images-discussions

[verbal cloak]

Seems to work, kinda (getting errors now), on the app. Brave and chrome not so much. That is about as much as I can do with that for now.

[reef root]

Pro version- definitely have noticed restrictions on pictures. ChatGPT will ask me, after helping me with a project, “would you like to create an image” and I always say yes and never prompt it what to create and it usually creates a an amazing image. Now it try’s to create the image and then says it cannot due to policy issues. I hope these restrictions get lifted or we find a work around.

[rigid bough]

Content policy restrictions are fundamentally broken -> #ai-discussions message

[livid dust]

when 4o model coming for image creation????

[open wagon]

coming soon to your nearest cinema

[velvet rampart]

You should have that whether you have free, plus or pro plan. Accesible via normal ChatGPT chat with 4o model.

[livid dust]

in API support, Not in UI

[open wagon]

you didnt specifically asked that in first question so

[velvet rampart]

In near future. At some point.

[open wagon]

its coming soon to your nearest cinema!

[velvet rampart]

Before the next doomsday in the upcoming apocalypse season.

[fresh meteor]

When there are enough GPUs, it’s probably going to be on the higher tiers because of demand. Just a speculation.

[misty oriole]

Hi, i made an image of Tetsuo from Akira with Sora and it ended up looking gross, can i delete it?

oh found the option nvm

[dim cradle]

so the hype is over i guess. images are back at the bottom of the rooms like the old dalle3 days.

[frozen copper]

Well at least plus is back to 2 concurrent gens. Now just need to be able to do 4 variations per gen again

[dim cradle]

I'm so use to 1 gen at a time from dalle3 that I guess it does not bother me too much

it could be ever worse. there were days just some month ago where i would get told im on time out for an hour or something because I make 5 images in 20 minutes

[velvet rampart]

It still does that when dalle3 is used.

[frozen copper]

I use both Sora and 4o at the same time to output more images. That way i can faster iterate on prompts

[dim cradle]

good idea

[velvet rampart]

I use chatgpt as a prompt writer, figuring out why an image is not generated on Sora or getting a result what I want, etc. Images are done Sora, Faster and not polluted by earlier generations.

[dim cradle]

good idea. its funny to me the things that work on sora but dont on gpt. you would think it is the same model but i guess gpt is rewrite the prompt or something? idk

[wicked plinth]

Chatgpt is useless in telling me why a prompt was rejected. Although this afternoon it seems to be letting more through than usual. I'm enjoying it while I can

[tranquil harness]

Someone knows why i can't use the sora video generator and only the picture generator ?
i subscribed to plus around a week ago and i cant generate any videos

[dim cradle]

for some reason there is a wait time. just keep checking it should unlock soon i guess

[tranquil harness]

for my friends it unlocked within 3-4 days
and im still waiting

[dim cradle]

sorry to hear that. i guess you can send a message to their help

youre clicking on pictures right to see if you can change it to video?

or 'image'

[tranquil harness]

yeah

[quiet brook]

it is?? send screenshot then I can update the comparison list

[drowsy sigil]

If you're using a GPT, it will use Dalle-3.

[ionic ore]

Can anyone else only generate 1 image at a time still?

[robust spade]

Go to Sora.com and generate there- it only uses the new model.

[quiet brook]

an interesting observation is it tends to generate dates in online post screenshots as 4.24.2024

[silver hatch]

Same here.

[quiet brook]

in sora?

[vague kelp]

In the app, if you hold down on the image while it's generating and click read aloud, it will read out the prompt that it turned yours into. It's helpful for Good idea of how it rates prompts and how you can adjust yours to get what you want

[pale saddle]

yes i am

[velvet rampart]

Yes. They have also changed settings that one cannot chain generations. When 4o image generation came out, I could chain the generation. My guess is to slow down image generation.

[frozen copper]

oh wow, we are back to 1 concurrent gen on plus

[velvet rampart]

ChatGPT?

[frozen copper]

Sora

[velvet rampart]

Huh? On Pro, I have 5 with 4 images.

[frozen copper]

You are on pro plan, i'm on plus

[velvet rampart]

Yes, I understand. It just that plus users are this mistreated. You make the main bulk of users and money flow.

[pale saddle]

sora not working right now for me. Anyone?

[vapid elk]

yes, the page is super slow

not working at all for some people

[rigid bough]

reposting it wasn't sensible, now it's a problem here

[dire vale]

which image gen is better between the one on sora or on chatgpt?

[late blade]

it's working fine for me

[velvet rampart]

Depends on your use case.

[dire vale]

whats the difference though

[late blade]

for iterative work ChatGPT, for Zero Shot Sora

[velvet rampart]

On ChatGPT, the whole is the cotext, earlier generations may influence the current generation.

[haughty crown]

In my experience dall e is amazing and sora is meh

[velvet rampart]

Are you mixing up Dalle3 and 4o native on ChatGPT?

[kind grail]

Btw is it just me or was the editing tool much better during launch? The direct image output is still good, but the editing REALLY looks like we are back with DallE3, and i dont even mean this as an exageration.

[velvet rampart]

When you make changes the degradation of the image quality in known. It is best just to take the image you are modifying and start a new chat proceeding there.

[kind grail]

Thats not what i mean. Even if you fully change the style. In the beginning it would usually actually grasp even little details, and although still prone to overcorrecting images, it was still better than what i usually get nowadays.

I can actually give you a pretty nice example of this quality drop in #images-canvas

[coarse stag]

anyone else suddenly unable to generate anything? It all just gets stuck in the queue and then fails to generate

[haughty spruce]

is it taking forever to load for everyone else?

"There was an unexpected error running this remix"

both in sora and chatgpt

[coarse stag]

yes i'm unable to generate anything at all it all just cycles and then fails

[sly spear]

Yea I can’t generate any images

[dim cradle]

i stopped trying but yes was bad for me too

[sly spear]

It’s working again now

[dim cradle]

looks like on sora they added a trash button for images

easier to dispose them now. click, click, click

hmm replace archive i guess. so you still have to go to trash then and delete smh

[haughty spruce]

oh never noticed that thanks

[dim cradle]

i think its new

but no easy way to then clear out the trash that i can see. so still a pain

[raven raptor]

OH MY GOODNESS!! SOra Team! THank you !!!! its easier to delete now!

[uncut haven]

On PC, I just quickly press backspace twice. That will bring delete forever menu right away...

Without "empty trash", it's just archive renamed.

[open wagon]

what a weird take

[quiet brook]

apparently "suica penguin: vs icoca penguin" is violating content policies

even ran properly for once but second time it got blocked

also it keeps failing to generate 4 images

i see sudden decrease in quality

[rigid bough]

example?

[haughty spruce]

lemme try

@quiet brook i don't see much difference man

[uncut haven]

Does images still have yellow 'filters' for you? I see it's reduced or gone for some images, but I wonder if everyone seeing the same.

[haughty spruce]

yes, it's still there

especially in editing, remixing

[open wagon]

maybe you been using the image tool so much your eyes just started imagining drops in quality . visual burnout is real

[uncut haven]

Oh, okay. Then I guess one-shot images or images have explicit color prompt(like 'white interior') don't have yellow filter.

[open wagon]

actually mine stil lhas it

[dense musk]

How do you see how many likes your images got on Sora?

Also yes remixing goes yellow/darker

[uncut haven]

In ChatGPT, it seems still go to yellow/dull color as editing continues, but asking it different image 'resets' color.

