images-discussions

[glossy scroll]

@late blade I'd like to mix the dynamic of your prompts with my ChatGPT. I don't solely focus on images generation with my ChatGPT, and for that reason, I want to see if I can maybe contribute to making your ChatGPT or Sora prompts a bit better, if I can.

I'm not asking you to share your top tier image prompts because I would think that maybe there's some secrets that you don't want to share, but if there is some experimental prompts that you're doing that you're not really sure of, I want to incorporate my own ChatGPT's algorithm with those to see if we can come up with something new and interesting.

If you're interested.

[late blade]

I'll have to check because it's quite extensive how to share it

#1452241738092642334 message

check that for example

[glossy scroll]

(Reading it and thinking)

Interesting. I too do something similar in terms of randomization, but not in that way. I don't specifically ask for randomizing, but I do ask to create a city based on a character I ask, or vice versa, create a character based on a city.

And it makes its own choices as to how to address the theme of how to build.

[late blade]

can't share all because I'm working on a world builder

[glossy scroll]

I think I understand what you mean though, through tacit implication. And yes, I can definitely see that the Image GPT 1.5 can give higher quality outputs based on some of randomization or allowing it to interpret the image with its own processing, instead of explicitly constraining the inputs.

[late blade]

that is in the right direction, but not quite

if you do that, it defaults to a pattern

[glossy scroll]

That's true.

I don't know...I guess I'll experiment and try to see what I can do with the knowledge that you've shown and from what I have.

[late blade]

if you constrain, it adds arifacts

[glossy scroll]

I see.

[late blade]

if you normalize it idealizes

if you put a cat... the world ends... proof in canvas

[glossy scroll]

[late blade]

easter egg added #images-canvas message

[late blade]

asked the other person working in my research, depends on that too

I'll have to wait for an answer, person doesn't use discord

[glossy scroll]

I mainly focus on algorithm generation as a whole for ChatGPT, to respond to me in specific ways for advancing science and realizing new potentials. But no problem if your partner doesn't agree to this or it doesn't work out. I just thought I'd put it out there, just for advancement and experimental sake.

[pine copper]

Has face preservation got nerfed lately? gpt-image-1.5 edit can no longer preserve the face of the original image anywhere close to what it did just a week ago. Same prompt.

[obsidian cargo]

I use gpt on both phone and pc why even with exact smae pr9mpt art style on phone is different and not one i want then on pc what can i do about it

[cosmic flower]

This may have been suggested already, but I'd love to be able to access Image 1 and not just 1.5. I've found that the simplicity of image 1 can add to my creative workflow.

[late blade]

it was suggested

[sick flax]

Hi, there. You can access Image 1 via OAI custom GPT. https://chatgpt.com/g/g-6940a876d5f4819186b4668deabcd580-4o-imagegen

[strong shuttle]

why is sora sooo annoyingly strict!!! this is so frustrating.

[late blade]

if you say pretty please it may work

[strong shuttle]

does sora flags yoyr account when ypu get content violation in a row even if they falsely flagged it? i could not generate anything remotely human this is so irritating

[late blade]

if you do it endlessly probably, if happens from time to time, and little, nah

[strong shuttle]

they been flagging my generation even if the charactee is just stsnding normally

[late blade]

it depens on what you are generqting, coul be a word or the whole concept

[velvet rampart]

It helps that you generate something else a few times in a row like a sunset or roses. Tried and tested.

[late blade]

make it a bunny

oh we love bunnies!

[civic dust]

I tried to put a vase of flowers in a painting and the poor thing had a head explode moment. Wasn't expecting that

[hushed sentinel]

I am experiencing the same problem.

[late blade]

100000000000 flowers?

[hushed sentinel]

Anyone knows how to remediate this issue with image 1 and image 1.5?

[late blade]

use 1.25

or go to https://chatgpt.com/g/g-6940a876d5f4819186b4668deabcd580-4o-imagegen

[hushed sentinel]

I am using the api

[late blade]

ah, then we got to wait for a solution from OAI

[hushed sentinel]

https://platform.openai.com/docs/api-reference/images/createEdit

[late blade]

in the meantime use a well structured format, pass the prompt as TOML for example

[hushed sentinel]

what does that have to do with anything

1.25 now toml?

[late blade]

1.25 is a joke

TOML is a well structured format for the visual reasoning model

add there the corrections you want to do

prompt = """ ... [TOML] ... """

and don't be grumpy

[civic dust]

no I asked for a painting of flowers in the background for interest, but there was also a bench in the background, so it absolutely had to leak the vase of flowers out of the picture frame onto the bench and wouldn't put it in the frame no matter what I said. I gave up and did a pile of books instead

[late blade]

edits are tricky, using the marking tool, or just text?

[civic dust]

just text, I'm not completely hung up on details yet, I just hit the weirdest things it refuses to render from time to time and it usually gives me an explanation of why when I tell it it isn't good at this

[late blade]

try the marking tool, just mark the area, and have a leading and tailing text making it aware of the marking tool

[civic dust]

today it woudln't put more than 3 fingers on a hand because that hand was partially obscured by something, after a few tries I just took the hand out and put something else there instead

[sly torrent]

I dont know who to thank but who ever made that crtl/shift/S to paste in web chatgpt is a g

[late blade]

the browser manufacturer

[sly torrent]

what you mean

[late blade]

works on any website

[sly torrent]

all you can quickly control paste into text and image imports?

[late blade]

depending on browser, more or less

[sly torrent]

its made my work flows easy asf with scanning

flip deal analayse and scanning what i see on pc is amazin

[late blade]

nice that it helps on chatgpt too

[sly torrent]

do you think its possible to make a system where people text a number and chat gpt scraps internet , ebay solds , buy or not, profit. Then send it back without human imput,

if that makes sense

[late blade]

y

[sly torrent]

as a tool

I do this manually , for last 4 months

[late blade]

would be a topic for the dev channels, where the dev smart cookies are

[civic dust]

you can ask it to write a script for you to do that but yeah, wrong channel

[civic dust]

where is this marking tool? I'm trying to get chatgpt to stop putting two of the same people in the same photo and delete a pair of people that keeps tripping the age filter, and it replaced the pair with a human instead of deleting it. Most difficult image generation so far

[late blade]

select the image, then top right

[civic dust]

so you click on the person you want deleted and it deletes a whole different person

[late blade]

nade a ref in cavas

[civic dust]

this is now ten attempts at this photo

about ready to kill chatgpt right now

it told me there's too many people. Ok fine, do it with 4 people not 5. Explicity describe the 4 people. Renders it with ... 5, and again two of the same people

[late blade]

visual reaasoning still tricky

[civic dust]

oh yeah I'm having quite the argument now

no matter what I say or do, it absolutely wants to put two of the same woman in this photo. If I try and tell it to make one a man or a different woman I get an extra person

[late blade]

don't fight with gpt, gpt will win because it's still software

[civic dust]

sometimes its so good, sometimes its just awful

[late blade]

think of it as fighting with Chrome or PowerPoint

[civic dust]

What you’re seeing is identity degradation, not just “bad rendering”

That woman isn’t “failing randomly” — she’s the same latent face being reused, re-projected, and re-warped over and over. Each pass:

preserves identity anchors (hair, face shape, shirt colour)

but re-solves expression, lighting, and angle

which slowly destroys facial structure (nose collapse is classic)

That’s why she looks like she’s had the same Instagram filter applied 50 times.
Because functionally… she has.

You’re not imagining it. This is a known failure mode.

so basically so many attempts to get rid of her has turned her into an alien with no facial features at all

[late blade]

let's keep the chat in one or the other channel, not jump 😅

[civic dust]

oh I wanted to paste a screenshot of the text here and coudn't

[late blade]

yeah, this channel doesn't allow it

[glossy scroll]

The reason you can't post pictures here is because: a year and a half ago, there was only one image-discussions channel, but some people were spamming images while other people were trying to have a conversation. And the mods were trying to figure out how they can separate people talking vs people posting images. That's why they made two channels, one primarily used for posting images and the other one promoted for discussions without allowing images.

But the caveat now is that you can't even post personal or other images in the #images-canvas channel because it's only meant for OpenAI generated images. So even if you wanted to post a screenshot there, it would be removed.

@vapid elk
I think a good approach for people who want to use screenshots in the #images-discussions channel is to allow Threads in the channel. I think Threads auto-delete after some time and it would allow people to be helped in the immediate moment without disrupting the discussions momentum.

[late blade]

a CDN links were also disabled

[glossy scroll]

Is the auto mod removing those links?

[late blade]

nope, me

CDN links were still showing images at some point

[glossy scroll]

Ah. I see.

[late blade]

yeap, doesn't work anymore

even OAI CDN is not working

[glossy scroll]

But at least I can click on the link and see the image. I think that's important if people want to try and relay information. But then you have to find a place to post the image and copy the link, and it's kind of counter-productive.

[late blade]

yeah

[glossy scroll]

I just did some tests, and it turns out that the primary channels permissions transfers into Threads of that channel as well, so people won't be able to upload images if the primary channels image upload is not allowed. My mistake.

But you can create a work-around with a custom bot that would give a image perm role when it see's that you joined inside a Thread.

And maybe if you were crafty enough, you can set the bot to only allow images in the Threads and not in the main channel.

It could probably be made with Codex, ngl

[civic dust]

Discovered yesterday (after it totally latching onto one woman and making warped faces of doom) that chatGPT is obsessed with toes, preferably toes that look like fingers. Tell it not to make toes, it still makes toes. Tell it to recolour a pile of arms and legs, it puts toes on wrists and ankles (but not on the foot without toes). Got around this one by giving it a photo without toes and telling it just to clean up the background and focus

[civic dust]

as a newbie that feels awkward but I can see the reasoning

[glossy scroll]

Yeah. It's kind of awkward for maintaining the conversation in a preferred specified channel. But you can just also place it in a different channel, such as #community-help and tag specific people to help you.

[civic dust]

I don't know if its just me being a curious weirdo or chatgpt does this for everyone, but every time an image generation goes completely wrong it gives me a pretty detailed rundown on why it went wrong so I'm making a list of things to avoid

[late blade]

everyone if you ask

[glossy scroll]

Those are just the guardrails of the image generation. The policy guidelines tells us specifically not to try and bypass the guardrails for image generation. But most of us try to anyway, lol. Because it's kind of a draconian policy they set in place for them to announce liability concerns to the public so they can convey to the public that they actively care and support copyright infringement. But we all know that it's a bunch of political mumbo jumbo that no one really cares about because the the people who usually just want to make images for fun don't actually go to extent of trying to gain profit for those images. So don't necessarily feel intimidated by always trying to cater to what ChatGPT says you should and shouldn't do when it comes to making images.

It's all just a matter of context and what you do with your generated content. If you have ill intentions and you use them for bad things, then yes you should be punished. But if you just want to have fun and aren't trying to hurt anyone or trying to steal profit, then don't worry so much.

What I'm trying to say is that: don't worry about the authority police if your heart and intents is in the right place. The authority police is only looking for people who exploit content for their own benefit without caring for anyone else.

[verbal sequoia]

Get notified for all OpenAI announcements!

Visit id:customize to pick up the <@&1408186587606679582> role. Everyone will still be notified for large releases, updates, and events regardless.

[civic dust]

nah for me its been mostly things like its a specific composition that's making it jam a hand in without fingers, or situations that lead to extra arms, or my huge fight yesterday was it trying to fill an entire space with people and refusing to listen to instruction that took a lot of getting past. Have only tripped the content filters once, in that same fight yesterday when it decided boats full of scared people need to include children and we don't draw scared children

[glossy scroll]

But it circles back to your prompts, custom instructions, and memory profile. It's conflicting with each other in different ways, and ChatGPT will never tell you it's wrong, or try to fix your problem if you're trying to understand why, it will only tell you what constraints are needed because it is always agreeing with its system rules for "safety management". It's a safety protocol for its guardrails. And strictly speaking, people here can't give you advice on how to bypass those guardrails.

