) > curate a training dataset (such as pictures of figurines) to feed a #Dreambooth finetune.
Once the model is ready, compare different generic prompts (e.g. "low-poly," "3D rendering“, "ultra-detailed," "pixel art," etc.)
2/ Once a prompt looks good, keep the modifiers (in this case: "3D rendering, highly detailed, trending on Artstation").
And start iterating around variations, such as colors (or pose). Don't over-engineer it, to keep the consistency. You should get this:
3/ A broader view of the red, blue, gold and green battalions :)
4/ Now you can get creative and add more colors (I just prompted "rainbow space marine, colorful" in the set below)
5/ It's not just about colors. The look or style of the models can be altered with other modifiers.
Such as "a dark Halloween zombie space marine". Now all the models have been zombified 🧟♂️💀🧟♀️. They could also melt or glow in the dark.
6/ I randomly found that one of the soldiers looked like a baby, so I prompted "a cute baby space marine" and made everyone look younger, in seconds:
7/ Because the shape/silhouette of the "baby space marine" was interesting, I kept it and used #img2img to generate pixel art variations, in different colors:
8/ img2img is very powerful when combined with Dreambooth. I went back to my previous tweet (link -
), and selected a "cyclopean golem", to resembling space soldiers, like this
9/ Same story below, but with a different (white) golem.
Also, I added "no gun" as a negative prompt. None of the soldiers has the machine gun they held before.
10/ As a bonus, you can have more fun with your finetune and generate other storytelling visuals.
For example, "urban warfare background" or "a group of space marines dancing in a night club"
11/ Or even find the image of a tank and then run #img2img to generate tank-like space robots/creatures 🪖
12/ I'm hoping this somehow demonstrates the power of having your own "finetune(s)"
Training lets you iterate around a specific concept, style, or object. You get consistent results, faster, and possibly using your training data. No more endless random prompting. What else? :)
end/ I've had some great discussions about it with the gaming community (artists or studios) over the past few days. If you have questions or would like to have a chat, please comment or DM me; happy to connect with innovative game creators.
The model was trained on just 11 pictures (!), with only 1500 training steps, which tuned out to be quick (20 min).
As before, the first step is to "explore“ the model with a few generic prompts. The goal is to find the modifiers that will keep a consistent style going forward.
Once the "stable modifiers" are found, it's time to select some of the best output and remove the background when needed.
"A dwarf, detailed, trending on Artstation, Clash of Clans"👇
As soon as the model was trained, the first step was randomly generating a large set of images w. a simple prompt ("golem, detailed, realistic, 3D rendering")
Each golem can be extracted by removing the images' backgrounds (reco: @photoroom_app).
Some of the designs are amazing. However, the golems all look very similar to each other. Let's separate them into categories.
I also used #img2img to instantly generate dozens of variants, "inspired“ by a single original photograph.
This provides consistent assets (similar shape, size, or materials) with some slight variations. It's up to the artist/user to select which one looks best.
Same thing here, using #img2img - however, I prompted "steel chest" instead of "wooden chest"
While it's not 100% perfect, there's still more steel in these chests than in the previous ones.
Also, some of the assets are disjoint or show anomalies. Some fine-tuning is necessary