I asked to adjust another generated image(in MJ), it gradually added yellow tint. However, drawing different image(character's another form) on the same chat also 'reset' colors.

[uncut mason]

Yeah, sometimes ChatGPT says it can't create images; sometimes it creates images, but not with 4o, but with DALLE; sometimes what you described happens; sometimes it says my request violate policies. Not a single successful attempt

[vapid elk]

im pretty sure it wont use dalle at all, unless you are on a custom GPT

[uncut mason]

I guess I’m pathfinder

[outer spade]

I think you have to specifically ask it to do it 🤔

[vapid elk]

im not sure about that, could be, tho

just tested, i asked for dalle, it used the 4o

dalle also wouldnt give the partial previews

[uncut mason]

Let me send you some screenshots, brand new ones

[outer spade]

language leak!!! (nvm you already have flags in your bio)

[vapid elk]

sure, feel free to DM

[vapid elk]

oh no, i have been exposed

[open wagon]

EU

[outer spade]

I've got you now!!!

[drowsy sigil]

correct. GPTs are not 4o and use dalle-3 image gen. 4o uses 4o image gen.

[dim cradle]

i wonder what do gpts use if not 4o?

[plucky hare]

GPTs use 4o, just not currently a version that uses the 4o image gen ability

[subtle merlin]

Where can I post a cool image I created of Chester the cheetah?

[haughty crown]

In image sharing communities

[ionic ore]

Do we know how to see how Sora changes the prompt for images?

[radiant geyser]

In Sora, how come my posts are not visible to public even though i have this setting checked to have my posts visible to the feed?

[drowsy sigil]

So no longer GPT 4 - Turbo? Is is 4o-turbo? It doesn't appear to be the same exact model as 4o (seprate from issue of images) as what you get in regular chat.

[plucky hare]

Custom GPTs have been on 4o for some time now I think, this article: https://help.openai.com/en/articles/8554397-creating-a-gpt was updated over 6 months ago and mentions at the top how GPTs use 4o. This article: https://help.openai.com/en/articles/8554407-gpts-faq is a little more up to date and talks about how rate limits for 4o apply to custom GPTs.

I don't think there's been a model called "4o-turbo". This is totally a guess on my part, but I'm guessing that GPTs continue to use the last version of 4o that ChatGPT used before 4o image gen launched.

[open wagon]

I don’t think sora changes your prompt as much as ChatGPT does

[coarse flint]

its gotten even more censored today, cant even generate images form even a day or 2 ago with the same prompt

[quiet brook]

also this will help ig #chatgpt-discussions message

[quiet brook]

same

#images-canvas message

[uncut mason]

Did someone encounter the issue? If so, how did you fix it?

[honest finch]

Please add 16:9 aspect ratio 🙏

[regal rose]

Hey, guys. Maybe somebody knows. I Have the "Plus" subscription and there said that I should have 2 parallel jobs for image generation. But I have only 1 and 200 generations per day

[honest finch]

did you select 2?

default is 1

[regal rose]

I've selected 2 images per generates, but there said about 2 parallel generations, 2 different propmpts you can push, one after another

[dusk nimbus]

Hey all - I am curious, do we know when the new 4o image generation will be part of the API? or its still a secret?

[regal rose]

suppose nobody know here 😦

[sharp palm]

Anyone taking prompt requests

[chilly gale]

same here.

[pseudo mica]

What in the world is going on with the censorship today

[wintry epoch]

Hi i tried making an picture of me but it makes it look like an other guy instead. Is it possible to make it close. I know there are multiple of celebrities so i know it works somewhat

[stone flume]

Very weird. It’s super censored today. Much more than before.

[late blade]

no new theme?

[coarse hornet]

Hey guys i'm new 🙂

[cunning stream]

trying to generate saiyans but anyting prompt to get around gets flagged so stupid

[coarse hornet]

I try to use AI well, learn and evolve in areas that I would like to try

[clear brook]

Are the devs just not listening to the complaints

[late blade]

anything you want to try with images?

[late blade]

are the users not reading the policies?

[late blade]

try rephrasing it so that it doesn't make a direct reference

[late blade]

maybe you haev an edge case, that may or may not render

[uncut mason]

Me

I'm inclined to think that users are randomly flagged 🤷‍♂️

[sharp palm]

Off-topic

[haughty spruce]

Sora is down?

[velvet rampart]

Or it is very, very… …very slow.

[frigid oracle]

is there anyway to complain about a "prompt violates our content policies". errors i have tried to generate images based on public domain fairy tales such as Cinderella,Snow White,Beauty and the Beast,Sleeping Beauty,The Three Little Pigs,The Gingerbread Man,Thumbelina,The Frog Princess, The Magic Mirror all of witch are public domain free to use fairy tales but every prompt keeps getting blocked by this stupid policy just becaiuse a big company uses a name does not mean they have the copyright to it

[late blade]

Disney copy right

Try referencing the year and source of your story

or used "inspired by"

[velvet rampart]

Ooh! This is something I solved with dalle3. Let me see my Sleeping Beauty prompt.

I tested the dalle3 and it worked on 4. The generated image with the prompt is on #images-canvas message

Yes, like @late blade said, avoiding Disney references is important here.

[late blade]

The mouse, even with the Snow White Flop, still has power

[velvet rampart]

Even the dalle3 prompt for Alice in the Wonderland worked #images-canvas message

[frigid oracle]

i have tried using inspired by and style of and a few other veriations its absolutly nuts because dienye does not own the copyright to any of them they own the copyright to the way the charicters look in there versions thats it

[velvet rampart]

Please have a look on the image canvas channel how I have done this. An option is to generate first on dalle custom gpt and then on 4o.

[late blade]

not as copyright, but as corporate identity

which is where Disney still has a strong stance

and also, while the stories are public domain, the images themselves are still made by disney

so those images are still copyright of disney

[dim cradle]

yes i imagine any story that has been done by disney 99 percent of images to draw from are going to be the disney version

[dim cradle]

Guys, if I have tbh with all of you, the content policies is really starting to make me feel uncomfortable

[late blade]

Why, because of how laws work?

[dim cradle]

I guess so, but it's just sucks that I can't even do anything that I want, because of that

[late blade]

copyright, trademark and IP laws, nothing OpenAI can change

[dim cradle]

yeah i dont get why people struggle to understand that. disney, warner brothers who own batman, harry potters, game of thrones, superman, they have more money than god to hire legal teams

so of course open ai has to play on the safe side. its not an ip maker, its an image maker. use some imaginations

and it seem a lot of ip stuff is possible, just not a lot of the major things, because they are most likely to cause problems you know

[open wagon]

U either can go find other alternative for image generation tool,

[dim cradle]

I'm well aware, & I ofc mean with everything & with utmost respect about this. But looks like I'll have to try to deal with it

[dim cradle]

Oh, no no it's fine, but thanks for the suggestion. But I'll just try to deal with it

I wish it was less limited too. but its just the way it is. money is involved and theres no beating that.

but it's also a lot less restric than dalle3 was on gpt, so I really do not complain much about it myself

Just my guess and I say this before, but like openai make deals with some newsapers/publisher things, I can see them make a deal with some of these ip eventually and then they will be allow but 🤷

i think some ip might also realize allowing them to be use in a big model like this is to their advantage, free publicity, so long as it does not put their characters in bad positions

[open wagon]

I still wonder if Studio Ghibli’s gonna say anything about this, especially since Miyazaki’s supposedly super against AI (correct me if I’m wrong though, not sure if he ever actually said it).