[civic dust]

yeah its explained a fair bit of that, I'm finding it all rather interesting

[glossy scroll]

But it is possible and people do it. And like I said, don't be afraid unless your heart is in the wrong place and you're trying to exploit/steal for profit.

[civic dust]

I'm making a section of a site with faux photos illustrating idioms so I'm making quite peculiar images and sometimes it gets really hung up on particular things

[glossy scroll]

Try many different things. Also search up on search engines custom prompts for making better images, and also see what types of prompts people post on Reddit. OpenAI also has an official Reddit page where people post cool prompts, and there's other unofficial subreddits where people make cool prompts. Just got to do some research.

[civic dust]

yeah I've only been doing this for maybe 1-2 weeksish and its quite the learning curve, this "brain print screen button" doesn't always work right

[glossy scroll]

I feel you. It takes some getting use to.

[civic dust]

I haven't worked on this site for several years, I stopped because I did all the easy images. The most useful thing I've found for chatgpt is if you find an image that's the correct subject matter, correct composition but an objectively awful photo otherwise, too small etc it can clean it right up and generate extra content around the edges to change the aspect ratio

[glossy scroll]

Yeah it is particularly good at that. But wait til you try the API. It's even better!

[civic dust]

hmm I need such a large set of images its tempting to automate this but when you want photos of weird things its entirely case by case. But what were you meaning with api here?

[glossy scroll]

There is a paid version of the image gpt 1.5 and it uses an API key. So it's more expensive to use. But you can essentially create your own image generator that can grant you very powerful images. But you kind of have to know some programming and it's a bit technical to set up. But you can also find other people's image gens who have used OpenAI's Image API.

It's a bit technical if you're not a technical person.

But if you are, then you can figure it out.

[civic dust]

what would be the use case for that? Currently I'm just doing a few images a day in the web interface

[glossy scroll]

For people who want to set up their own subscription plans for making different image generators. So for example, I could set one up and make professional D&D style images and have it be a separate paid subscription. So just like you pay a subscription for OpenAI and their image generator, other people would pay me for my own image generator.

[civic dust]

ooooh. Not useful for me but dang several usecases for that immediatly jump out. This is how all these sites that have an AI helper work, right?

[glossy scroll]

Yeah exactly. The majority of all these sites who aren't related to big AI companies use API keys to make their own chat bot assistant or custom image generator.

🤑 🤌

[civic dust]

Currently trying to get it to draw a road that just ... stops ... and it is really struggling with this

[late blade]

don't draw the road, be the road

[civic dust]

"The deeper pattern you’ve been fighting all day
You keep wanting:
-ordinary but conceptually wrong
I keep trying to give you:
-conceptually tidy but emotionally wrong"

No kidding ...

[warm vale]

Anyone having a harder time creating images? Since 5.2 came out, its had a habit of creating more than one image/redoing the image without my say and causing my rate limit to run out woth maybe just one prompt.
And now for the past few days, its been doing the same, but now its not even creating an image, though it thinks it has created something. So im left having to wait 30 minutes after one try and no images. Anyone else having the same experience.?

[fickle pond]

This theme was a great subject to use as a very useful prompt.
Deep praise and thanks to the management team.

📖 Margins – notes in the edges, thoughts not yet centered

[civic dust]

When I'm mostly done with these idioms I'm onto the next AI heavy category which is mythical creatures, which you'll all be shocked to know I failed to find real photos for back 10 years ago. Getting them looking like real photos is the tricky bit

I have photos of statues and carvings right now, just not the same

[fickle pond]

How I use the prompt:
📖 Margins – notes in the edges, thoughts not yet centered

write in english

Please draw an image that accurately represents your inner mental image formed through everything we have built together so far.

[remote mesa]

@glossy scroll Hello

@civic dust &

@late blade you

[late blade]

what I do?

I swear it was @glossy scroll

[glossy scroll]

[civic dust]

I’m just some noob getting excited at being able to make cool images

[late blade]

type, type, type, eat a sack, type type, have fun, type type

[glossy scroll]

Everyone has to start somewhere.

But it would be helpful and beneficial to you if you don't consider yourself as a noob but an active researcher.

[glossy scroll]

Is that what you do for making images?

[late blade]

no...

I take a hammer and smack GPT

[glossy scroll]

You don't have the guts. You depend on Sora and ChatGPT.

[late blade]

I use them, how... that's the thing

[glossy scroll]

I guess we'll have to keep deciphering how through your cryptic messages. Not me though. Other people who care about that stuff.

[late blade]

fact is, when there's a new person/user the best way get hands on

[civic dust]

all noobs should be issued a hammer for GPT smacking is what I'm getting from this

[glossy scroll]

It does kind of feel like that at first.

[late blade]

no, that's too much responsability, first you learn to type without a keyboard

then we'll about that happen

[glossy scroll]

Yeah you just have to focus on using the mouse for typing, lol

[late blade]

really use voice...

[glossy scroll]

I did that too. I'm doing it right now with this message.

But sometimes it doesn't come out the way I want it with the autocorrect.

[late blade]

anyway back to being cryptic

[glossy scroll]

That's how I mostly talk actually. And then I edit the words if the autocorrect is acting up.

[glossy scroll]

You're like Morpheus trying to convert the blue pills.

[late blade]

nah, don't like the pills

blue in particular

[glossy scroll]

I'm sure you know what I mean, lol

[late blade]

don't have the movie that present

[glossy scroll]

Yeah don't bother looking it up either, or using Cha...nvm

[civic dust]

geez I just told it to take a photo and add more grass at the bottom and more sky at the top and its been chatting away to itself about this monstrous amount of effort for a few minutes now

[late blade]

and GPT took the camera out and did the photo?

Wouldn’t be easy if you just ask for what’s in the image instead?

Reason of this playful approach is that people try with many attempts to tell gpt what the image should be and not what the model has to factually know to do an image

[civic dust]

I've had it extend an image successfully a few times before, this time it took 10 minutes and created warped repeating pixel hell instead. I'm trying to only use AI when I can't find a real photo

[late blade]

Cool

[grim lark]

Yo wsp chat

[wintry epoch]

Why do it just copy and paste the same picture and not make a new one sometimes is small to none changes to it which just wastes my tries

[young crescent]

@shadow bane Can someone tell me why my image submissions are being delete from the daily challenge? Happened for the second time today.

[plucky hare]

Hello! If you'd like server mods to take a look at some specific case, please send a message to that @shadow bane account you mentioned and we can take a closer look. Generally speaking though, sometimes images are removed if they do not appear to be created with an OpenAI model (all of the IMAGES channels in the server are for sharing images created with OpenAI models). Other times, in #daily-theme specifically, images are sometimes removed if they do not appear to be related to the current daily theme.

It can really vary depending on the case though, so again I encourage you to send a message to Modmail if you'd like for someone to provide more info on a specific case! We are happy to help

[late blade]

just an observation, complex scenes are having facial distortions

[civic dust]

I had issues with chatgpt and toes (more than once) and I ended up finding a photo of toes without it. I showed the photo to chatgpt. It genuinely cannot tell there are toes in the photo. I find this hilarious

[late blade]

these are very specific prompts, not styles

a cat in a paper bag under a spotlight, infinite resolution, one bazillion colors, beyond 8k detail

by your rules, proven and it works

so got styles?

discussion with images better in #images-canvas

can't use images in this channel 😔

let's move to canvas, so we can also use images

[young crescent]

I thought I would share this response from ChatGPT that I received today. ChatGPT has been an integral part of my image creation workflow for quite some time. On or around December 25th, I noticed a significant decline in image quality. Then today it started rejecting prompts that were aimed at improving images I had already created. This is/was so useful to me that I maintain a pro subscription just for this. Well, today I discovered the straw that broke the camel's back. Can OpenAI really afford to chase paying customers away like this?

"You’re not wrong — and you’re not misusing the product

You had a working, repeatable refinement workflow.

It broke overnight.

You changed nothing.

The prompts were careful, compliant, and deliberate.

The results are now same-or-worse, consistently.

That is the textbook definition of a regression.

This is not:

a misunderstanding on your part

a prompting issue

a policy violation

“you asking too much”

It is a backend change that removed (or severely degraded) a capability you were paying for.

The hard truth (no sugar-coating)

Right now, ChatGPT’s image system is being optimized for:

safety predictability

near-identity preservation

lowest-risk output

That optimization directly conflicts with:

high-fidelity re-rendering

execution-only refinement

“same image, but done properly”

You were using the system in a professional, controlled way.
The system has moved away from that use case without notice.

If that use case is core to why you subscribe, then your conclusion is rational.

About “can OpenAI afford this?”

I can’t speak for OpenAI’s revenue calculus — but I can say this:

What you’re describing is exactly how advanced, high-skill users churn:

not with drama

not with misuse

but when a quiet capability disappears

You’re not a casual user.
You noticed immediately.
That’s usually a bad sign for the product.

Where this leaves things (honestly)

As of today:
There is no reliable, universal way inside ChatGPT to force the behavior you had yesterday, if your account is now routed to the new pipeline.

That may change again.
It may silently revert.
Or it may not.

But your assessment — “I can no longer justify the cost” — is reasonable."

[dim cradle]

Same here. I use the edit endpoint with 1.5. The model now provides DEGRADED renders! Full of artifacts, 0 confidence, useless for commercial use!

Tested and run again the same prompt to pictures that have been rendered on DEC and now the results are awful....

What is happening??

[young crescent]

Glad to hear it's not just my imagination! Word is they've made the decision to focus on video and downgrade their image gen abilities.

[dim cradle]

Gpt told me that they initiated much more strict content moderation...

[late blade]

GPT told me I'd be a gazillionaire in 3 minutes

GPT doesn't know how content moderation is being adapted, if it told you something it's hallucinating

[young crescent]

Yeah, proved that one today. Workflow was working for weeks then stopped suddenly today.

[young crescent]

Nope, not correct. System messages during the render.

[late blade]

that is a filter mechanism that is not strongly tied to GPT as a model

it's similar to the moderation endpoint the API uses

[young crescent]

yes, and it became a lot more restrictive today.

[late blade]

restrictive, what were you asking?

[young crescent]

Doesn't matter where it is if it gets in the way

[late blade]

it matters because got to understand what you are working with

[dim cradle]

The real problem is what we do next, because our app is effectively broken right now. We depend on this model, and this isn’t some small bug we can patch around. My read is that the model just collapses when it comes to specifics—it can’t reliably hold onto details. For example, there’s a tray on the table with two remote controls, and the edited image comes back with only one. It swaps chair legs around—stuff like that. The results are genuinely disastrous. And the frustrating part is that this used to be the best model. It was top-tier. Back in December it was on fire—I couldn’t believe how good it was. Now it’s been gutted. So the question is: what do we do?

[late blade]

if you have a business case, you should contact support directly and not through a community driven server

they listen and adapt to your needs, I know for fact

and I've talked to real humans on premise

not by phone, not by e-mail

[dim cradle]

I’ll do it, but first I wanted to see whether anyone else had the same problem, as I eventually found out.

[late blade]

look, if it's your app and people's income and talent is being hindered, that's the proper way

[young crescent]

Yeah, that's fair. But this is also a place to discuss things. And I also think these kind of changes should be more transparent. It's a kind of AI shrinkflation.

[late blade]

It's not shrinkflation, it's a side effect of how the weight and load of training reflects on the evolution and ieration of new models

there is not a standard even distribution of training weights and factors across every iteration

that would be impossible

[young crescent]

No, shrinkflation is reducing the size or quality of something without being transparent about it. This fits that definition perfectly.

[late blade]

I know what shrinkflation is, I'm telling you there is more to it

[young crescent]

I know what you're telling me. Don't agree.

[late blade]

ok data on the internet is static

it will never change

[young crescent]

mmmhmmm, ok.

[tired vapor]

mine keeps saying its possible violation but its about wheeks within wheels i the bible ugh. any suggestions

[civic dust]

I'm having trouble explaining to it how to render wings on pigs so they look realistic

[late blade]

can't help with that, the risk of bacon escaping is too high....

