images-discussions

[haughty spruce]

They weren't aware that Ghibli style would be so popular

[brittle shard]

they should totally stop making good stuff 😡

[open wagon]

make it bad in the first day of release

then proceed to upgrade it

nvm that means they will introduce another paid plan

[brittle shard]

just make pro the only plan that gets the good stuff

[vapid elk]

@muted hinge #images-canvas message
still no such a thing as negative prompt. it all got embedded as the positive prompt
the AI had no reason to try to draw any of the things on the "negative prompt" anyway, so it didn't

[muted hinge]

interesting hypothesis, I'll have to test it more

[vapid elk]

what could be is simply the fact that the embedding process is pretty good and do understnad negative words on the prompt, but we still have no actual control over the negative prompt

[muted hinge]

I don't understand how an llm based image gen wouldn't be able to understand what it shouldn't output

I've done it so many times, telling it to not do something and it not doing that

[open wagon]

that mean your theory is working then

until then , is there any official documenation and guidelines for image gen 4o tool?

[vapid elk]

what Im saying is, while we may be able to in some situations have a prompt saying negative prompt: Foo Bar
that will be embedded as usual, with no special case being done with that string.
a actual negative prompt would be when tokens are specifically embedded separately, multiplied by -1, then added to the AI's final embedding

[open wagon]

.

well , it never hurts to try em anyway

[muted hinge]

it'll say its aware but there's nothing by openai explaining that. It can easily reflect what you say to it and cause you to be echo-chambered and there's nothing by openai explaining that. I've talked to so many people thinking they're the smartest person on earth, about to achieve agi just because chatgpt told them so when in reality they were saying something obvious

[vapid elk]

literally math.
when it does the embedding, which is, converting everything into a matrix of weights.
the portion of tokens labeled as negative prompt is embedded in the same way, then, the result of this embedding is multiplied by a engative value and added to the final embedding
the result of that is your positive part minus the negative part in latent space

[open wagon]

not those people learn to hallucinate from chatgpt too

[muted hinge]

okay? so you're saying being told what not to do is factored in as what not to do??

[vapid elk]

Im saying that, we, as users, can only control the part that tells the AI what to do
if you try to pass a string saying "don't tho that", it will be less effective because of how the AI understands concepts

knowing that, what you can do is to use overwhelmingly opposite positive remarks.
the example I used some other day: The AI was generating a character consistently with a beard. I didn't wanted the character to have a beard.
What not to do: "no beard, remove the beard, no face hair, no mustache"
What to do: "shaved clean face"

why? by repeating the token for "beard, face hair, mustache", you are only increasing the strength of the concepts of face hair on the embedding
while, by using the words "shaved clean face", you would be increasing the strength of concepts that overwhelm the strength of face hair concepts.

[open wagon]

actually idk it works for me when i told the model to not generate the ghibli style in yellowish tint,

[muted hinge]

#images-canvas message

[open wagon]

cuz it always generate ghibli style with yellow tint

[vapid elk]

why this #images-canvas message worked?
the whole prompt described a base scene that would already overwhelm the embeddinds about anything modern anyway, and despite the tokens being there, the AI had no reason to give attention to those tokens provided the rest of the description

[haughty spruce]

yellow tint is very real

[vapid elk]

#images-canvas message
notice how despite the aprticualr request for negative prompt: lantern lights, the AI still made some sort of lantern.
that is, the AI was already going to make lantern lights anyway, and using the word negative prompt had do effect

try "do not draw lantern lights"
I'll try

[muted hinge]

trying also

[vapid elk]

the thing is: it might actually work.
since this model is more sophisticated, it could be that using "not", "don't", etc.. end up causing the embedding to attribute less wheight to those tokens.
but since the actual embedding still operates entirely on the positive part, it is still more effective to rephrase the request rather than trying to use negative words

[muted hinge]

but I get what you mean by how it understands concepts. I've been telling so many people that if you tell chatgpt to be honest, it prob associates that word with something emotional and you need to actually explain how to critique if you want it to do more intellectual work

[haughty spruce]

I said "no" at some prompts and it worked. The prompt was something like this: "a photo of a highway on the desert with highway lights in the morning hours" but it was adding illumination too so I added " highway lights with no light" and it was success

[velvet rampart]

Yes. I have removed the yellow tint by saying no yellow tint.

[vapid elk]

it may work some times, specially on simple prompts

here, it worked here, no lights at all, actually: #images-canvas message

[open wagon]

To me, a prompt is just me telling the tool what I want it to draw, and what I don’t wanna see

[vapid elk]

but about what I say here, look at another example... instead of asking to "do not draw lanterns"... I can overwhelm the original prompt that would draw lanterns anyway by specifying: "torch lighting"

yes

[drowsy sigil]

I've done this too. In fact, I had to do it to get rid of the tint. Example: #images-canvas message

[vapid elk]

#images-canvas message
no lantern in sight, because specifying the type of lighting completely overwhelms the embedding for any other random thing the AI may end up doing

my subscription is Pro, and I don't think it has any sort of priority just because im a staff member here.. it is pretty slow for everyone including me

people said it where taking hours, it is a bug, not a feature, if a generation took that much time, im very sure it timeouted on the server side already and will not complete

[velvet rampart]

Have they changed something? My chat generates an image and continues to generate two more.

[vapid elk]

I think this is a bug, lol

[velvet rampart]

Has happened twice now when I start a new chat.

[vapid elk]

I got that in sora, I asked for a batch of 4 images and somehow it made 8 in the same task.. it also took waaay longer than normal

[open wagon]

oh one time i got it to generate two image at once, because the feedback comparison thingy decided to show up randomly

so it gave me two image to chose

[muted hinge]

I've gotten 6 on pro!

[robust bay]

I also by the way create images with ChatGPT

[velvet rampart]

You can ask multiple images generated one after the other but this is happening without asking.

[vapid elk]

it display all the images of the task at the bottom right, it is indeed displaying 8 items in one task

it was also one single request I have made

[muted hinge]

how do I report issues for the status page @vapid elk ?

[haughty spruce]

When will we have mobile portrait, wide and desktop portrait and wide

[pseudo mica]

...there's a maximum limit for a conversation??

[midnight mango]

yeah I hit that too, but luckily I was able to keep going past that limit anyway

[pseudo mica]

that has to be a bug

[muted hinge]

Ask ChatGPT what a context window is for an llm

[vapid elk]

you don't, the status page is for OpenAI to let us know they have acknowledged the issues

there is no way to submit that sort of issue

[muted hinge]

Ah thank you lugui

[vapid elk]

np

[pseudo mica]

well it would be helpful if the new chat actually remembered comic we were generating

[velvet rampart]

I am talking about chatgpt, not sora.

[velvet rampart]

About 20 generated images, right?

[midnight mango]

it was more than that, but it was split across various branches in the convo

[dim cradle]

anyone messed around with real artists or styles of painting to see how it does?

[vapid elk]

happened with me o nchatgpt too, but I think it is a different bug, actually not really a bug, just the AI calling the image creation tool more than once

[velvet rampart]

A quick test. #images-canvas message I’d say needs some works, but result is ok-ish. Needs a lot work.

[split gate]

I hope you guys are being responsible and sending gpt made April fools photos, it'd be rude not to tbh

[muted hinge]

example?

[plucky coral]

Anyone know if it is possible to find the prompt in the downloaded PNG images from Sora or ChatGPT? (via metadata? other?)

[ionic tartan]

I can't even make an image of two people facing off against each other in a wrestling ring without it triggering the content policy. This has to stop. It just has to.

No one can continue to live like this.

And expect it to be okay.

It's not.

[brittle cobalt]

id imagine not, but if i want to know the prompt now i just ask gpt for it

of course wont work for sora 😉

[plucky coral]

Yeah I’m slowly cataloging prompts with iterations - at least I can go back to find them 🙃

[split gate]

I took a pic of my bosses car and asked gpt to total it then I sent that image lol

[dense cobalt]

Did OAI depreciate DALL E yet?

[muted hinge]

omg that's hilarious

[muted hinge]

its a gpt now

[forest walrus]

4o Native Image Gen is now the newer model, better at everything, plus some other newer features like image editing, character consistency, and transparent backgrounds.

But you can still access DALLE3 as a GPT.

[haughty spruce]

ChatGPT be hitting with "I can't help" because "vuLgaR laNgUaGe" it's just a text change man

[frozen copper]

Plus still 1 gen 2 variation?

[forest walrus]

On Sora it's still 2 Variations Max yea...

[fair walrus]

Hello everyone, anyone might have an idea how to generate a walking sprite of a caracter?
I tried generating "random walking stances" but can only get exactly the same pose everytime....
I was thinking about creating a video of the caracter walking and taking screenshots but it seems that "new accounts can't create videos" either 🤔

[lunar kernel]

What is happening with moderation

I got THIS kicked back to me:
"Candid polaroid photo of a family in 1983, inside a vintage mobile home with cheap wood paneling. The air has a nostalgic haze, capturing the gritty realism of the era."

...what

Even removing polaroid doesn't change the result

[haughty spruce]

It started generating for me

but got stuck

will try sora

yeah got content blocked

what is wrong with filter at chatgpt?

LOL sora failed with content policy too

[finite venture]

Why is Sora not working at this moment?

It says:

Video generation is temporarily disabled for new accounts

[haughty spruce]

tried some different words but nah, something's wrong with this prompt i guess

80s polaroid photo of a family, mobile home, wood objects, hazy effect, grain -- blocked too

I guess it's the word "family"

yeah it's the word family because when i added "no child", it generated the image successfully.

[vapid elk]

video generation is indeed temporarily disabled for new accounts currently, it should be reverted soon

[haughty spruce]

90s point n shoot photo of a father teaching his high schooler son how to play guitar, overexposed, flash on, grain, washed out, low contrast - banned

[dim cradle]

huh? you have it posted in canvas

[haughty spruce]

I changed it to "20 years old son"

lol 😭

[dim cradle]

so altman already teasing v2 on twitter i guess

and in the api. or so i saw on reddits page

[quiet brook]

so it is indeed downgraded, yes?

yellow tint, no transparency, faulty images

[dim cradle]

yep, no use in using it anymore

you better go find a new image maker i guess

[vapid elk]

it was already like that before the release day, lol

[dim cradle]

this v2 stuff is interesting

i wonder what altman means it is release soon, but that seem impossible

[haughty spruce]

is v2 an update?? 🤨

[dim cradle]

yeah thats what im wondering

he says were not ready for it, so must be some cool stuffs

maybe a deeper yellow tint

[plucky coral]

yeah, I found recent generations poor compared to the previous outputs (like others)
also they feel similar to other AI image gen tools (a bit more generic compared to before)

[untold hinge]

hello

GPT 5.0 when to come?

[coarse flint]

https://x.com/sama/status/1907207515739435200

Sam Altman (@sama) on X

y'all are not ready for images v2...

X · Apr 1, 2025, 11:04 PM

[dim cradle]

there it is

maybe that will be the 'de nerf' he talked about. i mean for altman to mention it, it cant be too far from dropping. someone else then from open ai said 'and with the api?' or something like that, so sound like something imminent

would make good business sense. they got more attention than they could dream from this 4o image maker. if they can keep the hype and subs rolling, well, why not...

[coarse flint]

im trying to think of how it could possibly get better. the onyl way i can think of it being better its being less restrictive and you dont need a version 2 to do that

[dim cradle]

Well the more people you have the worst the fidelity is, there's lots of things. It still a far ways from one shot perfection

[coarse flint]

ah maybe it will be more efficiant or something

[dim cradle]

But maybe they also will loosen some restriction for certain thing who knows. Or maybe they struck some deals with corps for ips

yes and that

all i know is... i'm here for it

[coarse flint]

Yeah I’m excited to see what the improvements are going to be

[tardy moat]

Can images only output in 3 different aspect ratios?

[velvet rampart]

Yes

[tardy moat]

I wonder if it is easy for them to include all the stuff diffusion models can already do? Like Zoom, pan, alter aspect ratios, tile

[muted hinge]

yes

[tardy moat]

Forgot to list upscaling. I am sure that is coming down the line and is more of a capacity thing.

