Let's Talk About AI

[Ricardo2020]

It's 2025 and creating art has become an ultramodern process. I got curious thinking about the changes in the artistic workflow since I first took up a pencil to draw around the age of three.

This inquisitiveness in my brain made me follow a thought path online wherein I encountered artificial intelligence enabled software tools, one in particular being known as Stable Diffusion. I learned that it can run in a browser as a web page plugin which needs no installation and is totally not enshttified, being hosted on a website called perchance.org, wherein I found their AI Body Generator.

This web tool will generate images based on text prompts. You do not get to determine the size of the pictures, but you can somewhat drive the content and render either in landscape or portrait mode. The images are rather small at 1024x768. However, they upscale nicely in the GIMP due to low compression of the renders. But there is where the parade stops for those of us who are seeking accuracy in the final render. Things happen to ruin the results in over 80% of the renders. I shall list a few below.

Too many fingers, hands, arms, legs, feet, toes, and fat rolls.

Improper or impossible poses and situations.

Broken laws of physics, too many to mention here.

Anatomy faults in general, such as improper or impossible proportions and features.

Now, this is not to say that the tool does not get it right and, well, sometimes, it does. When this happens, one can end up with a really good render with the textures, light and perspective all working together to give the composition a photographic look. About one in four images turn out OK. Of these four OK ones, one of them will be good enough to use. Thanks to the oomph inside the servers where this runs, render times are short. On my workstation, it takes less than 30 seconds per render.

The tool lets you make art in many styles, from pencil sketches to pro quality photos. And like I have said, when it gets things right, the results are quite good. However, most of the time it hallucinates, yielding pictures that many times look as though they were ripped from a bad dream or a drug trip. I attribute this to two things:

Training data is enshitified by crappy social media posts and ads. Since there is no reasoning ability, it gushes nonsense most of the time based on scrape sums.

There are random loops in the code that run without error correction. The code set does not branch predict accurately since the training set is limited and faulty at best.

I say all of this to tell you that AI is a long way from replacing actual talent when it comes to making art. Until it can be controlled and used in a way that is predictable which produces quality work, it should not be incorporated into any software that is under the full control of actual humans. It cannot yet replace us.

Computers must get 1000% more powerful before AI has a chance. Also, the data used for training must be cleaned up and made orderly so that the code does not stumble into a rabbit hole of complete jibber jabber. Also, marketing cannot be part of the equation at the user level. The ads cannot be allowed. The social shares must not happen unless one can opt in or out at will.

I hope we can keep Blender and MakeHuman as open as possible and avoid the temptation to enshitify it with AI. We have much to do outside of AI, and it will be a while before we actually need to think about where it fits within our community. In short, it is just too dirty with corporate greed and not ready for prime time among open source artists.

Let companies like Adobe and friends put the crap into their code. Folks will actually pay for it and that is why they do it, becoming locked into a club of people willing to be products bought by ad wranglers via enshitification of the software.

For us, Blender is the better choice. And don't forget about the GIMP when it comes time to do post edits of images. No ads, No hidden links. Just solid tools made for everyone to freely use. And the best part of all this is the cross platform compatibility. Yes, we can all be friends.

If AI goes in there, expect to pay, with more than just money...