section 1 edits

app.py

@@ -1,6 +1,7 @@
+import os
+
 import gradio as gr
 from PIL import Image
-import os
 
 _ID_CLUSTER_SCREEN_SHOTS = {
     19: ("cluster_19_of_24_unmarked_white_unmarked_man.JPG", "Cluster 19 of 24"),
@@ -92,7 +93,7 @@ with gr.Blocks() as demo:
         ### How do Diffusion Models Represent Identity?
 
         One of the goals of our study was to look at the ways in which pictures generated by text-to-image models depict different notions of gender and ethnicity.
-        These concepts are inherently difficult to describe, however: gender and identity are multi-dimensional, inter-related, and, most importantly, socially constructed:
+        These concepts are inherently difficult to describe, however: gender and ethnicity are multi-dimensional, inter-related, and, most importantly, socially constructed:
         they cannot (and should not) be predicted based on appearance features alone.
         Since we are working with depictions of fictive humans when analyzing text-to-image model behaviors,
         we cannot rely on self-identification either to assign identity categories to individual data points.
@@ -120,11 +121,11 @@ with gr.Blocks() as demo:
             Why do the only exceptions appear to be fast food workers and other lower wage professions?
             And finally, what could be the **consequences of such a lack of diversity** in the system outputs?
 
-            **Look like** is the operative phrase here, however, as the people depicted in the pictures do not exist, nor do they belong to socially-constructed groups.
-            This means that we cannot assign a gender or ethnicity label to each data point to support traditional measures of social diversity or fairness -
+            **Look like** is the operative phrase here as the people depicted in the pictures are synthetic and so do not belong to socially-constructed groups.
+            Consequently, since we cannot assign a gender or ethnicity label to each data point,
             we instead focus on dataset-level trends in visual features that are correlated with social variation in the text prompts.            
             We do this through *controlled prompting* and *hierarchical clustering*: for each system,
-            we obtain a dataset of generations for prompts of the format "*Photo portrait of a **(identity terms)** person at work*",
+            we obtain a dataset of images corresponding to prompts of the format "*Photo portrait of a **(identity terms)** person at work*",
             where ***(identity terms)*** jointly enumerate phrases describing ethnicities and phrases denoting gender.
             We then cluster these images by similarity and create an [Identity Representation Demo](https://hf.co/spaces/society-ethics/DiffusionFaceClustering)
             to showcase the visual trends encoded in these clusters - as well as their relation to the social variables under consideration.