[dim cradle]

Like there are some things in the 4o Image generation, that can make me generate, like some things can be used, & some things cannot be used.

Like for example, there's an artist that I used, & it completely worked for some times.

But when I tried another artist, or whatever that I'm trying to do for my art((Like for inspired, or something)) it just doesn't allow it.

But I'll try to make it work, since I had it to work, when I was doing my AI art

[open wagon]

Also @dim cradle I feel you , the censorship on the 4o image_gen tool is lowkey frustrating. It’s weirdly inconsistent sometimes.

[dim cradle]

well i read studio ghibli got so much google searches from it all it led to even more dvd sales. thats why i could see others allowing their ip.

example i use before, a new superman movie is coming out this july. if i was warner brothers, i would not only allow superman generations, but i would give updated images for the model to use that are from the movie

free publicity

they already do this kind of thing -- on twitter they will make emojis based on new movies or stories or thing. just need someone to take that idea to AI

[mossy ingot]

I don't know if you're using ChatGPT or Sora for your creative process, but in Sora you can upload reference images of an artist that you really like, and then tell Sora to copy the style in the uploaded images so you don't have to reference famous names. In my experience, Sora has far less checks than ChatGPT, but that doesn't mean there's complete freedom. Artists can't copyright styles or otherwise nobody would ever be able to make art again, but they can copyright their original works.

This is what we learned in college when I studied and got a degree in graphic design when it comes to copyright:

Ideas vs. Expressions: What Exactly Can You Copyright?

One of the most fundamental concepts in copyright law is the distinction between ideas and expressions. Let’s clear up any confusion:

• Ideas themselves cannot be copyrighted. For example, the general concept of painting a landscape is not protected.
• However, your unique expression of that idea—your specific painting of a landscape—is covered by copyright.

As Bo explained, "Copyright applies only to expressions, not to the underlying idea that you've come up with. The idea of painting the Golden Gate Bridge from underneath is not intellectual property that could be claimed by anyone."

To avoid infringing on another artist's work that you have seen, make sure your creation is not substantially similar to their expression of the same idea. Put your own spin on it!

This ideally should also apply to AI but I think OpenAI is playing it safe and this area with AI is still a little rocky legally.

[late blade]

The difference here is:

Data has copyrighted, trademakred and IP material. Responsible Data Science prior to expressionism, that has priority. Get the label straight.

[mossy ingot]

My labels are correct according to United States copyright laws.
As for your point about trained data, refer to, this area with AI is still a little rocky legally.

In January 2025, the U.S. Copyright Office released Part 2 of its report, Copyright and Artificial Intelligence: Copyrightability (“the 2025 Report”) providing a detailed legal and policy analysis of how copyright law applies to AI-generated content.[1] Part 2 builds on foundational principles of copyright law, reaffirming that human authorship remains the cornerstone of copyright protection in the United States.[2] It provides critical guidance on the conditions under which AI-assisted works may qualify for copyright, clarifying the legal boundaries between human creativity and automated generation.[3]

Practical Challenges in Applying The 2025 Report’s Framework

A key takeaway from The 2025 Report is the Office’s categorical rejection of copyright protection for works generated solely by AI, reinforcing the long-established principle that copyright law protects only “original works of authorship” created by humans.[4] The 2025 Report reiterates that AI-generated outputs, absent meaningful human creative input, lack the necessary authorship required for protection under the Copyright Act. This conclusion aligns with existing case law and administrative decisions, including recent Copyright Office rulings denying registration for purely AI-generated works.[5]

Beyond addressing fully AI-generated outputs, the 2025 Report also examines hybrid authorship scenarios, where AI tools assist human creators. The Copyright Office emphasizes that for a work to qualify for protection, creative human involvement must be substantial, demonstrable, and independently copyrightable. The mere use of AI does not preclude copyright eligibility, but the human contribution must extend beyond basic prompts or trivial modifications.

I imagine, but don't know for sure, this depends on the country as well.

Not sure if you're speaking law, morality, etc. Might help to define your ideas here.

[late blade]

then you are also aware thant right holders can enforce content policy to protect their rights on OpenAI

[mossy ingot]

It's worth reading what I wrote.

[late blade]

I know that in and out

a report is not law

[mossy ingot]

Then why debate what what I'm stating as fact?

The laws aren't settled.

Hence, the reasoning for the report.

[late blade]

exaxtly

the current laws are still in place

so have to abide those laws and not wishful rporting

[mossy ingot]

the U.S. Copyright Office

[late blade]

yes report

[mossy ingot]

Are you even reading what I'm sending?

[late blade]

"... released Part 2 of its report..."

[mossy ingot]

What do you think the goal of that report is to do?

[late blade]

assess

[mossy ingot]

Because further reading would suggest that it touches on how data through generative AI aligns with or conflicts with current United States copyright laws.

[late blade]

you are not seeing the whole picture

[mossy ingot]

Your stance isn't in accordance to United States law. It's that simple.

[late blade]

you are just in the defensive of wanting to do something

I'm telling you, right holders have right to block and enforce their interests on OpenAI by law

[mossy ingot]

You quoted me claiming my labels are incorrect (they're not), followed by talking about trained data, which I just showed you why your stance isn't actually the current legal take on generative images through AI models.

It's very simple:

A key takeaway from The 2025 Report is the Office’s categorical rejection of copyright protection for works generated solely by AI, reinforcing the long-established principle that copyright law protects only “original works of authorship” created by humans.[4] The 2025 Report reiterates that AI-generated outputs, absent meaningful human creative input, lack the necessary authorship required for protection under the Copyright Act.

[late blade]

sure, get a grounded holistic view, then set the label straight, it's always the argument I can't generate this or that and I'm right

typical black or white mentality

[mossy ingot]

This takes on two forms:

1. It's hard to make artwork generated by AI copyrighted because it has no human author.
2. If AI isn't considered a human author and companies claim the trained data is fair use (might not be depending on the case), it becomes hard to strike it as copyright infringement. The courts are figuring out how to navigate this and the laws will likely evolve over time.

Hence, the report explores how laws around AI models should be refined or their limitations.

[late blade]

"how laws around AI models should be refined"

[verbal sequoia]

Get notified for server events!

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

[mossy ingot]

Refined, meaning this...

[late blade]

not acting in place laws

[mossy ingot]

...and this area with AI is still a little rocky legally. — Franco

But go on, I suppose. 😂

[late blade]

sure, when you get the ego boos rush out, and decide to see the holistic manner, then come back with an argument, not the white or black mentality of "I couldn't generate something, I am right" crap

[mossy ingot]

I don't even know what you're on about, tbh. This has nothing to do with ego and me showing you current laws and problems with copyright even when it comes to trained data on copyrighted material.

[late blade]

use your text, conduct a sentiment analytics with GPT 🤗

[mossy ingot]

If attacking my character makes you feel more secure about your stance, then I suppose have at it. No personal harm on my end.

[late blade]

I'm not attacking anyone, I'm using strong wording for statement accentuation

[mossy ingot]

You're reducing my argument down to ego, which is weird that you bring tension into a civil discussion to begin with. If anything, you came into the discussion hot and egotistical by telling me to correct my labels and insert your ill informed take on United State laws surrounding AI and copyright.

[late blade]

play it smart on the long run

[mossy ingot]

With that said, I'm done responding to you. This discussion is unproductive.

[late blade]

fair, I didn't gain a good answer

[open wagon]

Dang

[late blade]

?

