May 2026 — ChatGPT / API Image Gallery, Prompt Tips, and Help: Generative Art Theme: Science

_j:

Blame it on the creator

if it comes off that way that’s the exact opposite of what I’m putting forward!

it’s more like attempting to merge concepts that don’t actually go together…

like photorealistic painting

how does a model rectify natural texture with a brushed texture being forced into the same domain within the image gen?

we get a mess of patterns across the image - which might be outdated as the model seems to be stabilizing that sort of thing already…

so it’s not ‘blaming the creator’ but rather trying to draw awareness to the domains that don’t logically work together…

the previous model would resolve pretty much anything but then loose coherency with deformations and such…

And I noticed this bit -

_j:

You create the idealized rendering of this concept

That prompt ensures that whatever the model deems is the most pleasing to a broad audience is what it will push towards which is always going to be ‘good’.

I’m coming at image prompting from the angle of finding new art styles that haven’t even been imagined yet so I come across this fail mode more often while learning how to rewrite the logic of what is being rendered…

It’s way different now than it was last year…

I could rewrite the refraction parameters within the system through its interpretation layer(s) (whatever that may be technically) of yarn in about an hour… but its fail modes until success looked way different then…

So i’m not trying to cast shade or blame but rather try to share my understanding of what’s causing that sort of thing…

And also took the moment to distinguish the important difference between how we’re going about trying to make the prompt thing easier and more fun.

We’re always going to be limited to ‘what the model deems most liked or popular’ when we ask the model for an image…

Which is why I diverged to dynamic prompt steering…

Which… Well… the bags under my eyes went away last week, but I wore them for almost 2 months straight while working out as much as I could about the failure modes and how to pull the model to specific points created by the user rather than the model…

We might speak the same language but I often feel like I’m trying to translate latin when adressing you, please don’t take it the wrong way!

^.^