«   |   »

Before You Were Born

So I’ve been having a certain amount of fun, suffering (“suffering”) a certain amount of boredom, doing various things. The weather was lovely yesterday! But also I was oncall at work. :) But also I was in Manhattan! It all balances out. I’ve been posting some quick chatty things on the schlaugh, which is fun (I don’t think I said anything yesterday) and completely non-algorithmic, which is refreshing.

I’ve been making the usual dozens / hundreds / thousands of pictures on Midjourney, saving a tiny fraction of favorites to my Windows wallpapers folder (currently just over 300), and posting a fraction of those on PixelFed. So many pictures!

The U.S. Copyright Office is seeking public input on this whole “AI tools and copyright” thing that I used to think about so much, after finding yet again against Thaler‘s claims that his software does it all by itself. I think Thaler is really an annoying distraction here; it gets everyone thinking in terms of “art created by an AI” rather than “art created by people using software tools that incorporate AI”; the latter seems far more accurate.

And Midjourney has added some features! Perhaps most notably, you can now indicate an area of an image, and point the engine at it with an optionally-modified version of the text prompt; which means that if you get say a great image except for the hallucinated copyright / watermark / signature and some horrifying hands, you can indicate those areas and have it try again.

Here, for instance, is a fun image created from the prompt “Before you were born”, which has exactly those two problems:

A blond baby standing next to an interesting scaly monster with weirdly-placed eyes and many sharp teeth. The baby is holding an orb and some sea plants or something. The baby seems to be holding the things with three different hands, and there is a vague watermark saying something like "OOSAND jlok.com" neat the center.

Note the rather scary third hand, and the annoying hallucinated watermark.

I used the new “Vary (Region)” thing in Midjourney to draw a rectangle around the watermark, and added “–no watermark,signature,text” to the prompt, and got four new images which didn’t have the watermark, and had four slightly different fillings-in of the place where it used to be; I chose my favorite one of those. Then I used the lasso-select tool in “Vary (Region)” to select a small irregularly-shaped area around the extra hand, didn’t change the prompt at all, and ran that. Of the four results, my favorite was this:

A blond baby standing next to an interesting scaly monster with weirdly-placed eyes and many sharp teeth. The baby is holding an orb and some sea plants or something. The baby has only two hands, and there is no watermark evident.

If you know where the extra hand used to be, you can still sort of imagine that some pink stuff there might still be a bit of it, but I think without knowing, the viewer will just see that as more of the strange sea-stuff around the orb.

So that worked pretty well!

And now we know where we are before we’re born. :)

Update: The pictures above are from Midjourney 5.1, which is my favorite version 5; but I gave the same prompt to 5.2, and I thought this idea of the pre-birth state was also notable…

A baby standing up in some kind of futuristic environment suit, looking upward cutely. Behind the baby and stretching off into the distance are many shiny orbs, some with orange-white sparks some also with some kind of filament leading to a clump of something. (Spines and brains?)

Posted on 2023/09/01 at 2:35 pm in ai, quotidia, tech   |  RSS feed |   Reply   |   Trackback URL

Tags: , , , , , ,

5 Comments to “Before You Were Born”

    
  1. Karima Hoisan
    2023/09/01 at 2:59 pm

    How nice you can kinda edit these images without using new prompts. My brain sees the left over pink stuff but that’s because it had already seen the third hand ✋🏽. These new tools seem important. By the way, where I was before I was born had mostly blue including my adorable outfit. 👶

    Liked by 1 person

    Reply
  2. mattwassel
    2023/09/01 at 4:22 pm

    Why do AI people so often have extra hands or extra fingers? I asked this of ChatGpt and got a wishy-washy at best answer. So if more and more of these images appear, won’t AI learn from them and make it even more likely that an AI person will have extra hands and fingers?

    Liked by 1 person

    Reply
    • ceoln
      2023/09/02 at 12:49 am

      There are all sorts of theories about hands. Mine is that it’s the main thing where we care a lot about an exact number, and the exact number is greater than two (five, in particular). And then that the models aren’t good enough at numbers to get any number greater than two reliably correct. Although they are getting better!

      And for sure, there is a general worry about what happens when AI models are trained on data that includes increasing amounts of stuff produced by AI models. There’s a whole thing called “model collapse” that people study, even.

      Liked by 1 person

      Reply
    

Hm?