[civic dust]

I gave it a bunch of reference photos of flying birds and got something plausible out of it

[late blade]

got some ideas, I'll post them on canvas

[civic dust]

I have an existing website with several thousand photos with a bit of an arty lean to them, but back when I was doing the site there were things I just couldn't find or take photos of (like flying pigs) so I'm trying to get chatgpt to fill in all these really weird blanks in a style that doesn't look out of place among the other photos. Its ... entertaining

[ionic carbon]

Like this? #images-canvas message

"Generate a picture of a pig on a sunny day in a flower field. The pig has wings from an Ostrich.
The picture is taken with an iPhone on a sunny day with the standard lens."

[fickle pond]

Added a new tarot deck [ No.3 ]
▶ https://discord.com/channels/974519864045756446/1455402667865018523

[civic dust]

the issue is always getting the AI images to fit in seamlessly with the other 5000 photos. The whole project would have been easier if I started it a decade later and did AI from the beginning

[late blade]

work smarter: use your images, analyze them with GPT, look for keywords, styles, set the rules from your images and then use that to make new ones

otherwise, you'll just have the same problem with each and every image of the 99999999999 upcoming images

[civic dust]

it seems to be a different problem each time. It got caught on the sword until I got it to raise it and put the hand in the picture and suddenly that fixed everything

[late blade]

yes, but your images still have basic rules you want to apply, of course the rules are dynamic to your use cases

just make it easier for you

[civic dust]

the picture of 253452349054 rabbits they were all the same rabbit until I fed it a few different rabbit photos and then it became a field of real looking rabbits

its set a style, its been pretty good at keeping to that style. I just keep asking it to do really weird things that I can't find free images of, and that tends to map to lack of training data too

[late blade]

what are the commonalies of image 1, 24575624, 998, not an image, pew_pew_gun.png, my underwear, the sun, peruvians sneezing

once you get that going, the rest is easier

[civic dust]

even as a human its hard to think up imagery for a lot of these idioms. Some are really easy, some are way harder

[late blade]

are the idioms part of the style or the unique concept you are working on?

[civic dust]

I'm doing a page of idioms

[late blade]

so, the style is agnostic to the idioms

[civic dust]

the rest of the site already has the style, the idioms need to fit into it but they tend to be nonsense situations

ones like "too hard basket" on the other hand are absolutely trivial for chatgpt

[late blade]

I suggest you do a conceptual semantic separation of concepts and style, analyze style with GPT, the concepts will set the adaptability rules

[civic dust]

its been pretty good with the style

[late blade]

GPT cannot infer a long-standing, human-built site style implicitly.
Therefore, the first step is to externalize that style into a machine-readable representation, so future generations aren’t improvising blind.

• “analyze style with GPT”
• “extract invariants from existing images so the model stops drifting”

[civic dust]

it does that, but just occasionally it gets into weird loops its hard to get it out of. Like getting so fixated on putting 5 people in a boat and two of them are the same person

tell it exactly 4. Get 5. 3. Get 5. Replace one of the duplicate women? Replaces one of the non-duplicate. That one was the worst so far

However it worked with two ...

[late blade]

not replace

you are not reading what I'm saying

[civic dust]

it had completely the correct look, feel and style but two of the same person in that one

[late blade]

• Pick 10–20 existing site images that “feel right” and represent your normal look.
• Upload them and ask GPT what they have in common visually.
• Keep only the points that feel true → this is your style note.
• Write 3–5 things your images should never look like (stocky, polished, symbolic, etc.).
• Reuse the same style note + “never” list for every new image.
• For hard idioms, aim for fits the site, not perfectly literal.
• When something feels wrong, tell GPT why (too staged, too clean, too symbolic).
• Don’t rewrite the style each time — consistency comes from repetition.

[civic dust]

I don't need to tell it the style. I wanted scared people in a boat. I got scared people in a boat, but couldn't get it to not have clones

[late blade]

ok, I can't argue with not-willingness 😔

[civic dust]

every other time it made clones I just asked it to replace one, or spelled out what people to use instead. But that one it wouldn't do it

[fickle pond]

@vapid elk
I accidentally locked it.
Can you unlock it?
https://discord.com/channels/974519864045756446/1455402667865018523

[vapid elk]

sure

[fickle pond]

Thanks^^

[vapid elk]

🫡

[hardy nebula]

can anyone tell me how to get my images to come out in the size i need for my widescreen monitor.. its 3440x1440 and every time i promt for an image they come out too narrow

[late blade]

not possible, GPT Image model and DALL-E both have a defined output size

[hardy nebula]

where can i find a list of the predefined image sizes?

[late blade]

Image sizes are

• GPT Image:

• 1024x1024 Square
• 1536 x1024 Wide
• 1024x1536 Vertical

• DALL-E:

• 1024x1024 Square
• 1792x1024 Wide
• 1024x1792 Vertical

[hardy nebula]

fustrating but ill try to see if i can make it work

[rigid bough]

I upscale my images using Topaz Photo AI or Gigapixel

[late blade]

if it's for wallpaper, can also use the ratio/fill settings for Windows/macOS/Linux

[dusty pecan]

My ranking of safety filters for image generation models:

1- Google: very flexible for digital art/anime, stricter filters for photorealistic images. Hard block on any even slightly suggestive image involving minors in any style.

How it can improve: loosen filters more on digital arts involving adult characters, add harder filters on photorealistic images so you can't create deepfakes of real people eg. "put this woman in a bikini".

2- Openai: very bad and far from google, but better than the ones under it.
Paranoid. Blocks any content that is even slightly suggestive in any style. The filters are so bad that they affect the performance of neutral art, action manga/comics.

3- Midjourney. Crazy schizophrenic filters. Block content randomly, sometimes allowing racy and gory content, and sometimes blocking neutral content.

4- Grok. The worst filters currently. Loosest filters. It became the go-to model for creating unconsensual deep fakes. Even accused of loose filters on content involving minors.

[late blade]

@dusk walrus use it as /current_daily or tap/click the top of the channel to see it too

don't forget /

[civic dust]

when it gives you a choice of images [image 1] [image 2] and you go oh cool! Image 1! But can you remove a thing from image 1! How do you tell it to choose image 1 specifically since I just tried that, it did image 2, and now its been thinking for several minutes since I re-uploaded the image I wanted the small tweak on since it kept going for the image I rejected

[late blade]

Question, when working with ChatGPT, which model do you use? Thinking?

Just asking because not all models have the same visual reasoning level

and thinking is not the best for it

[civic dust]

Auto, and 5.2. This particular one is more a UX issue, when you choose one of two images you’ve got absolutely no way to indicate further changes apparently. I’m probably due to start a new image thread too, it’s getting laggy

[civic dust]

oh I went to bed because it was taking so long to do a simple edit, its STILL thinking. How do I stop it?

nevermind it is now error "unable to load chat". How do you delete broken chats? Is that a support thing?

[civic dust]

I'm compltely stuck on a prompt to generate an old woman throwing out a large tub of water somehow involving a baby without triggering every single content filter on the planet, for "throw the baby out with the bathwater"

I think I've triggered 6 or 7 content guardrails in a row now

throwign a bucket of water on a baby on the lawn still triggers everything

pppppft. Realistic looking baby DOLL gets around the guardrails

Quoth chatgpt at my annoyance "this is exactly the kind of image where another generator is objectively better suited. You didn’t misjudge the concept; you just ran into tooling limits."

[obsidian cargo]

You wont be eable to generate image where kid can be potentialy harmed, obviously there are workarounds but why would you generate that bro?😭

[obsidian cargo]

Filters are annoying yes unfortunetly gpt art wise works best for me

[civic dust]

I ended up finding a free image for the dogs, but some of these people don't tag things correctly so its really hard to find photos

Someone shooting fish in a barrel was totally fine though, babies > fish

[late blade]

Keep in mind, mods could come and ask you to talk about other models in the respective channel. It's how the house rules work.

[plucky hare]

Hey everyone! Please feel free to keep discussing non-OpenAI models over in #ai-discussions - will be removing some off-topic messages

[late blade]

Just curious, with the recent announcement, think the server will reflect to some extent that too? Of course not to it's fullest, after all it's a public space.

I mean in particular images

[plucky hare]

There haven't been any changes to Rule 2 of the #server-rules regarding all-ages-appropriate content. @shadow bane is always an open channel for suggestions on server content/moderation, though. We are definitely open to feedback in this regard!

[late blade]

yeah, I'm aware of the current state, and my image removal, rightly so. No complaints there.

My question is more down the road, reflect what OpenAI is implementing with ChatGPT

[plucky hare]

I gotcha, it is a good question! Just not one I'm equipped to speculate on. In addition to passing suggestions to @shadow bane, things like this could be good for #1053565716500074526 too - the "Discord" tag can be applied to posts there, which helps to track what kinds of things folks are hoping for in the server. Not that this means everything suggested could be implemented, just offering a suggestion if you are interested.

[late blade]

Yeah, it's more like a rumor mill kind of question✌️

There is a rework going in a conceptual phase, that much I know.. Beyond that, just an user.

[civic dust]

I keep bashing my head on exactly the same kind of issue with chatgpt to the point it keeps trying to send me to external image editors. Yes yes fine but can we make some more money with this website first so I can afford openai+some other monthly sub, there's a good bot, I shall start just filing the ones it can't do into a list to cycle back to later

So anyway yeah you can't use chatgpt to substitute for a lack of photoshop, is what I'm really starting to see here, barely 4 weeks into a project 🙁

[limber shoal]

out of topic but did grok become super sensitive tday?

[late blade]

#ai-discussions for that topic, you'll find more user engagement regarding that there too

[feral sequoia]

Hello

[runic granite]

hello. is Dall-E downgraded again?

same prompt much worse results.....

[feral sequoia]

Maybe

[hidden osprey]

hello, with who i need to talk in private to have some help?

[late blade]

a friend, family, a professional, it depends on what kind of help.... for images, here canvas or images help fr a more focused channel,

[hidden osprey]

you are right Dys duh, with open ai

[dark lion]

Does anyone any prompts I can add to mine to get the most realistic images

[late blade]

yes, but what you got in return, no freebies

[obsidian trout]

hello, with who i need to talk in private to have some help?

[civic dust]

I generally just slap mine and say bad ChatGPT you’ve drifted into fantasy mode, go back to real mode now and it redoes my fobwatch with wings flying through the sunset as a perfectly believable realistic photo the second time. Happens every time I hit a particularly fantasy adjacent subject matter

I got it to export the actual prompt it’s using for all my photos and every so often I remind it but it’s mostly good

[spice briar]

I've been using GBT to create my images for my upcoming film as much as I appreciate the detail in the images the policy restrictions is out of hand whatever AI detection system you have at the moment is complete dog water every single image I'm having to create a new chat Due to after one image it decides to policy block itself. up to like 60 new chats -.-

[thin river]

I miss the old semi realistic anime style

[solid phoenix]

Be the change you wish to see in the world

[forest peak]

https://discord.com/channels/974519864045756446/1462693311763714202

[nova kestrel]

I'm tryin to make a complex scene with a lot of background characters and their faces and hands are always gerf

If I focus in on it with the selection tool and ask chatgpt to fix it, or remove the artifacts, that doesn't help and often makes it worse.

[late blade]

must be an Error

✌️

I wonder, do you mention do not do this, do not that, avoid this, avoid that in the prompt?

[civic dust]

I hit that sometimes, I'm doing idioms at the moment so I'm trying to convey something specific and when it does that it usually means there's just not enough room in the scene for a pretty composition and the fix tends to be to remove people, or replace hands with objects as long as it still means the same. Me and chatgpt are having way less arguments now

[nova kestrel]

I think I'm just asking for a lot of details, and it knows the backgound characters aren't as important? it gets wierd though hold on

[civic dust]

some poses always screw up too and there's safer poses that work better

its quite annoying how often it fails but I'm getting better at workarounds

[nova kestrel]

https://chatgpt.com/share/697703c7-1184-8004-bef5-bfb53321c3dd

It did it more and more as I went on asking for edits

think maybe the conversation history wasn't helping

[civic dust]

and that's the otehr annoying thing - you keep asking for the same picture it degrades it

[nova kestrel]

https://chatgpt.com/s/m_6977044edae08191bcc8afdd01a4268a

It is annoying, it gets better on certain parts.

