"Confliction"

This image is about the conflict that resides in all of us between the light and the dark. In this image, she straddles the boundary between good and evil and feels the pull in both directions.

She is torn between two opposing and irresistible forces. And yet she resists! She wears her confliction on her face. Her body is an outward reflection of her inner turmoil. She yearns for peace that will never be found.

Prompt Strategy

I felt strongly I'd never be able to realize my vision of nested split images through a simple one dimensional prompt. I wanted to have our subject with one foot in a world of darkness and the other in a world of light. And I wanted herself to be the the inverse of that - her dark half is in the world of light and her light half in the world of darkness.

This seemed like the perfect opportunity for me to learn how to use regional prompting in SwarmUI. My idea was that I could create a prompt for the left half of the image of a white angelic girl in a dark abyss, with tears of despair flowing down her cheek. And I could create a separate prompt for the right half of the image of a dark demonic girl in an ethereal place, complete with a wicked grin of delight.

The way you do regional prompting in SwarmUI is with the <region> modifier. With that, I can define the following prompt structure:

global prompt that gets applied to the entire image. I put in here image descriptions that complements the 2 regional prompts I'll use.  I avoid adding anything that contradicts the regional prompts.

<region:0.5,0,0.5,1,1>Prompt for the left side of the image.  Here I describe the angelic girl and the dark space.

<region:0,0,0.5,1,1>Prompt for the right side of the image.  Here I describe the demonic girl and the light space.

<segment: face | hair>describe the face and hair, copied from the global prompt for the auto-segment detailer

What happens during generation is that basically all 3 prompts get fed into the model. For this generation I gave the regional prompts 100% of the weighting so in theory the global prompt does not matter. But in practice, the model uses the global prompt to coordinate the 2 regions so that they generate a cohesive image. This is why I use tags in the global prompt to reinforce the dichotomy. I want the global prompt to understand and expect the girl and scene to be divided.

Prompt

Here is the actual prompt I used. Note it is missing the quality tags, because those are in the 174 presets for the different models. You can download the attached presets.zip if you want to see what all the settings and tags were for each model.

underlighting, 1girl, solo focus, split-color hair, split-color clothes, split-color background, heterochromia, glowing eyes, glowing hair, messy hair, drill sidelocks, skin tight evening gown, evening gown, thighhigh front slit, glowing evening gown, panty & stocking with garterbelt, standing, spread legs, cowboy shot<break>
crying with eyes open, tears from one eye, evil grin, mismatched horns, broken halo<break>
straight-on, facing viewer, outstretched arms, palms up, (holding orb:1.5), glowing orb, purple aura, white aura, chest tattoo, glowing tattoo, right bat wing, left angel wing, ethereal background, white background, abyss background, black background<break>

<region:0.5,0,0.5,1,1><var:q><break>underlighting, 1girl, solo focus, glowing purple eyes, glowing black hair, glowing purple hair, messy hair, drill sidelocks, evil eyes, evil grin, skin tight evening gown, black evening gown, thighhigh front slit, glowing evening gown, black panty & stocking with garterbelt, standing, spread legs, cowboy shot<break>
demon horns, straight-on, facing viewer, outstretched arms, palms up, (holding orb1.5), glowing purple orb, purple aura, black chest tattoo, glowing tattoo, bat wings, ethereal background, white background

<region:0,0,0.5,1,1><var:q><break>underlighting, 1girl, solo focus, glowing blue eyes, crying with eyes open, tears, sad, tears from one eye, glowing white hair, messy hair, drill sidelocks, skin tight evening gown, white evening gown, thighhigh front slit, glowing evening gown, white panty & stocking with garterbelt, standing, spread legs, cowboy shot<break>
halo, straight-on, facing viewer, outstretched arms, palms up, (holding orb:1.5), glowing white orb, white aura, white chest tattoo, glowing tattoo, angel wings, abyss background, black background

<segment:face | hair,0.3,0.5>portrait, 1girl, heterochromia, split-color hair, glowing blue eyes, glowing purple eyes, crying with eyes open, sad, tears from one eye, evil eyes, evil grin, glowing white hair, glowing black hair, glowing purple hair, messy hair, drill sidelocks

Results

Here's a link to a google sheet where you can view all of the gory details, including list of models with links to CivitAI model cards.

Now I'm going to go through each of the realism groups from 1 to 5 and display the top 3 ranked images from each realism group. Some groups had some ties so you may see more than 3 images for that group.

After the 5 realism groups I'll show you which image was best overall. Then there will be an honorable mention section where I'll paste mini-grids of all other images that didn't make the top-3 for their group. I encourage you to browse through them because there were a hell of a lot of good images that came out of this grid.

CivitAI shrinks these images down to where the labels aren't really legible. If you right-click an image and View Image in New Tab you can see the full-sized image and read the label to find out the model name, scores, and my notes about the results.

Realism 1

There were 31 images I graded into this bucket. Two images tied for the top spot. And two images tied for the rank 3 spot.

1-03-4.50-Hyphoria [Illu NAI] - v001.png

Realism 2

30 images fell into this bucket. Here are the top 3.

2-01-4.80-Little Monsters - Illustrious x4.png

Realism 3

This was the biggest bucket with 57 images. It also had the highest average image score of any category. There was ALOT of competition in this group.

Here are the top 3. Actually two of them tied for rank 2.

3-01-4.90-Cyberpunk Art Illustrious - 1_0.png

Realism 4

38 images fell into this group. In general the overall quality of the images take a noticeable drop in this group. Seems like the more real you go, the more trouble the model has at following your prompts. But there are still some notable exceptions. Here are the top 3. There was a 3-way tie for rank 2.

4-02-4.20-Little Monsters - Illustrious x2.png

Realism 5

This group, the smallest with 18 images, really struggled to deliver for this prompt. Here are the top 3.

5-01-3.30-Nova Reality XL - Illustrious v40.png

5-02-2.30-Illustrious Renaissance - V2 Beta A17.png

The Winner

And which of the images got the best score overall? It was the Cyberpunk Art Illustrious model. It achieved 98% on prompt adherence, getting only a single detail wrong.