Proof that AI is stupid

[MCDesigns]

Was working on generating images at Krea and trying to be clever with prompting to get the results i wanted. There were plenty of moments I was able to fool the AI, but also plenty in which I wasn't, but this one instance takes the cake.
I created a scene with 4 secretaries in a conference room and it generated fine, multiple times. I then changed the number of secretaries and it gives me a NSFW warning. I tried changing the number of secretaries from 5 to 20 and each time, a warning, only 4 worked, WTF!? LOL

 

[Badwolf666]

I want to know if you ask for one female, that's fine, then add in prompt add two more females to scene, [so this gives me three right] but AI add six females making a total of seven? it never gives me the number i ask for? So i take it it can't do Math neither?

 

[Nyghtfall3D]

I know nothing about zimage, so I asked ChatGPT how many characters it can render in the same frame. Here's what it said:

Short, honest answer: there’s no hard public limit—but in practice, zimage is happiest with 1–2 characters, tolerates 3, and starts to get flaky at 4+ in a single frame.

Here’s how it usually shakes out in real use

What zimage can technically do​

• 1 character – rock solid. Full control over pose, anatomy, expression.
• 2 characters – generally reliable if they’re not physically entangled.
• 3 characters – doable, but you’ll see compromises (pose drift, face swaps, cropped limbs).
• 4+ characters – the model prioritizes composition over accuracy; individual anatomy and intent degrade fast.

What actually limits it (more than a “character cap”)​

It’s not the number of people—it’s the complexity density:

• Shared vertical plane (people stacked front-to-back) → harder
• Physical interaction (touching, lifting, restraining) → much harder
• Precise anatomy demands (feet, hands, necks, faces all specified) → hardest
• Camera constraints (low-angle, macro, forced framing) → compounds everything

So two characters touching can be harder than three characters separated.

Practical sweet spots (based on how the model behaves)​

• Cinematic realism with control:
  → 1 protagonist + 1 implied / partially visible second figure
• Full-body interaction:
  → 2 characters max, simplify one of them (cropped, obscured, off-frame limbs)
• Crowd or group scenes:
  → 1 detailed subject + others described as background figures only

 

[MCDesigns]

Yep, not buying it ChatGPT. Just tried again with 3 instead of 4, still an error, yet with the magical 4 number it sometimes gives me 5 figures and I have other scenes with 8 of more figures. On the plus side, I was able to use Nyghtfalls/ChatGPT's modified prompts for large breasts in all of my images in this series, so not a complete failure, LOL (I found it really depends on where it is placed in the prompts on whether you get an error or not, LOL)

The ChatGPT explanation reminds me of trying to do large crowd scenes in Daz Studio, LOL

I have had plenty of other times where changing one word gave me an error, so I am sticking with stupid, LOL

 

[TailsWin]

The nsfw filters are trained on images and captions, just like the image generator.

Here's my guess about what's happening: You ask for a topless lady in a picture with 3 people, the nsfw rating goes way up, because, well, a lot of images with 3 people are nsfw.

Similarly, if you add too many people, the filter may be thinking of something like this:

But search for images of 'work meeting'. My first result... Four people:

So since images with 4 people are so much more likely to be office meetings or whatever (these are secretaries after all), the nsfw filter be like