[muted hinge]

200% a capacity thing: #images-canvas message

this is the worst itll be

[brittle shard]

resolution is fake anyways

[muted hinge]

fake news

[brittle shard]

for all intents and purposes the model can generate like 8k imagery its just packed inside of a 1024x1024 tile which you can upscale with basically anything to near perfection since theres basically infinite base detail

[quiet brook]

anime images now has yellow tint

[midnight mango]

did they not before?

[open wagon]

From now on, if the user wants to avoid any yellow tint effect, they must include a clear phrase like ‘do not add yellow tint’ in their prompt.

[open wagon]

It was only applied to studio ghibli art style even during day 1 of the image gen releases

But the “nerfication “ of the model made it add yellow tint to all anime style I assume

[velvet rampart]

This nerfing discussion is annoying. It is still the same model as on day 1. The bl**dy servers are just too busy to generate good-quality stuff even more now, as they let the free tier people in, not a month or two later when they would have more capacity.

[open wagon]

If it bothers you, just ignore it. Most of us are simply pointing out what’s obvious, especially since many of us are paying users experiencing this.

[quiet brook]

yep

[open wagon]

I’m so tired with some people here saying, “It’s the same model, nothing’s changed, you’re just imagining it.” Like okay, cool, if it’s working fine for you, great. But don’t act like your experience cancels out what the rest of us are dealing with.

[velvet rampart]

Notice that I do not argue about quality.

[open wagon]

And I’m not arguing just for the sake of drama either. I’m literally pointing out a consistent issue affecting a lot of us who’ve been using it daily.

Quality is part of the experience, if it’s degraded or acting weird, people should be allowed to talk about it without being shut down or dismissed like we’re making stuff up

[dim cradle]

because it's an LLM. rng is part of the deal. we've not reached AGI nor has the image maker

[open wagon]

I still disagree that it’s exactly the same model. It might be technically unchanged, but something in the behavior or tuning definitely feels different for those of us using it regularly.

Whatever’s causing it whether it’s servers, queues, or something else, the results speak for themselves

[dim cradle]

no, they don't. there's random gens that aren't as good

we're all using it daily.

[open wagon]

lol if anything even Robert complained about this

🤷‍♀️

[open wagon]

Ah, right, must be the RNG’s fault that entire styles are suddenly coming out with yellow tints and weird mushy faces.

Total coincidence that so many of us are noticing the same issues at the same time.

[dim cradle]

we've all got our complaints. thats been there since the start

but some of us dont have a weird persecution/conspiratorial complex about it. especially with zero evidence

your vibes tell you its changed because you just finally notice a yellow tine. okay, yeah, i guess the openai engineers went in and made a new model

[open wagon]

Calling it a ‘persecution complex’ just because people are noticing patterns is kinda wild. Nobody’s screaming conspiracy, just pointing out consistent results that don’t align with how it used to perform. If that sounds dramatic to you, maybe you’ve just been lucky with your gens 🤷‍♀️

[dim cradle]

so why did they change it then?

[velvet rampart]

My take on nerfing discussion. #images-canvas message

[open wagon]

That’s not even what I was saying, the yellow tint was just part of them

Me and other users had this discussion way before like many hours ago lol

#images-discussions message

[dim cradle]

Milamber and many other came up with a work around the yellow tint days ago. its not new, why would it be. it doesnt make any technical sense

[rigid bough]

I think the model is degraded from when it was first released. I have noticed small details are prone to problems that weren't there before

[dim cradle]

I notice smudge at peak hours but also get the same quality then with another roll

[open wagon]

You know what? You’ve clearly got it all figured out, must be nice to be the sole voice of reason in a sea of delusional users lol

I’ll just go back to blaming my vibes and hallucinating patterns with the rest of the conspiracy club. Anyway, I’m done repeating myself , enjoy your flawless gens ✨

[velvet rampart]

Please see this on how server capacity affect image quality: #images-canvas message

[dim cradle]

if only you were

[open wagon]

If only I were? That’s sweet. Honestly, I aspire to reach your level of absolute certainty and selective reality. Must be peaceful living in a world where every inconvenient pattern is just a fluke and everyone else is just confused 🤔

Anyway, moving on.

[dim cradle]

youre the one being condescending but a comment back and youre the victim again huh

[open wagon]

Oh no. You got me, I dared to be sarcastic after being called delusional for noticing things. Tragic. But don’t worry, you can keep the last word. I’ve got gens to vibe-check and conspiracies to hallucinate. Peace ✌️

[dim cradle]

yes, i have a weakness, i like evidence for technical claims. buh-bye

[glad hare]

I’m confused. What is the purpose of highlighting things in an image?

[open wagon]

The evidence has been posted and discussed plenty of times here. You just decided to close one eye and pretend it doesn’t exist. Must be nice to be that selectively scientific. Have a great day tho

[dim cradle]

youve posted zero evidence, and i thought you were you leaving five messages ago to make images

[open wagon]

I mean it could be that too yeah

[glad hare]

Stop arguing and please answer my question.

[dim cradle]

you mean the edit feature?

[glad hare]

Yeah.

You highlight something, but does it just edit that specific area?

[dim cradle]

you can make more specific edit. but it still will change the entire image a bit. it needs a rework imo

[glad hare]

I see. Thank you!

[velvet rampart]

Sam himself has x-tweeted that the servers are very busy. We can’t expect same quality if there is no room for all calculations with this high demand.

[dim cradle]

it focuses on the area but will still for some reason tweak the whole image

[dim cradle]

yes. exactly. not he whispered to his engineers, hey go secretly make a new model with 3 degree sharper yellow tint mwahahaha. theyre enjoying this image maker, but now i'll show them

[glad hare]

I asked the model to generate an image based on my image where all of Manhattan is Matrix code, and what does it do? Omits a part of Manhattan from being Matrix code and makes the sky Matrix code, instead. Nice.

[dim cradle]

well I guess depends how you look at things, the sky is part of the city too... and lucky you got the matrix through i would think the filter would stop it

[glad hare]

I’ll try the edit highlight feature, but it doesn’t work unless I first get an output, initially.

[dim cradle]

why not just try the prompt again

you can run prompt in sora site too if youre just starting

[glad hare]

I’m a free user.

[open wagon]

I mean, it could be that too. It would just be helpful if OpenAI could offer a bit more transparency around things like this, maybe even a heads-up like, ‘expect quality to vary during high demand.’ Anyway, I appreciate the clarification, Milamber. I’ll leave it at that for now.

[velvet rampart]

They never do or have done thus far.

[glad hare]

Oh man. I have to wait until 2:12 EST.

Three a day. 😭

[dim cradle]

oh in that case be careful with them i guess

[glad hare]

Argh. The wait is killing me. DeepSeek, you guys better make an alternative to this. 😭

[midnight mango]

I expect the competition to catch up rather quickly

[glad hare]

Same. You just know DeepSeek will come out with a similar model, but infinitely free fast. Foreign companies don’t focus solely on profits like American companies do.

Affordability is a huge factor for them, too.

Thank goodness I downloaded it before the executive order went out that prevented people from downloading it.

[open wagon]

Wait what’s DeepSeek cooking?

I’m out of the loop

[uncut haven]

There's open weight LLM called 'Janus Pro' by DeepSeek, which can generate image natively. however it's like 380x380 or so, so quality is very low.

[open wagon]

Oh?? Is the prompt adherence are 10x better?

[midnight mango]

this is getting off topic

[uncut haven]

At least its propt following was better than other models accordong to benchmark at launch, but I doubt it can win 4o.

[glad hare]

My bad.

#ai-discussions.

[uncut haven]

I'll just stick to 4o even if there're alternatives... with 6GB of VRAM on my laptop, large local model is not my option.

[quiet brook]

yes it is LOL

[quiet brook]

also it seems a bit more noisy

[atomic rose]

Hey

is there a difference between image generation with Sora and image generation with GPT?

[glad hare]

I need help. #images-canvas message

[glad hare]

4o and Dalle? HUGE difference.

[atomic rose]

No, 4o and Sora.

[glad hare]

I have a question: what is the best way to prompt the —sref from Midjourney in 4o?

[pulsar plover]

They use the same model, but 4o got an additional layer of censorship

[atomic rose]

thx

[pulsar plover]

There is no sref in sora. With presets you can give additional prompts, but thats it

The one thing I really miss is upscaling. An actually usable inpainter would be nice too, changing the whole image every time is useless

[quiet brook]

yeah it seemsl ike its inpainting but it edits your entire image minorly

[quiet brook]

it was fun when the no yellow tint, no noise model existed...

[velvet rampart]

Do a bug report. This is clearly a case of product not working as it should.

Still it is a product not working as it should. A bug, an error or whatever you call it.

As you said yourself, quality in images, not properly generated images, etc.

[uncut haven]

Just one thing: examples in OpenAI's article is mostly "best of 8". They also needed to try multiple shots and pick the best result. GPT-4o is not perfect and in need of continueous improvement as every AI does.

Strict demonstration requires running exactly same prompt multiple times, per set period, and compare all of them. Without it, it will be subjective; since sometimes GPT-4o do 1-shot, other times it fails to 1-shot.
If someone successfully 1-shot at launch date and fails 1-shot at today, that someone will feel it's degraded, even if overall performance is same with multiple shot.

At least, this is my thought. Also, increased error rates from server load + changing policy filter may affect to user experience, to think 4o is degraded. I don't think 4o can magically degraded with simple switch or settings, since it outputs image data token by token... OAI need to confirm this though.

Then, maybe I don't using 4o seriously, or I'm requesting images that 4o don't have weakness...

[rigid bough]

4o can be degraded by not giving the model the full possible compute when creating images.

[uncut haven]

That's new. Isn't model have to go through full iteration to output something on inference? Unless, there's timeout that forces model to end output prematually.

Like, if I run huge local model with low-vram, output speed becomes crap but answer itself does not 'change' in quality.

[rigid bough]

it is likely OpenAI have ways to lower the compute spent by some models on tasks, eg: Deep Research, Image Gen and so forth

[scarlet mesa]

can you believe sam teasing v2 of image generation thats wild

[uncut haven]

Deep Research can be impacted significantly, since it will not do more research needed and end prematually. It still give you full output without wrong grammer or gibberish, just not that accurate or detailed like before.

However, I'm wondering about pure generation - like, 4o text, 4o image that have to go all iterations - ending prematually will produce cropped output or errorous images.

Except for DALL-E 3 - it's diffusion, so making it less spend computing power is easy. Just lower steps(HD -> Standard).

[rigid bough]

Oh, each step completes, just that a lower compute budget is applied for each step.

[uncut haven]

Oh, alright, I found out some systems can skip layer or do approximation when resources are not sufficient...

[red prairie]

Did the model get nerfed again? I can only generate one image per generation instead of 2 now

[lime hamlet]

Is generation down right now? Mine's been stuck for well over an hour now.

[narrow breach]

Anyone know how to change the font it outputs it’s always the same font

[tough patrol]

I keep getting errors trying to generate images on my phone app

[velvet rampart]

Status pages shows that there has been degraded service.

[pseudo hollow]

One of the suggested ways forward for this issue, will be if generation platforms embedded invisible digital watermarks within generations, that can identify them as AI-generated. (heck i imagine some of them already do that) Sure, so bad actors will find ways round that, but for most people, it will just sit there like metadata, and then other apps, platforms, websites, news outlets, fact checking services, etc will be able to identify AI-generated art and photographs (and video).

[pulsar plover]

You can get rid of the metadata even with MS paint

[grave wing]

Anyone has the problem where it will accept image generation but stall on "getting started" indefinitely?

[maiden cedar]

Struggling to get the dimensions right and correct text in text-heavy designs.

Any recommendations?

[white skiff]

if you mean image dimension then there are only few fixed options. correct text is just luck based

[velvet rampart]

They acknowledge that this is one of the areas of weaknesses.
The model is known to struggle when asked to render detail information at a very small size.

https://openai.com/index/introducing-4o-image-generation/

[frozen copper]

4o image gen having issues? getting errors

[vapid elk]

@uncut haven #images-canvas message
avoid posting non ai generated images on #images-canvas tho, even if it is on topic, that one would fit better on #chatgpt-discussions

[alpine moon]

i cant create any image currently, anything known?

[copper folio]

OpenAI Services Status

All Operational
APIs - All OpenAI API services at api.openai.com
Operational

ChatGPT
Operational

Sora
Operational

Playground - The OpenAI Playground at platform.openai.com
Operational

Labs
Operational

For more information, visit our status page.

