Hey everyone, I've become very excited about Stable Diffusion this past couple weeks so I have built a website to generate images. The server is written in Nim of course. Check it out at https://theartbutton.ai
Examples:
If you haven't heard of Stable Diffusion, it's a trained machine learning model that can generate art based on a prompt. It was made freely available which is really cool. There's a huge amount of enthusiasm for this and it can generate some really cool stuff if you put together a good prompt. See examples on https://lexica.art
After generating an image you can share the link so that others can try out the prompt, or download and share the images. Please give it a try and see if you can generate something neat.
Here are some funny results I got with the website (exact prompts in the website):
Russian building https://theartbutton.ai/image/iSsWgGXfI14fKtCgLcmNo90a04qCjmt2
Di Caprio, art by Jeff Koons https://theartbutton.ai/image/hgu7ktzukFgz2KMuqEg29YXLPXlDuiF1
Fireworks above Lake, black and white photograph https://theartbutton.ai/image/1vIL7Rlz2OSWbdPxK0xsPhVXAElw9zhj
New york, photograph, black and white, art by Félix Nadar https://theartbutton.ai/image/lw23MDiRoyvlYgVadQQJGRKk4MiRS5BQ
Colored rounds, art by Paul Klee https://theartbutton.ai/image/6lRnuoVvoqpFiCXaAG0f8DpoX5lheU3x
In general, A.I. (at least this model) has difficulty with three things:
- hands. So many portraits have too much fingers which are too long or awkwardly shaped. Di Caprio got three legs but I swear that is better compared to all the weird hands I got.
- format. Many portraits are cropped in the default square format. I have not tested other pictures formats.
- perspective. The buildings lines aren't lines, they are wavy.
My curiosity has been picked by data style transfer, which looks quite amazing too. Maybe we could create some styles, acting like instagram filters, to add as overlay on resulting images ? This would give us a little more control on the picture/reduce resources for frequently asked art styles (Picasso, Warhol, etc …)
I made a video showing how to use AI to make 2D assets:
https://www.youtube.com/watch?v=81phsupsoZE
AI has completely changed the way gamedevs should create game assets. Now you can generate amazing assets with a simple workflow.
In this video I show how I'm genrating some cool 2D space ship assets. The same workflow I show in the video will work for all kinds of things.
Some key points to remember!
- A good prompt is key. It's ok to improve it as you're working.
- Generate lots of images! You may have to go through many to get 1 good one.
- Tweak the one you like to make it better and generate a lot more!
That's the basic concept!
And yes! Everything on the site was made with Nim except for the Stable Diffusion AI itself.
It's very impressive and frightening.
If only I wasn't so terrible at describing the images I want it to create. :-)
The one thing to be wary of with these composing systems is for example the golden emblem at 4:40 that seems to be very close to a direct rip from the overwatch ranked system. (Even has the token stable-diffusion blurred text at the top). This isn't super important as you can continue to iterate ofc, but it is interesting to be wary that just because it is new to you, and reverse image search does not find results, does not always imply its uniqueness.
Here's a comparison between an original art on the left, and some possible outputs.
Just a heads up for anyone using it in production. And seeing as my avatar is also from stable-diffusionMK1, uniqueness is still possible ofc. (Not strictly Nim related I apologize, but you could always throw in a vague analogue about compositionality and language design if desired.)
Not strictly Nim related I apologize, but you could always throw in a vague analogue about compositionality and language design if desired.
No worries, still roughly on topic.
Just a heads up for anyone using it in production.
Can you really? If it's not made by a human, it can't be protected by copyright law, so you can't claim ownership, including derivative works. Am I right?
- I wasn't using a disposable email, it's just that I use DuckDuckGo's email service. Turns out duck.com is in the list of disposable email list.
Can you really? If it's not made by a human, it can't be protected by copyright law, so you can't claim ownership, including derivative works. Am I right?
Why would it matter? Any game built with these graphics would still have plenty of human work inside and would be protected by copyright law.
I made a new video tutorial here: https://www.youtube.com/watch?v=erU-DMOC4k0
One thing that can be difficult when generating assets for a game is keeping the style consistent.
In this video I show how I generate multiple variations of an asset while keeping a consistent style between them.
The key to my workflow is:
- I start with an image I like and want more variations based on.
- I use an image editing tool or TheArtButton.ai's image editor to create examples of the general idea I want. I call this image my template.
- I generate a bunch of images using this template and pick the best ones!
You can repeat this process as many times as you want to get more variations or continue to refine the style.
Ask me anything about AI art and will try to answer.