July 22, 2021

Steam hovertrain departing from the station

This an experiment, where I did the same promt + same seed BUT each run tried different variables.
Each run been doing just 400 iterations, around 10 minutes. Not super-detailed, not blurry mess, possible to see the difference.

Link on collab

Lets begin from the initial image.

At this point I had all the parameters set to default.

And then, changes began to happen.
MSE Epoches: 5 -> 30

MSE Epoches: 30 -> 128

MSE Epoches: 48

At this point felt like this parameter should not be adjusted anymore.

Next thing to adjust - Cutout Value. Default value is 64, but I was running with 48 all these samples.

Cutout: 48 -> 10

Cutout: 10 -> 90

I like this result the most. The top left corner seems the only thing I'd fix in photoshop, but the rest looks almost like a complete image. People are gathering at the station where the levitating train is about to depart, they walk at the sandy road along the fences in fields; at the left it is possible even to recognize the sheep, at the background right area - some kind of buildings or city structures. You could even see the outfit of people who walk on the train: rich people in dark blue suits with etches of orange, some rich madame in fluff red dress and large hat. Amazing one.

Cutout: 90 -> 150

Image became very readable, yet it began to feel like ugly photo-bash without having solid structure.

Cutout: 76
MSE Epoches: 25

Time to play with image modifiers.
Flip Horizontal: 0.4 -> 0.1

Padding: 0.8 -> 0.5;
Affine random degrees: 30 -> 45;
Affine translate: 0.1 -> 0.4

Flip horizontal: 0.1 -> 0.95
Affine random degrees: 45 -> 75
Affine translate: 0.1 -> 0.55

Okay, it already looks the best for now. Consistent structure and composition, proper foreground and background, clear and readable details, I could even tell there's a horse-drawn carriage at the left area. Buildings doesn't look like a mess anymore, no flying islands of image in the sky. Yet, this steampunk shoe doesn't look like a train haha!

Affine random perspective: 0.2 -> 0.5