An AI's Paintbrush

A blog for my experiments with generative AI

[aispaintbrush]

Hello all, welcome to my little corner of the internet!

This blog is a place where I share my experiments with various generative AI, such as VQGAN + CLIP created by Katherine Crowson, CLIPIT PixelDraw made in a collab with Dribnet and several others.

The whole purpose of this blog is to share interesting/beautiful/terrifying/etc. images made from this AI. If you have a request for a generation (requests are free!), please include the phrase you want to generate (no NSFW), along with the AI you want it generated with (More information behind the Keep Reading link!).

VQGAN+CLIP (or as I like to call it, Painter) is an AI that takes any phrase given to it and attempts to make an image out of it. It can also accept images to build off of, or you can make it attempt to recreate the image in its own style. I have recently discovered that adding the " | " character in between the objects you want generated usually results in better generations. Otherwise, the AI will try and merge all the words together as a single object and try to generate an image with it. I recommend 200 generations as a minimum. The lower the number, the less detail Painter generates. Conversely, please keep the number of generations under 5000 - I could theoretically do that amount, but it risks the AI crashing, not to mention that more generations means a longer time to make the final image (1000 generations takes up to 30 minutes!). If you want to generate something using VQGAN+CLIP, please include the number of iterations you want it to run for, otherwise I’ll just default to 1000 generations

CLIPIT PixelDraw is similar to VQGAN+CLIP in that if you give it a phrase, it will generate an image from it. However, instead of making high-quality images with more abstract themes, CLIPIT makes low-res pixel art. Ironically, almost all generations using CLIPIT will be recognizable as the phrase you gave it. It still has a difficult time generating faces, however, it does far better compared to VQGAN+CLIP when it does. CLIPIT can generate images in two aspects: Widescreen [500 x 280 pixels] and Square [375 x 375 pixels] . Widescreen will typically try to encorperate as much as it can into the image, even if you specified for there to be one object in the prompt. Square typically only tries to include one thing as the object, however, it is not impossible to get multiple objects from this size. If you want to generate something using CLIPIT PixelDraw, please specify the image aspect: Widescreen or Square. If you want to generate more than 500 iterations, feel free to add the amount you want in as well

[aispaintbrush]

“digital art of a wooper”

50 generations

[aispaintbrush]

“an oil painting of a jagged mountain range”

50 generations

[aispaintbrush]

“an oil painting of the sun in the void of space”

50 generations

[aispaintbrush]

"a tear in time and space"

1000 generations

[aispaintbrush]

"a depiction of a pokemon. unreal engine"

1000 generations

[aispaintbrush]

"a geode. unreal engine"

[aispaintbrush]

"pokemon"

1024 generations

[aispaintbrush]

"a neutron star. unreal engine"

900 generations

[aispaintbrush]

"a clockwork dragon breathing fire by ross tran"

1000 generations

[aispaintbrush]

"SCP-096" 1000 generations

Decided to edit it for spooky month:

[aispaintbrush]

“Undyne”

500 generations

[electricturtle]

polar bear

i always like the first attempts of the model... they look ethereal

[aispaintbrush]

tried the same prompt - got a weird looking sheep:

1000 generations

[aispaintbrush]

"antimatter" 1000 generations

[aispaintbrush]

"the butterfly effect"

1000 generations

[aispaintbrush]

"an old well in a fantasy forest"

1000 generations

[aispaintbrush]

"an industrial ocean"

1000 generations