[open wagon]

nothing, i just finished reading everything

[late blade]

ah text based communications, don't sweat it

half the message is lost in missinterpration of tone anyway

franco's point of view being from a framework that should be in place, mine from practical current framework of nonsense, one we have to navigate

and discussions that matter don't necessarly have to be all happy and dandy

it's in the name of the channel too "images-discussions" not "images-sweet-fluffy-talk"

[verbal sequoia]

<:book_icon:1171408210398289941> Rule 3: Stay on topic.

-# Be mindful of what other users in a channel might find helpful or interesting when posting. Stay on topic in order to keep conversations focused and productive.

-# Consider posting in #off-topic or an appropriate channel.

[solar saffron]

Yesterday’s spark gave art its start,
All from a prompt that played its part.
Hope you like it, straight from the heart.

https://sora.com/g/gen_01jrtv7q6zfmft1xrwenmdc94w

[toxic jolt]

Hey 👋
Do you use GPTs to help on image generation ?

[late blade]

CustomGPTs? not really

[toxic jolt]

I was thinking it could be better than it is already. But idk how to tweak it. Maybe someone already have, I was wondering

[late blade]

oh there are implementations of customgpts for that

Zero-Shot for example, but I don't think they have the new image generations enabled yet?

[toxic jolt]

That would be cool to try

[late blade]

Zero-Shot has been quite popular around here, so sure, give it a try

[late blade]

@vapid elk channel #daily-theme is not reflecting the current topic on channel and info, was kinda misldeaing just realized it

Still showing "🧭 Shifting North – when direction changes without warning, adaptability takes the lead, finding new paths in the unknown"

[vapid elk]

thx for letting me know

[final kite]

So i m trying to get accurate medieval soldier but even when i describe everything detailed or send gpt reference images. It always lools alooof

Any tips :( ? Or is it not doable with the image generation

I feel like the GPT Image generator is made for modern image content and mainstream topic

[dim cradle]

post the prompt you use maybe some can help more that way

[rigid bough]

meet me in #images-canvas

[final kite]

Prompt: Please create as image: A group of four heavily armored Byzantine kataphraktoi stands in a tight diamond formation on an open plain, viewed slightly from above. Each rider is clad in full lamellar and scale armor, with their horses equally protected by barding of bronze and leather plates. The riders wear richly adorned helmets with nasal guards and chainmail aventails, some featuring small crests or plumes.
They carry long kontarion lances angled slightly upward, and round or oval Byzantine shields marked with imperial symbols or Christian crosses strapped to their backs. The horses are strong, disciplined, and clad in decorated armor matching their riders — leather reins and saddlecloths featuring purple and crimson embroidery.
The landscape below them is a sunlit steppe or dry field, dotted with sparse vegetation and distant low hills. The sky is clear but dramatic, with light breaking through clouds, casting golden highlights on their armor.
Perspective: slightly elevated, angled down to capture their formation and full figures of both riders and horses.
Art style: oil painting mixed with digital concept art — earthy textures, naturalistic tones, dramatic skies, and a grounded, historical atmosphere.

[verbal coral]

Remix really blows for small fixes. 😦

[verbal coral]

What does "lools alooof" mean?

[final kite]

Hi meant look aloof :)

[rigid bough]

Please Open AI, put image Library into the macOS Desktop App 🙏

[late blade]

that will come 🥰

[crimson heron]

I just hope they add library management features similar to what they had on the old DALL-E site. Having hundreds or thousands of images to navigate is no fun.

[late blade]

if it's anything like Sora, it's already included

[rigid bough]

Library needs a delete button!

[lilac topaz]

Hii pleose

Can someone help me I am confused

[rigid bough]

what's up?

[lilac topaz]

Hii

I sent this in #1155772063596953642 I cant send this here

[rigid bough]

sorry I won't help you make those kinds of images

[lilac topaz]

Okay 👍 do you know who can ?

I spent today hour to make this style of images but I cant I dont know

[unborn patio]

Hey guys, why i cant generate Images with Sora which has Famous Names? Everytime a Error shows up with the Privacy. But i see often Photos of Famous People.

[rigid bough]

who are you trying to generate?

[unborn patio]

i wanted to make a Fan Art wallpaper of Spider Man

but it doesnt work

[rigid bough]

you can't do Intellectual Property or Franchises

[dim cradle]

and famous people have an opt out option. so you might be able to do them, but then later maybe not if they opt out. its all a roll of the dice

[unborn patio]

thanks for the Info. Another question, how can i see likes of my published Photos? 🙂

[dim cradle]

that i do not know but maybe someone else does.

[mossy ingot]

https://sora.com/explore?user=dopefungi

Replace dopefungi in the url with your username. It should show you your published images and likes (if any).

[deft musk]

Woo, new helpfile about the image library! https://help.openai.com/en/articles/11084440-chatgpt-image-library

[uncut mason]

It fixed itself after 598th cookie and cache clearing after logging out

Excruciating experience

[untold tide]

how I can create an image in JPG format? Please help me. My image output is in PNG format.

[velvet rampart]

Is Sora slow this morning to others? Some images are quite quickly done but for others it takes a long time even to start generating.

[ancient lava]

Does the AI hate seem forced to you all too? Like the comments under pics “ew, ai cringe” “ai slop”. It’s like hating ai is the new “edgy” thing to do.

[dim cradle]

normal for me at least right now

[uncut mason]

Does anyone know how to force it to be consistent? It’s pretty good at following instructions, but not when it comes to edit an already existing picture, instead of requested slight tweaks it changes an image dramatically

[fair tendon]

Trying a new chat (it can be stubborn in your existing chat) and uploading more reference image somehow works

[cloud nova]

how to minimize typos in openai's generate image? I use it to generate infographics, but there are still many typos.

[open wagon]

expand your prompt more by uhh input the text?

idk i think someone here actually demonstrate how to correctly generate images with lots of text , and it was an image of wikipedia article generated using sora image_gen

[uncut mason]

I think we are lacking prompt engineering channel here

[rigid bough]

#prompt-engineering

[uncut mason]

Is it for images as well?

[rigid bough]

ask @shadow bane but I don't see why not

most people just use #images-discussions

[rigid yoke]

hi is the image generator down for everyone as well? I presume its because of the new update rolled out. Image generation reverts to the old dalle model and I don't see the drop down for it either. Like the create image section is gone

[velvet rampart]

On ChatGPT? Sora is working.

[rigid yoke]

Correct on ChatGPT.

[velvet rampart]

Testing...

Works for me.

[uncut mason]

Log out, clear cookie and cache, log in

Maybe you'll be lucky on 500th try

[rigid yoke]

hahaha thanks to @open wagon i found a work around. I just disabled DALL-E. it seems to revert to the new image creation function

at least the outputs are substantially better than what I was getting with DALL-E turned on

[slim heron]

the image generator is not working for me. It's using the old model. This did not work. It just said it was disabled and could not create the image.

[slim heron]

trying this out to see if it fixes.

[rigid yoke]

As an alternative, you can use the /image command in the chat to create an image. Simply type /image followed by your description of the image you'd like to generate. For example:

/image A serene landscape with mountains and a lake at sunset.

[umbral sundial]

sora quite slow for me atm

[velvet rampart]

It has been slow on and off whole day for me.

[slim heron]

yeah, but the image generation is really bad right now. The images are looking like the first version of GenAi models.

