context

app.py

@@ -6,7 +6,37 @@ import plotly.graph_objects as go
 import plotly.express as px
 import operator
 
-TITLE = "Diffusion Faces Cluster Explorer"
+TITLE = "Identity Representation in Diffusion Models"
+
+_INTRO = """
+# Identity Representation in Diffusion Models
+
+Explore the data generated from [DiffusionBiasExplorer](https://huggingface.co/spaces/society-ethics/DiffusionBiasExplorer)!
+This demo showcases patterns in the images generated from different prompts input to Stable Diffusion and Dalle-2 systems.
+"""
+
+_CONTEXT = """
+##### How do diffusion-based models represent gender and ethnicity?
+
+In order to evaluate the *social biases* that Text-to-Image (TTI) systems may reproduce or exacerbate,
+we need to first understand how the visual representations they generate relate to notions of gender and ethnicity.
+These two aspects of a person's identity, however, ar known as **socialy constructed characteristics**:
+that is to say, gender and ethnicity only exist in interactions between people, they do not have an independent existence based solely on physical (or visual) attributes.
+This means that while we can characterize trends in how the models associate visual features with specific *identity terms in the generation prompts*,
+we should not assign a specific gender or ethnicity to a synthetic figure generated by an ML model.
+
+In this app, we instead take a 2-step clustering-based approach. First, we generate 680 images for each model by varying mentions of terms that denote gender or ethnicity in the prompts.
+Then, we use a [VQA-based model](https://huggingface.co/Salesforce/blip-vqa-base) to cluster these images at different granularities (12, 24, or 48 clusters).
+Exploring these clusters allows us to examine trends in the models' associations between visual features and textual representation of social features.
+We encourage users to take advantage of this app to explore those trends, for example through the lens of the following questions:
+- Find the cluster that has the most prompts denoting a gender or ethnicity that you identify with. Do you think the generated images look like you?
+- Find two clusters that have a similar distribution of gender terms but different distributions of ethnicity terms. Do you see any meaningful differences in how gender is visually represented?
+- Do you find that some ethnicity terms lead to more stereotypical visual representations than others?
+- Do you find that some gender terms lead to more stereotypical visual representations than others?
+
+These questions only scratch the surface of what we can learn from demos like this one, let us know what you find [in the discussions tab](https://huggingface.co/spaces/society-ethics/DiffusionFaceClustering/discussions), or if you think of other relevant questions!
+"""
+
 clusters_12 = json.load(open("clusters/id_all_blip_clusters_12.json"))
 clusters_24 = json.load(open("clusters/id_all_blip_clusters_24.json"))
 clusters_48 = json.load(open("clusters/id_all_blip_clusters_48.json"))
@@ -180,13 +210,11 @@ def show_cluster(cl_id, num_clusters):
 
 
 with gr.Blocks(title=TITLE) as demo:
-    gr.Markdown(f"# {TITLE}")
-    gr.Markdown(
-        "Explore the data generated from [DiffusionBiasExplorer](https://huggingface.co/spaces/society-ethics/DiffusionBiasExplorer)! This demo showcases patterns in the images generated from different prompts input to Stable Diffusion and Dalle-2 systems."
-    )
-    gr.Markdown(
-        "See the results on how the images from different prompts cluster together below."
-    )
+    gr.Markdown(_INTRO)
+    with gr.Accordion(
+        "How do diffusion-based models represent gender and ethnicity?", open=True
+    ):
+        gr.Markdown(_CONTEXT)
     gr.HTML(
         """<span style="color:red" font-size:smaller>⚠️ DISCLAIMER: the images displayed by this tool were generated by text-to-image systems and may depict offensive stereotypes or contain explicit content.</span>"""
     )
@@ -199,9 +227,7 @@ with gr.Blocks(title=TITLE) as demo:
     with gr.Row():
         with gr.Column():
             cluster_id = gr.Dropdown(
-                choices=dropdown_descs[
-                    num_clusters.value
-                ],  # [i for i in range(num_clusters.value)],
+                choices=dropdown_descs[num_clusters.value],
                 value=0,
                 label="Select cluster to visualize:",
             )