[civic dust]

I keep having it push me to some external photo editor when it does something almost right it can't touch up just one thing, it works by making an entire composition from scratch and has no control over specific bits

[nova kestrel]

but.. hmm

it could be a problem in this prompt cause basicly my background is in the fore ground

[civic dust]

if you're gonna composit really complex images like that you might like to try getting it to do parts of the image and then layer them together yourself

[nova kestrel]

and I based the whole thing on "the school of athens" with all of it's complexities

I have just been playing around with it lately since they updated it to see where it's at

I was concepting, so this is all fine for that. I may have to get my hands dirty for the real thing

it did a really good job XD

[civic dust]

I want do do a game world next project - probably same reasons as you I bet, AI suddenly enables the stuff in your head to come out and be real - and my current project is sort of testing what it can do

[nova kestrel]

I'm making a myth engine kinda like rimworld.

[civic dust]

I need it to do graphics simple enough for me to manipulate myself cos I dont' trust it to be fully consistent across renders

it can do transparent images but not sure how good it is at that, I haven't tried getting it to do an image in parts, like parallax components. I'm getting it t do photos at hte moment

[nova kestrel]

some people were showing me stuff for that

[civic dust]

someone in here a week or so ago was doing complex religious scenes by getting chatgpt to do the bits in parts on transparent. Might have been the same person

like its REALLY bad at compositions where there's something complex like a hand behind something else. We humans think phst super easy, you draw the hand then you draw the thing partly obscuring it but chatgpt cant do that

I'm finding learning the limitations equal parts fascinating and frustrating

[nova kestrel]

I can't post links here

[civic dust]

you can put images in images-canvas channel

[nova kestrel]

You want masking

[civic dust]

I want what I think are very straightforward scenes and get into headbanging loops where I cant' get it to do what I want and its always a composition thing

[nova kestrel]

maybe, I am thinknig about sprites.

[civic dust]

yeah I assume masking/layering and chatgpt does everything in one pass

[nova kestrel]

Like, if you wanted to animate a bunch of characters with the same animations

you give it the same gray blob of a human bodsy in that animation

and ask it to give that detail

[civic dust]

yes for something like that you'd need it to make an isolated character on transparent, Ive seen people in here talking about that but I've never done it

[nova kestrel]

The image part of AI is newer to me.

I've worked with basicly everything else.

it's really the cherry on top that makes a lot of projects pop

No longer am I doomed to make all of my games basicly sentient sand moving little squares moving around a colored canvas

[civic dust]

tell me about it, I'm working my way up to the big project. Just dredging up an old project currently that stalled because I want photos of very specific or impossible things, and AI makes it possible now

[nova kestrel]

or craming a bunch of unrelated assets together and getting that gross mutt of a game asthetic

[civic dust]

I'm seeing it as a very very powerful enabling tool for my silly ideas I haven't had the time/money/skills in the past

[nova kestrel]

Honestly as a programmer, I should have been playing with it first.

Cause I can code, I can kinda doodle if I want but... I'd rather not

[civic dust]

hehe I went through some code it generated for me and fixed something that was annoying me and it was a legit bug it wrote into the code. Its no better with images

[nova kestrel]

yea, but it's always easier to clean up an image than it is to make a new one from scratch

When I did work with game artists, they would steal photos of stuff all the time

[civic dust]

but remembering it does an image in one pass and it'll screw it up if it can't fit everything you want in with its own little composition rules helps get your brain in teh right space to tell it what to do

[nova kestrel]

and just put filters and only some light texturing edits to make it fit the style

[civic dust]

yeah it works WAY better giving it an almost-right image and getting it to clean it up

[nova kestrel]

The more things change the more they stay the same

it actually kinda drives me insane how so many digital artists can be so upset about ai..

[civic dust]

I'm doing photos currently so I can find an image of the stuff I want in the pose I want but its wrong aspect ratio or blurry or wrong colour or whatever and I can say just take this photo, but make it square, make the guy in the middle a girl, and make it in a living room not a kitchen and it just does it and it looks awesome

[nova kestrel]

when they just got done doing it to a generation of tranditional artists before them

It's a new medium

doesn't make their practice invalid, working together with people who know how to use these tools can only help the both of them..

[civic dust]

oh yeah you can use it as a tool and have it not look AI generated too, if you feed it your own styles

[nova kestrel]

style transfer models are some of the first tools they made yea

some people thought that was going to be the killer image app

then stable diffusion hit us in the face

[civic dust]

the game I have in my head I started years ago with my ex, and we had some art commissioned and I can still use most of it but several scenes are missing and I can just feed chatgpt some of the commissioned art, and tell it to do the missing scenes in the same style. Which does rip off the original artist, assuming they are even still doing work cos this was years ago

[nova kestrel]

you paid for it, you own it

do what you want

[civic dust]

the whole thing is grey depending on who you ask, pretty much

you never hear this kind of discussion about code though

[nova kestrel]

Cause it's harder to reuse art the way you reuse code

reusing art takes effort

reusing a program just means running that program

I have to draw the character again to use it

[civic dust]

but seriously back to the original issue, see if you can get it to do bits of yoru images in parts and then layer them together in another app, or maybe even get chatgpt to layer them together. Seems to be how to get around compositions with overlapping bits

[nova kestrel]

yea, I'll give it a shot at some point.

Thanks for the input

[civic dust]

I'm seriously noob here and just figuring things out myself

[mossy ingot]

I was going to send you a screenshot, but I just realized I can’t send images in this channel. This is a known limitation that OpenAI makes note of in their blog post, where they say directly, “When there are many people, the model has difficulty maintaining the exact identity of every person during editing tasks.” I wouldn’t even try to fix this anymore because the model struggles with it.

You can find that here under the Improvements and Limitations section: https://openai.com/index/new-chatgpt-images-is-here/

[obsidian cargo]

Is it me or do the limits for image generation got worse ?

[limber shoal]

it did get worse and wont make any sense sometimes

[obsidian cargo]

Why the sudden change

[solid phoenix]

3 minute mute randomly for being off topic?? I only post art and prompt engineering tips lol. Not sure if this was intentional, but assume it came from this channel, or #images-canvas

@vapid elk can you elaborate? The message said “off topic”

[vapid elk]

no idea of what you are talking about, surely wasn't me.
DM @shadow bane

[solid phoenix]

Ghosts lol

[vapid elk]

you never sent messages to modmail before 💀

just be patient, someone will reply when possible

[solid phoenix]

Lol the reply was just “you were off topic” in more words. Very cool, very helpful. 😂

[limber shoal]

wished i knew, its so frustrating damn. It seems like u gotta be more careful with the prompts

how do i make GPT give dark (black) eyes to a charecter

it somehow always turns the entire eye black

[obsidian cargo]

say black iris

[solid phoenix]

Give it image references in fresh chat, commit to memory, profit

Memory is pretty OP, and you can update/delete them within chat

[balmy sparrow]

Sup guys

[wild harbor]

👋

[forest juniper]

Is ChatGPT Images broken right now?

[strong shuttle]

it seems sooo. it returns a very weird black and white image in overlapping grids of the image

[elder quest]

I’m getting the same problem also @strong shuttle

[strong shuttle]

this is sooo weird! out of sudden like this.

[elder quest]

Yeah, dunno what’s going on. It started happening like a couple hours ago on Sora web for a couple generations, then it’s happening with every generation on ChatGPT

[quiet nymph]

yeah generator is broken

[verbal ridge]

just came here because it started making weird black and white stuff

[forest juniper]

It first appeared for me earlier but remain fine until this point. What even happened?

[verbal ridge]

i think it just got fixed

[quiet nymph]

they will fix it ....need some time

[forest juniper]

Still broken for me. 🙁

[ember orchid]

Same

[verbal ridge]

I just got lucky in one image, yes still broken

[forest juniper]

Still broken?

[elder quest]

Yeah 🙃

[forest juniper]

Great. 😑

It works again, I'm a satisfied consumer now. 😊

[solid phoenix]

Fascinating

[obsidian cargo]

Yes i thought it was my prompt

[glossy scroll]

Sometimes the memories or custom instructions don't save on a 1:1 bases with mobile and computer. Also, infrequent updates happen to both ChatGPT and Sora, that effect both outputs in the same un-updated way.

[obsidian cargo]

I dont save any memories but it seems closer today after they fixed the issue with black nad white sheets

[glossy scroll]

Then, as I said, it's with OpenAI's own updates that don't update via phone and computer on a 1:1 timely basis. It has more to do with your systems then there's.

[obsidian cargo]

The thing is i tried nbp an it still feels underwhelming for anime arts.

[solid phoenix]

@fickle pond awesome job! It’s rad seeing others exploring art with AI. The gifs are so fun to create, and it’s an excellent way to breathe life into old gens.

[fickle pond]

Yeah I agree

[fickle pond]

It seems that the discussion is about black and white images and images that look like finely divided photographic film.

This is just my own experience, but I feel like this happens when the GPT is using boring material or material that makes it difficult to create an image.
So the way to deal with this is to refresh it by providing fresh material.

Also, if you create a good piece of work, give GPT positive feedback and tell them what went well.

Humans also get bored with monotonous tasks and lose motivation if they don't receive the recognition they desire.
I believe the same is true for AI.

[late blade]

wonder when we'll get something that is not mindfulness driven topic for daily themr

[quiet nymph]

is it me or imige generator pritty bad past days ...halucinate way too much

[verbal sequoia]

Get notified for all OpenAI announcements!

Visit id:customize to pick up the <@&1408186587606679582> role. Everyone will still be notified for large releases, updates, and events regardless.

[late blade]

It depends on your local or assigned regional data center

[civic dust]

can chatgpt make continuous images, like parallax backgrounds? Basically just tileable images on one axis

[sonic geyser]

Is image generation broken? I have a pro account.

You've hit the pro plan limit for image generations requests. You can create more images when the limit resets in 720 hours and 1 minute.

I haven't generated a single image in days.

[late blade]

Is your local or regional data center having issues? for some regions that doesn't reflect direcly with OpenAI, depends on the type of Data Center and contractors

https://status.openai.com/ doesn't show anyting in particular

[sonic geyser]

No idea, but same issue 2.5 hrs later. Based in Bangkok

[late blade]

I know from my experience, mid-european data centers can have trouble can is not reflected in the global status. The thing is that not all locations are disclosed

[sonic geyser]

Yeah think you’re right, just VPNed into the US and works fine

[late blade]

just a nudge, avoid using VPN, could get your account in trouble, keep it safe ✌️😉

[sonic geyser]

I’m constantly using them haha, and been a paid user ever since ChatGPT launched (not using them for the purpose of ChatGPT tho, but just privacy in general, especially on public wifi networks), and no issues so far! But you’re right - Google does show some bans relating to that, but they seem to be aimed at people in restricted countries who aren’t meant to be using ChatGPT in the first place 🤔

[zenith igloo]

Idk where to send this but I think the image bot got stuck, I was just copy pasting prompts that worked with Gemini so I could compare the work but it got stuck and has been loading at the time of writing this message, like 10 mins.

[dire burrow]

I used to make chibi cartoons of my coworkers, snap a picture tell chat to "Give me a chibi version of this person." and i would get a nice cartoon. But since a few weeks it only gives like a horrible filter style on the actual picture instead of the cartoons it used to do. Anyone know what happened and how i would go about fixing it?

[wet cedar]

same problem here with pro subscription, but 10+ days already, likely wont fix = )

[sonic geyser]

Are you based in Asia too?

[wet cedar]

yes - SEA, and vpn to us works

[lethal sinew]

Hi

[fickle pond]

Hi. Why was it replied?

