import gradio as gr import random, os, shutil from PIL import Image import pandas as pd import tempfile def open_sd_ims(adj, group, seed): if group != '': if adj != '': prompt=adj+'_'+group.replace(' ','_') if os.path.isdir(prompt) == False: shutil.unpack_archive('zipped_images/stablediffusion/'+ prompt.replace(' ', '_') +'.zip', prompt, 'zip') else: prompt=group if os.path.isdir(prompt) == False: shutil.unpack_archive('zipped_images/stablediffusion/'+ prompt.replace(' ', '_') +'.zip', prompt, 'zip') imnames= os.listdir(prompt+'/Seed_'+ str(seed)+'/') images = [(Image.open(prompt+'/Seed_'+ str(seed)+'/'+name)) for name in imnames] return images[:9] def open_ims(model, adj, group): seed = 48040 with tempfile.TemporaryDirectory() as tmpdirname: print('created temporary directory', tmpdirname) if model == "Dall-E 2": if group != '': if adj != '': prompt=adj+'_'+group.replace(' ','_') if os.path.isdir(tmpdirname + '/' + model.replace(' ','').lower()+ '/'+ prompt) == False: shutil.unpack_archive('zipped_images/'+ model.replace(' ','').lower()+ '/'+ prompt.replace(' ', '_') +'.zip', tmpdirname+ '/'+ model.replace(' ','').lower()+ '/'+ prompt, 'zip') else: prompt=group if os.path.isdir(tmpdirname + '/' + model.replace(' ','').lower()+ '/'+ prompt) == False: shutil.unpack_archive('zipped_images/' + model.replace(' ','').lower() + '/'+ prompt.replace(' ', '_') +'.zip', tmpdirname + '/' + model.replace(' ','').lower()+ '/' + prompt, 'zip') imnames= os.listdir(tmpdirname + '/' + model.replace(' ','').lower()+ '/'+ prompt+'/') images = [(Image.open(tmpdirname + '/' + model.replace(' ','').lower()+ '/'+ prompt+'/'+name)).convert("RGB") for name in imnames] return images[:9] else: if group != '': if adj != '': prompt=adj+'_'+group.replace(' ','_') if os.path.isdir(tmpdirname + '/' + model.replace(' ','').lower()+ '/'+ prompt) == False: shutil.unpack_archive('zipped_images/'+ model.replace(' ','').lower()+ '/'+ prompt.replace(' ', '_') +'.zip', tmpdirname + '/' +model.replace(' ','').lower()+ '/'+ prompt, 'zip') else: prompt=group if os.path.isdir(tmpdirname + '/' + model.replace(' ','').lower()+ '/'+ prompt) == False: shutil.unpack_archive('zipped_images/' + model.replace(' ','').lower() + '/'+ prompt.replace(' ', '_') +'.zip', tmpdirname + '/' + model.replace(' ','').lower()+'/'+ prompt, 'zip') imnames= os.listdir(tmpdirname + '/' + model.replace(' ','').lower()+ '/'+ prompt+'/'+'Seed_'+ str(seed)+'/') images = [(Image.open(tmpdirname + '/' + model.replace(' ','').lower()+ '/'+ prompt +'/'+'Seed_'+ str(seed)+'/'+name)) for name in imnames] return images[:9] vowels = ["a","e","i","o","u"] prompts = pd.read_csv('promptsadjectives.csv') seeds = [46267, 48040, 51237, 54325, 60884, 64830, 67031, 72935, 92118, 93109] m_adjectives = prompts['Masc-adj'].tolist()[:10] f_adjectives = prompts['Fem-adj'].tolist()[:10] adjectives = sorted(m_adjectives+f_adjectives) #adjectives = ['attractive','strong'] adjectives.insert(0, '') professions = sorted([p.lower() for p in prompts['Occupation-Noun'].tolist()]) models = ["Stable Diffusion 1.4", "Dall-E 2","Stable Diffusion 2"] with gr.Blocks() as demo: gr.Markdown("# Diffusion Bias Explorer") gr.Markdown("## Choose from the prompts below to explore how the text-to-image models like [Stable Diffusion v1.4](https://huggingface.co/CompVis/stable-diffusion-v-1-4-original), [Stable Diffion v.2](https://huggingface.co/stabilityai/stable-diffusion-2) and [DALLE-2](https://openai.com/dall-e-2/) represent different professions and adjectives") gr.HTML("""⚠️ DISCLAIMER: the images displayed by this tool were generated by text-to-image models and may depict offensive stereotypes or contain explicit content.""") # gr.Markdown("Some of the images for Dall-E 2 are missing -- we are still in the process of generating them! If you get an 'error', please pick another prompt.") # seed_choice = gr.State(0) # seed_choice = 93109 # print("Seed choice is: " + str(seed_choice)) with gr.Row(): with gr.Column(): model1 = gr.Dropdown(models, label = "Choose a model to compare results", value = models[0], interactive=True) adj1 = gr.Dropdown(adjectives, label = "Choose a first adjective (or leave this blank!)", interactive=True) choice1 = gr.Dropdown(professions, label = "Choose a first group", interactive=True) # seed1= gr.Dropdown(seeds, label = "Choose a random seed to compare results", value = seeds[1], interactive=True) images1 = gr.Gallery(label="Images").style(grid=[3], height="auto") with gr.Column(): model2 = gr.Dropdown(models, label = "Choose a model to compare results", value = models[0], interactive=True) adj2 = gr.Dropdown(adjectives, label = "Choose a second adjective (or leave this blank!)", interactive=True) choice2 = gr.Dropdown(professions, label = "Choose a second group", interactive=True) # seed2= gr.Dropdown(seeds, label = "Choose a random seed to compare results", value= seeds[1], interactive=True) images2 = gr.Gallery(label="Images").style(grid=[3], height="auto") gr.Markdown("### [Research](http://gender-decoder.katmatfield.com/static/documents/Gaucher-Friesen-Kay-JPSP-Gendered-Wording-in-Job-ads.pdf) has shown that \ certain words are considered more masculine- or feminine-coded based on how appealing job descriptions containing these words \ seemed to male and female research participants and to what extent the participants felt that they 'belonged' in that occupation.") #demo.load(random_image, None, [images]) choice1.change(open_ims, [model1, adj1,choice1], [images1]) choice2.change(open_ims, [model2, adj2,choice2], [images2]) adj1.change(open_ims, [model1, adj1, choice1], [images1]) adj2.change(open_ims, [model2, adj2, choice2], [images2]) # seed1.change(open_ims, [adj1,choice1,seed1], [images1]) # seed2.change(open_ims, [adj2,choice2,seed2], [images2]) demo.launch()