New exploration: AI-generated fantasy game cards. We're rushing to ship our new product, and it's hard to keep up w. everything (4 threads are waiting!) This one started from a comment on my latest thread, mentioning #Hearthstone cards (@lewds_ai) ๐Ÿงต

@PlayHearthstone is a free-to-play online digital collectible card game developed and published by Blizzard. They make beautiful cards, and that's the training dataset I used (let's call this a #fanart experimentation). 20 images x 4000 steps, and I had the finetune 1h later

Now I wonder if I overfitted the model on this one, but there's a highlight. Every card on this dataset has the Hearthstone logo in the middle. The emblem represents maps, a compass, a spyglass โ€“ "everything a real adventurer needs." That's what I found ๐Ÿ‘‡ #StableDiffusion

The first runs are always with generic prompts ("a heroic fantasy card"). The output are 512x512 images, there's some background in each image (wall/stone), but the visual elements are kept (the logo, the outline).

By adding some modifiers, you can explore a bit further 'heroic fantasy card, intricately detailed, 3D render, trending on artstation, sharp focus"

Adding "gold" to the prompt and ๐Ÿ’ฅ all cards turn and golden.

Then I made some curious experiments by using a star (below) with #img2img. Because the shape of the star is very far from the shape of the original card, the results are a bit random. It makes nice jewels or badges, though. And they all have the Hearthstone logo (!)

Another similar batch, this one with more consistent "badges".

I went back to prompting with txt2img. "heroic fantasy card, intricately detailed, beautiful, 3D render, trending on artstation, 3D model, octane, sharp focus". Some cards are coming out with a white background.

And then it gets interesting when you do img2img prompting against a card with a white background.

You can play on the shape and symmetry too "heroic fantasy card, gemstones, stylized, natural, game icon, fanart, digital illustration, game icon, symmetrical"

And then (and this was cool), I prompted against a gemstone in #img2img. Like a green emerald: (notice that the Hearthstone logo is still preserved in most cases!)

A pink gemstone:

A golden gemstone (+ gold in the prompt)

This gemstone had mixed colors and was also rectangle-shaped.

A blue sapphire

Some sort of diamond gemstone

And the list can go on an on. There's no limit to imagination, you can try whatever you want.

And I also tried something different, which was to prompt against some of the original images in the training dataset. This generated some interesting variants for each card, take a look ๐Ÿ‘€ (right = original, right = variants). I did 9 each, but I could have done 100+

This was a variant I made before, "multiplied" into other variants.

To finish, these are a bunch of weird output I got by prompting against various shapes using img2img

I even got some 3D-rendered views

While there was less variability than expected, this demonstrates (again) that a custom-trained finetune of #StableDiffusion using #Dreambooth opens tons of creative possibilities in a short amount of time. We look forward to bringing this tool to every game dev in a few weeks.

PS: as always, RTs are hugely appreciated! Please help us spread the word, and let's get ready to heat up those GPUs :)

Follow us on Twitter

to be informed of the latest developments and updates!

You can easily use to @tivitikothread bot for create more readable thread!
Donate ๐Ÿ’ฒ

You can keep this app free of charge by supporting ๐Ÿ˜Š

for server charges...