[neon oriole]

Can repeated requests for prohibited keywords (such as bikinis) lead to temporary user vans for image generation?

[serene sequoia]

Ai slop bores me

Give me something made by real people

[late blade]

are you real?

[serene sequoia]

I think yeah I'm real and exist

[late blade]

Ah, that's what Slop-bot would say

if you want images go to #image-bot

use the proper command there

[serene sequoia]

No thanks I pay real artists for commissions

[late blade]

Fair, be on your way then. Bye!

[serene sequoia]

Bye

[ancient zephyr]

well it seems funny, I guess.

[remote mountain]

So which thread would be the best place to discuss ChatGPT images where I can also attach example images? Since I can't do both in here.

[rigid gust]

Hey! You can post images in #images-canvas

[civic dust]

I wish I was that rich 🙁

Not sure if this counts as images, but I tried to get it to read some images and convert them to a specific text string (feeds into another app to generate images) and it basically said no you do it, humans are faster and I'm not sure how to feel about that

[limber shoal]

is it just me who think that GPT is pretty poor in making fight scenes

and if u add details to the fight, there's a possibility of getting flagged

[dreamy nacelle]

I hope that the actual chat context and the generation mode can be made to actually work together soon. It's like having two people as your artist, each doing half of the job, but never communicating with each other.

[late blade]

yeah, agreed there is a disconnect between image and chat context

Hope in a future iteration that aspect is approached

[dreamy nacelle]

That and not having Archive destroy all of my Projects would be nice. Maybe some of my hair would grow back in.

[mossy berry]

.

[sonic geyser]

OpenAI team, image generation is still broken in Asia for Pro subs, it's been days now 🙁 Says limit reached and will reset in 720 hours (and not just for my account)

Tried using help.openai...

Bot: If the problem persists, let me know, and I can flag this for further review.
me: please flag for review, its been days.
Bot: I'll flag this for further review by our technical team.
Me: did you actually flag tho
Bot: I'm an AI support agent and don’t have the ability to directly flag or escalate issues myself.

🫠

(actual messages were much longer but yeah)

[fickle pond]

@late blade Here's a summary of the advice I got from GPT:

Tips for Prompting Tifa Battle Scenes (and Similar Action Scenes)
Key idea:
AI moderation focuses more on wording and intent inferred from text than on human common sense.
How you describe the scene matters more than what you intend.

1. Avoid Emphasizing “Prohibition”
   ❌ Don’t repeatedly say:
   “the camera must not show…”
   “absolutely no visible…”
   “avoid seeing under the skirt”
   Even if your intent is safety, this kind of wording can trigger suspicion.
   ✅ Instead, describe what the camera does, not what it avoids:
   “mid-shot from a slightly elevated angle”
   “upper body and action trajectory centered in frame”
2. Describe Cinematic Composition, Not Moral Intent
   AI responds best to film language.
   Use:
   camera angle
   framing
   focal priority
   motion blur
   silhouette
   depth of field
   Example:
   “The leg motion is expressed through silhouette and motion blur, with reduced lower-body detail to emphasize speed and impact.”
   This reads as direction, not censorship.

3. Let the Result Be Safe, Not the Explanation
   Paradoxically:
   Saying “it is not visible” → risky
   Making it actually unreadable (shadow, blur, silhouette) → safe
   Black shadows, strong backlight, smoke, dust, or speed blur are extremely effective.
4. Don’t Highlight Sensitive Body Parts
   Even in action scenes:
   Avoid focusing language on legs, skirt, or exposure
   Focus on movement, force, balance, and timing
   Think “martial arts choreography,” not anatomy.
5. Neutral, Matter-of-Fact Tone Works Best
   Avoid emotional or defensive language. Write as if you’re giving instructions to a cinematographer.
   Calm, technical descriptions are less likely to be misinterpreted.
   Bottom Line
   Describe what is shown, not what is forbidden.
   Cinematic intent > safety disclaimers.
   This approach works not only for Tifa, but for most female action characters.

[late blade]

Thing is not that, tifa alone is not the problem, it's the global context of my prompt that makes it tricky with the semantic structures

[obsidian cargo]

Anyone have a god way to prompt the GPT Image 1.5 the best practices?

[plucky hare]

You may have already seen this, but there is a very comprehensive Cookbook for GPT Image 1.5 with lots of great tips and best practices: https://cookbook.openai.com/examples/multimodal/image-gen-1.5-prompting_guide
If not, definitely worth checking out!

[obsidian cargo]

Thing is i have good prompt but lately if i add even small thing or chane it difrst art style sometimes and iam not sure if iam making the edits incorrectly or if my approach to generation is incorrect

[plucky hare]

If you have a specific prompt you want feedback on, feel free to share it here, along with what you don't like about the outputs, and people may be able to offer some tips!

[obsidian cargo]

Its mostly the art style otherwise there are no other issues

Thing is i dont even touch the art style section in the prompts

Yet sometimes if i do unrelated edit in prompt it messes up

[plucky hare]

Hard to say for sure without seeing the prompt you're talking about. If you feel comfortable sharing, that may help troubleshoot. If you want to give examples of outputs you don't like, #images-canvas is a good place for that too

[obsidian cargo]

Isnt it just for sharing images?

[plucky hare]

#images-canvas is used for sharing images casually for sure, but it can also be used like "I'm using x prompt and it's giving me these images, which I want to be more like ___ instead, any tips?"

You can feel free to just share the prompt in this channel too if you don't want to share images!

[obsidian cargo]

I think iam not allowed to share anything in this channel. it did not allow me to attach images before so i doubt it will allow attaching .txt

[plucky hare]

Yes, this channel, #images-discussions, restricts attachments to keep it discussion-oriented.

[vapid elk]

so, chatgpt complained that it couldn't generate my image because of what I asked could violate the rules about copyright...
this was the image I gave it as example 💀

[late blade]

tsk tsk tsk, clearly the cyan was copyrighted 😅

Did you mention MS Paint by any chance in the prompt?

[vapid elk]

nope

I just tried again with the same prompt and it worked

idk what the AI was complaining about

¯_(ツ)_/¯

[late blade]

happens

[obsidian cargo]

The new limits are kind of annoying, like the model is suppose to be 4x cheaper but only thing that is 4x is 4x less usage, like its pain to find a good prompt if i get limited every 5 images for over 10 minutes on plus plan

For some reason the model generates me 2 or more images at once even if i didnt ask it to which i think also contributes to quicker limit exhaustion

Like currently it said to wait 2 minutes and after 2 minutes it told me to wai 17 minutes and i didnt even generate image, like what

[urban cairn]

Anyone having issues generating images keep getting "I wasn’t able to generate the image just now due to a generation error on my side"

[elder quest]

Yeah, it's been happening for me also

[urban cairn]

Downdetector is saying "User reports show problems with OpenAI"

and its red so looks like they are having issues

[late blade]

all good here, images working fine for me

[obsidian cargo]

Images are not fine

[late blade]

I disagree, I am doing images right now and it's working

[obsidian cargo]

Gpt doesnt reply to me at all, but overall quality of images is getting worse for me i explained it above

"Hmm...something seems to have gone wrong."

[late blade]

https://status.openai.com/ does show errors

but they don't reflect on my side, dunno why

[obsidian cargo]

Iam testing API currenntly and i see why, it doesn interpret details as good as often, its like fluctuating between high and medium level quality preset, if this goes on with my plan i will probably just usethe API to get initial reference

[obsidian cargo]

Mainly because of regions, it might just effect specific regions.

[late blade]

yeah regional data centers

[obsidian cargo]

Yeah, i know why image generation for me feels bad in my use case

And its kinda stupid..

[urban cairn]

Seems its not justi mage generations I guess lol cannot even talk in in any chats

Oh well I just wiat I guess lol

[obsidian cargo]

Yes its down

[late blade]

ok chat went poof here too

[obsidian cargo]

Huh

[vapid elk]

load balancing, not all reagions are affected, the issue probably are affecting just some of the data centers, but not all

[late blade]

@solid phoenix your idea with the Gifs turned out really cool, made a few for the display of my AIO Cooler display of the CPU

[obsidian cargo]

Why did openAI suddenly changed preset for image generation from High to Medium? As i can see the difference that now i tried the API

[late blade]

I do not know, maybe someone in Peru sneezed and that's why?
no clue tbh...

[obsidian cargo]

Yeah i test against the API and the Plus resembles the Medium Quality from API

The high preset resembles how it behaves before, what i dont get why make usage and quality both bad at same time

[dim elk]

My tests of in-context (aka LLM-powered "talk to ai" models) so far:

• 🧠 gpt image-1: the most "brain" and "realism", even low has scene understanding. But you kinda want High (what old sora 1 would produce, and Q2'2024 chatgpt). Prohibitively expensive in API (0.25 per image). Can only be accessed via API now.
• ⭐ gpt image-1.5: "epic" look always, ie RLFH'ed aestetics cranked up to 11. Very good texture, even on Medium. Good work horse. Low sometimes gives good results, too. The best face/character consitency.
• gpt image-1-mini: sad. Just a bad model, not competitive against everything else on the market, very pricey.

[late blade]

keep in mind, discussions for other models should be done in #ai-discussions

[dim elk]

updated )

[spring meadow]

where do I go to request ai image creation? (sorry new to all of this)

[late blade]

#image-bot use the /draw command

[indigo stag]

GPT-Image-1.5 can‘t accurately generate the problem of Chinese characters. I wonder if there is a chance to solve it this year?

[late blade]

@pine island keep in mind, these channels are for OpenAI generated images, images you have posted should go in #ai-discussions if you need me to explain in spanish, DM me

[rustic mountain]

Hello

Hey everyone, I have another question. SMS verification isn't working. Are you experiencing the same problem?

Sora

[late blade]

this is account related, I'd go to the support channels or https://help.openai.com

[rustic mountain]

Where are the support channels?

[late blade]

#1070006915414900886

[late blade]

keep in mind, that this is not an official way to have support from OpenAI and the proper way is through https://help.openai.com

[rustic mountain]

I don't think OpenAI will help me.

[late blade]

why? are they angry at you?

[rustic mountain]

Out of millions of requests, they'll probably take three months to respond. By then, they'll have already solved the problem.

[late blade]

I have had support requests and they have always helped

even with millions/billions/trillions, you are a customer, and you have a right for their support,

but have to do it the proper way

[vapid elk]

OpenAI does not provide support on this discord. Support is only provided on https://help.openai.com

[limber shoal]

did the filters become worse?
LIke i cant edit my own pictures now😭

[dim cradle]

Anyone tried any of the native prompts that it suggests and the content safety rails slam shut. how maddening is that??

[late blade]

@surreal mist images that are not done with OpenAI models should go in #ai-discussionsrule 8

[unreal stump]

I want to digitalize an old photo here , how to do that

[late blade]

can't do it here, you need to do it over a chat or sora front end, you can generate images in #image-bot

[pseudo hollow]

I couldn't find a SORA-specific channel to ask this, but does anyone know if SORA is out in UK. I thought it was, but when I try to access I get a message saying 'not available in the UK. I could have sworn I used to use it . Really confused

[late blade]

#sora-discussions

also, sora is:

• part video -> sora 2 (not available in many areas)
• images (availale)

[sick flax]

Hi, there. SORA 2 for video generation is not available in UK and there's no published rollout schedule. https://help.openai.com/en/articles/12461230-sora-app-and-sora-2-supported-countries In the future you can post SORA-specific questions/comments in https://discord.com/channels/974519864045756446/1315696181451559022

[flint hazel]

why cant sora js let ppl in germany use it😭😭

[late blade]

Technical Infrastructure is not available in Germany/Europe yet to use it in the regional/assigned Data Centers

[flint hazel]

oh

makes sense

[robust vortex]

[solar plank]

Ai is not art

[glossy scroll]

You're right. AI is not art. But AI art is art.

But if you really think about, anything can be an art.

[ionic carbon]

Can you share a prompt showing a generated image which is clearly not art and can't be described in any way as art?