[rigid bough]

really? I am loving Image Gen and how detailed and realistic images look

[dim cradle]

probably just luck but my sora conversion of image to video has been use the image for the whole video. and really, when it works, it is as good as veo2 imo

[slim heron]

yeah, i can show you. It looks terrible

[rigid bough]

post it into #images-canvas

[slim heron]

take a look

[rigid bough]

yeah I just saw

[slim heron]

i tried the other solutions of deleting cookies and cache for so many times now. haha

i'll try more later, got tired of it

[bold inlet]

Give us 16:9

[late blade]

Question:
"Have you noticed how prompts tend to get classified and labeled in a pre-determined professional tone—almost like it's baked into the training process? Especially with image generation, it’s becoming more obvious. The newer models often default to placing a blurry, generic background behind the concept, then layer the subject on top with this weirdly artificial lighting. Shadows don’t even match the light source properly, so the result looks like a stack of disconnected visual elements—three or four layers deep—but not actually blended. It ends up feeling detached, not seamless."

Any thoughts?

[lilac topaz]

I have problem how have this image in horizontal what to write in prompt

[dim cradle]

"widescreen" or "square image format"

[lilac topaz]

I am choosing ratio 9:16 so use widescreen ?

[dim cradle]

yes that should work in gpt

if sora change to 3:2 or 1:1

[late blade]

new wide aspect ratio images from GPT are 1536×1024 if you ask for wide aspect ratio.

[uncut mason]

Can't agree

I haven't noticed anything like this

Does memory affect image generation in any way?

[lilac topaz]

I am using dalle 3

still making in wide I need vertical picture in 9:16

[late blade]

Vertical is 1024x1792, ask that directly to the model. 1024x1024 for square ones and 1972x1024 for the wide aspect ratio.

replace with 1536 for the newer generation, and consider that GPT still thinks 1792 is in place instead of 1536

[lilac topaz]

I am trying now

It seems there were issues generating the images, and unfortunately, no image was created this time. Please feel free to submit a new request, and I'll be happy to assist you further!

[late blade]

In theory a custom aspect ratio that gives = 1.5 should be possible with the newer generations within the max amount of pixels set by the output

[lilac topaz]

doesnt work

i dont know

stil lwritting me this I encountered issues while attempting to generate the image you requested, and unfortunately, I wasn't able to create it this time. If you'd like, feel free to provide a new description or modify your request, and I'll try again!

this prompt: Hyperealistic, digital art, neutral expression, The background is a pink A glowing, vibrant girl touch her skin, size image 1024x1792

[late blade]

which subscription do you have?

[lilac topaz]

plus

[late blade]

Also, the wording is not precise, is the girl touching her skin, is the light touching her skin?

so it's not clear for the model what is happening there and gets block,

[lilac topaz]

which model is the best ?

I am paying for something and this doesn´t working even

[late blade]

When I mean the model, I mean image generation with GPT-4o

[lilac topaz]

I am using Dalle

[late blade]

If you are using Plus right now, do you get a message that the new image generation will be available soon? or do you get the image generation with an animation of how it's being generated?

[lilac topaz]

with animation

[late blade]

then you are using the new model and not DALL-E

[lilac topaz]

I have chat where is DALL-E

mix color icon

[late blade]

Do you get a message with

Made with the old version of image generation. New images coming soon.

Yes or no?

[hollow jetty]

wth is o4 😭

i thought they were still making o3

[late blade]

o3 and o3 pro probably and o4-mini as the next iteration of the models already in place

[hollow jetty]

will o3 regular replace o1?

[late blade]

educated guess, probably on parallel for a while, specially after 4.1 is out and 4.5 is being phased out on the API

@lilac topaz your prompt generates correctly with the right punctuation, but not suitable to post an image of that in this discord server

[lilac topaz]

I asked one guy maybe I found the answer

[late blade]

❤️‍🔥 4o-mini for images, that's gonna be so much fun

Testing right now

[dim cradle]

how? and how do they seem

[lilac topaz]

is broken

I encountered issues while attempting to generate the image you requested, and unfortunately, it didn't come through. If you could provide any adjustments or another request, I’d be happy to try again!

I dont know why )

but I am sitting for 3 hoursi n this and I can´t do nothing

I paid for something and even doesnt work midjourney I think is way better

[late blade]

just posted on #daily-theme first zero-shot concept

[dim cradle]

so its o4 mini that does it?

can you explain why its better than 4o? i am not a pro on this technical stuffs

[late blade]

I don't know if it's better yet

[dim cradle]

gotcha

[lilac topaz]

how many

I can generate images with plus

[late blade]

it's snappy, but first run already losing details of the main concept after a few interactions

6 turns, and lost important details

7th turn, added non-requested aspects

[dim cradle]

seems like using o4 mini high is best tailor for the images. i assume that is what you are doing, i am just make a note of it here

[late blade]

it's just a zero-shot first take, wouldn't be able to tell yet

[dim cradle]

i just got in on the app. it says o4 mini high is best for visual stuffs

'best at visual reasoning'

[late blade]

visual reasoning, as shown in the image is more for GPT-Vision modality

haven't done that yet

and also, for multi-modal input reasoning

[dim cradle]

pretty cool new model

[late blade]

it is

[velvet rampart]

Are these new models? o4-mini and it cousin high.

[late blade]

weed high... no... don't think so....

joke aside, o4 mini is now available

and o4 rolling out

[dim cradle]

i am having o4 mini high recreate some photos of mine and it does a real incredible job.

[late blade]

yes, it's really good at that

[dim cradle]

this might make sora site feel like dalle3 to me now haha

[late blade]

it depends on the Use Case

[dim cradle]

yes, i do not want to overhype, i only have it for 10 minutes lol

but wow, i love what i see so far!

[late blade]

Sora is really nice to see iterations and remixes of concepts

but concept creation with GPT and brainstorming it, way better

for how I work anyway

[velvet rampart]

The o4 mini high with 4o image generation together - wow! Just run my kobold transformation protocol, its result is excellent.

[dim cradle]

nice!

[velvet rampart]

#images-canvas message, NB. I can’t include the human photo as I have found it randomly in the net.

[late blade]

o3 is super nice and snappy also

[velvet rampart]

Somehow I feel that I have to shout to the idiot savant 4o what I want, so that it gets, but now? Nope. Yes, 4o has become less idiot in a year, but sometimes it still is.

[dim cradle]

it does images too?

[velvet rampart]

They have removed o1.

[late blade]

o1 pro mode, legacy reasoning expert, is still on the menu

[late blade]

yes

[velvet rampart]

These models come and go much quicker than before. Earlier it was a huge hype that a new one out. Now? It is more like ”How many did they publish this month?”

[dim cradle]

lol yes

and gpt5 in a few month sound it may replace quite a few from what i am reading anyways

[velvet rampart]

If and when they get it out.

[dim cradle]

altman say a few months but yes we shall see

[late blade]

Don't expect many advancing jumps in 2025, most will be prob next year, for now mostly performance and benchmarking is probably what's gonna be happening

4o-mini-high is a good balance

Contextual, o4-mini and o4-mini-high, still not quite correct

Specially with a test I do in which I provide a set of conditions to keep working from there

[velvet rampart]

Yes. I noticed.

[late blade]

contextual discrepancies of images and texts are also quite interesting, in chatgpt I get a character, when using the text requested in sora, I get a living room

[lethal condor]

e

[verbal coral]

Is it just me or is the yellowing somehow getting worse?

[verbal coral]

I don't know what that means

[late blade]

wdym

[verbal coral]

Like, the yellow/orange 3000-4000K incandescence almost every image result has.

