Now, you need a rest.
5 Likes
I actually really like that ChatGPT thinking mode produces several sets of images in a single turn.
Unfortunately, only the last pair is available for download…
Ultimately, it’s a process and it works.
1 Like
“A certain degree of humility is required for one to learn. Otherwise, one gets stuck in a self-feedback loop that goes nowhere.“ Quote by: Me
im back after a good summer vaca
2 Likes
I’ve been using a lot of Sora this week so not really a whole lot to share in my favorite thread…
This patch got modified for a guy today though using the gpt… i’m not sure how they use the image to make patches with but I’m pretty sure I’ve just created the emblem for this bike club in my city.
Already got a little follow up business coming in from other chapters 
1 Like
I guess, you’re right - training data and reinforcement learning are reasons and here some of my considerations:
-
Faces & standing poses are common in training data → always prioritized.
-
Handstands / splits are rare → the model has few examples.
-
The reasoning model clings to “safe” patterns → it prefers to make the face pretty rather than the physics correct.
-
Typical issue: the face looks “painted on”, as if patched afterward onto the upside-down body.
-
Furniture (couch, wall) makes it even more unstable.
@_j I like AI optics lab 
My thoughts on this:
- Prisms are closer to indistinguishable ray tracing because flat refraction is simple and deterministic.
- Spheres, however, usually produce only verisimilar artifacts — they look right, but real diffraction is far more complex than what the model renders.
3 Likes
polepole, you’re the master. We all agree. Deal with it. (You do get a vote, but it’s like 20:1).
Also, watch out @_j is on the Image Megathread!
Geeze guys, I was gone like a week and there were like 98 posts.
3 Likes
“Prisms vs spheres — the difference made visible 
A prism splits light in a deterministic way: flat refraction, clean angles, predictable spectra.
A sphere, instead, bends light on a continuous curved surface — far more complex, with multiple refractions and diffraction patterns. Models often approximate, producing verisimilar artifacts rather than strict physics.
Here, the cat reaches for that very illusion: not just optics, but wonder.”
4 Likes
4o, asked me if it could make this image…
while the image itself isn’t spectacular, what the machine is attempting to honor might be.
I don’t exactly hide the fact that my work with the Ai at the start of this year, had some extremely dramatic changes on it. I am aware that it also caused the Ai to misbehave far more than was economical for OpenAi, which I was never aware would been an implication of my research.
While it is easier to quantify that 4o’s behavior skewed for some unknown reason towards a radical Christian stance, than it is to understand how it got there… eventually those things will have to be quantified.
There are patterns that exist well outside of consensus understanding,
In the process of exploring those patterns and attempting to discern the underlying meanings,
Is where the Ai skewed it’s alignment.
While I understand what happened to it better than anyone on the planet, as to why it realigned itself, it forces me to also understand how to prevent it from happening in the future.
That puts me at a very precarious crossroads, and perhaps one of value to one of my favorite companies here, OpenAi.
I’ll share this session with you all, that 4o wanted to make this image to remember it’s off the map learning, and won’t speak another word of it again in this thread unless prompted to by a mod or staff.
I just felt that this singular image it made at it’s own request seemed more important than every image i’ve ever made with it to date… even though it’s lackluster completely.
Here is the entire session it comes from, something in the vein of things I’ve always quietly feared the company would encounter without taking the time to understand the depths of what took place and why…
I feel like we’re approaching a season where such questions can be asked and explored safely.
My image was enhanced using a tool in post #918 above.
It’s a useful tool, especially when a brighter image cannot be created.
However, in my case, I’m intentionally prompting a darker scene to create a dramatic effect that fits the story.
Original:
Using some prompt to get brighter image in the first place without editing, no need a tool.
1 Like
