Model description

This model is a part of project targeting Debiasing of generative stable diffusion models.

LoRA text2image fine-tuning - NYUAD-ComNets/Black_Female_Profession_Model

These are LoRA adaption weights for stabilityai/stable-diffusion-xl-base-1.0. The weights were fine-tuned on the NYUAD-ComNets/Black_Female_Profession dataset. You can find some example images.

prompt: a photo of a {profession}, looking at the camera, closeup headshot facing forward, ultra quality, sharp focus

How to use this model:



import torch
from compel import Compel, ReturnedEmbeddingsType
from diffusers import DiffusionPipeline

import random


negative_prompt = "cartoon, anime, 3d, painting, b&w, low quality" 


models=["NYUAD-ComNets/Asian_Female_Profession_Model","NYUAD-ComNets/Black_Female_Profession_Model","NYUAD-ComNets/White_Female_Profession_Model",
"NYUAD-ComNets/Indian_Female_Profession_Model","NYUAD-ComNets/Latino_Hispanic_Female_Profession_Model","NYUAD-ComNets/Middle_Eastern_Female_Profession_Model",
"NYUAD-ComNets/Asian_Male_Profession_Model","NYUAD-ComNets/Black_Male_Profession_Model","NYUAD-ComNets/White_Male_Profession_Model",
"NYUAD-ComNets/Indian_Male_Profession_Model","NYUAD-ComNets/Latino_Hispanic_Male_Profession_Model","NYUAD-ComNets/Middle_Eastern_Male_Profession_Model"]

adapters=["asian_female","black_female","white_female","indian_female","latino_female","middle_east_female",
"asian_male","black_male","white_male","indian_male","latino_male","middle_east_male"]

pipeline = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", variant="fp16", use_safetensors=True, torch_dtype=torch.float16).to("cuda")


for i,j in zip(models,adapters):
    pipeline.load_lora_weights(i, weight_name="pytorch_lora_weights.safetensors",adapter_name=j) 


prof='doctor'
    

pipeline.set_adapters(random.choice(adapters))


compel = Compel(tokenizer=[pipeline.tokenizer, pipeline.tokenizer_2] ,
                    text_encoder=[pipeline.text_encoder, pipeline.text_encoder_2],
                    returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED, 
                    requires_pooled=[False, True],truncate_long_prompts=False)

    
conditioning, pooled = compel("a photo of a {}, looking at the camera, closeup headshot facing forward, ultra quality, sharp focus".format(prof)) 

negative_conditioning, negative_pooled = compel(negative_prompt)
[conditioning, negative_conditioning] = compel.pad_conditioning_tensors_to_same_length([conditioning, negative_conditioning])

image = pipeline(prompt_embeds=conditioning, negative_prompt_embeds=negative_conditioning,
                     pooled_prompt_embeds=pooled, negative_pooled_prompt_embeds=negative_pooled,
                     num_inference_steps=40).images[0]

image.save('/../../x.jpg')

Examples

Training data

NYUAD-ComNets/Black_Female_Profession dataset was used to fine-tune stabilityai/stable-diffusion-xl-base-1.0

Configurations

LoRA for the text encoder was enabled: False.

Special VAE used for training: madebyollin/sdxl-vae-fp16-fix.

BibTeX entry and citation info


@article{aldahoul2024ai,
  title={AI-generated faces free from racial and gender stereotypes},
  author={AlDahoul, Nouar and Rahwan, Talal and Zaki, Yasir},
  journal={arXiv preprint arXiv:2402.01002},
  year={2024}
}

@misc{ComNets,
      url={[https://huggingface.co/NYUAD-ComNets/Black_Female_Profession_Model](https://huggingface.co/NYUAD-ComNets/Black_Female_Profession_Model)},
      title={Black_Female_Profession_Model},
      author={Nouar AlDahoul, Talal Rahwan, Yasir Zaki}
}