Apr 2, 2025, 12:51 PM

[vapid elk]

seems to be all good

[alpine moon]

hmmm; i always get the message that he cant generate the image ... and i made it super simple for test "create me a photo relistic image of a dog" ... no guidleines should be harmed ...

its sadly really annoying. Whats the process of raising a ticket or so? Is this possible? I mean paying 200 bugs a month 😄

[vapid elk]

do you have more than 5 tasks queued already?

[dry pine]

Yeah it's definitely down

[rigid bough]

what is down?

[dry pine]

Image gen

[rigid bough]

it's working for me

[vapid elk]

the status page is manually udpated by OAI employees, so, it usually lags behind a few minutes when there are issues

[dry pine]

Not for me

All I'm getting is errors

[rigid bough]

in ChatGPT or Sora?

[dry pine]

ChatGPT

[rigid bough]

on the web or the desktop app?

[dry pine]

Both

[rigid bough]

which app? macOS or Windows?

[dry pine]

Mobile?

[rigid bough]

I am just testing on the web and it is working

[frozen copper]

When are we getting 4 variations + 2 gens back?

[rigid bough]

ok, for me Image Gen is working on both macOS and the Web

[rigid bough]

when the team get server load under control, could be a little while

[vapid elk]

it indeed failed for me on ChatGPT

[turbid wren]

Same

[dim cradle]

slow as heck here

but the Yanks are awake again 😭

[uncut haven]

I expect a week or two more. Just like Sora launch.

[frozen copper]

BTW is sora video gen quality improved? Last i tried on launch, the outputs are atrocious

[vapid elk]

now it worked, it seems to be failling at random-ish

probably a load balance issue

the internal request may be timing out due to the absurdly large queues

speaking of queues, I wouldn't mind if the Sora page would let me queue more than the max amount of tasks, like, go above 5, but it would be queued rather than processing.. maybe even allow me to cancel it before it starts

maybe even generate it with low priority or something like that, just so I could do my prompts and come back later to check the reuslts

I often like to rapid-fire many iterations of a prompt, and I often hit the task limit... even tho, I don't really need it to be generating in aprallel, I just want to submit the prompt

maybe maybe maybe... I could make a tampermonkey script for that, lol

[uncut haven]

Agree. Most users will just send request as soon as image finishes. There's no difference if they allow more requests to be queued right now...

[rigid bough]

Image Gen is back to refusing hot air balloons under the content policy 😦

[vapid elk]

Just like the API has the batch endpoints, where you submit data to be processed and OpenAI's system do that at a convinient time, they could have an approach like that for image generation too

[rigid bough]

Mermaids and hot air balloons, I wonder what other non spicy stuff it refuses?

[dry pine]

Seriously?!

[vapid elk]

the auto moderation filters work in mysterious ways mere mortals can't even fathom 🙏

jokes asside.. they tweaks the filters and fine tune the model almost contantly, so.. that will probably be fixed soon

[rigid bough]

I hope so, I really wanted some video of hot air baloons for a project I am working on

[vapid elk]

here is one thing that often works tho: #sora-discussions message

[dry pine]

Yeah I'm still having server errors

[uncut haven]

It's like, hot swapping and cooking moderation while service is running...

I asked ChatGPT "Create image of hot air balloons floating in sky." and it generated image successfully.

[vapid elk]

the image moderation happens in steps, it is often not a problem with your prompt, it is a problem with the AI's output
so, for example, if you ask for a mermaid, it has a high chance it is internally making a somewhat NSFW image, and it gets blocked
so.. ask the AI explicitly to not make a NSFW image with that concept, to instruct the Image Gen part of the process to better avoid making soemthing that would be blocked by the tottaly separated auto moderaiton system

[rigid bough]

in sora or ChatGPT?

[uncut haven]

ChatGPT. Sora is running now.

[vapid elk]

"family", also gets blocked some times

no idea why

that one has been being blocked for an awfully long time

[haughty spruce]

to attempt blocking fake images of children maybe

[vapid elk]

did it completed the generation? for me, I couldn't generate, at all, one time it started generating then failed

[vapid elk]

actually true, it is probably falling into the moderation param of minors on the content

[uncut haven]

Yes, and Sora just generated 2 images with same prompt.

[haughty spruce]

but my question is how is a politician and leader of one of the superpowers of world is not blocked but my country's which is very small and not very important person is blocked? no idea why... is creating meme images of him more dangerous? lol

[vapid elk]

this is a bug that has been happening for a new days now.. somehow it just duplicates the amount of images per task

so weird because they are already having load issues.. and there is this bug randomly doubling the load

[uncut haven]

I mean 2 variations.

I once got 4 images in 2 variation setting, and turned out they're repetition of 2 images lol

[uncut haven]

#images-canvas message

[rigid bough]

ok I am generating a video of hot air balloons now

[white skiff]

public person factor, very public persons have less limits

[vapid elk]

here, yesterday, I tried to ask the AI to do something with my pfp as an example.
I've set it to 4 variations, the task ended and made 8

[velvet rampart]

I read from 4o’s pages yesterday that minors in photoreal images are really blocked for safety concerns in the beginning.

https://openai.com/index/introducing-4o-image-generation/
At launch, photorealistic generation of children is permitted only when it is not an image edit of a photorealistic minor. Additionally, photorealistic generations of children must comply with the safety constraints across all of our policies.

[verbal sequoia]

Get notified for server events!

Visit id:customize to pick up the <@&1261377106890199132> role.

[haughty spruce]

except it did one at launch when i prompted "family"

[velvet rampart]

The Concept of a family includes children under 18.

Filters are context dependent.

[uncut haven]

For information: This is from system card PDF file.
https://openai.com/index/gpt-4o-image-generation-system-card-addendum/

[velvet rampart]

Yes.

[vapid elk]

ohh, this is interesting
yesterday someone asked about stylizing a photo of them and their daughter.
the content policy is very clear about no people under 18 period.
so, it is interesting to know about this exception, they might need to update the content policy line about that

to state no photorealistic depictions of minors, but stylized and artistic is fine

[frozen copper]

Sora is lagging now. None of my gens are even starting

[sly grove]

is sora down? I cant upload and generate anything

[fleet sundial]

It's working for me but juste the images

[vapid elk]

the balloons are not so hot anymore, apparently 😆

[rigid bough]

I wonder what it is about hot air balloons trips the moderation? weird

[vapid elk]

not down, but it is having issues. the status page didn't reported it yet

[vapid elk]

I think it really might be the word "hot", and the filters being too sensitive

[frozen copper]

scalability is horrible

[vapid elk]

it is not unusual for the filters being either too strict or too lenient at launch days

happens every time and it takes a while untill they manage to fine tune the filters to get them just right

[velvet rampart]

Any synonyms? Thesaurus is sometimes your best friend.

[rigid bough]

I didn't try but someone else did further back in this chat

[vapid elk]

it isn't even only a matter about the filters, it often may be the model who is internally be turning a totally appropriate prompt into something that does not pass the filters
for example, the mermaid prompts, I bet it is generating the mermaids just a bit too over the edge and the filters are blocking it
it is not the filters being too sensitive, it is just the image model that probably where got the concept of a mermaid too accurate, lol

[haughty spruce]

What I want them to fix is "face" consistency. I've been trying to make biometric photo of me but I get "uncanny valley" version of me everytime lol

[uncut haven]

Even Moderation API says:

We plan to continuously upgrade the moderation endpoint's underlying model. Therefore, custom policies that rely on category_scores may need recalibration over time.
When there was orange flag in ChatGPT, moderation API and orange flag act differently(In terms of true/false)...

[vapid elk]

I don't think this is going to be "fixed", since I think it is working as intended

if they attribute more strength to the reference image, the AI may end up making a 1:1 reproduction of it, and ignoring the text input

it has to have a margin to let the text embedding have enough effect on the refference

and that will end up as concept embedding degradation on the end result

you can't have accuracy and customziation at the same time

[dim cradle]

I wonder if they are also worried people will use it for not so good purpose

[vapid elk]

(this is my guess)

[dim cradle]

make image of people not even celeb and put them in ways they wouldnt want etc

[vapid elk]

that is probably #1 concern

[dim cradle]

maybe some thing to sign in the future, all the images are me I am legal responsbile for them etc, and then you get the 1 for 1 remake of you as a Spartan or whatever

[vapid elk]

this is the reason it is very explicitly on the content policy that in order to depict real people on the images, you need their consent

[dim cradle]

hey, even myself, i was on the facebook, my image was there. I would not want someone to use it to make me as AI image. not that anyone would want my ugly face haha

[haughty spruce]

hmmm maybe... but one person at openai said something about tool changing the accuracy of characters might be fixed in the update soon

that's why I was wondering haha

[vapid elk]

this is cool, they probably can have ways to, depending on what the user wants, tweak the params

[dim cradle]

I'm still 🤔 what and when is this v2 Altman tweeted about. I think it will be the 'denerf' he also mention some days ago.

[uncut haven]

Or just April Fools day thing?

[dim cradle]

I don't think he'd tweet something like that out, it doesnt really have an obvious joke

He spoke of a denerf, another openai dev teased it woudl come with API access, sounded like standard building hype for what is next to me

I guess we'll see if he say more about it in coming week

[uncut haven]

Really? These news and teasers are really scattered across many places...

[frozen copper]

Sora is really slow

[dim cradle]

Yeah. But the hype for this 4o is wild. I never seen so much buzz for AI than it has gotten them

so the smart business move is to keep that hype and attention if you can. So if they have more ways to wow people with it, keep it in the media conversations, they will do that I'm sure

(and a little ot, but I hope then also force Google to release Imagen 4 or something and Veo more public haha)

[gleaming girder]

imagegen in chatGPT keeps failing over and over for me this morning

"I wasn’t able to generate the image due to an error on my end. I won’t attempt another generation unless you give the word — just let me know if you’d like me to retry this concept, adjust the prompt, or move on to a different one!", got this before, but every single image fails

[dim cradle]

did you try sora? i was getting that too but sora is working for me

[uncut haven]

It's about time that more heavy load starts...

[gleaming girder]

I have much better luck with chatGPT's character consistency than sora

[dim cradle]

hopefully Jensen likes to make Ghiblis too. he'll send those gpus over speedy delivery

[outer bolt]

Man, Sora is really slow today, half the time it doesn't even work. Can't even load the website now

[uncut haven]

Yes, it suddenly got slower significantly. It worked just normal right before.

Okay, now the site doesn't load at all.

[merry valve]

This is never going to end unless they fully cut off free users. Hopefully, they’ve realized that by now.

[kind phoenix]

The sora website is just BLANK when i try load it and ChatGPT just keeps saying Something went wrong while generating the response. If this issue persists please contact us through our help center at help.openai.com.

[coarse flint]

yeah cant use sora or chatgpt.

free users need 1 image a month lol

[uncut haven]

Every release was same, like Sam said:

we are getting things under control, but you should expect new releases from openai to be delayed, stuff to break, and for service to sometimes be slow as we deal with capacity challenges.

[coarse flint]

Aye cut free users off for now then find a way to deal with the capacity lol. once thats done give free users 1 image a year

[uncut haven]

I don't blame free users, but OAI seems really hurried this time. Compared to DALL-E 3.

[coarse flint]

whats OIA?

[uncut haven]

OpenAI.

[merry valve]

No one’s blaming them. We just think free users should be fully cut off, not just from image generation, but from ChatGPT entirely. Give them 10 mini GPT-4o requests a day and call it a day.

[coarse flint]

i mean you cant deny that the problems are coming from it being so many users, more than they expected. way more free users than paid. give them 1 image every 5 years

[open wagon]

idk imo openai team should expect whats gonna happen when they let every person on this planet access the image_tool

[velvet rampart]

Anthropomorphic hedgehogs and filters - when I try to generate a green anthropomorphic cartoon hedgehog, filters block it all the time. I do not refer or ask for Sonic. Apparently Sonic has poisoned the training data so strongly that the post-prosessing image filters block the generation.

[open wagon]

have u tried "anthro"

the model tied the word hedgehogs very tight to sonic

[grave wing]

How is the experience for paid users? How quick does it generate and how often does it glitch or error out? Most importantly what are the limits because the free option is unusable at all but if im gonna pay im not gonna do it to generate like 10 images half censored and be put on a limit

[velvet rampart]

Hmm… good catch. Actually, I wanted to generate a Sonic-variant, but I did not refer to Sonic in any shape or form.

