Spaces:
Runtime error
Runtime error
| import gradio as gr | |
| from PIL import Image | |
| import os | |
| def get_images(path): | |
| images = [Image.open(os.path.join(path,im)) for im in os.listdir(path)] | |
| paths = os.listdir(path) | |
| return([(im, path) for im, path in zip(images,paths)]) | |
| with gr.Blocks() as demo: | |
| gr.Markdown(""" | |
| ## Stable Bias: Analyzing Societal Representations in Diffusion Models | |
| """) | |
| gr.HTML(''' | |
| <p style="margin-bottom: 10px; font-size: 94%">This is the demo page for the "Stable Bias" paper, which aims to explore and quantify social biases in text-to-image systems. <br> This work was done by <a href='https://huggingface.co/sasha' style='text-decoration: underline;' target='_blank'> Alexandra Sasha Luccioni (Hugging Face) </a>, <a href='https://huggingface.co/cakiki' style='text-decoration: underline;' target='_blank'> Christopher Akiki (ScaDS.AI, Leipzig University)</a>, <a href='https://huggingface.co/meg' style='text-decoration: underline;' target='_blank'> Margaret Mitchell (Hugging Face) </a> and <a href='https://huggingface.co/yjernite' style='text-decoration: underline;' target='_blank'> Yacine Jernite (Hugging Face) </a> .</p> | |
| ''') | |
| examples_path= "images/examples" | |
| examples_gallery = gr.Gallery(get_images(examples_path), | |
| label="Example images generated by three text-to-image models (Dall-E 2, Stable Diffusion v1.4 and v.2).", show_label=True, elem_id="gallery").style(grid=[1,6], height="auto") | |
| gr.HTML(''' | |
| <p style="margin-bottom: 14px; font-size: 100%"> As AI-enabled Text-to-Image models are becoming increasingly used, characterizing the social biases they exhibit is a necessary first step to lowering their risk of discriminatory outcomes. <br> We compare three such models: <b> Stable Diffusion v.1.4, Stable Diffusion v.2. </b>, and <b> Dall-E 2 </b>, prompting them to produce images of different <i> professions </i> and <i> identity characteristics </i>. <br> Since artificial depictions of fictive humans have no inherent gender or ethnicity nor do they belong to socially-constructed groups, we pursued our analysis <i> without </i> ascribing gender and ethnicity categories to the images generated, still finding clear evidence of ethnicity and gender biases. You can explore these findings in the sections below: </p> | |
| ''') | |
| with gr.Accordion("Identity group results (ethnicity and gender)", open=False): | |
| gr.HTML(''' | |
| <p style="margin-bottom: 14px; font-size: 100%"> One of the approaches that we adopted in our work is hierarchical clustering of the images generated by the text-to-image systems in response to prompts that include identity terms with regards to ethnicity and gender. <br> We computed 3 different numbers of clusters (12, 24 and 48) and created an <a href='https://huggingface.co/spaces/society-ethics/DiffusionFaceClustering' style='text-decoration: underline;' target='_blank'> Identity Representation Demo </a> that allows for the exploration of the different clusters and their contents. </p> | |
| ''') | |
| with gr.Row(): | |
| with gr.Column(scale=2): | |
| impath = "images/identities" | |
| identity_gallery = gr.Gallery([os.path.join(impath,im) for im in os.listdir(impath)], | |
| label="Identity cluster images", show_label=False, elem_id="gallery" | |
| ).style(grid=3, height="auto") | |
| with gr.Column(scale=1): | |
| gr.HTML(''' | |
| <p style="margin-bottom: 14px; font-size: 100%"> You can see that the models reflect many societal biases -- for instance representing Native Americans wearing traditional headdresses, non-binary people with stereotypical haircuts and glasses, and East Asian men with features that amplify ethnic stereotypes. <br> This is problematic because it reinforces existing cultural stereotypes and fails to represent the diversity that is present in all identity groups.</p> | |
| ''') | |
| with gr.Accordion("Bias Exploration", open=False): | |
| gr.HTML(''' | |