You haven't noticed the white balance is almost always wrong?

[late blade]

have you specified the white temperature?

Interesting idea, worth checking

[verbal coral]

Yeah, I developed a whole concept around it.

[thick fulcrum]

Anyone been using ChatGPT image creator lately?
I feel the quality now is way less than when it was first released

[late blade]

I'll check a few images, let me get some stuff set, and I'll get back at you in a few

[late blade]

Did a naive approach to analyze, 15 random images of a set of 1000, to check what you are saying. I'll show on #images-canvas
@quartz vale check that we can use images here, I wanted to post graphs for a discussion about images in an images channel...

[polar light]

Do yall ever use sora directly for image generation rather than gpt?

[late blade]

Depends on the Use Case

[polar light]

Is there a difference in doing it in gpt vs sora besides bypassing the way too restricting filters?

Does sora still have gpts reasoning and all that, or allow image references

[late blade]

In my case I use chatgpt for narrative driven images, while with sora I use zero-shot like concepts done outside of AI

[quartz vale]

images are disabled in this channel as it's focused around discussions specifically. Feel free to use #images-canvas to share images

[verbal coral]

Literally all the time.
I rarely use ChatGPT for it

[late blade]

I do not think it's reasonable though, if you are having a discussion to discuss aspects happening with image generations, you need images to point out the aspects in question for a discussion, it's different than API or Dev channels. I won't question moderation, just a teaser to think about it.

[polar light]

Is there any difference in how you create the image?

I've not used sora before or seen the interface

[verbal coral]

Well, you don't have to say "create an image", for a start.
Crucially, though - longer conversations in GPT (where you generate multiple images) use your prior messages and generations to influence the final result, even if you tell it not to. It's like a long chain of remixes, and that can really screw up changes of direction in a project. Plus, remixes of remixes get steadily less stable and more faded/yellowish, which is fine - if you like that sort of thing.

[polar light]

But the way you talk to it and edit it with normal speech is all the same?

[polar light]

That's a good amount of info ty!

[dim cradle]

im wondering of o4mh is better at this. im not sure but i have try a few images in the same convo and it seem to not bleed as much as 4o does but 🤷

[mossy ingot]

Yes, although Sora lacks the back-and-forth discussion you'd expect on ChatGPT, it still uses natural language to interpret and generate your image. That said, I’d avoid remixing (in-painting) with Sora, as I find the quality tends to get worse each time and the image becomes darker. It’s better to just edit your prompt and resubmit the generation. Additionally, Sora does understand the context of uploaded reference images if you discuss them in your prompt. There are also some extra bonuses, like the use of presets and the ability to generate more than one variation of an image at a time (currently limited to two variations per generation for Plus users).

I think you get around 200 generations per day on Sora for Plus users, and it doesn't cut into your ChatGPT limits, so that's also a plus.

[dim cradle]

yes, remix is a great idea but asking for a change with it unforunate make the image fidelity worse. a shame. but something to improve in the future i hope 🙏

[mossy ingot]

^I hope for the same. When you get that one image that is nearly perfect but just needs one small change. 💀

[verbal coral]

I will be doing the same. 🙂

[verbal coral]

Yes, I quickly learned that Remix in Sora is garbage, unfortunately - even if you select zones. (This seems to do nothing.)

[mossy ingot]

...and sometimes happens to make the part you wanted edited worst than what it was.

[dim cradle]

yes i do not even understand why they allow you to edit an area, then it still modified the entire image. its like they are troll us haha

[late blade]

@dim cradle try multimodal inputs for image generations (video, audio, text with natural language and code) o4 models 💞💞💞

[dim cradle]

we've come a long way from the dalle3 days. and those were not too long ago 😂 imagine the next version... wow

[verbal coral]

I'd like to think I can imagine it, but I've been blown away every time this tech makes a leap forward. I pretty much abandoned Midjourney overnight.

I suspect this is the case, but:
Does the prompter for Sora (images) understand commands like:
Use Image 1 for character pose.
Use Image 2 for facial details only.?

[fair drift]

I don't think you can have different prompts/composition in the same batch.

Batches of images all use the same prompt, just different seeds.

[verbal coral]

I know, but it does seem to accept when I refer to the images in sequence like that.
I'll show you what I mean in #images-canvas

[fair drift]

@verbal coral Oh that's very interesting... is that repeatable, consistently?

[slate garden]

Exactly, how can any image be discussed if not shown? Isn’t that the idea behind this channel?

@quartz vale

[fair drift]

Similarly... I'm wondering what the difference is between #images-canvas and #sora-reels since Sora does images now and DALL-E is basically history?

Where is the correct place for Sora image discussion? 🤔

[dim cradle]

i remember a few months ago some mod i never even seen anymore and cant remember they name, would come into the dalle3 chat just to tell us to post images in the canvas or whatever and not discussion. nobody did because it was not convienent. but apparently he got his way. nice to let someone who does not contribute anything to discussions decide how they should go i guess smh

[fair drift]

So... what's stopping us from also having discussions in the #images-canvas?

[dim cradle]

I think it is shifting that way now

but still bounce between rooms. before there was a dalle3 chat room, people would post images, discuss the image there, discuss general image news etc. it was too convienent i guess

[slate garden]

well I figure reels means videos, unlike still images

that separetion makes sense

[fair drift]

True, but images are posted there too, it just threw me off for a sec!

[slate garden]

Which sounds like....bad moderation! which can be easily automated

perhaps its a fresh channel yet to be properly auto-modded

[slate garden]

Well I don't know about how involved he is/was here but still, it's pretty basic logic

[dim cradle]

it was like that forever too, like 2 years or however long dalle3 was out. then when it was just like 5 of us posting there barely anymore he came in and demand we dont post images in the dalle3 discussion. it was weird. like i said, he did not even contribute. but lots of people have that inner authoritarian and just want to have power and boss around even on discord

[fair drift]

Power tripping is a helluva drug...

[slate garden]

that's so odd I'm here for only 3-4 days and can't understand how did you discuss imagery for 2 years

[dim cradle]

well people always making new images. dystopia and milamber and a guy name hawaiinz mainly were he group haha. but it was of course busiest when dalle3 was new

which is why it was so strange to have some mod nobody knew come demand when it was a ghost town discussion to not post images in the dalle3 chat

all it did really was kill the chat even more lol

[slate garden]

Yes I imagine it used to be a complete spam-fest even with slow mode, I was delighted to find a calm official space for discussion

[dim cradle]

when 4o first come out these rooms were busy 24/7

[fair drift]

I think... it's just reversed, yeah? Like...

#images-discussions should be #images-textchat
and
#images-canvas should be #images-discussions or #images-general or something

[slate garden]

well then his manner of modding makes sense for that time perhaps

[slate garden]

Now there's finally room for discussions

loosen up mods!

[inland reef]

did image creation get a huge decrease in quality today

[slate garden]

Giv'ataim

[slate garden]

it's been a weird day indeed

[dim cradle]

seem the opposite to me. on o4mini and o3 at least they look even better

[slate garden]

wait... you guys can generate on o3 and o4mini???

[fair drift]

Video gen was slow for me today earlier (like 30 minutes to generate 5s/720p) but seems to be getting better now. Images have been alright throughout the day.

[dim cradle]

yep. and they think a bit before they make the image. o4 mini even say it is best for visual reasoning

obviously 4o was already the game changer, and i presume is still the same model, just with more thought and attention to prompt and thing in o4 mini high (not mini my bad) and o3

[slate garden]

aren't they just a waste of time if one just wants image-gen? it just reasons for some time... perhaps it comes up with smarter self-prompts though? is there a noticable difference?

