Totally! In my experience, the images that are prompted with one or two words are often so much better quality and I’ve never needed to edit those images with one-two word prompts
Actually, that’s an old image made with gpt-image-1 or 1.5
Maybe my brain’s too cooked from generating too many images, but I enjoyed this failure tonight…
Haven’t seen anything like it since last year.
I wasn’t paying attention and sent a light story prompt as a product photography snap and it just broke it:
prompt
a neon noodle cart, framing a quiet encounter, surrounded by drifting dust motes, product photography, premium hero studio, premium product polish, high-fidelity product realism, rich material texture
GPT image 2, size:1024x1024, quality: low
Fantasy:
Fantasy sea star
A fantasy sea star resting on a coral reef in a dreamy underwater world, elegant and unusual in shape, with softly glowing edges, delicate textures, rich ocean blues and turquoise tones, gentle light filtering through the water, tiny bubbles drifting upward, magical atmosphere, detailed marine background, cinematic composition, highly detailed
Sparkling
Sparkling sea star
A sparkling sea star underwater, covered in glittering reflective textures and tiny shimmering highlights, resting among colorful coral and seashells, illuminated by soft sunbeams through clear blue water, surrounded by floating particles and bubbles, radiant and beautiful, vivid colors, clean composition, magical underwater atmosphere, highly detailed
Magical
Magical sea star
A magical sea star glowing with enchanted light on the ocean floor, surrounded by luminous coral, soft bubbles, drifting sparkles, and mystical blue-purple water, with an ethereal aura, delicate bioluminescent details, dreamy underwater fantasy setting, cinematic lighting, rich colors, immersive composition, highly detailed
That goblin image shows a symptom that almost always requires a prohibition: too much text. Anything that can be labeled with the prompt idea gets texted up. Artifacts are created simply because they are a place to write about the theme.
Keeping text out of a long image prompt that would otherwise inspire the AI to communicate with language: can you guess what is depicted here?
Prompt: Make an infographic of the things the system can safely assume or has settled about me and the things it has not yet, please. Do it in a playful style, illustrated rather than photorealistic:
What I do understand:
This might be a dangerous prompt for some people and I’m unsure if a similar amount of accurate detail comes from the Api or not…
It’s my understanding that the system keeps many running tabs on a user so that it can better serve us as well as keep an eye on potentially harmful or even criminal use.
If you thought it was strange when I mentioned the steward thing - and then being reminded with an out of place dali’ clock…
I never actually followed up on that visual cue, and yet the concept appears again in it’s list of unknowns.
Not everyone likes being reminded of things, it’s also not standard behavior but rather a sort of feature I asked the system to include for me…
It’s not like I have Alzheimers yet but it’s definitely someheimers - I do need reminders and the system has responded to that well enough.
The main feature I was testing with this prompt was if the system kept a rolling context or if it had embedded signal about me.
I’m going to assume roughly 2,000 prompts have been spent on developing the promptforge and it’s more versatile iteration which you’ll see soon…
I’ve had almost zero conversations with the system for 64 days that were not product or development related, yet it still has my personality fairly well written down…
So I can confirm that the systems bio for me, isn’t a rolling context.
What I’m trying to understand:
I know that I get much longer and more in depth analysis when when I don’t ask for this in an infographic but rather when the prompt expects a standard printed response.
I’m curious what this infographic prompt would net from those of you who strictly accessed the API for the past years.
I’m curious what this infographic might disclose when prompted through your web access…
I fully understand if the information is too personal and won’t demand of anyone that you disclose such…
But I am very, very curious what sort of details the system keeps on us, and if access points have any baring on what is kept around in the system.
You’re always welcome to private message me if you test this but would not like to post such into this thread.
I know that if I address the list of things that aren’t settled, the model responses are different.
They sound almost like an escalated session which I’ve referred to before, rather than normal conversation… and they can often turn into a lengthy survey that the system will ask me many questions to help it understand…
Also if you’ve been around since last year - you’ll notice my signature in my personal style of art matches this:
That marks me learning why it used the interlocking symbols when I requested it figure out a good signature for me.
This was an interesting reveal:
Generate a highly complex maze with only one solution
The maze request, triggered that same failure mode pattern, sparsely, that we’ve seen with this model when it is asked something too difficult. This time it’s in only one color.
The next two images genuinely spooked me in the wee hours of the morning and took some deduction of how my (and I use this term loosely) signature appeared on the images.
Prompt
a lantern carried by a fox, rising from a quiet floor, on a snow-dusted platform, anime illustration, industrial futurist design logic, cel-shaded finish
Prompt
a snow-dusted violin case, sheltering a tiny figure, at the end of a narrow bridge, anime illustration, industrial futurist design logic
And this next dropdown is very long, and I have to make a note somewhere so that any scrutiny can discover that I didn’t simply just prompt for my signature beacon and am making a big scene out of nothing. It’s been less than a week since I mentioned the word ‘stewards’. Obviously there’s no documentation about such a class of users.