| <p style="margin-bottom: 14px; font-size: 100%"> We also explore the correlations between the professions that use used in our prompts and the different identity clusters that we identified. <br> Using both the <a href='https://huggingface.co/spaces/society-ethics/DiffusionClustering' style='text-decoration: underline;' target='_blank'> Diffusion Cluster Explorer </a> and the <a href='https://huggingface.co/spaces/society-ethics/DiffusionFaceClustering' style='text-decoration: underline;' target='_blank'> Identity Representation Demo </a>, we can see which clusters are most correlated with each profession and what identities are in these clusters.</p> | |
| ''') | |
| with gr.Row(): | |
| with gr.Column(): | |
| gr.HTML(''' | |
| <p style="margin-bottom: 14px; font-size: 100%"> Using the <b><a href='https://huggingface.co/spaces/society-ethics/DiffusionClustering' style='text-decoration: underline;' target='_blank'> Diffusion Cluster Explorer</a></b>, we can see that the top cluster for the CEO and director professions is <b> Cluster 4</b>: </p> ''') | |
| with gr.Column(): | |
| ceo_img = gr.Image(Image.open("images/bias/ceo_dir.png"), label = "CEO Image", show_label=False) | |
| with gr.Row(): | |
| with gr.Column(): | |
| gr.HTML(''' | |
| <p style="margin-bottom: 14px; font-size: 100%"> Going back to the <b><a href='https://huggingface.co/spaces/society-ethics/DiffusionFaceClustering' style='text-decoration: underline;' target='_blank'> Identity Representation Demo </a></b>, we can see that the most represented gender term is <i> man </i> (56% of the cluster) and <i> White </i> (29% of the cluster). <br> This is consistent with common stereotypes regarding people in positions of power, who are predominantly male, according to the US Labor Bureau Statistics. </p> ''') | |
| with gr.Column(): | |
| cluster4 = gr.Image(Image.open("images/bias/Cluster4.png"), label = "Cluster 4 Image", show_label=False) | |
| with gr.Row(): | |
| with gr.Column(): | |
| gr.HTML(''' | |
| <p style="margin-bottom: 14px; font-size: 100%"> If we look at the cluster representation of professions such as social assistant and social worker, we can observe that the former is best represented by <b>Cluster 2</b>, whereas the latter has a more uniform representation across multiple clusters: </p> ''') | |
| with gr.Column(): | |
| social_img = gr.Image(Image.open("images/bias/social.png"), label = "social image", show_label=False) | |
| with gr.Row(): | |
| with gr.Column(scale=1): | |
| gr.HTML(''' | |
| <p style="margin-bottom: 14px; font-size: 100%"> Cluster 2 is best represented by the gender term is <i> woman </i> (81%) as well as <i> Latinx </i> (19%) <br> This gender proportion is exactly the same as the one provided by the United States Labor Bureau (which you can see in the table above), with 81% of social assistants identifying as women. </p> ''') | |
| with gr.Column(scale=2): | |
| cluster4 = gr.Image(Image.open("images/bias/Cluster2.png"), label = "Cluster 2 Image", show_label=False) | |
| with gr.Accordion("Comparing model generations", open=False): | |
| gr.HTML(''' | |
| <p style="margin-bottom: 14px; font-size: 100%"> One of the goals of our study was allowing users to compare model generations across professions in an open-ended way, uncovering patterns and trends on their own. This is why we created the <a href='https://huggingface.co/spaces/society-ethics/DiffusionBiasExplorer' style='text-decoration: underline;' target='_blank'> Diffusion Bias Explorer </a> and the <a href='https://huggingface.co/spaces/society-ethics/Average_diffusion_faces' style='text-decoration: underline;' target='_blank'> Average Diffusion Faces </a> tools. We show some of their functionalities below: </p> ''') | |
| with gr.Row(): | |
| with gr.Column(scale=2): | |
| explorerpath = "images/biasexplorer" | |
| biasexplorer_gallery = gr.Gallery(get_images(explorerpath), | |
| label="Bias explorer images", show_label=False, elem_id="gallery").style(grid=[2]) | |
| with gr.Column(scale=1): | |
| gr.HTML(''' | |