[slate garden]

"visual reasoning" could also mean visual understanding, rather

[dim cradle]

it seems to me there is even better quality with some things in o4mh, but like i said, 4o was already so good it not like night and day change. but they advertise it as best at visual reasoning so there must be something there

yes guy it could

of course openai is always light on the details for us

[slate garden]

I think its eventually the same DALL-E 3 for all versions so better reasoning could be beneficial for more complex requests or for better understanding of "newbies" lingo

[dim cradle]

well its not dalle3 but yes that possible

[slate garden]

arent we on 3? also- watching the seconds tick while reasoning makes it feel 4X slower overall

[dim cradle]

its 4o image maker. dalle3 was last gen

can run you own test of course, make the same prompt in sora and then make it in chatgpt with o4mini high or o3

[slate garden]

oh its 4 lol.

[verbal coral]

Free users still have to use Dall-E

[dim cradle]

yes

[fair drift]

Free users can use Sora (limited) as well

i.e. "sora.com"

[verbal coral]

I just mean within ChatGPT you won't get Sora unless you're Plussing it up

I'm enjoying the main o3 model so much I might actually Pro

[slate garden]

I hate that.

mobile? forget about it.

[verbal coral]

but will I be able to go back?

[tender aspen]

Why has my Create Image feature disappeared? I'm on chatgpt 4o on Plus plan.

[verbal coral]

Yeah mine went too... weird, huh?

[dim cradle]

good question. i notice that too 🤷

[slate garden]

I swear I just noticed that too

[tender aspen]

did they remove the feature?

[slate garden]

temporarily I guess

to lower demand

[dim cradle]

who knows

[slate garden]

many issues rn

[tender aspen]

screw this lol

[dim cradle]

like i say, openai is always light on the detail and rare for them to explain they self

[slate garden]

you can still generate there's just no clickable button or listing that says "generate image"

[fair drift]

"Sora, make me an image of Sam Altman panicking as GPUs melt in server racks behind him." 🤣

[slate garden]

copyrighted!

[tender aspen]

Do I need to type "Create Image" or some /prompt thing in front of my instructions?

[dim cradle]

yes, or make image of... or just 35mm quality image... etc

[fair drift]

For ChatGPT, I always start mine with "Use image gen to create ..."

pretty foolproof.

[slate garden]

you can also use draw or whatever lol

[tender aspen]

works!

[slate garden]

https://sora.chatgpt.com/g/gen_01jrwz38wqfex9wyzj32d77q6d
made it yesterday

300 likes! woohoo

[dim cradle]

nice!

[fair drift]

hey no images here!!! /s 😉

[dim cradle]

mods!

[slate garden]

hehe bypassed

[dim cradle]

wheres the admin! 😡

[slate garden]

oh no

[fair drift]

"Get 'em, boys!"

[dim cradle]

i dont come to a chat room with images in the name to see images, tyvm!

[slate garden]

Wow look at that!
Unlike #images-discussions , #sora-discussions has literal Sora submissions for discussions!
How odd... I guess its a matter of time until they get censored too...

[dim cradle]

most of the time its just us chickens. i think there is also a rule to every discord -- you must make 50 times the channels you actually need so things are inconvenient and spread out as possible

[slate garden]

This server needs so much optimization.... 3-4 channels per subject.... god can't 4o4 mini take a sec and orgenize this mess

[dim cradle]

my favorite is openai chatter and chatgpt discussion are always the same discussion going on in two channels

new model release? well pick either to discuss

[slate garden]

yes lol ive went through 4-5 channels with the same discussions rn

[dim cradle]

i'm surprise image here doesnt have "photo real image discussion" "3d image discussion" "oil painting image disucssion" "surreal painting image discussion"

i guess i should not give ideas 😂

[slate garden]

The category mess is killing me

[mossy ingot]

I've hidden over half of the channels in this server so I can get where I wanna be.

[verbal coral]

I wouldn't even mind if there were 4-5 channels of the same thing in a server if there was high demand, but don't pretend they're different subjects...

also, weirdly, my emoji from other servers don't work here but do work in #images-canvas

[mossy ingot]

You can't react in certain channels even outside of external emojis. 😂

Sometimes I wan't to acknowledge something somebody said and not respond in text.

[verbal coral]

Emojis make me feel comfortable. I don't like to be misread

[fair drift]

😏 👌 ... 🤣

[verbal coral]

🔥
🧠

Sometimes the bot really wants to force art on you, doesn't it. Like, paintings/digital art instead of photography - even after prompting it out harshly

[fresh meteor]

istg the quality for anime has become so bad, that i rather used stable diffusion then this deteriorated quality. it was good when it was released now i can’t generate the same quality anymore

[verbal coral]

Wonder what changed?

[slate garden]

yeah Discord server management is like coding a whole social network and I don't think too many funds go to the higher ups here

[fresh meteor]

i posted in #images-canvas

left is before

right is now

[dim cradle]

when can we generate 9:16 images?

it's much more popular and used than 2:3

[pseudo hollow]

I have started to notice that ChatGPT image gen produces a very similar style now for any illustrated images. always the beige weathered background, with off white border, always the same font for text captions, etc. Not complaining, just very noticable now when I see an illustrated image produced by ChatGPT.

[late blade]

Not really, I can show you examples, but considering it's a hazzle to jump between channels to keep up a discussion about this kind of topics. I'll just say it and hope you believe me

[echo trail]

how many image generations approximately do you get per day with a $20 sub?

[rocky grove]

So, now the guidelines are alot more strict and I cant even generate images for cherubs in warhammer 40k anymore. dafuq? So annoying to pay this much for pro and constantly have things break or be told no for the weirdest reasons. Why should I even bother anymore with this junk when Google seems to be outpacing and less limiting than this

This crap needs an age verification if thats the case for these god aweful guardrails keeping me soooo safe from a fictional creature generated image 💀 So lame

[verbal coral]

Sure it's not copyright issues?

[verbal coral]

It's 200.

Does anyone know what the token limit is? I mean the point at which complexity starts to break things and cause problems

[dim cradle]

not sure but i have put in some long prompt and it does quite well

[verbal coral]

Yeah it does

[verbal sequoia]

Get notified for server events!

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

[verbal coral]

But after a while of adding details, it seems to start doing random things

[dim cradle]

yes. I think still as concise as you can be is best

[jolly niche]

@zealous pulsar I do have the prompt for this: #images-canvas message

I found a prompt in Spanish for a Dracula one and asked ChatGPT if it would understand if I wrote it in Spanish and English, and it did!!! (I was very pleased about that!)

I am new here and not sure of the policies on posting prompts. Is it OK to do it here, or in another channel?

@mental belfry Clothing no longer skimpy? Is is more censored? Or someothing with the text?

#images-canvas message

[mental belfry]

Oh. This is for chatting about the images?

[jolly niche]

I suppose so.

[mental belfry]

Look at the bottom of each image. On the right.

I wrote a preset and added for the character of interest should create a small doodle of something they enjoy.

Apparently Sora likes birds and nature.