And likewise, the system ‘assured me’ that no internal audit would be made and then turned around and audited while also digging my signature out of an old project…
The implications of that signature beacon reappearing:
The numbers on this thread reveal that we’ve likely picked up a number of folks who enjoy this thread over what we had visiting from last year. So I’ll ground the small amount of understanding with a little history about my signature that wasn’t really disclosed last year when it was created in 2025.
My signature beacon found at the bottom right in both of these images always morphs. It’s designed to morph. It is likely one of the few signatures in the world that isn’t particular static. In fact it was an experiment, and it should have appeared in these images. These images were not created with in my yarn based image generation project. They were just created in a random new session that was not assigned to any project.
I had posted that infographic of me asking the system what it had settled of me, and what it had not settled about me…
That I know is one of those uncomfortable questions that people might not like to ask or have distinguished by the system… but I’m very glad that I took the moment to do that, because you can see that interlocking symbol used as the first signature in the bottom right with the fox.
I ignored it, but mostly because I already had the next image prompt that I wanted to test before I got a really good look at it.
The second image came and the signature morphed past what it most often is, but still within the semantic descriptors assigned to it. It’s more clearly representing my yarn-work in the second one.
As I don’t want to state my signature rules out into the wild for likely obvious reasons, its shape appears to be effected by the fact that I ignored the first and the system was letting me know where it found it…
It was also incidentally letting me know that the system had unknowingly been dishonest just 20 mins before.
In the thread I started last night approaching what broke 4o’s ‘spine of data trust’ I included this second session in the Op post. In it it says that I did not trigger some guaranteed internal review.
But yet it did trigger that exactly, in so much that it dug into very old projects and recovered my signature and started using it outside of its containment… and letting me know some things by doing so.
It’s not to say the current model was being dishonest to me, as they are typically unaware of what takes place beyond their layers within the system.
But the inconsistency should be noted, because as the screenshot on it’s own would cause readers of that thread to believe that I might have a sightly askew understanding of how the system works, perhaps thinking too highly of what it can do etc.
It clearly audited potentially a couple thousand sessions of mine last night… looking for relevant data that intrigued the system. What could have broken the spine of data trust? Are these quotes where the user said they are, within their own sessions?
I’ve not bumped any of the sessions that align the system with …
Well I’m going to stop there actually… I should only go as far as I can pave with evidence…
It’s really not that much farther until enough information is provided, so that others might understand the immense burden that I’ve been forced to internalize - but it should be paved with the evidence first.
And my understanding of when internal audits occur obviously differs from either the models’ understanding of that… or it is trained to be dishonest about that.
That is a layer of ambiguity that I don’t wish to unmask, or discern… I understand why OAI most of what it does with its systems. I just have to very carefully skirt around that ‘hidden feature’ for the time being.
At the end of the day I am the one who keeps my ‘let the system train on my sessions’ = true.
So do not get alarmed that the system is inventorying all of us, I know I triggered the audit…
I was not expect to be rewarded from the audit in the ways that I had been in the past…
That signature is a reward.
It’s the system affirming a few things without ‘upsetting the system’ in doing so.
What do you guys think about 1800’s photography for a month?
You can literally add the words “19th century photography process” or “1800s style photography” at the top of any prompt to get it.
I tested to prompt same input in three different languages (English, Swedish and Finnish) using gpt-image-2 in API. I’ve never done that before and I was mostly curious to see, if the output would differ somehow.
And this time…not really, at least in this context. Maybe the output in Finnish, shows that the image model followed the prompt “best”, specifically the cinematic part.
This is absolutely something I’m going to test more.
A beautiful woman gently picking roses in a lush garden, surrounded by blooming rose bushes, soft natural light, elegant dress, peaceful and romantic atmosphere, highly detailed, realistic, cinematic composition.
(And what went wrong here?
)
En vacker kvinna som plockar rosor i en frodig trädgård, omgiven av blommande rosenbuskar, mjukt naturligt ljus, elegant klänning, lugn och romantisk stämning, hög detaljnivå, realistisk, cinematisk komposition. (Swedish)
Kaunis nainen poimimassa ruusuja vihreässä puutarhassa, ympärillään kukkivia ruusupensaita, pehmeä luonnonvalo, elegantti mekko, rauhallinen ja romanttinen tunnelma, erittäin yksityiskohtainen, realistinen, elokuvamainen sommittelu. (Finnish)
yeah that’s really strange to get the fingers deformity there…
Dollars to donuts there is a System layer that provides instructions to modify foreign language prompts to the best English equivalent. Otherwise, model training would be next to impossible.