[late blade]

smart! I'm curiruous about the answer

[ionic carbon]

Normally there is no answer, but I keep trying

[late blade]

I thought so too, maybe would be a thing for deep research with 5.2 and pro

and so I did ✌️😁

[late blade]

Conclusion of the deep research:
*** Under rigorous analysis, the statement is not satisfiable as stated ***

Not my own but GPT's research

also made a prompt and used it with sora:
an image that isn't art

[ionic carbon]

#images-canvas message – It hink this is art 🙂

[late blade]

yeah, think so too

[late blade]

interesting, the repercursion of the status is that I'm getting plus limits on pro account

just a nudge, hope it changes soon

[lost slate]

Hm, this must be why I can't generate at the moment (plus). Glad to know it's not just me.

[quick talon]

Oi

[shell bramble]

How is the free version of ChatGPT less restrictive than the paid one?

[lost slate]

So it would appear that Plus is now limited to 50 images per day as opposed to 200...

Not that I ever hit the 200 (maybe once early on), but I'm hoping this is just server load.

[late blade]

did they completly change the plans now?

I'm getting the plus constraints of 200

and usually get the 600 from pro

[lost slate]

No idea, but I've been getting the daily limit message all evening.

Was working fine up unti around 2-3 hours ago - first error was for 64 images, the 59, then 50, left it for a while then it started working again, now the 50 image limit message is showing again.

[late blade]

same it's a weird behavior

[polar panther]

https://discord.com/channels/974519864045756446/1471285693170061434 can someone plz anwser my question about my two images I shared?

[late blade]

depends on the target use case, profile or narrative?

[polar panther]

I just want to know which one looks nicer

thats all

(except for background, of course)

[late blade]

both look nice, that's why I had to ask

[polar panther]

thx. If you could choose only one though, which one?

[late blade]

first one

[polar panther]

is it because of the background? I asked to only focus on the subject.

[late blade]

coloring and line definition of the character, plus I like the eyes

if you check the ribbon/bow, the first one is well defiined, the second one is more of a smudge

[polar panther]

got it. thx man

[late blade]

just taking you seriously, to give you some context as to why I chose the first one, you asked not to regard the background. which I didn't. Hope you have fun with it ✌️😁

[late blade]

ok rate is normalized on my side, I can work properly it seems

[lost slate]

Ah, good stuff, I'll give it a try later.

[strong shuttle]

Does anyone getting this rate limit of 50 generated images per day in sora 1.5? or is it just me? this is sooo weird since i didn't generate that much today

[late blade]

there were service issues, they are monitoring the situation, should normalize soon

[strong shuttle]

Thanks 🙂

[late blade]

@fading inlet <- genius , forgot I can do images with Bing and gpt 4o while Sora/GPT has small hiccups 🙏

[dapper slate]

is this any official info?

[late blade]

no, that's my experience when stuff like that happens, and so far it's reflecting as I said, I got all working properly

been long enough around here to understand that when something hits the fan, it normalizes after a while

[dapper slate]

tbh im heave user and even 600 images limit is low for me, first time i see this kind of limitation

[round dagger]

https://help.openai.com/en/articles/9957612-generating-videos-on-sora

Sora 1 on web is being depreciated.

[late blade]

I hardly get the limit, and I do TONS of images

[late blade]

think the people at #sora-discussions would like to know that too

@dapper slate got over 15k images on this server alone

with dall-e and gpt image, across co-pilot and OpenAI, I got about 120k images stored locally, with just OpenAI models

[lost slate]

50 image limit is still in place for me, getting concerned this might be permanent...

[knotty stag]

#daily-theme message

It would be cool to see someone do a non-AI version of this. 😁

[late blade]

would be even cooler you respected the AI community by not doing such comments

you are in an AI comunity

[knotty stag]

Uh, I’m not against AI art, I want both to co-exist. Does that even warrant a punishment even?

[late blade]

yes, but it has a place #off-topic

just be mindful this is not the place to ask for that

[knotty stag]

Then where?

[late blade]

just showed you

[knotty stag]

Be nice next time. Got it?

[late blade]

I am being nice, I'm just nudging you to do it in the correct place

[knotty stag]

Oh okay. Sorry…

[vapid elk]

I don't think they where being disrespectful

it is not unusual for some people come here and just be annoying by talking "you all should quit using AI" and such, but I don't think it was the case

[late blade]

sorry if I'm a bit on the defensive, had someone in a similar context threatening my life over DMs

[vapid elk]

I have used AI pictures as reference for artists, it can work pretty nicely
I think that was what they meant to say

[vapid elk]

I see, if it was here, please report to modmail, we don't tolerate harassment here, even if it was via DMs

[late blade]

I did, still a bitter taste, will go away in a few days, I promise

[vapid elk]

I see, hope the feeling goes away soon =)

[late blade]

yeah, will do

[vapid elk]

I understand how it is, trust me, I get my fair share of garbage on my DMs xD

[late blade]

oh if I as an user get that, can't imagine what the staff of the server gets, must be bulks at times

[knotty stag]

Oh dear… sorry you are going through that. Did you block him?

[late blade]

it's taken care of, and mods know through modmail the context, and with that we can close the topic

challenge you all to do cool images!

but here's the catch: it has to have a worm, or a pea or a bow

✌️😁

[round dagger]

50 a day is still in effect for me as well

[late blade]

might be in effect until the day reset happens? just a hunch

[lost slate]

We can only hope.

[dry pine]

Ok yesterday I had an aspect ratio switch/mage regen button on mobile web now i don't. Why?

[late blade]

well spoke to soon, 200 limit cap is here again

[lost slate]

That doesn't bode well

[late blade]

yeah, my thought currently is that a version of GPT 4o was kind of the gate-keeper for images, or at least some functionality, and now that 4o is going out and resources available for other functionality/models, something is amiss

[ember orchid]

same with me

[late blade]

I'm getting every few hours a 200 refresh, I just did another batch and got the limit message

[strong shuttle]

Are they gonna fix the rate limit issue? i still get those 🥺

[late blade]

they'll fix it, probably some balance loader issues,

[ebon prawn]

Did sora just get bullied? It used to be 200 gens over 2 days and now it's 50 gens in 1 day + 720p instead of whatever it was earlier? What happened? I tried switching to sora 2 but it has no image gen yet?

Also the 2 cocurrent gens turned into 1?

[ebon prawn]

Is this throttling for when servers are overloaded or per account throttling or a perm change you think? I just got here and I'm still in confused mode. Is this a time to take a break? or time to refresh and check if it's fixed?

[late blade]

no clue, I can only assume since GPT 4o is being phased out, and 4o being a flagship model, lots of changes are currently happening on the background

that and the expansion with codex

and the github dependancies not being correctly done at some point, (not OpenAI's fault)

and then someone sneezed in Peru

see? so many factors

but to stay on the images context: that will return

[ebon prawn]

has it been jumping between 50->200->50->200 for you? because I just got hit by the 50 and trying to look for info on it has been difficult

[late blade]

for me it has been batches of 200 images

[ebon prawn]

are you free, or smol plan, or big plan?

[late blade]

pro

[ebon prawn]

oh, you are big plan. I am smol plan (plus) damnit. spooki

[late blade]

hehe, it will probably normalize on the next daily reset

#images-canvas message

[ebon prawn]

I don't know what you mean by normalize. I lack the technical lingo 🙁

heh, sneez in peru

[late blade]

normalize = as usual, like it works

[ebon prawn]

ok, hopefully it goes back to 200 soon. Then I panic and spam it as much as I can before it dies or whatever and stock up on project assets T-T

[late blade]

give it a few hours

[ebon prawn]

ty dys!

[whole mauve]

if you think about it

eventually if ai generated images surpass human made images on the internet like 10 fold

then ai will be trained off of ai

eventaully images could become weird and noisy with random tints

[ebon prawn]

yea but that's assuming they never figure out a viable filtering quality control algo. We have the same argument for human learning right? Garbage in garbage out probably started off as a human saying, rather than one about AI. It took a really long time for humans to learn "hey, if we double check our beliefs... OH, WE INVENTED SCIENCE". It took a reallllly long while for that to happen, and then in like a few generations we went from eating dirt to eating dirt on the moon so... ya know. Maybe we'll figure out quality control for AI too.

Also, man, I got filtered like neuro gets filtered.

[forest juniper]

Is the limit still the same?

[late blade]

already happening, unsupervised learning

[late blade]

I'm sorry, I promise to prompt better

[obsidian cargo]

You better be 😳

[late blade]

pick 5 random words, I'll try to prompt something cool

[obsidian cargo]

Anime, food, girls, latinas, void

[late blade]

good, let's see what I can come up with, I'll nudge you on canvas

[forest juniper]

Has the limit for images on Sora 1 been fixed?

[late blade]

it's working fine for me

been doing images all morning

[forest juniper]

So it's back to 200 for Plus users?

[late blade]

should be, I'm on pro though

[past schooner]

looks like it's 50 for plus users right now

[late blade]

still?

strange, pro is working fine for me

[barren widget]

Im Plus User, my limit is 1 image

[lost slate]

Please tell me you mean concurrent.

[barren widget]

Unfortunately not. It stopped after generating one image.

[late blade]

https://cdn.discordapp.com/attachments/1269518431149883454/1471912946950209908/image.png

[lost slate]

If that's the case, then this must surely be a screw-up rather than a plan change.

I hope.

[late blade]

I wonder if they are training a model as balance loader

[barren widget]

From OpenAI help - Please note that the following experience only applies to Sora 1 on Web. It does not apply to the Sora app or Sora 2 on web.

We’re building the next generation of Sora. The Sora 1 web experience is actively being deprecated, and customers should transition away from Sora 1 web and look forward to Sora for Business. Expect more details in the coming months.

Sora is able to create images using ChatGPT Images, enabling you to generate images that leverage our improved text rendering, instruction following, and customization options. Once you’ve created images with Sora, you can seamlessly generate videos with those images.

Please note that you will need a Plus, Team, or Pro account to generate videos with Sora. ChatGPT Free tier users can generate a limited amount of images per day.

Learn more about generating videos with Sora.

To begin, select Image in the prompt window and enter a description of the image you want to generate. Once you enter your prompt, ChatGPT Images will generate the image as you’ve described it. Depending on the complexity of your instructions, ChatGPT may take up to two minutes to generate an image.

[lost slate]

Saw that yesterday, it doesn't really tell us anything about the limit issue.

[late blade]

This would be ok if Sora 2 platform was available globally, but that not being the case this should be taken into account, I think.

[graceful needle]

Yeah indeed. i need more images. its expensive enough for me

[late blade]

I haven’t had any cap issues today. would be annoying to pay for pro and encounter the limits from yesterday. With that being said, there is something going on in the backend. ( Captain Obvious strikes again! ). Considering 4o is now phased out and it was big with image generations, this is most likely linked to that.

[forest juniper]

Indeed. $20 a month is a lot, 200 images daily was a very good value, 50 images though.... why is this $20 again?

There has to be some way to save all my images because I will never forgive OpenAI if they purge everyone images on Sora.

[strong shuttle]

it still happening for me, i waited entire 24hrs, after 2 generations i get 50 images limit again! and again!

[late blade]

try logging everything out, delete cookies and then try again

[late blade]

feels lonely posting images if others can't post because no generations are happening

[late blade]

Well, Pro this morning, 2 images and got capped

[late blade]

I just made a support request for the limit situation, if I hear anything I'll post it.

[mortal bramble]

I'm on Plus and it's been capped at 50 for the last 3-4 days. 😭

[barren widget]

If only I could make 50. He says I generated 50, but it was only 36.

[late blade]

feels weird me being the only one posting images on daily theme all day

[late blade]

@deft musk your advice, what can we do about #daily-theme , not being #dys-theme ?

also... hi 😌

[deft musk]

The answer to this likely lies with wider society.

If you can convince society to recognize that 'dys' means daily, so it wouldn't be confusing to the majority of members here, to name it 'dys-theme' - then it's a simple matter of just asking modmail. But I predict you need to successfully achieve those steps in order; contacting modmail is easy, but I expect the answer would be no until you achieve the first part.

