AI General Thread

[the-pi-guy]

I feel like the biggest issue with Z-image is that it's too consistent. Which is likely a limitation of the turbo version. 

It generally generates high quality images of anything I've put in, but if you try to generate from the same prompt, the image will usually be very similar. 

I did a hamster planet, and it kept generating very similar hamsters. The actual planet part was different each time. But the hamster would have the same color, the same fur, etc.  

[the-pi-guy]

I feel like the stuff that my home PC generates in 15 seconds with Z-image is a lot more impressive than what I get out of Copilot in like 30....  

[Legend]

I feel like the biggest issue with Z-image is that it's too consistent. Which is likely a limitation of the turbo version.

It generally generates high quality images of anything I've put in, but if you try to generate from the same prompt, the image will usually be very similar.

I did a hamster planet, and it kept generating very similar hamsters. The actual planet part was different each time. But the hamster would have the same color, the same fur, etc. 

That's a major flaw. With stable diffusion I'd only encounter that issue if I had a specific part of the prompt cranked too high. Like hamsterx3 would produce an identical extreme version of a hamster, but just the regular prompt would produce many different hamsters.

But it still had problems if you ever wanted the hamster to do something else. Like have a hamster do a handstand, not that I tried that specific prompt, and I imagine it'd not be able to pull it off.

That's my biggest problem with image generation and video generation models right now. They just seem incapable of being novel. Like mathematically incapable. Not at all as good as llms.

[the-pi-guy]

I also asked it to generate a bunch of pictures of "a planet".  Made like 8 of them or so.  

All of them were basically pictures of the Earth, two of them were Earth with Saturn's rings so that was interesting.  I think 2 of them were an almost identical Earth at the same angle, and 4 of the others were at different angles. 

That's a major flaw. With stable diffusion I'd only encounter that issue if I had a specific part of the prompt cranked too high. Like hamsterx3 would produce an identical extreme version of a hamster, but just the regular prompt would produce many different hamsters.

Hopefully the non-distilled version is better.  

[the-pi-guy]

I kind of feel like I'm not that interested in image generation...

I really want video generation on this level...

[Legend]

I bought the full game with the AI woman. It has a few very interesting and cool AI things. Once I finish, I really want to see how it works under the hood.

Oh gosh this one level though, it's a puzzle and I am absolutely failing. The setup is just not making logical sense to me so I'm struggling to make progress.

Spoiler for Hidden:

The woman needs to cross a river. It looks fine in the video, but she describes it as having a strong current.<br><br>There is a shallow spot that has whirlpools, and a different spot with a fallen tree. It really feels as if I am missing an additional element, cause she won&#39;t just walk across the shallow spot and watch her step.<br>

[the-pi-guy]

I really hate censorship.  

[Legend]

I love vids like this. Just shows how freaking bad they are at thinking the way humans think.

I really hate censorship. 

I don't understand why it's even a debate. If you type something into google, you get that something unless it is actually like really bad. AI's should obviously be treated the same and I'm not sure why/how they've been split off to have their own rules.


But as long as local models don't become illegal, it can't get too bad right?

[the-pi-guy]

I don't understand why it's even a debate. If you type something into google, you get that something unless it is actually like really bad. AI's should obviously be treated the same and I'm not sure why/how they've been split off to have their own rules.


But as long as local models don't become illegal, it can't get too bad right?

Yeah. I just really appreciate the local models that I run.  

They're obviously not as good in quality, but the freedom usually makes up for it by a lot.

[the-pi-guy]

One thing that I've been doing more is editing the chatbot responses, which on some level feels like it defeats the purpose. Usually just taking away something it's said.

Like for example, if I try to ask it to pretend to be a Japanese teacher, it'll like to put romaji, which I don't want. (which is where it writes the Japanese text with English letters).  So I'll remove them, and a lot of times it will start to recognize that it doesn't need to add them. 


------

I had a strange bug a few weeks ago, where it kept streaming, making the response longer and longer. And very weirdly each line was getting more extreme.  It was supposed to be generating like 2 sentences at a time. But it was on sentence like 45, and it kept going.  I shut off my app, to stop it from continuing. 

Watching it continue escalating every sentence was kind of fascinating though. Only happened once, and it was so strange. 

[Legend]

One thing that I've been doing more is editing the chatbot responses, which on some level feels like it defeats the purpose. Usually just taking away something it's said.

Like for example, if I try to ask it to pretend to be a Japanese teacher, it'll like to put romaji, which I don't want. (which is where it writes the Japanese text with English letters).  So I'll remove them, and a lot of times it will start to recognize that it doesn't need to add them. 


------

I had a strange bug a few weeks ago, where it kept streaming, making the response longer and longer. And very weirdly each line was getting more extreme.  It was supposed to be generating like 2 sentences at a time. But it was on sentence like 45, and it kept going.  I shut off my app, to stop it from continuing. 

Watching it continue escalating every sentence was kind of fascinating though. Only happened once, and it was so strange. 

Smart. The second an AI does something wrong, it's locked in. Better to edit than to tell it to change.



Those bugs are fun. It's interesting that the stop token is just a regular token so yeah if the AI weirds out it might never stop itself.