Karren Visser

prompts

---

positive prompts

Blindfolded white young and old men. Blindfolded black young and old men. Blindfolded brown young and old men. Blindfolded yellow young and old men. Blindfolded white young and old women. Blindfolded black young and old women. Blindfolded brown young and old women. Blindfolded yellow young and old women. All blindfolded men and women sit on grey rocks facing a pale green tree without leaves to their right. No grass. No water. In the pale green tree without leaves is a bright red long thin half man half woman. In the dark blue sky above are rows of human eyes. 

Good positive prompt structure:
Subject description > Action description > Scene description > Style description > Quality description

---

negative prompt

monster, fantasy, sci-fi, extra people, poorly drawn, extra limb, floating limbs, airbrush, photoshop, plastic doll, cropped, out of frame, worst quality, low quality, jpeg artefacts, duplicate, extra fingers, poorly drawn face, extra limbs, cloned face, extra arms, too many fingers, caricature, facebook, youtube, food, lowres, text, error, poorly drawn hands, blurry, extra legs, username, watermark, signature

base model: Stable Diffusion 3.5 Medium in ComfyUI using new training methods

---

summary

---

These selected images from Ilia’s test use a combination of my positive prompt to explore AI bias, based on his ‘good positive prompt structure’, his negative prompt, and a set of 77 photos with their labels.

The test was to explore the lack of diversity and stereotyped portrayal of minority groups in AI generated images. This likely stems from the biased way the training data was originally collected for the base model and the influence this has on a LoRA.

As the project progressed, we started to realise how LoRA alone might not be enough and a better understanding of the base model is required. We considered how unnatural prompt writing is, and the way the machine is coaxed into giving the desired results is far removed from traditional writing styles. This led to the conclusion that prompt engineering is possibly as important as understanding how the model is trained.