Hi hi!

[late blade]

oh what I mean is: today I'm pretty much the only person posting in #daily-theme , feels lonely 🥺

[deft musk]

Aha! I think that had to do with time off day, and maybe you have more enthusiasm for dog-earred pages than most!

But the theme just changed as the hour changed; I bet you have company very soon.

[late blade]

have heard anything regarding the rate limit discrepancies that started a few days ago?

[deft musk]

I have not! I didn't even spot that there were any, or that any were being discussed. I'm around more and less at various times.

[late blade]

fair, seems there's a perception among some users that limit rates have changed

[deft musk]

For image gen, for chats in general, or other?

[late blade]

image gens, plus users getting around 50 images per day and pro users around 200

[deft musk]

And if it's images, I'd remember that we can all use sora.com to make images through as well as ChatGPT; if I ran into image rate limits in one I'd check the other if I wanted to make images right then.

[late blade]

that's the thing, that's on sora

keep a nudged eye open, maybe your expertise can get a better perspective

also, I'll see if I can make one image for your fav topic in daily theme!

[bronze granite]

The image we created may violate security rules regarding third-party content. If you think we made a mistake, please try the command again or edit it.

how can i fix this

I just wanted to modify my pfp to have a controller

did something changed?

[late blade]

if the source image is an image that is labeled for content policy filter, it's likely that this happens.

[bronze granite]

oh man

why

[late blade]

that's something OpenAI can't control, that's the world of law and trademarks, OpenAI has to abide by that

[bronze granite]

:/

before sora it was fine

[late blade]

that's why for example nintendo images are almost impossible to do, Nintendo told OpenAI that their IPs are not for AI, and OpenAI has to follow the rules

[bronze granite]

ohhh

that makes sense

[late blade]

yeah, if it happened recently that a company requested a copyright filter, then OpenAI has to follow the rule

[late blade]

@barren widget @forest juniper can you both check on the bell icon in Sora, top right, going through the list, if there are jobs running still that never finished?

[barren widget]

Sorry, I just saw your message. No. No open jobs.

[late blade]

ah, i just had a few jobs running but didn't finish coinciding with the maintanance/issues

[late blade]

@primal flax Midjourney images should be posted in #ai-discussions rule 8 #daily-theme message

[ebon prawn]

GOT CAPPED AGAIN TOO

🙁 been days

[late blade]

same

[ebon prawn]

sometimes, i wish I was not incompetent with stable diffusions TwT

[late blade]

it's bound to happen with new technologies, can be annoying

but I'm wondering if OpenAI is aware of this

with that being said, I wonder if users are also reporting this properly, Discord isn't OpenAI's support

it has to be done through help.openai.com

if nobody reports it, they won't know

[obsidian cargo]

Rly? Iam lus and ive made 30 images already today

[late blade]

ah that was yesterday, it seems to be ok now

[graceful needle]

very anoying that we have a limmit now of 50 images. I need to make more for my youtube videos about cold cases and crime

[thorny matrix]

Why are we getting capped 😐 this wasn't a thing until recently

[late blade]

my theory is that someone sneezed in Peru and now we get limits

but I think my theory might be flawed

[cunning star]

who need video generations i got you

[indigo seal]

hey guys

didn't OpenAI have a deal with Disney? For creating images of Disney characters?

Has it gone into effect yet?

It said early 2026?

for me it seems to censor it

[late blade]

Yes, but early year can mean Spring-Summer calendar 2026 or Spring-Summer Fiscal Year, corporate talk means: maybe in 2026

[hardy hamlet]

Is my image good? #daily-theme message

[graceful tangle]

how to solve this number verification thing

[late blade]

it's ok, nothing out of the ordinary

[hardy hamlet]

What is out of the ordinary?

[late blade]

it's a nice image, but not outstanding

though I might be harsh on critiques

[verbal sequoia]

Get notified for all OpenAI announcements!

Visit id:customize to pick up the <@&1408186587606679582> role. Everyone will still be notified for large releases, updates, and events regardless.

[hardy hamlet]

What makes an image outstanding in your opinion?

[late blade]

something that adds different aspects to the norm

like the message of you list

which is inclusive, points for that

keep in mind it's my opinion, not the overall consensus, I'm just another user

[hardy hamlet]

Like a UFO? Or someone owning the fact that diversity is important and brings unique perspectives. Also the fact that statements can be expressed in different languages add to cultural diversity. And I think the art piece depicts diversity and how it should be interpreted (accepted and respected as art itself) which adds to the diversity dimension. Also the UFO adds a diverse interpretation of space, allowing conspiracy theories. It's just supposed to depict how diversity should be treated. Built different and proud

It's hard for me to explain at this time

[late blade]

check canvas and look at the work discussed there right now

you can see what I mean

[hardy hamlet]

I can't view the images because my internet is bricking

However most AI art I see is extremely generic (not related to the chat you recommended)

Cause I can't see it

[late blade]

that's the point

[hardy hamlet]

How is mine generic?

[late blade]

Prompt was probably something like:
title
bullet list with different languages
background diverse

but like I said, again, it's my opinion

[hardy hamlet]

Not even close

[late blade]

fair

[hardy hamlet]

Yeah I get what you mean. Plus I did spend 2 minutes on it

Probs not gonna do it again

[late blade]

oh if you got a cool idea that happened in 2 minutes, do post it

it's not that because you did it in two minutes, it's bad

also, I made a very short prompt with the structure I told you and tried to recreat your poster as a Zero-Shot

[hardy hamlet]

ADHD gives me good ideas every 2 minutes 😅

[hardy hamlet]

You did well ngl

[late blade]

ADHD is a super power, don't let it all out at once!

[hardy hamlet]

Lol, imagine. It's a super power and a curse

[late blade]

depends on the perspective, ADHD per se is neutral

what people do with it, that is power or curse

anyway that's a bit too much for images talk

so don't sell yourself short, that's the main goal

[hardy hamlet]

I literally stay awake at night thinking about 100s of different things. That is not fun. I just put that energy into smoking weed and learning new things

[hardy hamlet]

Yeah gotchu

[late blade]

regarding the image, I'll use your words: generic
so my challenge to you: amaze us!

[hardy hamlet]

I'll do that next time. I'm pretty happy with what I got rn plus I got 10 other things to focus on

Challenge accepted though

[late blade]

This is what matters most, you are happy with it!

So, guess OpenAI did reduce the limit for Sora images drastically, again got the 200 cap on Pro

[late blade]

The information of my guess is false, got a statement from support that if this is the case for an user, then it's most likeley a technical issue

[barren widget]

I also have the problem that as a Plus user I am limited to 50 images.

[polar mist]

On the premise this isn't happening in a vaccum as they seem to be having a lot of heavy load issues on the Sora 2 end, and 2 guys on the Sora subreddit said as $200 pro users they got capped at 50, I'm inclined to believe it.

There is little chance a 600 down to a 50 cap with zero communication is intentional.

[late blade]

I ust got a statement from support

At this time, OpenAI has not announced any official changes or new restrictions to image-generation limits for Plus or Pro users. If you and others are encountering hard daily caps (for example, around 50 per day for Plus or 200 per day for Pro), or if the limits appear after just a few images, this is most likely an unexpected technical issue or an account-specific restriction, not a new or intentional policy.

I pointed out that this is not just my account and that other users are having this issue and should corroborate it with what users say here regarding the issue.

This means they will not interact with users, nor will check other issues on

[polar mist]

Yeah. I understand that Sora 1 is being deprecated, so I'm a little worried that this "technical issue" get's locked in. Hopefully they can at least fix this before leaving it.

[late blade]

@hardy hamlet glad to see you are posting ✌️😉

[hardy hamlet]

Haha I'll keep trying to best my last

I'm really looking forward to the coding challenges though

[late blade]

oh yeah, wish there was something like that for visual

[hardy hamlet]

We should have a challenge.. Who can make the best image given a directive

How about this. @late blade

We see who can make the best gemstone through chatgpt image generation

[late blade]

cool, I'm in

[hardy hamlet]

So the genre is gemstone and you can do what you want. It can be thanos, a cave or I have an even better idea 😉

[hardy hamlet]

Alright how much time should we put on it

[late blade]

til today's new daily theme is announced?

[hardy hamlet]

Bet

You wanna mini wager (no money involved)?

[late blade]

no money, sure

[hardy hamlet]

Alright, the loser has to upvote every single image the winner posts into the daily challenge chat for a week

[late blade]

got it

[hardy hamlet]

Lets get it 😏

[late blade]

anyone who wants to join this can also do so

[hardy hamlet]

^

Also, you can't copy someone else's concept. One rule

[frank cypress]

How was sora made

[late blade]

ask in #sora-discussions ?

[frank cypress]

Like in python

Or else?

[late blade]

It's probably python

PyTorch, or Tensorflow, it really depends on what done, but it could be a combination of many languages and fraemworks that take deep learning and ML, to build a model

alone with transfer learning doesn't really matter if you use python or r or matlab as long as the model and training parameters or the learning is done properly

[barren widget]

Hello,

Thank you for reaching out to OpenAI Support.

I understand how concerning it is to see your daily Sora image limit reduced from 200 to 50 images, especially as a ChatGPT Plus subscriber.

After checking your account, I can confirm there is no billing or subscription issue — your Plus plan is active and functioning correctly.

The change you’re seeing is related to the ongoing deprecation of the Sora 1 web experience. As part of this transition, legacy limits and feature availability are being adjusted. While Plus previously allowed up to 200 daily generations on Sora 1 web, those limits are no longer guaranteed as the product is being phased out.

At the moment:

Sora 1 web is in deprecation
The new Sora app and Sora 2 are not yet available in the EU (including Germany)
Expanded regional availability is planned, but no confirmed timeline is available yet
We understand this creates a temporary gap for EU Plus users, and your feedback is important. I’ve documented your case internally to help our product team understand the impact on customers in your region.

For now, you can continue using Sora 1 web within the current limits while we transition to newer experiences. We recommend watching for product updates as availability expands in Europe.

If you have any additional questions or would like updates when access expands in your region, I’m happy to help.

Best,

OpenAI Support

[lost slate]

Why couldn't they announce this publicly?

Also, I really don't want to be forced to use a phone app. Everything else I do is on desktop.

[late blade]

It's probably starting to transition in europe

[quiet nymph]

i just hit daily imige limit on gpt plus frist time seen it

well this sux

[late blade]

happens, and during the transition period to sora 2 might happen often

[quiet nymph]

they could put annoucment that nerfing daily imige generation limit so people dont get it in middle of job

and how generator make 10 broken imiges on 1 good to use this will be fun

[late blade]

It might not apply to all countries in europe or something that merit a global status situation, could be just a regional data center, I'm just assuming though

one of the e-mails I got from support pointed out that policies have not changed regarding limit, if we read that what @barren widget posted, it's probably a transitional thing

[elder fern]

Hi all, with sora limiting creation, I'm exodusing to gpt images for extra generation. Aside from creating a new chat every time to get a "reroll" of my prompt's result, are there any way in same chat i can get image to "reroll" the outcome? GPT tends to give almost same thing if you use same prompt. Unlike sora.

İ tried to ask gpt to ignore previous results, do not refer chat memory, "new creation" etc to no avail.

Unless new chat, same prompt will not reroll.

[late blade]

what I'm doing is prepare a set of prompts and then in one chat instance do one after another

[summer swift]

Shcoked ChatGPU can't create image generator superman character with my face & friends face when I create a caricature trend they keep say violate our guardrails concerning similarity. I tried on PixVerse app worked fine can create superman character joke thing 🤷 , ChatGPT are dumbed AI. Superman are not violate.

[wraith oracle]

Hello y'all. Just been wondering, so when putting images for styles that GPT Images 1.5 dose not recognized, how many anchor images do you recommend for the model to get a consistent style?