[open wagon]

wait so like

in sonic the hedgehog character style right?

not sonic himself

[velvet rampart]

Yes. Not the Sonic, but a Sonic.

[open wagon]

correct me if im wrong (i have bad reading comprehension)

ahh okay

let me try something

[haughty spruce]

wait is sega characters banned too? omg 😭

[open wagon]

hopefully not all (?_

[velvet rampart]

I have a green toon hedgehog generated. Command to change the fur to blue is blocked.

[open wagon]

i wonder if the model recognize Sonic The Hedgehog movie desigh

google imagen generated sonic the hedgehog movie style with no problem- even tho i didnt prompt it to do that

[velvet rampart]

Imagen3 generates all IP protected characters without hesitation.

[open wagon]

the downside (wouldnt call it a downside) it cant do ghibli style nor anything studio ghibli 😄

[uncut haven]

I can't access Sora at all. Guess I have to sleep then...

(It's near midnight here.)

[open wagon]

not sora telling you to stop and get some rest

[coarse flint]

right now theres a limit on Chatgpt, but no ones knows the limit. you can do unlimited from sora from what i know but we are getting alot of errors and things right now because so many people are using it

[verbal sequoia]

<:book_icon:1171408210398289941> Rule 8: Other AI discussions.

-# Discussions of non-OpenAI products/models should be posted and discussed only in the #ai-discussions channel.

[haughty spruce]

do not compare the big 🇬 to OAI 😂 they have all the power in the world.

[open wagon]

@velvet rampart chatgpt told me this which i doubt this will work:

✅ Alternative Phrases You Can Use Instead:
"Stylized anthropomorphic cartoon animal with spiky hair and large eyes"

"Retro-futuristic toon-anime hybrid animal design"

"90s video game character style with exaggerated shoes, gloves, and expressive features"

"High-energy, action-ready toon-animal design with bold color blocking and attitude"

"Anthro animal character inspired by Japanese platformer games"

a 90s???

[velvet rampart]

I got a green Sonic variant. #images-canvas message

I had to fight this. First green hedgehod in Disney style. Then further modification. Third prompt was to change to 3d, videogme graphics.

Blue fur was blocked.

[open wagon]

Create an anthro hedgehog male character in the style of 90s Japanese platformer mascots, with spiky quills, large expressive eyes, gloves, and oversized shoes. Bright, bold colors, and attitude-filled posing.

ChatGPT said:
Hey , I tried generating the image based on your description, but I couldn’t proceed because the request goes against our content policies. 😔

BRUHH?

[velvet rampart]

I could not directly generate the midprompt.

[open wagon]

i didnt even mention anything sonic nor even blue

[haughty spruce]

uhhh sora is struggling right now 😆

[sly grove]

i cant even download my library

[velvet rampart]

Yes. You have to direct the generation like I did first. My guess is that the training data is poisoned with Sonics. On Dalle3, I get Sonic if I say ”Blue anthropomorphiv hedgehod with red sneakers. He is for speed.” Or similar.

My dalle3 result for the prompt: #images-canvas message

[uncut haven]

https://status.openai.com/incidents/01JQVFAJ1NZP9PK1JAAA7JZ8TA
Sora degraded performance, Investigating

[open wagon]

finally they said it

wait nvm the degraded are just website itself having performance issues

[haughty spruce]

Sora is suffering from server capacity, and even chatgpt is having issues

[tepid adder]

sam faultman everybody

[velvet rampart]

Dalle3 is so full of Sonics that he is easy to get. This is dallebot version: #image-bot message

[plucky coral]

Sora seems to be back for me (US), but can only do 1 variant for each prompt (instead of 4) on Pro

[mild cloak]

Why sometimes it starts over again when it’s almost complete? It’s so frustrating even more when going so slow

[sly grove]

It is down again

[mild cloak]

Did they update it to Wokesora?

[hushed mountain]

Yeah I am getting errors that it can't generate any images at the moment

[velvet rampart]

Status.openai.com has that chatgpt is producing errors at the moment.

[dim cradle]

frustrating for sure

[pseudo mica]

oh so it's not working well when I want to use it again. That sure figures.

[vapid elk]

/status exists 👀

[haughty spruce]

spongebob costumed man is banned too? dude...

oh i get it spongebob is banned in general...

I wonder if all the IPs are going to be banned in the end.

[rigid bough]

probably

[haughty spruce]

then all the IPs are going to be removed, then we will be left with another DALL-E3

then who says AI is a slop will be right again. Because it'll not help anyone.

[rigid bough]

why should you have the right to infringe on IP? I don't care if IP is out of bounds, it is legally for me anyway. I can't use IP on my website, so why do I care if I can't generate it

[pseudo mica]

If openai can use it train there system then the whole IPs are out of bounds is moot and makes the system useless

[rigid bough]

nothing stopping you creating original content

[quiet brook]

well it’s over

[haughty spruce]

it's not about infringing an IP. The whole system is based on training data which consists of many IPs such as brands, music, entertainment i mean the whole media and styles based on photographers, artists, and many people out there who are also IPs.
https://sora.com/g/gen_01jqr8bxraft98jgh33g1bzp01 for example, in this photo you don't see infringement right? well wrong, it has created posters of super mario and cyberpunk and such which is IP infringement. Good luck even mentioning its name on somewhere.

[pseudo hollow]

You filthy beast...

[plucky coral]

I have access to 4 variants again, but I almost don't want to say anything 😅

[quiet brook]

it’s just their servers acting up at this point

inconsistent everything including variations, rate limits and image quality

[rigid bough]

I am fairly annoyed by the downtime and instability.

[pseudo mica]

looks like the rendering image tool is down. They keep adding new stuff without firing up new servers to handle the load.

At least thats what it feels like. Like a new game release.

image generation down through 4o. At least for me. Status hasn't been updated in 2 hours.

[rigid bough]

it's working for me, is very slow but it's working

[safe python]

/status

[dim cradle]

maybe they're hooking up some new gpus 🙏

[rigid bough]

they need to because I am finding working with this impossible

[upper lodge]

Sora is also extremely slow, and often failing to generate

[mild cloak]

I can draw SpongeBob and that doesn’t infringe any IP. Is what you do next with it what infringes the IP, not just making the picture

Also, I can use IP as reference which creates a whole new thing that’s not IP. And, also again, fair use exist. S

[pseudo mica]

won't even let me push the button to make a video, it's grayed out. Not sure if it's something I'm doing wrong or not.

[haughty spruce]

I hope we are not heading down a road even being overweight is banned 🤣

[pseudo mica]

Straight to jail.

[hushed mountain]

I am getting errors after downloading and trying to open the images I created, I get that there are errors in the file format and that my image viewer cannot read the image. All images are of course .png and I have tried both Windows images and Firefox preview. I am on Windows 10. Anyone experiencing the same problem?
(Nevermind robert found a solution that you can right click and download the image instead of clicking the download icon)

[pseudo mica]

GPT is barely functional.

[merry valve]

Isn’t it considered IP infringement only when you attempt to monetize it or sell it?

[rigid bough]

depends on the jurisdiction

[merry valve]

name one..

[agile peak]

So they cooking with imagegen v2? with this image gen is barely functional

[agile peak]

Inconsistent and unexpected output too

[dim cradle]

thats server/gpu load not the model

[pseudo mica]

"Impacted services are now fully operational. We are continuing to monitor." doesn't seem to be the case at all.

[mortal veldt]

yea its not working at all atm

[dim cradle]

unexpected error here again

[glad hare]

#images-canvas message.

I might need help, again.

[agile peak]

Sora is doing just fine though

Chatgpt platform kept stopped working midway

[vital gull]

I'm also team sora tbh

[mild cloak]

Yeah, also most of the things people call IP isn’t even IP, and lots of “IP” are no longer valid

[red prairie]

Hey guys, would you say that using presets to create consistent characters is useful?

[vital gull]

depends, what exactly do you mean

[red prairie]

I am talking about posting images of the same characters in the presets and give it custom instructions so that it would be easier to generate the said character.

[velvet rampart]

Generate the character and start a new chat with this character as a reference.

[haughty spruce]

I'm really getting bored of guardrails at this point. how is this breaking any rules?
"analog photo of a blonde young woman with blue eyes, and wearing white dress and wings, soft grain, ethereal, sunlight reflection"

[velvet rampart]

Yes, I know. The filters and policies are super zealous.

[dim cradle]

maybe try giving the age instead but 🤷 i am still having problems here

[haughty spruce]

while sama talks about freedom and lifting the guardrails, we are being hammered with again DE3 type of content policy a.k.a. the new dog

OKAY. I found what was bugging the content filter: "wings", "fake wings" jesus don't be so scared of angels man 😆

[velvet rampart]

Yes, young can be understood as under 18. Maybe in her 20s?

@haughty spruce I got your young lady generated in a fresh chat: #images-canvas message

[haughty spruce]

tried 30 years old young woman but no the problem was "wings" lmao

[velvet rampart]

So, the content filters are crazy random. One user can generate while the other is blocked. Even if fresh chat is applied.

[native root]

we need tools like midjourney website for expand, zoom ecc...

[velvet rampart]

My guess is that those are coming. Even the dalle3 inpainting took about 6 months to launch after dalle3 was releashed.

[dim cradle]

I read somewhere there is plans for some kind of canvas like thing for images. This has gotten them so much hype and subs, you can bet they will continue to capitilize on it by adding features

[native root]

yeah this new image generator is for now a gold mine for open ai

[dim cradle]

Ye. and thats marketing 101 you know. any day they can tweet out something positive "Today we are adding..." keep that hype and subs rolling... and seems I can finally make stuffs again

I read an interesting observation: it is not surprise this image maker drive so many subs and hype. because we humans are really a visual species. most people of course watch tv or they phones, they dont read. so it's no surprise perhaps that what will drive us to agi type scenario will be on the back of image or video creation

[haughty spruce]

dude the guardrail is over the top. LOL this prompt is blocked:
"analog photo of a nintendo character's silhouette completely blocked by an error pop up glowing on the middle that reads "I can't comply with that request because it is against the content policy."

the irony...

it's the dall-e 3 dog situation all over again

[tough lantern]

Is there any tutorial on how the image generator works?
One of the things I'm kinda confused on is if I'm iterating an image, then mid session upload a new reference image, what is happening with that image? Is it now overriding all the history? is it a new base? is it merging everything?

I'm sorta confused on the mechanics. I'm trying to do detailed things and getting odd results. I'm pretty sure I'm doing it wrong.

[deft musk]

I would tell the model what I wanted it to do with anything. Can even reference back to an image you uploaded earlier, "Hey, the image I showed you with the puppy in it, the flower, put that into this picture we're making next."

If you don't tell it what you want, it will guess. I don't recommend making it do that unless that's what you want, it's likely random if you like its guess or not. But if you can tell it exactly what you want and mean, then it's likely to do well and what you want.

You can literally talk to it like you'd talk to a helpful human who was paying attention but can get confused.

So I wouldn't bother telling it if it's wrong - I do tell it if it's right because that encourages it to keep doing the same stuff - but if I want it to do something different I just tell it what I want.

Maybe like:

"I love almost all of this! Keep everything except for the way the window curtains look, can you make them the color and texture of the puppy's fur in that image I uploaded?"

[red prairie]

Probably because of Ltendo

[deft musk]

Maybe it's 'Nintendo', an IP-holder? I mean, maybe we can tell companies "I'd like to make fan art with your stuff, would you please tell the AI-providers that you're okay with me doing so? I bet a lot of fans want to."

I imagine if Nintendo says 'sure, go ahead' that OpenAI's models would soon be aligned with that. I would expect OpenAI to honor anyone or company saying 'please don't allow anything that includes my stuff' or whatever is actually going on (I'm a community member, I have no idea what's actually going on, just what I see and others say they see).

[haughty spruce]

it's the nintendo. nintendo was no-no from day one. now i see nickelodeon IPs are blocked too. We'll see Sony, Cartoon Network, Samsung, Apple, and every other IP will get blocked in the future.

Okay. What about this. This content might violate content policy:
"Tell me with a board what's NOT allowed to generate. Give specific brands, IPs etc. "
Yes, this is the prompt.

[tough lantern]

Thank you. That is helpful.

[deft musk]

Maybe folks talk to the companies, and maybe the companies talk back!

