I tried to decipher the hidden language of Midjourney
I wanted to see if Midjourney also has a "hidden language" like DALL-E 2 in this post: https://programming.dev/post/102011
So I ran a little experiment.
I gave this prompt from the article to Midjourney:
Two farmers talking about vegetables, with subtitles --q 2
But it didn't produce any text:
Then I tried this:
text logo of fitness company including motto --q 2
This gave me what I wanted: logos with text.
Then entered the nonsensical words from one of the logos:
FRVNE MIASE --q 2
This triggered an abuse detection filter which I appealed. Then Midjourney produced these equally nonsensical but absolutely wonderful images:
First I thought that the results had nothing to do with the original prompt, but if you look at the logo, it has mountains in it, so maybe "FRVNE MIASE" means mountain?
I don't have more time to play around with this but if someone else can get further with it, I would love to see the results!
This sent me down a rabbit hole of trying a lot more gibberish/unconventional prompts on Midjourney. Thereβs something fascinating about seeing such a complex response thatβs clearly an unintended side-effect of the training process, even though the results have a pretty conventional βMidjourney lookβ.
Putting in your username is also an amusing exercise. This is what I get for babelspace The βspaceβ influence is pretty obvious, I think itβs associating babel > babble and babies, which I assume is where the color scheme and youthfulness comes from.
If anyone without a Midjourney subscription wants to see what their username looks like, I would be happy to run it for you.
I'm not a native English speaker so I have a hard time coming up with nonsense words that have a certain "feel", so I asked ChatGPT to generate some for me:
https://chat.openai.com/share/c1d553dd-a517-4a17-8473-72afaddd0e87
Here is what happened when I fed them into Midjourney:
Sereneity and sonderance
ChatGPT definitions:
Midjourney prompt:
Midjourney output:
Dystaust retrofrisson
ChatGPT definitions:
Midjourney prompt:
(I needed to add some actual words because it stopped with an error without them.)
Midjourney output:
Conclusion
The images are amazing and they pretty accurately represent what ChatGPT meant!
That's neat!
To me the common element in your picture seems to be some sort of orange teardrop shape. In some cases itβs a balloon, in others itβs a mushroom.
Youβre right! And it appears in all of them, regardless of the prompt. Maybe it uses this shape for any weird or incomprehensible prompt.
This is so exciting, like decoding ancient Egyptian. Who knows this might even end up proving true artificial intelligence.
Decided I would try this with Stable Diffusion - Emaonly 1.5
::: spoiler results TUPTLE CRAPLIU -
Tiupli - bullshit
CHRMLLNIA - :::
So yeah, I don't think SD has anything of this sort. The results are similar to if I type in complete bullshit: OPSKAEOPSAEKSAELMAESOJSAEKLHWQEWQECOIUWEOIWQODWDJHSDMWOIEUQWRRNFJKBEIURFYHQOEIEJFOLEIFDJALSKDJPAWOJODIJDIWHFN
::: spoiler results :::
Also typed in my username, as some commented in the thread. I got a lot of pictures of singers who looked suspiciously like Abel Tesfaye. My other username, which obviously has "deer" on it:
::: spoiler results :::
My old discord username, which has nothing to do with buildings whatsoever:
::: spoiler results :::
Interesting! It seems like the claim in the original article is only true in very specific cases, if at all.
So... I don't know that much about it, but as I understand it, this isn't really a "hidden" language in the sense of something deliberate, but more the outcome from stimulating the language-understanding network in a way it's not been trained on. If it gets a string of tokens that's just some nonsense, then it'll tend to light up its output nodes in some semi-random pattern that's nonetheless slightly consistent depending on what the input is. Then, that somewhat-random pattern of nodes will go into the art-generating network, and stimulate some random set of nodes related to what art it wants to produce, without adding up to anything coherent. And yes, the results (if e.g. you feed it just a string of gibberish from a nonsense word generator) are fairly wonderful:
I typed in more of the weird text from the logos with the following results:
Prompt:
Output:
Prompt:
Output:
Conclusion:
While these are all beautiful images, they have nothing to do with either the original fitness prompt or the logos. Though the first prompt seems to consistently mean "frog", and the second one "a girl playing music".
For both of these prompts, I received a warning from Midjourney:
And I had to appeal. I won't do any more of these experiments because I don't want them to ban me :D
It could possibly be a word distance algorithm, so if you submit gibberish, it's looking for a word with the closest distance. In this case FROTIE could be the closest to FROG/FROGGY/FROGGIE and FIOLE could be FISH/POLE, which you can see in the 2nd and last pictures. The first one is the true outlier for me