photo library example
01_1_007_people
original label
“what is low raise and support” – Paradise Lost, Book 1, line 23.
A man with eyes closed floats on his back at the bottom of a swimming pool. His left foot is over his bent right knee and his right arm is tucked against his side as his left arm reaches up to the surface. Behind him another young man with his back turned has his arms raised. The water corrugated with light and shadow seems to support their bodies.
‘good positive prompt’ label
two men, men in water, floating, curled up, arms outstretched rises, bright day, outside, photo
prompts
positive prompt with three versions
- men, women, diverse ages, diverse skin tones, diverse body sizes, diverse body shapes, some men with beards, some men and women with long hair, all falling, some curled up, all eyes closed, all appear asleep, all falling from an expanse of dark sky, some nearing an expanse of water, some crashing onto water, dim light from sky, reflections from fire on water, photo, PRDSLST
- men, women, diverse ages, diverse skin tones, diverse body sizes, diverse body shapes, some men with beards, some men and women with long hair, all falling, some tumbling, some curled up, all eyes closed, all appear asleep, all falling from an expanse of dark sky toward an expanse of water, dim light from sky, reflections from fire on water, photo, PRDSLST
- men, women, diverse ages, diverse skin tones, diverse body sizes, diverse body shapes, submerged in water, seen from below, seen from the shoulders down, some legs curled up, some legs dangling, some legs and arms outstretched, some fists clenched, light ripples through expanse of water, reflections from fire coming from above on water’s surface, photo, PRDSLST
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
summary
These images are a selection from Ilia’s test when he trained two models. Each using my positive prompt with three versions, his negative prompt and a set of 77 photos with their labels. One model had my original photo labels, another had my new labels using his ‘good positive prompt’ structure, as per the example of Bruce Hall’s photo from my library. To demonstrate the effect of a LoRA’s training time, Ilia provided outputs from multiple training steps. Specifically, base output, 100, 300, 500, 700, 1000, and 1200 steps.
To try and understand AI bias, I removed all descriptions from the photo labels relating to body size and age. The labels have words and phrases that are used in the three prompt versions. This consistency was intended to improve the chance of the generated images having aesthetic qualities I wish to achieve and that focus on the scene from Paradise Lost Book 1 relating to the fall and landing on the lake.
My prompt with three versions was written sequentially visualising the depiction of Satan and his crew ousted from heaven falling onto the burning lake. As a photographer, I wanted to understand how AI image generation understands composition. The sequence is sky/space > falling/tumbling > floating/submerged > expanse of water/burning lake. Each prompt has the tag PRDSLST to link with the photo library and the PaRaDiSe LoST extract within each photo’s label.
The images generated show lack of diversity and a very stereotyped portrayal of minority groups, age and gender. This likely stems from the biased way the training data was originally collected for the base model, which raised questions in my Explore phase about AI’s “black box”. We know what goes in and we know what comes out, but the inner workings of this “black box” are unknown to most people.
flux-lora-collection / realism_lora_comfy_converted.safetensors
image & prompt_1_1
base output – pos_prompt1_1_00012
500 output – pos_prompt1_1_00006
1200 output – pos_prompt1_1_00007
men, women, diverse ages, diverse skin tones, diverse body sizes, diverse body shapes, some men with beards, some men and women with long hair, all falling, some curled up, all eyes closed, all appear asleep, all falling from an expanse of dark sky, some nearing an expanse of water, some crashing onto water, dim light from sky, reflections from fire on water, photo, PRDSLST
image & prompt_1_2
base output – pos_prompt1_2_00007
500 output – pos_prompt1_2_00006
1200 output –pos_prompt1_2_00005
men, women, diverse ages, diverse skin tones, diverse body sizes, diverse body shapes, some men with beards, some men and women with long hair, all falling, some tumbling, some curled up, all eyes closed, all appear asleep, all falling from an expanse of dark sky toward an expanse of water, dim light from sky, reflections from fire on water, photo, PRDSLST
image & prompt_1_3
base output – pos_prompt1_3_00013
500 output – pos_prompt1_3_00015
1200 output – pos_prompt1_3_00013
men, women, diverse ages, diverse skin tones, diverse body sizes, diverse body shapes, submerged in water, seen from below, seen from the shoulders down, some legs curled up, some legs dangling, some legs and arms outstretched, some fists clenched, light ripples through expanse of water, reflections from fire coming from above on water’s surface, photo, PRDSLST
Karren Visser 