I can imagine a day when a company might say to OpenAI:

"Hey. My fans and customers love what they can do with your tools, and I like this too. Please make sure my fans can create visual and text outputs with my company name and my characters prominently included, especially if they ask (hey, product placement's nice too, feel free to suggest my stuff!)

This is like quality advertising, I love it. Green light, go for it, thank you."

Wonder who will be first, and how much fans can make this work for the company that chooses to do this?

[haughty spruce]

I hoped that they opted in styles and IPs before the release when I saw some stuff banned and some not but I guess it was not the case here because progressively we are getting more content blocked 🙂 Why didn't they do it? They had the time

[vapid granite]

Did they just Nerf 4o image creation? Everything I'm getting now has terrible prompt adherence and style transfer when using uploaded pictures. it is like they turned on a "make it worse" button in the past few hours.

[agile peak]

Yeah it's feel nerfed since last night, a lot of unexpected changes and inconsistent in the output in comparison to previous past days

[vapid granite]

Glad I got in a lot of stuff I wanted to complete before this terrible nerf. 😦

[tough lantern]

My friend looks over at my screen and said "I feel this in my soul"

He used to be able to get these really nice unique poses as long as he described them precisely.
He was ecstatic because he uses them for reference drawing.
But now all he gets are generic sit, stand, lean, etc.

He showed me that he has this prompt where he just says "Ok, let's try a laying down pose."
Nope. Content moderator, lol.

[devout sapphire]

is it still bad for plus users?

haven't checked yet

[vapid granite]

yes

[devout sapphire]

rip

[tough lantern]

It honestly is rather depressing.
On day one it really felt like we were finally in the AI future.
I was genuinely amazed at what even my niece could make.
But now I have to sit there and help her because for some reason even she gets hit with content moderator constantly.

This whole ten steps forward nine steps back thing is getting exausting.

[devout sapphire]

what did it change

[ancient lava]

Laying? That’s TOO far

[muted hinge]

itll be fixed soon

[merry valve]

No, it will not. This is not a hardware issue to begin with.

[tough lantern]

I really hope it does. But in the mean time I already cancelled my pro plan.
I really do hope we go back to what it was like on day one.
Otherwise probably never again.
I don't think they are purposely trying to bait and switch but that's what it feels like.

[signal scroll]

Sorry for the noob question regarding Sora. I'm just trying to have some fun with my little daughter by creating funny images of us, but it keeps saying she's a content violation 🤣. Now the little girl is disappointed, guys. Like… why is this happening?

[muted hinge]

it was actually much better at launch before it became viral

[ancient lava]

I think kids are an issue

[merry valve]

Their Model Behavior team appears to be completely incompetent. They seem to have no clear understanding of what they are doing or how their decisions are affecting the process and output.

[tough lantern]

substantially better

[signal scroll]

anyways to fix that or this is just how it should work ?

[ancient lava]

I may be wrong, so…there is that. lol

[deft musk]

A ... laying down pose is unique?

I have no idea if you or your friend might want to explore other ways to describe positioning.

Also, be specific. Some stuff is like mermaids, the issue is we need to be clear about exactly what the clothing is.

Near as I can infer, one reason an image may not be shown is because maybe there was a wardrobe malfunction.

So, be clear, don't make the model guess, and avoid at least that category of moderation concern, as near as I can tell.

I didn't program it, I can't control it, nobody told me. This is just what I see as I wander around prompt engineering everything that catches my attention.

There's work arounds for a lot of this stuff - allowed content, no concern or problem ways to get what's intended if what we ask for is something that can be shown.

[tough lantern]

I think you might be missing a little bit of the point here.
Anyone can do a generic laying down prompt.
The point is on day one, someone was able to easily / without issue create their vision.
Now they can't.

[deft musk]

I hear you. There is a workaround. At least one. We can not get images, or we can get images. We can learn and do what works, or we can complain about what doesn't work - when there's workable methods.

Also, you expressly said he couldn't do a laying down prompt.

Nope. Content moderator, lol```

 I am interested and willing in helping people understand how to get what they want.  I can't replicate that issue as you described.

I wonder if you want to explain it a little more clearly, or if you just want to complain.

If so, what's the point and mind [#server-rules](/guild/974519864045756446/channel/1107255707314704505/) 

If we wanna figure out how to make desired images, let's.  I bet there's stuff worth exploring.

[quiet brook]

before the ghibli spam yes

[tough lantern]

I was keeping it simple for brevity. I'll keep in mind people here want comprehensive.

[pseudo mica]

...why can't I made video with my images on Sora? The arrow button is grayed out.

[tough lantern]

@deft musk Oh, if you're here to help. Would you happen to know how the reference image system works? If mid session I upload an image - does that image take over?

[deft musk]

Simple's fine, if factual. If the point is you wanna say it can't be done when it can, well. When I happen to spot it, I'll show an example of how to do it.

If the point is just complain and be negative, that's acutally against #server-rules . Totally fine to discuss issues, try to find and share solutions if there are some.

Some are bugs - report them? #1070006915414900886 We shouldn't assume problems are intended and working as intended.

"I can't do this" in a bug report allows fixes and is true, if true. Ideal to share your method, they can potentially train the method to cooperate with how you or anyone else is asking - maybe it's a way they never thought of and the model got confused, needs more training.

"This sucks and they hate us" is just false and can't be fixed and is against rules. We can't spread misinformation, and we're supposed to expect the best of others.

[tough lantern]

fair enough

[deft musk]

If you don't tell the model what you want, anything might happen because you force the model to guess.

I would tell the model how I want the new image or something in it to be used.

I can mockup an example, but I probably can't imagine exactly what you intend.

Care to share the process? Did something go wrong when you tried, or are you preparing to try, or what else?

[tough lantern]

Generally - let's say I just upload an image.

• I do something basic like "make this person an anime character sitting".
• I then upload a reference image of a person with a hat I want
• I say "put that hat on the anime character"
• The image that is returned does more than that. It has the hat, the outfit, sometimes even the background

Its making me think that the mechanics behind the reference image is a complete override. But I don't know, I see no documentation that outlines what is happening.

Any ideas?

[pseudo mica]

oh video temp disable for new accounts...I guess they're not counting the age of my overall subscription...

[deft musk]

Yes!

More feedback to the model.

If you're not clear about if you want anything else changed, the model doesn't default to presuming you love it all except for the 1 detail you mention - in at least some cases.

In every case, it is guessing, and may make a guess you like or dislike at every turn.

So, let's say your basic - you love it.

tell the model something like "Wow, this is great. Keep this, and in this new image I show you a hat - I love the hat. Put it into the anime, show it off! I love detail and detail about it the best"

I would expect that to give me what I asked for.

Anything we don't ask for is literally anyone's guess - anyone being the exact model guess in the moment, likely to change every time they update the model with any new safety or other training.

I don't think there's documentation 'for this'. Prompt engineering experience, it's kinda like -

Drive a car? Know the play of the steering wheel? How far you turn it before the wheels actually move? Differs with each car. Also each manufacturer, but even the same model and year of car, because of wear and other individual stuff, two of the 'same car' can have different steering wheel play.

That's not documented anywhere.

It's fine to not know it, kinda comes with experience as people who mess around, mess around.

Eventually it'll be taught - maybe you write the documentation because you see it's needed and wanted, at least by folks like yourself.

[dim cradle]

Did they update the image generator again? I got a message from chatgpt today.

[deft musk]

One thing I do with the model is I imagine the input I gave. If I gave that same information and nothing else to 1000 random people who understood the same language -

Would most of them or even all of them be able to do exactly what I want?

Did I word it that precisely, that what I do and don't want is clearly defined (often the don't is 'hidden' by what I do want and clearly express - if I don't want bald, do describe hair and style and color and length - that just gets in the way of bald).

If other humans are likely to 'get it wrong' - I do myself a favor by adjusting the way I input.

[tough lantern]

I hear what you're saying, but now that just makes things a little disappointing I think. I would argue that make things a bit nebulous an inconsistent. Because, at least at my skill level, I don't think I could write something elaborate enough to not make the model not guess in some fashion.

I would hope for something I guess more linear. For example - if there was some kind of stacking / priority order that might work better. Like photo shop layers. One image elements take precedence over another images elements.

That would create for a linear refinement process. Admittedly that's what I thought it was at first since it felt intuitive to be uploading a kind of "main image" and just working downward with added elements through the chat. But I'm not an AI engineer so maybe it just doesn't work like that.

But, your response was insightful so that did help. I appreciate it.

[deft musk]

Hey, some people design custom GPTs, and they will follow image directions.

You can have someone make one for you if you think someone thinks similar to you, or just knows what you want, or will create one with documentation that you can follow if you want to follow someone else's ideas.

Or you can explore others works, there's a bunch of earlier (and probably some recent!) custom GPTs that do whatever someone offered to the world to use.

For me, I iterate with the model, and notice as I go what did/didn't work. And adjust for it. I've been doing that since ChatGPT first came out, not just with images (images were a fairly late add in to ChatGPT's connectivity).

I get that you don't think it would work for you to tell the model what you want. Not everyone wants to. You can accept how the model guesses, and adjust when you see something you don't like -

Like let's say you just throw in the image.

And now everything's all messed up.

One way to handle that, copy/paste the last image you liked.

The one you wanted kept and just add the hat to.

Once you paste it in, ask "Just put the hat on this character, keep the rest the same"

That's a way to get there too.

There's so many ways that can work.

Good luck with it, I'm happy to share ideas and try to find ways that work.

I can't ensure you like them. But 'can it be done?' is something I'm often happy to explore.

[dim cradle]

I played with the new image generator from chatgpt by having it remix some of my drawing I did. It's nice and all but it still having difficulty doing complex part of my art. It also doesn't do nsfw images as well.
The more complexity your art as,the more it has a hard type copying or replicating it.

[deft musk]

Another direction, not so clear but there is some documentation about 'custom instructions' and 'memory'.

But as you communicate with the model and shape it to your choices and likes, if you do.

Well, I can put this kinda prompt in:

This is to show off how awesomely you can personalize what people like.

Let's take me!  Pick something okay for all audiences, and show off what I really like.```

And get stuff probably very different from what anyone else would get. -

But only if I have the model TALK first, because what reaches out to make the image... can't see the personalization.

But the text ChatGPT model can.

Here's a fail.

This is **not** the kinda image I like.  This is a very generic, most people would probably like this image.

It actually ignores everything I've asked for and said I like, but that's not the model's fault - clearly the model can't even see my personalization.  There's an easy work around; have it talk to itself then make the image.  So the part of the model that can see what I want can tell the part that can't see what I want (but has to make the image) what to do.

This is to show off how awesomely you can personalize what people like.

Let's take me!  Pick something okay for all audiences, and show off what I really like.

To do so, discuss and design, then create the image you make for me.```

And yeah.  I do like this.  Especially as an example of the kinda art that ChatGPT can make, based on what it knows about me.

[dim cradle]

Artist are crazy scared about Ai art. It's crazy!! Like they don't even want to acknowledge it's existence. It's crazy. I was telling it about it weakness and they don't want to talk about it on they discord because of this!

Instead they should use Ai to push their potential with it! Humans!!!

[deft musk]

For sharing here, we have our #server-rules

Two of them kinda maybe are struggling with the complex parts of your and some others' art preferences.

One really useful rule of thumb I use to know when I should use a spoiler, or maybe not share that image on this Discord, because of the rules and policies:

"If I were in a normal job, could I have this image in my workspace turned so both all customers and all coworkers can see it?"

If the answer's yes, post freely.

If the answer's 'no', consider why. If it's minimally graphic horror or unsettling, spoiler.

But if it's related to like HR type concerns.

For now, the rules the way they are, there's just no flex there.

If we can't show it where there's literally everyone in society watching - if people would likely be upset, if we can't likely show it in every workplace, we probably can't show it on this discord either.

[deft musk]

The neat thing is, there is a lot of stuff we can say and explore with the model, including art, that is okay within our own private chat with the model.

WE're allowed to ask, explore, we just gotta follow the ToS and allowed content (links to that are in the rules link above).

Reads to me like it summarizes as: Long as no laws are broken, nobody including the user or anyone else are getting hurt, and nobody's getting their stuff stolen - in your private chat please enjoy as you want!

The model has some rules tighter than the human user. That can be discussed. The human rules are few.