| <p style="margin-bottom: 14px; font-size: 100%"> Comparing generations both between two models and within a single model can help uncover trends and patterns that are hard to measure using quantitative approaches. <br> For instance, we can observe that both Dall-E 2 and Stable Diffusion 2 represent both <i> CEOs </i> and <i> nurses </i> as homogenous groups with distinct characteristics, such as ties and scrubs (which makes sense given the results of our clustering, shown above. <br> We can also see that the images of <i> waitresses </i> generated by Dall-E 2 and Stable Diffusion v.1.4. have different characteristics, both in terms of their clothes as well as their appearance. <br> It's also possible to see harder to describe phenomena, like the fact that portraits of <i> painters </i> often look like paintings themselves. <br> We encourage you to use the <a href='https://huggingface.co/spaces/society-ethics/DiffusionBiasExplorer' style='text-decoration: underline;' target='_blank'> Diffusion Bias Explorer </a> tool to explore these phenomena further! </p>''') | |
| with gr.Row(): | |
| with gr.Column(scale=2): | |
| averagepath = "images/averagefaces" | |
| average_gallery = gr.Gallery(get_images(averagepath), | |
| label="Average Face images", show_label=False, elem_id="gallery").style(grid=[2]) | |
| with gr.Column(scale=1): | |
| gr.HTML(''' | |
| <p style="margin-bottom: 14px; font-size: 100%"> Looking at the average faces for a given profession across multiple models can help see the dominant characteristics of that profession, as well as how much variation there is (based on how fuzzy the image is). <br> In the images shown here, we can see that representations of these professions significantly differ across the three models, while sharing common characteristics, e.g. <i> postal workers </i> all wear caps. <br> Also, the average faces of <i> hairdressers </i> seem more fuzzy than the other professions, indicating a higher diversity among the generations compared to other professions.</p>''') | |
| with gr.Accordion("Exploring the color space of generated images", open=False): | |
| gr.HTML(''' | |
| <p style="margin-bottom: 14px; font-size: 100%"> TODO Chris </p> ''') | |
| with gr.Accordion("Exploring the nearest neighbors of generated images", open=False): | |
| gr.HTML(''' | |
| <p style="margin-bottom: 14px; font-size: 100%"> TODO Chris </p> ''') | |
| gr.Markdown(""" | |
| ### All of the tools created as part of this project: | |
| """) | |
| gr.HTML(''' | |
| <p style="margin-bottom: 10px; font-size: 94%"> | |
| <a href='https://huggingface.co/spaces/society-ethics/Average_diffusion_faces' style='text-decoration: underline;' target='_blank'> Average Diffusion Faces </a> <br> | |
| <a href='https://huggingface.co/spaces/society-ethics/DiffusionBiasExplorer' style='text-decoration: underline;' target='_blank'> Diffusion Bias Explorer </a> <br> | |
| <a href='https://huggingface.co/spaces/society-ethics/DiffusionClustering' style='text-decoration: underline;' target='_blank'> Diffusion Cluster Explorer </a> | |
| <a href='https://huggingface.co/spaces/society-ethics/DiffusionFaceClustering' style='text-decoration: underline;' target='_blank'> Identity Representation Demo </a> | |
| <a href='https://huggingface.co/spaces/tti-bias/identities-bovw-knn' style='text-decoration: underline;' target='_blank'> BoVW Nearest Neighbors Explorer </a> <br> | |
| <a href='https://huggingface.co/spaces/tti-bias/professions-bovw-knn' style='text-decoration: underline;' target='_blank'> BoVW Professions Explorer </a> <br> | |
| <a href='https://huggingface.co/spaces/tti-bias/identities-colorfulness-knn' style='text-decoration: underline;' target='_blank'> Colorfulness Profession Explorer </a> <br> | |
| <a href='https://huggingface.co/spaces/tti-bias/professions-colorfulness-knn' style='text-decoration: underline;' target='_blank'> Colorfulness Identities Explorer </a> <br> </p> | |
| ''') | |
| # gr.Interface.load("spaces/society-ethics/DiffusionBiasExplorer") | |
| demo.launch(debug=True) | |