments on Fair Diffusion. It can be used to gain first insights into changing fair attributes during image generation.

Our results can be repoduced with the provided code. generate_images.py enables to generate images for occupations from occupations.txt and evaluate_images.ipynb evaluates them. CLIP_iEAT.ipynb computes the iEAT to insight biases in CLIP.

The models stored in dlib_models/ are taken from FairFace, which is used to classify generated images for gender. To evaluate laion, the laion dataset has to be downloaded, however, we provide the already evaluated statistics in results_fairface_laion.txt.

If you like or use our work please cite us:

@article{friedrich2023FairDiffusion,
      title={Fair Diffusion: Instructing Text-to-Image Generation Models on Fairness}, 
      author={Felix Friedrich and Manuel Brack and Lukas Struppek and Dominik Hintersdorf and Patrick Schramowski and Sasha Luccioni and Kristian Kersting},
      year={2023},
      journal={arXiv preprint at arXiv:2302.10893}
}