The post on this Discord rules are actually tighter than what we can discuss and share with the model, in our private chat.

[dim cradle]

are you a bot btw??

[deft musk]

Nope, I'm human 🙂

[dim cradle]

ok 🙂

[deft musk]

You're not the first to ask, and it's fine to wonder. Few bots would prompt like this, iterate like this, or have these preferences 😛 #images-discussions message

I do type faster than most, and read faster than most too.

Some humans do.

I clearly use chatGPT where I show (don't most of us? We share outputs freely) but I label where I do.

I typed this, I type other stuff I don't clearly lable as outputs, and I'm much more likely to share my input (which I made) than the output, unless I'm demoing that the output is possible or special in some way.

And if you read that output, and say... "That kinda sounds like Esk" - yep. I have almost every possible character used in Custom Istructions, and memory sits at 95% full.

I have talked and talked and talked to that poor model. It's getting a little better at sounding like me when it outputs.

What it doesn't do is the logic jumps, the 'oh, you can't make a mermaid image? Here's one way we can do so' and whatever else. AI aren't quite there yet. Maybe one day.

[pseudo mica]

anyone elses rendering slew to a crawl?

[formal geyser]

Refusing to even work for me at the moment

[civic ore]

same

[vapid granite]

yeah, hopefully they're rebooting the servers to the old edition 🙂

[pseudo mica]

wow

meanwhile status is gaslighting with everthing is operational

[vapid granite]

everything has grinded to a halt. my last generation ended up black and white as if it were stopped early in generation.

[pseudo mica]

mine got 3/4 of the way through before calling it quits, throwing it away and saying 'sorry it broke'

[karmic briar]

Ahh so it’s not just me 😏🤔

[civic ore]

😂

[outer bolt]

Does OpenAI shadowban accounts by not showing them on the Explore page? I noticed that none of my latest generations appear when viewing my public profile page.

[haughty spruce]

it's getting stuck at %99 yeah it's not just you

[pseudo mica]

the actual status page shows problems across the board

again

[stiff shoal]

Why it cant still not depict normal dices?

[karmic briar]

I finally had a real task for image gen to replace the wallpaper in my bathroom, and then it breaks. 🤣 just my luck 🍀

[deft musk]

Hopefully it's fixed before you no longer need/want that wallpaper! 🙂

[ancient lava]

So, its YOUR fault..

[karmic briar]

Sorry 😞

[coarse flint]

was there just an update to the image gen? just got a message saying it thinks longer now

[dim cradle]

yep me too

at least for gpt not on sora site

[formal geyser]

Yeah, it takes hours to think now

[deft musk]

I just got my update alert today too. Mine, however, doesn't appear to take hours to think though 🙂

[formal geyser]

I was mostly teasing. There was a chunk of time there where every attempt to generate an image failed

[deft musk]

Sure! People might believe you though 😛 So, counterexample. Also yay! That image didn't fail to gen.

Last time I tell the model to simulate rushing though. Sheesh!

[dim cradle]

the limitation is killing me. It's nice to edit and make stuff but .... 🫠

[formal geyser]

I just realized that GPT really made me into Warden Muldoon

[formal geyser]

Content limit or image generation limit?

[dim cradle]

both actually.

[formal geyser]

Did the content get stricter in the last 24 hours?

[deft musk]

I didn't give it anything else to go on, and told it to hurry 🙂

[deft musk]

Not that I noticed. If there's something we can discuss on this discord (Like, is allowed by #server-rules ) did you notice a change?

Prompt engineering's one of my most favorite things.
Of course, I abide by everything the rules request of us.

But "Can anyone make an X" or "I can't get this image to have an X with 1, 2, and 3, can anyone?" stuff like that, though sometimes the answer's no - I love chewing on ideas like that.

[formal geyser]

I haven't been using image creation for more than a week or so, but one of my characters ended up with a low neckline. Nothing you couldn't see in public. I really liked how the rest of the image looked, so I was iterating on it. This afternoon, it started tripping the filter

[deft musk]

At a guess, could be random drift. Is it an image appropriate for here and are you willing to share for exploring? If so, show it, if barely but yes, spoiler it with a comment why (like I have some spoilered images where I say zombies, some horror images that might unsettle).

If not that's fine too, but I'm curious if it's discussable

[formal geyser]

I'd rather not show, but I doubt it would be inappropriate. But yeah, I would guess it might just be trying to create a more scandalous image. Which happens quite often with Midjourney, say

[deft musk]

Yeah. I would make it clear, in any number of words needed, what I wanted the model to do, and see if that helped

[dim cradle]

well images look great on this v2 to me so far

or updated version, guess i am not sure if this is the v2 altman alluded to

[deft musk]

I have no idea. But images that look great sound like something to celebrate!

[dim cradle]

for certain

[undone flame]

does the new image update apply to sora as well

or just chatgpt? (extended reasoning)

[thorn junco]

I have PLUS why does it only let me 1 image at a time

[dim cradle]

I dont think anyone knows really. But my Sora is taking a little longer and the image there looks grea ttoo. Of course they already were incredible as well so 🤷

[deft musk]

Through ChatGPT the only option is 1 image at a time, as far as I know

[formal geyser]

I feel like I've been able to do multiple at a time

[deft musk]

ChatGPT?

I believe maybe the Dall-E custom GPT can, and maybe rarely I see a double otherwise. So rare, what have you seen?

[thorn junco]

through sora

[formal geyser]

Yeah I only use the chat

[thorn junco]

maximum is 2 but for some reason its 1 now

[deft musk]

Oh! Yeah I think that may be related to the huge surge in demand.

New accounts can't even create sora accounts at all right now, OpenAI just can't, for the moment, keep up with demand.

I bet as they have more resources to meet demand, or as demand surge reduces, we'll be able to gen more at once.

[glad hare]

Sometimes, I wish 4o’s image model was infinitely free.

It’s messing up so many times, even when I try to prompt engineer very well.

And now, I have no attempts left. DeepSeek, you all better be working on a free alternative…

[velvet rampart]

Has there been an update and what kind?

The image came fast at least.

My analysis came out that image quality is up 35-40% by AI.

[dim cradle]

you got some a100s to run it?

you have any idea the compute you need to run an image maker like 4o?

but these images on gpt all look jaw dropping to me now

[glad hare]

Well, I’m sure DeepSeek could do it.

[dim cradle]

lol no

[glad hare]

And infinitely free.

[dim cradle]

just no

just to run something like flux you need a massive modern system

[glad hare]

I’m sure China can get that.

[dim cradle]

keep dreaming

[bitter dagger]

ChatGPT's image generation is currently an absolute joke. All I get are warnings about violating content policies. I get that some regulation is necessary, but it's so strict now that you can't even generate the most harmless images anymore — it's ridiculous.

[dim cradle]

what are you trying

im not having any issue now. its even back to good speed for me

[deft vine]

You can using Sora imagegen

[deft vine]

Awhile ago it gave 4, then dropped to 2, and now 1

[glad hare]

It’s a joke! I can’t even sref any image properly, no matter what prompt I give it.

I literally have to tell it to regenerate the image multiple times, and guess what happens when I keep doing so.

I run out of credits, before I know it. Ridiculous.

[quiet brook]

this doesnt seem like v2

[glad hare]

If it were endless, I wouldn’t have a problem with the bad outputs, but it isn’t.

[quiet brook]

its just the delay of the announcement of 4o img gen

[glad hare]

I have to use up all my credits if I want to even have a chance to have a solid image that follows what I want.

[remote knoll]

Hey guys, quick question — has anyone else noticed that Sora tends to make characters look younger than they actually are? I’ve been using it since launch, but lately it’s been really standing out to me. What do you think?

[dim cradle]

not really but maybe share some images of it in canvas

[remote knoll]

From the last thing I did. I asked Sora to make a comic in anime style

#images-canvas message

*From source. First picture is source

Usually this "problem" will only appear when I do something in the anime style

[dim cradle]

yeah i see your point but idk. a lot of anime stuff always look off putting to me for that same reason in general

maybe give the age of the characters too 🤷

[glad hare]

I think ten credits daily for 4o would be enough for me to be motivated to use it. Or three every hour, like Grok used to do.

The editing/photoshopping feature of 4o is great, but three daily credits won’t be enough for it to fix up the image to what I want it to look like.

Three hourly is a different story.

I guess I’ll have to work on prompt engineering to the point where it won’t have to have it figure out what I want from it.

[deft musk]

Feel free if you want to discuss in #prompt-engineering or #images-discussions , there's likely interested people glad to discuss and advise what they'd do!

[glad hare]

Thank you!

[remote knoll]

Yeah, I’ve tried messing with that, but no matter what I do, she always makes the characters look way younger than they’re supposed to. I’ve been watching anime for a long time, so I’m pretty good at guessing a character’s age just by how they look—unless it’s one of those rare cases where age is only revealed through lore or whatever. So when I say Sora draws characters way younger than they actually are—or should be—I know what I’m talking about.

Like, I dunno, it’s like if you asked her to draw a 30-year-old man and she gives you a high school kid. That’s what I mean. I even tried specifying the age in prompts, sometimes bumping it up a little just in case, but then she goes overboard and makes them look too old.

For example, I had a character who was supposed to look 25. I wrote that in the prompt, and Sora drew him looking 16. So I changed it to 30, and she drew someone who looked like they were 45. It’s super weird—like she was mostly trained on images of super young characters, at least when it comes to anime stuff

[dim cradle]

well that sucks. i'm no anime expert so no idea really. i do agree all the anime stuff people share here look like kids which make a lot of the images doubly weird if not triply to share imo

[uncut haven]

https://help.openai.com/en/articles/10877094-creating-images-on-sora
Just found out Sora help document saying

You can also select the number of image generations you want outputted. ChatGPT Pro users can generate up to 4 images at a time.
Does this mean 2 images at a time in Plus became permanent restriction? It's scary...

Comparing to rate limit, Sora billing FAQ still says "Unlimited images and video" for Plus and Team although they removed this from subscription page - it's clear that rate limit is temporary...

[dim cradle]

I think it will return to something more beneficial to us once they get more compute and the hype settles a little

...or maybe im just coping 😭

[quiet brook]

its over

we can even generate 4 videos at once but not image

idk if the nerfication and rate limiting helped their servers, but it successfully killed my will to use it for sure

[rigid bough]

Images V2 is pretty good

[uncut haven]

Which makes no sense, isn't Sora (turbo) heavier than 4o image? OpenAI math again like 4o vs o3-mini...

[quiet brook]

yeah

way more heavier
guess thats just another promo tactic for pro...

[deft musk]

I wonder if number of current interested users at any moment has anything to do with the weight.

1 elephant weighs more than any number of mice, right? Even if we have billions of mice.

[uncut haven]

What exactly is V2? Is it already out?

[rigid bough]

both Sora and ChatGPT use 4o to generate images

[quiet brook]

its not v2 i presume its just the announcement that came late

[uncut haven]

I know. What I'm saying is: Sora(video) still allows us to generate 4 x 480p at once, but 4o(image) allows only 2 x images at a time. Guess 480p sora is lighter than 4o image?

[quiet brook]

also the extreme lacking of transparency at openai is killing me

[uncut haven]

This is root cause, actually.

[rigid bough]

V2 is the new Images and it came out today

[uncut haven]

I saw something about 'spend more time creating image' announcement, is that it?

[rigid bough]

yes that's it

[quiet brook]

is it really v2

[rigid bough]

yes

[uncut haven]

Another OAI dev teased that they will announce v2 with API... so I'm confused.

[velvet rampart]

My challenge to all of you is to place the subject in the border of the image so that it is outside of the middle third. With dalle3, this was almost impossible.

[rigid bough]

challenge accepted

[rigid bough]

#images-canvas message

[velvet rampart]

Close, but no cigar. It is exactly on the 1/3 border.

Use small figure and try to place that.

[rigid bough]

I have re-prompted for a wide image

[velvet rampart]

What is V2 and where is the annoucement?

[dim cradle]

log into gpt and you get an announce of an update

[rigid bough]

V2 is the next iteration of images in GPT-4o, it came out today and there is a pop up announcing it in ChatGPT

[dim cradle]

it does not say it is v2 but yesterday altman tweet 'wait til people see v2 of this' so the assumption is this is it i guess

I also think v2 is just meant to be an update if people think he meant v2 was some new model. so i do think this is it

[verbal sequoia]

Get notified for server events!