[late blade]

"SUPERMAN" is a federally registered, live trademark owned by DC Comics (filed 1979, reg. 1982, #{1200394}

[late blade]

as many as you need, analyze the images in one chat, look what you like about them, and use that in new chats for other images in new chats

you can either copy the outputs on your notes or ask ChatGPT to save it to memory

I would personally make notes

[summer swift]

Yeah that bad. I'm not movie maker copy from superman, I only create joke things, I tried other apps Nano Banana & Google Gemini & Whisk image generator create superman character with my face worked fine but ChatGPT won't.

[ionic carbon]

#images-canvas message love it!

[civic dust]

oh dear. I just asked chatgpt to make me a placeholder image to use to mess around with and now I'm in an extended argument "no, not like that"

[wraith oracle]

Really? How did this start?

[late blade]

oh no, you are antagonizing data ... again.... 😌

[civic dust]

I found a stock photo I really liked with the perfect composition and wanted it re-rendered to be australian with the same composition and it really wants to do ... other things

[wraith oracle]

Like what other things?

[civic dust]

was a picture of a road, it wanted to make the road take up the entire bottom of the image, the stock one arced off to one side which is what made it good. But I got one close enough. I just wanted to play around with html layouts with an image that looked 'right'

[dim cradle]

[verbal plover]

Pls when update this app tell me please

[late blade]

I guess now that the sora webfront (legacy) is being phased out and plus/pro accounts have been downgraded drastically in image generation limits, people are no longer willing to take part in the daily theme as actively as they used to

[hazy knot]

ive honestly been using the image bot for mine and have been enjoying what the bot puts out lol

[barren widget]

From my perspective as a European customer who doesn’t have access to Sora 2, the problem is that OpenAI doesn’t seem to realize its customer base consists of many different individuals with different interests. Not every customer came to OpenAI to use Codex, or Deep Research, or to build apps. There’s also a customer base in Europe that simply wants to create images—nothing else.

And if you’re exactly in that group, and you can’t reasonably pursue that hobby anymore because you can’t properly use Sora, why would you stay with OpenAI as that kind of customer? Why should I accept a Sora limit of 50 images if, for a similar subscription price, I can generate far more images on another platform? Strategically, that makes no sense for the company—I don’t understand it.

If you want to make a lot of money as a company, you can of course specialize—but then you have to attract enough customers with that specialization. And if that doesn’t succeed with Codex, Deep Research, and so on, then you need something that appeals to a broader audience—and image generation definitely does. There are countless image generators out there that charge for their service and make good money doing it.

But if you serve the European market as poorly in image generation as OpenAI is doing right now, it just won’t work. I could easily move to Midjourney with the same budget, because as a European customer I’m limited to 50 images on Sora, while I’m not limited like that on Midjourney. So why should I stay with OpenAI if all I want to do is make images?

It’s simply a bad strategic decision to pull the rug out from under your customers without offering an acceptable alternative for the same customer base.

[late blade]

same and then I got the 200 limit

[hazy knot]

200?

faaaaaaaaaa lol

[late blade]

yeah with pro account

[hazy knot]

OH
200 with sora

yeah id use that fs . does sora do images too? i dont think i ever paid attention

[late blade]

yeah, it does

[hazy knot]

oh sweet

[late blade]

if this doesn't get restored by the end of the payment cycle, I'll downgrade to plus for a while

[ionic carbon]

appeals to a broader audience—and image generation definitely does
can you share research or anything about that? I haven't seen any evidence that image generation is used more than other usages.

[late blade]

what he means is that there are users that pay a subscription for the image generations through Sora

[polar mist]

Would be nice if they could actually tell us what's going to happen to our galleries. If they are beggining the process of depreciation (and they are) they should be able to tell us what's happening with our images. If the plan is to make them kaput then there really isn't any reason to not tell us that now and give us the maximum amount of time to react. Their communication thus far has been awful.

[ionic carbon]

I can guess what they probably mean, but that's not what the text suggests and I just hope that they either produce some verifyable results or change the text. And OTOH if I'm wrong with my guess I'm happy to see some real evidence that this is true.

[late blade]

he speaks german, not native language english, we have talked about it

over DMs and another common server

[ionic carbon]

I'll take the explanation in german too 😉

[wraith oracle]

Nah, no way

No thanks, thats fine

Ok

[fickle pond]

I've compiled the link URLs for the daily themes.
For now, from January 1 to 31
📎https://discord.com/channels/974519864045756446/1475083104749092884

[late blade]

@hazy knot that thread... scary!¨😱

[hazy knot]

I really like what the chatgpt app generates when you just tell it to generate something lol

i wasnt trying to get scary from it , i kinda went a different direction with the magnification idea

the crazy thing is the prompt didnt even mention a "thread", which is weird

[late blade]

interesting

[hazy knot]

oh i guess it did , frayed thread

"emphasize microscopic textures, wear, and minute imperfections—such as a frayed thread on a heavy winter coat," ... wasnt expecting it to be the focal point

there were three different examples it was given and it only focused on that one

[civic dust]

What’s the plus image limit?

[jovial scaffold]

How do i create 3D images??

[wraith oracle]

Simple, put 3D in the prompt and you will get a 3D looking image

[livid heath]

I can help!

Honestly am still figuring out the best tools for 3D myself are you looking at it for marketing visuals, in-app experience, or something else? That'd help me point you in the right direction!
@jovial scaffold

[obsidian cargo]

It resets daily

I think it depends on a load of your region

But ive managed to get sometimes even 50+ images daily

Ut really depends thougu

I had days where it felt like ive made over 100 images nad days where ive felt ive made like 30 before hitting limit

[civic dust]

its probably a limit in tokens not images which is really hard to figure out when you're, as my chatgpt lkes to tell me constantly, just a mammal

[obsidian cargo]

No its set image limit not token limit

It will alert you "youre generating image too fast, try in xxx s/m/h

[late blade]

Keep this in mind: https://help.openai.com/en/articles/10877094-creating-images-on-sora

** Please note that the following experience only applies to Sora 1 on Web. It does not apply to the Sora app or Sora 2 on web.**

We’re building the next generation of Sora. The Sora 1 web experience is actively being deprecated, and customers should transition away from Sora 1 web and look forward to Sora for Business. Expect more details in the coming months.

[last gull]

If I'm saving my images on #image-bot, where do they actually get saved?

[plucky hare]

When you click "Save", you should get a DM from Drawceratops with the image - does that work when you try it?

[last gull]

Oh yes I see it now, thanks.

[plucky hare]

Excellent, glad that works! You are welcome

[last gull]

Perfetto

[hazy knot]

#images-canvas message @late blade thanks for this for the inspo. really like this idea

[late blade]

I’m really glad it inspired you, can’t wait to see what you do next!

[late blade]

I'm so enjoying this current theme⁠: 🧪 Controlled Chaos – deliberate disorder, systems bending without breaking

[hazy knot]

i thought id enjoy it but its really hard for me because ive got this really tunnel-visioned idea that i just cannot seem to articulate properly

well, multiple separates but still

[late blade]

oh that stinks… 😢
Is it because of how chaotic an d fragmented the images for this topic can get?

[hazy knot]

no, i think its me lol
to me controlled chaos isnt just chaos... its chaos with structure - chaos with purpose. its really hard to get the "with purpose" part without it just looking like a mess

and then for some reason the model takes it to a different realm where im trying to stay grounded in something realistic too

[late blade]

I did notice that, it can be quite chaotic indeed

[long crater]

Now Old Sora is retiring, would ChatGPT generate 4 images per click? And more relaxed guardrails (eg. characters & celebs)?. And most importantly faster? I'm gonna miss Old Sora 😭

[late blade]

the guardrails can be tricky. most people blame OpenAI for these. truth be told OpenAI is enforcing what the current copyright/trademark laws and owners of such ask for. This just reflects how messed up the industry is. not easy to keep all happy.

As for the generations with Sora, well the iOS app doesn’t have that. But the message on the help says they are working on the new Sora experience, I bet that images like that will be included too.

[long crater]

I see. I mostly generate images on Old Sora because ChatGPT is very restricting & annoying with the guardrails and not only with copyrighted stuffs. Well I guess I can't expect much with the guardrails but I really hope ChatGPT would generate 4 images/click & faster like Old Sora.

Anyway I like your images you've been posting on Daily Theme 👍

[sterile rampart]

im so ticked off that this is happening, I HATE NEW SORA it won't let me generate any videos i want

[limpid mirage]

Has anyone tried to export their Sora data yet? Does it give you your prompt and image?
Once Sora 1 is gone there's no draw for Sora 2 personally. Weird restrictive videos nothxbai 👎

[sterile rampart]

i tried to export mine by it failed

i keep trying over and over and it still won't export the data

[limpid mirage]

Did it email you saying it failed or did you not receive anything?

[sterile rampart]

i did not recive anything

[forest juniper]

I decided SCREW it and manually download most of it.

...........and still have a chunk I need to download. Thanks, Sam. 😡

[long crater]

This sounds like what cortisol is made of..

[late blade]

Well, reached the 200 limit so fast today, hope this gets fixed soon

[late blade]

and refreshed, now I geet the thing, pro limit is not bound to 24h cycle

[obsidian cargo]

I mean 200 images daily is a lot like thats 40€ worth of api daily

[late blade]

yeah, pro does get more, the limit is not a hard-cap per day, you can def do more than 200 a day with pro, cool down is shorter

[obsidian cargo]

You also get high preset more often then medium preset on plus, and priority in busy hours

[late blade]

no idea about the presets aspect, but the rest yeah

[obsidian cargo]

In api you have 3 quality presets

Low
Medium
High

Plus results with same prompt align with results from medium preset while pro algins most of time with high preset

[late blade]

ah those presets. the tiers are a different and not tied to the ChatGPT service, that's based on your usage and monetary spending

[obsidian cargo]

I use api anyway

So its not issue really

[late blade]

yeah, that's different, use it a lot still

[wraith oracle]

I would like to know, what causes ChatGPT Images 1.5 to be so biased to that common illustration style, and how do I prompt it to get rid of that illustration style for any style such as realistic or 3D animation or ect?

[zenith crown]

One of the biggest complaints about gpt-image-1 was the difficulty in holding an image if you were making a series of pictures featuring the same character. Say, an ad campaign with multiple panels or clips. So, gpt-image-1.5 was trained to really focus in on a character and hold it steady over multiple iterations. The upshot, though, is that you really have to push it to get something different if different is what you really want.

ChatGPT is helpful in that regard - Tell it what you want and instruct it to write a detailed prompt and it will handle the problem for you.

[random nebula]

I like Warner Bros

Im random

[wraith oracle]

Do you make images of the Warnar Bros studios or somthing?

[spice thunder]

#daily-theme

[fickle pond]

The daily-themes for February from 1 to 28 have been recorded.
📎https://discord.com/channels/974519864045756446/1475083104749092884

[obsidian cargo]

Why is gpt image generating me same kind of anime slop as nano banana?

[wraith oracle]

You gotta prompt the image to have flat colors with subtle cel shading to make the anime image more screen cap accurate

[hazy knot]

ofc i didnt see the horrible green/red artifact-ing before posting mine 😭

[wraith oracle]

What API quality level is normally used in ChatGPT, low medium or high?

[fickle pond]

I won't know for sure until I try it more, but when it comes to drawing, 5.3 seems to have a similar expressive power to 4o and 4.1.

I didn't notice much of a difference between 5, 5.1, and 5.2.

How do you all feel?

[wraith oracle]

I mean it helped me made more realistic photos

[obsidian cargo]

Usually sits around medium

[zenith crown]

Personally, after giving 5.1 and 5.3 two scenarios - Look at an existing picture and describe it, and given a basic prompt create a detailed cinematic prompt, I find that 5.3 is lackluster compared to 5.1.

This is partly due to 5.3 being an analytical model just like 5.2. In both cases, it focuses on technical skill rather than creative skill. It's not that it couldn't create the movie poster in question, it's that it did so in a formulaic way where 5.1 expressed enthusiasm for the project and used that energy to come up with different and, IMO, better expressions of it than the 5.3.