There are many like this but this one is mine
I started and couldn't stop myself from doing more. These took maybe 7 hours of generation and tweaking in total (I can't bring myself to call it 'work').
Full album (but without commentary):
https://imgur.com/a/I7rOEs7Baird: Not too hard, glasses took a bit, and I think it captured her well.
Sebestyen: Was tricky to keep his expression, lost a bit of it. He's a little too saturated but overall it captures him.
CEO Sun: Eyes and making her not completely Caucasian were kinda tricky, otherwise fine.
Young Andrada: I spent AGES on Andrada and this is the best I could do. The AI really really wanted to make him look like a Civ leader. Something about his vanilla sprite.
Not thrilled with this one.
Current Andrada: Same as above.
Daud: Doesn't perfectly capture him but I was getting tired.
Cotton: Making photo-realistic ones is hard when they have sci-fi things that aren't really in training models. Rather than try for photoreal, I left it at this, with most of the vanilla features (headset band, scar even though it's awful) intact.
Coureuse: Same as above. Had to balance realistic with the green doodad.
Zal: Fairly happy with this one.
Gargoyle: They are a really hard one because of all the gadgetry and atypical style, so I'm proud of how this came out. Still a bit stylized, but it keeps almost all of their features intact while being a bit creative (like the visor shape).
Jaspis: Fastest of them all to make, one or two minutes. I'm sure I spent over an hour on the Andradas.
Kanta: Similar to Gargoyle, very tricky subject and I "rerolled" individual parts a lot to get it to where it is. Pretty pleased with the result.
Loke: Scars are hard. The face shape and expression was hard. One of the worst of the bunch.
<SUPER REDACTED>
Volta: Astonished that it could do this at all.
<redacted> captain: This one was a lot of fun. Ended up with a lot of cool-looking images, but this had a lot of floaty energy and doesn't show too much face.
MethodologyI am by no means an expert on this and have taken the "screw around and find out" approach to learning.
Ran on my own PC using Automatic1111's Web UI.
Model: Mostly Deliberate v2. Used Liberty just for the Andrada ones because Deliberate was making them even more video game looking.
Hypernetwork: none for most, Midjourney for the zig captain.
Sampling Method: usually DPM++ SDE Karras. Euler a was sometimes better for initial upscale.
25-50 steps for each. 7 or 8 CFG. Using img2img, start with a low Denoising Strength as an upscale, then increase the denoising on the result. Used inpainting for tweaking.
Prompts almost always included "8k, ultra realistic" and negative prompts were "cartoon, 3d, blurred".
I did no manual editing or post-processing (so imagine what an artist could do).
This was a lot of fun to do and brought the NPCs to life in a different way, like seeing the movie after reading the book. I've used it to to guided upscaling of older game sprites, namely from EV Nova, and it's quite satisfying. Not useful in any way, but ...satisfying. For older game art, it sort of creates the image that we remember it as; back in the day it didn't matter that it was 64x64, that was the norm. Seeing it today, it looks dated, but upscaling it bridges that gap of years.
BONUS: Kanta's TriTinder pic.
I don't know if I have any rights to these images but if I do, I give them all to Fractal Softworks. Welcome to the new world of AI, where the rules are made up and copyright doesn't matter. They're coming for my job, too.