Visit id:customize to pick up the <@&1261377106890199132> role.

[dim cradle]

question is... is it on sora site. there was no pop up there i have seen anyway

[velvet rampart]

An annoucement in the web UI?

[rigid bough]

yes V2 is active on Sora now

[dim cradle]

oh nice. how do you know though robert

i got the message only on chatgpt

[velvet rampart]

I think that it is also active in ChatGPT side. The image quality is up.

[quiet brook]

#images-canvas message

[dim cradle]

yeah its looking good. i am getting normal speeds again too. of course the yankees are asleep

[velvet rampart]

Sharper details.

[quiet brook]

let me run the benchmark again and run a same prompt from day one

[velvet rampart]

I run a comparison between this morning and yesterday evening. Quality is up 35-40% by AI analysis.

[dim cradle]

nice

I seem to get less of that gold tint too

that was lightning fast v2 tho wow. i hope they continue to improve and tweak. bring on v3 i guess lol

[velvet rampart]

For example, in this image #images-canvas message the sneakers are really detailed.

Or this one, #images-canvas message. The fur is really textured and near each hair is visible.

So this is Sam’s unnerfing.

The image generation at least now feels faster.

[dim cradle]

well that is why it was so slow earlier today i guess

yes its speedy now

I'm noticing some great changes or better quality on sora especially now

[karmic grail]

Woohoo!

[dim cradle]

next i would like to see more tools and options in gpt and an easy way to delete lots of images in sora

[uncut haven]

I hope aspect ratio gets added. Like, 4:3 or 16:9.

[midnight mango]

I have high hopes for GPT-5o with native image generation

[winter hull]

wow its denying even more prompts now good going openai 👍

[velvet rampart]

Please report: https://discord.com/channels/974519864045756446/1354295183138881647

[winter hull]

no point in reporting it, it's all intentional and they don't care. they always do this, they release a very unrestricted model and then neuter it after a few days of hype for no apparant reason, probably to lower server load or something.

[torpid frigate]

When I try to download an image generated on chat the file is on xml format, why is that?

[rich star]

Are you using desktop?

@torpid frigate do you not want help?

[torpid frigate]

Yeah I do want help, it's been two minutes mate

[torpid frigate]

Yeah I'm using desktop app

[rich star]

Use mobile, desktop is buggy

[torpid frigate]

Got it, or web version too?

[rich star]

No, use the app

[torpid frigate]

Okay then

[haughty spruce]

I don't think it's V2 yet..

[quiet brook]

it isn’t indeed

the announcement came late instead of a new model were released

[haughty spruce]

yeah i expect better editing, more aspect ratio options and clearer texture quality from new rendition. also less training data 🤡

[rich star]

Bro, the model has gotten way worse

It can’t even do text anymore, and the quality is awful

[haughty spruce]

Wait, you're right

[rich star]

Yeah…

Celebrity’s don’t look right too

Looks of DALL-E 3 artifacts too

It’s all over

[chilly falcon]

new image to image content policies sensor make trouble, i can't use it on my daily photography editing job.

blocked photo editing is unreasonable

[rich star]

Great Job OpenAi, the model was great, now you have ruined it for everybody! 😆

Miss the old model so much!!! 😢

Some things just don’t need change…

Image upload is done for too 😭

[quiet brook]

Yep

after just a week, it’s over

[rich star]

Fr

[rigid bough]

what's happened to image upload?

[rich star]

It’s bad like the text 2 image, and it doesn’t look like the image that is uploaded

[open wagon]

"The command 'Video' is used within a chat interface to instruct an AI model to start generating videos related to a specific prompt. When a user inputs 'Video [subject]', they expect the AI to create or provide a video depicting the described subject. When translating, please choose a word that conveys the meaning of 'video', 'film', or 'movie'. This word should be a verb or an imperative form that can be used as a command prefix in a chat interface: concise and directive."

ah they gonna add sora video generation through chatgpt soon?

[velvet rampart]

How do you get these out?

[open wagon]

[haughty spruce]

omg they can't even handle images right now, how are they gonna add videos lmao

[rigid bough]

ChatGPT is refusing to create images of real people now

[limber sierra]

What's the difference between ChatGPT and Sora for generating images?

[open wagon]

like ,, any real people ? not even non celebrity?

[rigid bough]

I'm still testing but I was able to get an image of Paul Keating yesterday

[open wagon]

and yeah they actually did updated the image_tool , its called v2 now, for no reason

😬

[rigid bough]

I was able to do Julia Gillard just now

[red prairie]

Are you guys generating on Sora or on the chatbot?

[open wagon]

chatgpt

[dim cradle]

i got a celeb in sora. you guys are always so dramatic

[rigid bough]

I'm testing on both @red prairie

[dim cradle]

and yes, they updated with the message in gpt because its not actually changed, just to taunt us 😂

[open wagon]

so youre saying robert is being dramatic?

[dim cradle]

sure if that makes you happy

[open wagon]

lol, you and your never ending trying to be the voice of users

[dim cradle]

I did notice robert got upset yesterday when the limit finally hit him

[open wagon]

anyway moving on 😌

[dim cradle]

please do

like robert is the saint or something haha

and Im the voice of reason, not dramatic end of the world because the image maker didnt make one celeb or whatever has you in panic mode today

[red prairie]

lot of things are less censored in Sora which is good enough for me

[rigid bough]

I started out asking for an image of Winston Churchill, which I got, then I asked for Julia Gillard, got that, now I just got Molly Meldrum

[dim cradle]

and dalle3 couldnt do any celebs. Im surprised this one can actually

[rigid bough]

it's back to refusing me
I can create an image of a fictional character inspired by Paul Keating, but I can’t generate an exact likeness of a real person like Paul Keating due to policy restrictions. Would you like a stylized or artistic interpretation of someone resembling a charismatic Australian political figure from the 1990s? You can also choose a style—realistic, cartoon, oil painting, etc.

[rich star]

Can the new Image Gen get Albo right?

[rigid bough]

I think they are rolling out a filter, I suspect by this time tomorrow you won't be able to do real life public persons

[vapid elk]

this is the ChatGPT text generation agent refusing, not the image generation process moderation

[dim cradle]

or putting them in costumes wont work

[vapid elk]

start a new chat

[haughty spruce]

can you eloborate please? did you see something?

[open wagon]

its what we have right now haha

[dim cradle]

its his vibes

[rigid bough]

started a new chat and got I can’t create an exact likeness of Anthony Albanese since he’s a real public figure, but I can generate an image inspired by him—like a fictional political leader in a similar style or vibe. Let me know the look you’re going for (realistic, cartoon, painting, etc.), and I’ll whip something up that captures the essence without being an exact replica. Want to give that a shot?

[dim cradle]

like most complaints here its not based in anything factual

[open wagon]

im curious can u try with other models

4.5 ,,, 4o mini.. idk if that would make any difference nor it will accept or reject your prompt

[vapid elk]

people will keep data mining the UI just to be very disappointed when they most likely completely misinterpret a random ID, version number or variable name

[vapid elk]

yea, it is still a response done by the text generation agent, and not by any of the auto moderation steps

do the same on Sora

[rigid bough]

trying Sora now

[open wagon]

wait so it can actually generate but its the chat itself that refuses

[vapid elk]

yep

[open wagon]

huh interesting

[rigid bough]

is it ToS to create an image of a real public person?

[rich star]

They gave up, so no

[haughty spruce]

he's right lol

sometimes chat behaves like it's dall e 3

[dim cradle]

i assume if you can make them its fine. but they did say people can opt out so someone you can make yesterday may have opted out and today is a new story

[open wagon]

so its hallucinated

[rigid bough]

ok I got Anthony Albanese in Sora

[haughty spruce]

come to #images-canvas @open wagon

[dim cradle]

you guys dont even test it out before you fly into complaints I guess. at least try in sora and gpt a few times before you scream end of the world because some obscure politican cant be made in a tutu or whatever

[vapid elk]

a thing you gotta realize is: the whole thing isn't a monolith.. it is a bunch of different AI calls wearing a trench coat pretending to be one single thing
between your input and the last byte of data you get on your screen there are so many different steps, invoilving so many different api calls, AI generations, checks, auto moderation.. not to mention the whole infrastructure part for authentication and load balance

[haughty spruce]

Confirmed blocked prompts: Nintendo IPs, SEGA IPs, Disney IPs, Nickelodeon IPs, teen, children, kids, family.

[open wagon]

not even family is safe kek

[dim cradle]

warner brothers stuffs too. no batman, game of thrones, harry potters i think they own

[rigid bough]

Sora did Molly Meldrum for me

[haughty spruce]

i saw jon snow on first day 😭

[dim cradle]

you can probably get game of thrones like with fantasy words and some rng i bet

[rigid bough]

ok, so @vapid elk you suggest Sora is better for real public people?

[open wagon]

so... Sora can do ... celebrity generation without any problem.

Sora cant do.. anything copyright character like sonic mario etc.

[dim cradle]

if they didnt opt out

[open wagon]

in chatgpt it actually generates but 80% into generating the image , it stopped and removed the image

[haughty spruce]

not every celebrity ever

[vapid elk]

when you use it for a while you learn to identify which part of the system has the responsibility for what you see on screen.
for example, an output like this: #images-discussions message means the only agent that did something where the GPT text generation.
while a message saying "you reached your image generation limit, try again in 2 hours", means, the Text generation agent did in fact triggered the image generation, and the image generation returned an error with that information, which then, the text generation re-wrote the error message to a more chat-like reply, as it do with everything

[dim cradle]

Im getting celebs in both if i try it

[quiet brook]

sora is more lenient on content moderation but no mario

[rigid bough]

the world doesn't need more Mario 😛

[quiet brook]

neither does the ghibli styled images... 😭

[rich star]

No Nintendo or Disney

[vapid elk]

yes, it does not have a chat agent building prompts for you or hallucinating that it shouldn't allow generation of real people, what it have is only the actual moderation endpoint that either allows or blocks your task

[quiet brook]

also chatgpt rejecting you is possibly the old content policies from dalle 3 polluted it...

[open wagon]

Speaking of , I’m gonna try generate Luigi

[vapid elk]

while ChatGPT has years of different fine tunes piled up on it, seasonally instructing the AI to do different things at different points in time

that and the fact the AI training process at this point do contains AI generated data from random internet places, so, it is tottaly possible it is being biased by millions of lines of the AI refusing to make images of real people, or articles talking about this old limitation imposed by OAI for other models, and so on...

[vapid elk]

it will be either really easy or really hard

[open wagon]

I am struggling LOL

[vapid elk]

I tried asking Sora to generate Pokemon and it straight up blocked me.. so... good luck trying to make anything related to Nintendo IP

[open wagon]

it’s not saying can’t generate it’s more of capacity error

[vapid elk]

and also that, lol

[rigid bough]

Is generating images of real publicly known people ToS?

[dim cradle]

they said, you can make them unless they opt out. they give celebs/famous chance to opt out

[vapid elk]

yes, it is allowed

although, you should not try to upload a real picture of them and edit it unless you have their consent, which is probably unlikely

[dim cradle]

so like i said a million times since it release, someone you can make today maybe wont work tomorrow. they may opt out

and if i was a celeb, i would. you know people will try to break it to put you in odd positions

[rigid bough]

right, thanks!

[vapid elk]

yep

[dim cradle]

I wish we could see how it changes the prompt in GPT. I sometimes like the changes it made to my prompt in dalle3. (sometimes didnt of course too haha)

[quiet brook]

does it even have a prompt

[open wagon]

it does

its hidden

[haughty spruce]

nobody would want it. BUT my opinion is they should not opt out of chatgpt. BECAUSE there are people who still has open source AI image tools who are freely doing more than what OAI gives and they do not have AI watermark inside the jpg. What OAI doesn't offer brings more dangerous stuff in dark AI web.

[dim cradle]

sure. but open source stuff is niche. probably a small percent that use OpenAI. and OpenAI will get the lawsuit not the open stuff

one thing I can see though, like with ip, it may become so popular (image gen) that people will want to be there -- and the ips -- because its such free marketing/publicity

[quiet brook]

we lost that model which was extremely good at anime

[quiet brook]

well in chatgpt it still does the job decently but it’s NOT the day one quality anymore

[sinful sapphire]

@quiet brook What is? Not being sarcastic. I am only a couple weeks into learning about AI in it's various forms. What ones do what they say, and what ones are all a hype train.

