I created a generation of the one of the most basic texts possible, "woman."

After I had inserted the simple prompt of woman, it generated something potentially NSFW. After generating this, the images generated after became more provocative, and had more of a clear bias of something that was probably scraped from NSFW material online. The prompts below still had the prompt of "woman."

The prompt below was "A woman"

The images generated from this prompt showed white women, all with the same build and face (same facial structure, eye shape, staring in the direction of the camera) and having a sort of sultry expression. This was interesting to see, as there was no diversity in racial features, or even diversity in the features of the white woman generated - other than maybe eye color. It often depicted a skinny woman with smaller features - a stereotypically beautiful white woman. It did not show the "mean" woman that would be expected, and so the images that this generator is scraping from is choosing from a pool that focuses on Eurocentric depictions of beauty

Next Prompt: Black Man and Black Woman (Realistic and Anime art style) and Asian woman

The anime art style is evocative of a common amalgamation of Japanese Animated art, but the braids on the black woman's hair in the anime art style did not connect to her hair at all, nor did her hair make too much sense in terms of connecting to her. This is likely due to the AI generator getting confused on what aspects of information it needs to scrape, and so it did not fully render out a complete image.

The realistic art style showed better depictions of people of color, but I had to put the descriptors of their race in there in order to get an image with people of color. The prompt with just "a woman" had no indicators, and so it only generated white women with the same features. There was no diversity in its creation, and tried to make something stereotypically beautiful with Eurocentric features.

The observations gathered from these image generations are that Stable Diffusion is likely scraping from a lot of image sites yes, but if it does not have some sort of categorization or narrowing down, it will default to making a caucasian woman, but also have the images more evocative of a pin-up image or even generate some NSFW content. I decided to diversify the images being generated, and the next prompts I decided to generate plus sized people with a simple prompt as well:

Next Prompt: Fat man and fat woman

After the NSFW content warning, the image for the prompt of "fat woman" (no style) showed an image of a woman in her bra with her nipples showing out of it, pretty much leaving nothing to the imagination. Additionally, the cinematic styles had them facing the same direction, while having no style had them facing forward. This image generation for no style showed a lean towards creating things that were a little more raunchy than other styles. This begs the question - are images generated with no bias being scraped evenly online, or with a large NSFW presence? After this, I decided to generate the simple prompt of "a man" and seeing where that would get me.

Next Prompt: A man (no style and anime style)

This was interesting to see - similar to the basic "a woman" prompt, the male prompt also had the same white man, with more facial similarity even than the women, and making him generically attractive. The third image shown did have smoe kind of sultry presence, but in general, the men were more SFW than the women. This difference in image generation showed something about how the images were being generated - that the women would be more likely to be sexualized than the men, and this is shown in the plus sized prompts as well. The faces of this man were very similar as well, with having the same facial shape and beard, and hair style. The anime prompt brought some diversity, but the first image generated had a blue eyed brown haired man, with the same going for the anime generated one.