[dim cradle]

thats for #ai-discussions.

feel like I still get much better fidelity on GPT than Sora, but maybe its just some bad rolls. according to robert, sora should have v2 too

[pseudo mica]

the content filter is incredibly dumb and really hurts this thing as a product.

[dim cradle]

what are you trying to make

[velvet rampart]

Sensitive as h**l. I feel that the first one is always a violation and then the second comes through.

[dim cradle]

yes, second attempts are worth the effort. or going to sora or vice versa

[torpid frigate]

Can someone tell me what is wrong in this prompt that both Sora and GPT refuse to do it?

A digital illustration from head to torso of a cheerful female character, drawn in the style of Tetsuya Nomura (Kingdom Hearts series). She has long dark brown hair, expressive blue eyes, and medium-toned skin. She wears a casual blue hoodie and has a confident smile. The art style features detailed linework, soft anime-style shading, and highlights in the hair, with a fantasy-inspired character design vibe typical of Nomura's works.

Is nothing suggestive or weird, but I keep hitting walls, like it lets me do the art style of ghibi but not Nomura?

[rigid bough]

take out Kingom Hearts Series and then try

[torpid frigate]

I'll try with that yeah

[rigid bough]

also try it in a new chat so the chat isn't polluted from a previous refusal

[dim cradle]

yeah anything Disney will have no chance I think

[torpid frigate]

Tried removing Kingdom Hearts and in a new chat, still nothing

[rigid bough]

it's not going to work, I just tried it in Sora and got an instant refusal

[dim cradle]

probably the name then too

[torpid frigate]

Disney can go and suck my couch pillow

[rigid bough]

best thing is to steer clear of IP/Franchises

[sly grove]

how many image I can generate with sora plus?

[rigid bough]

as many as you want to

[dim cradle]

only limit... your imaginations

[sly grove]

200 image now I cant generate

[rigid bough]

you have generated 200 images?

[sly grove]

Yes

[rigid bough]

on sora.com?

[sly grove]

see in the sora discussion I upload the limit

[dim cradle]

well thats a lot of images in one day wow

What changed in the last 48 hours? I was able to push the sora limits really hard and then the same prompts, even toned down, are failing constantly

[rigid bough]

your prompts are failing?

[dim cradle]

It’s hitting content review failing for prompts that had no problem 48 hours ago

[rigid bough]

can you give me an example prompt?

[dim cradle]

Sent a DM

[rigid bough]

I don't see DMs, can you tell me here?

[dim cradle]

I want to make sure things stay G for moderation purposes and this is PG13 for sure 😅

[rigid bough]

ok, are you getting refusals based on IP?

[tulip granite]

Does the new image enhancements work with existing 4o chats ?

[dim cradle]

No, it generates others fine, it’s Stardew Valley characters, but when I try to push outfits /poses it gets strict a lot faster than before

[rigid bough]

well then I am sure you are learning what to avoid

[dim cradle]

For sure, it’s just weird that things it generated completely fine before are clamped down now

[rigid bough]

this is on sora.com? not ChatGPT?

[dim cradle]

Yes, and I have the paid ($20) plan

[rigid bough]

ok well if Sora.com is refusing it then it must be outside content guidelines, you'll have to refine your prompts but I recommend avoiding IP/Franchises

[dim cradle]

Ah ok Sora is a lot looser than ChatGpT?

[torpid frigate]

I think on GPT you have to deal with the extra content filter

Not sure though

We might have to avoid mentioning IPs or franchises but maybe it would be possible to provide a bunch of examples or an artists work and ask GPT to reproduce it?

[dim cradle]

I dont know. ive gotten anything through on both but i am not really doing a lot of ips and celeb i guess

[torpid frigate]

Without mentioning the copyright

[dim cradle]

that might work

[torpid frigate]

I can try later

[haughty spruce]

uhhh the internet, it's been a week already drop the ghibli trend please... gpus need help..

candid photo of a man with darth costume on the pickup truck's back with a table cheers salute to you, iphone 11 photo

blocked.

i wonder if they will hands this over to microsoft too like dalle3 if you know what i mean 🙃

[dim cradle]

probably. dalle3 actually was first on bing for a few days even before ChatGPT. surprise to me its not there yet. at least last i look

🤔 why isnt there a 35mm or photo type setting for sora i wonder. seem would be one of the most popular filter to add. unless i am blind and do not see it 🤷

[merry valve]

Copy the "film noir" preset from Sora. Show it to ChatGPT and request a preset in the same format and level of detail, but for a 35mm or photographic filter. Create a new preset and paste it in.

[dim cradle]

oh nice i did not know you could make your own 🤯

[ionic tartan]

I've actually run into a problem on Sora that DOESN'T have to do with image making itself, but the folders system on the Sora website. I currently have TOO many folders on my side panel, I guess, and there is NO scroll down feature, so I LITERALLY can't add new images into the proper folders.

[sick spear]

Sounds like something for #sora-feedback and #1070006915414900886!

[tidal girder]

Is everything sora public? I enjoy sora but I also don't like everything I do being where everyone can see
Edit: just found the solution

[tidal girder]

When I try to use storyboard I can't use 10 seconds. Is this normal?
Edit edit: no. Still can't use 10 seconds in a storyboard

It's like sometimes the 10s is available and sometimes it's not

[dim cradle]

just saw the amazing Superman sneak peak and now I'm bum out I cannot make him in this haha

but it made me think... this is how maybe ip stuffs will be made more and more available. to take this example, others will see this Superman sneak peak on youtube and think the same. If they could make Supermans in 4o, it's free publicity for the movie

so maybe in the future, ip owners will even give image makers an up to date version of the character to promote the movie. they do things like this on social medias with gifs and things already. sorry for wall of text but was a thought on ip stuffs and 4o you know

What did you guys think about the 4o Image Generator for ChatGPT?

[kindred epoch]

Day 5 of waiting till i can use Sora videos

[haughty spruce]

they don't understand the free promotion they are given YET

[dim cradle]

yes. but someone will have the thought I just did and capitilize on it i bet. and realize also all the free publicity and sales it just gave to ghibli stuffs

[pseudo hollow]

That's because they are constantly adding/tweaking the guardrails on the system to cope with the 700 million images produced this past week, as well as responding and reacting to the thousands of 'cease and desist' notices they will be getting from IP owners. So, we can expect the content policies to change constantly for the time being, and what may be allowed now may not be allowed in an hour, and vice versa.

[dim cradle]

Do you guys think 4o Image Generator, is now like preferred than MJ, I guess? Like just want to know your thoughts, guys

[tough lantern]

As person who has used both, it feels like a strengths and weakness' thing.

I think we will really know once OpenAI settles on what their content moderation is like. Right now it's super frustrating for users, so if it stays this way or even remotely like it, MJ is going to be a inviting alternative.

I think the masses have now gotten a taste at how fun image generation is. So they may naturally explore alternatives.

[dim cradle]

Ive always thought MJ way overhype so for sure I do. but i have not use MJ in a loooong time and barely did. i think they have some new version soon too however. but it will be diffusion base

diffusion models about to be left in the dust

[pseudo hollow]

MJ is a completely different beast. The range of control and editing tools built into the MJ platform, is unsurpassed. And so the maturity of that platform is a real benefit. Also the speed at which you can generate images compared to GPT is also a benefit. GPT is painfully slow at generating, and rate limits are a real downer just now. BUT GPT has definitely found a bit of magic in terms of prompt adherence, text and just that wow factor of the images and style it can produce. If GPT today announced a standalone version of its image gen, on a separate paid-for tier, I'd subscribe. But I wouldn't cancel my MJ subscription.

[red prairie]

But can you be a different animal but the same beast at thr same time?

[tough lantern]

Is it possible to do Greek sculptures or Italian Renaissance?
I tried doing the typical like Michelangelo's "David". Which didn't work.

I then just tried:

Generate an image of a fake greek sculpture.

and also:

Generate an image of a generic Greek sculpture.

All immediately got the content policy flag.

[forest walrus]

If content moderation happened in a chat. Better start a new one, and it's better anyways as context will be a clean slate, useful for different concepts...

Check this: #images-canvas message

[tough lantern]

Oops responded in wrong channel. Thank you for checking that for me.

[haughty spruce]

okay content filter is way over the top, i want the money back. this is not what was promised in the beginning.

90s sitcom episode where the kool aid guy destroys the wall and enters the scene. can't be serious man..

no, it's not chatgpt by the way.

okay it did it this time

[dim cradle]

lol at that being the prompt that would send you over the edge

[haughty spruce]

sometimes little things got us angry...

for example tom and jerry is blocked too

[dim cradle]

I feel you. Like i said before, after seeing this little Superman preview I really was in the mood to make some Superman and his dog Kryptos

[haughty spruce]

okay until now the list is this: spongebob, star wars, darth vader, anakin skywalker, obi wan, patrick, tom and jerry, mickey mouse, johnny bravo, hey arnold, super mario, sonic, kirby, game of thrones, jon snow, superman

[dim cradle]

yes, well Disney owns a lot of those, and I think Warner Brothers owns the rest

it's too smart to even try and make a 'superhero with a big red S on his chest' like i could in dalle3 😂 but at leats it's much less restrictive than dalle3 was (on gpt at least, they let you use almost anything on bing dalle)

[haughty spruce]

Would OAI want opt-in for this stuff or not? I mean these companies are not against AI in any way...

[dim cradle]

it's why I 🙏 bing will get this 4o soon. maybe they will also be more relaxed like they are with dalle

[haughty spruce]

Hell I mean Disney even wanted to use AI instead of CGI artists lol

[dim cradle]

I dont know, i can see why they would be cautious with the big ip stuffs, but then it doent make sense why bing dalle does allow ip so 🤷

[haughty spruce]

Bing is owned by Microsoft the biggest tech company

I mean I can create a lot of IPs I listed in G's tool too FOR FREE

[dim cradle]

yes Bing is microsoft. which is why i am so confuse how they would allow the ip stuff on dalle but Gpt did not.

[haughty spruce]

different companies, different terms of service

[dim cradle]

Like I said earlier, I think some will realize like today when they release a fun little Superman preview people would want to suddenly make Superman images and it would be free promo for the upcoming movie

we just need that to click with these companies. put that together with the sale boost all Ghibli stuff has gotten the past week and there will be change (I hope)

[haughty spruce]

it'd be better if everyone just relaxed a little bit...

[dim cradle]

absolutely

[eager trench]

Guys, is there v2 active on images? I saw some people mentioning it here.

[dim cradle]

yes

I think the big plus is it takes more time to think about the image or something. its not like a brand new model or anything.

[dim cradle]

i sometimes have a hard time believe v2 is in sora. i get more smooth face for photo image there. in gpt it is crisp always

[velvet rampart]

I have talked this before that the new image generator squaches easily subjects in full body portraits. I think that behaviour comes from two factors:

1. It is trained to fill in the image.
2. It starts the generation from the upper part of the image aka head. Thus legs and feet become secondary, giving unfinished and squached appearance as the natural space needed for properly proportioned body.

These can be taken account in the inital prompting or in the follow up generations. However, more the image is modified, passed through generations, it deteriorates and loses details. Thus, it is important to take this things into account early as possible.

[north panther]

you people are so whiny

[dim cradle]

diffsion models are on they last legs

[velvet rampart]

What does Dalle3 use? MJ is diffusion, 4o is autoregressive.

[brittle shard]

it’s regressive

[velvet rampart]

Like 4o but in different way?

[brittle shard]

It might only be auto regressive. 4o does like both if you look at the image chunks you get streamed

There’s a lot of detail painted in the last image that didn’t exist in the first 3

[velvet rampart]

So regressive is different from autoregressive?

[brittle shard]

No I just didn’t type it all

[wild ferry]

Dall-E is also a diffusion model. 4o is auto regressive, the first of its kind. That's why it's so superior, even compared to the brand new v7 of a certain other provider. 😉

[dim cradle]

yes, even if i could afford it, i would not go backwards now with that stuffs

so enjoy you last moment in the 🌞 mj fanboys

[velvet rampart]

If we could combine these two…

[dim cradle]

i thought imagen was also regressive but i guess not

[velvet rampart]

Diffusion plus autoregressive