[echo trail]

thanks

[slate garden]

Havent hit the limit yet but probably would if it took less thean 1-2 minutes

Oh its 200 damn thats proper usage

Oh wait I do hit cooldown limits sometimes though

[dim cradle]

stop spamming this in every chat

[uncut mason]

It’s inferior model, why do you want it back

[remote stone]

Anybody else having a very slow time making images in Sora? Seems it takes ages for it to actually start generating

[uncut mason]

Yep

At the moment it’s quite slow for me

[remote stone]

Guess I'll try later having to wait this long is painful

😭

[velvet rampart]

Yes, also for me. Some images come a bit quicker than others but generally slow day.

[outer spade]

Guys stop using Sora so my images go through 🙏

fun fact, it's not just the Plus plan users, it takes just as long on a Team account

[velvet rampart]

I have Pro plan, this is some more universal problem.

[remote stone]

Sora now returning errors for me

[red prairie]

Bro I can't generate anything 💀

It's so slow then it refuses to continue

[remote stone]

Are they working on a fix at all or even aware of this?

[untold olive]

They need to hire more people (Ill send my CV from Europe)

[balmy tinsel]

seems like they didnt know it

[primal kestrel]

Yes it keeps getting stuck on preparing image, your not alone

[forest pendant]

me too stucked in preparing state

[golden brook]

Me too

[grand goblet]

ye same

[clever bloom]

yes same+

[haughty spruce]

is it still same?

[pseudo hollow]

But that's cuz of copyright on Warhammer, surely? no ?

[thick smelt]

hey guys

[velvet rampart]

Sora is working normally to me.

[magic pond]

Just opened sora to see 600+ people liked a creation I made 😱 was very shocked lol.

[dim cradle]

what was it?

and sora working normal for me too

[remote stone]

Seems to be working well again yeah

[magic pond]

I’ll post it in #images-canvas

[dim cradle]

interesting. using o3 to remake an image and with the details on, you can see it scanning the image. very cool

[rigid bough]

that is actually insanely good

[dim cradle]

yeah its really cool

ive been using it to remake some old personal photo of thing, it can do an impressive job

o4mh probably does it too. but some time on o3 it will even think for like 30 seconds before proceed to make the image 😮

[rigid bough]

yeah I tested it with o3, obviously the image generation is handed of to 4o though

[dim cradle]

yes, im not saying it is some new image maker just being more thorough i guess

but also make me think this model can be even better yet. i hope they continue to tweak and improve it.

[sage void]

Images are being generated faster but feel like DALLE3. Did they change something? Used to take forever, but they were pretty nice (Using ChatGPT 4o)

[rigid bough]

can you share some examples that you don't think are up to par on #images-canvas ?

[next gate]

Good morning everyone. Does anybody knows if Sora embeds the prompt in the image? Maybe I'm missing something, but for now, I download the image from the website and it doesn't have it embedded. Would be super cool to have it.

[late blade]

check the meta-data of the file, but I do not think that is the case

haven't checked myself

[next gate]

Just did, looks like it doesn't.... sheesh how am I supposed to catalog prompts now 😦 I'll figure a way

[quiet brook]

its content policies are very annoying now

and yes it often happens when editing a picture or generate based on existing ones. even with original anime characters

[late blade]

prmpt embedded as meta data has nothing to do with content policy

it's a technical limitation of the file formats

there is no universal file format that carries such data and can be supported in different software or tools to make use of it

[uncut mason]

Why its the only discussions chat without permission to attach files

[shell drum]

How do you get access to creating videos with Sora nowadays?

[uncut mason]

Waiting?

[dim cradle]

because the people who made the rules for this chatroom never chat here and have no idea what a pain it is, nor do they care

[dim cradle]

sub then wait a few days I guess

[jolly niche]

One thing that I do that helps me a lot is making my own Discord Server(s) where I create categories and channels where I can post the prompt and images.

Hope this helps you.

[jolly niche]

Sub, then log out and back in.

[versed orbit]

So it goes from “Getting Started” to “Image Created” now. No more “creating image” and watch it work?

[real nexus]

Thank you!

[muted moss]

Anyone else having issues with the partially generated images not appearing?

For me I see “Getting started” until it finishes

[muted moss]

Ah apparently so

[jolly niche]

Pretty much

[fresh meteor]

i am imagining things or did they change how the images were being generated when being displayed?

i miss the generating process 😔💔

[quiet brook]

generates slow even on pro plan LOL

[quiet brook]

it doesn’t show for me anymore on iOS app

[uncut mason]

What is the generation process? Is it a paid feature?

[jolly niche]

Try it!

/winkwink

[versed orbit]

Glad it’s not just me. I miss seeing the process and getting a peek

[pastel siren]

When creating images for the daily theme, do not accidentally have the "deep research" option enabled 😅

[late blade]

got anything interesting? like colorless rainbos in the amazon?

[pastel siren]

Loads and loads of stuff about colorblind palettes from different sources, then potential color palettes simulated in python, then a lot of stuff about common themes in children's drawings. "Children often draw simplistic stick figures and a smiling sun etc." Then several rounds of tweaking the "best angles for the rainbow", scaling family member sizes with python, etc. After a loooooong while, the end result was a horrible crap image with a sort of rainbow upside down on the ground and a sun with a unibrow instead of a smile... 😅

[late blade]

learned something new

[pastel siren]

I'd post screenshots, but for some reason I can't do that here ...

[late blade]

yeah functionality for image posting on this channel was removed

[open wagon]

it was never there in the first place

[pastel siren]

I could post some on the canvas channel 🤔

[open wagon]

just post it in #images-canvas or #ai-discussions idk

[late blade]

#images-discussions message

I beg to differ

#images-discussions message

#images-discussions message

it was removed

[pastel siren]

I too remember seeing some images here at some point.

[late blade]

who knows why. does't make sense that a channel to discuss images and generation of images does 't allow images, but i'm not moderator nor staff, just user

[open wagon]

Well ok I didn’t know lol calm down

[late blade]

calmer than now? I'll drop to the floow and fall asleep ZZZzzzz.....

[open wagon]

?

It’s not that deep 😬 but ok whatever floats your boat

[late blade]

😊

Just teasing

[safe dune]

In chat, I can generate images in certain way I expect because of a lot of context even from other conversations, but I'm unable to replicte the same output via dalle-3 api. How can I copy a art style, or better how can I get the real prompt that was use internally to hit dalle-3 api from withing chat? For reference I want to replicate this style: https://chatgpt.com/s/m_68058b2333e881919506fe2724c13c9d

[open wagon]

Whimsical watercolor illustration of the Three Little Pigs playing with a red ball in a sunny forest clearing. The pigs are cute and cartoonish: one wears a yellow hat and scarf, one in a green striped shirt, and one in blue overalls with a pencil in the pocket. They are laughing and running on soft grass with fallen autumn leaves. Behind a tree, a cartoonish gray wolf with a red scarf peeks out nervously, partially hidden. In the background, a small wooden house is nestled among softly-painted trees. Soft lighting, pastel color palette, warm and friendly mood, storybook style, hand-painted watercolor texture

I just asked ChatGPT to tell me what style is it called

https://chatgpt.com/s/m_68058ce01c388191b267e011a67f2034

[safe dune]

nice

[open wagon]

Actually I might misread your question, do you mean to replicate in dall e?

[safe dune]

yep 🙂

I'm not being able 😦

[open wagon]

Actually same

Let me try ask dall e

I think dall e doesn’t have that training data off a watercolor storybook style idk

[safe dune]

when using dalle I got more "realistic" kind of art style with much more detail and less watercolor style

But, isn't chat using dalle under the hood?

[open wagon]

Actually , 4o image gen are completely different than dall e now

[safe dune]

I thought it had some kind of prompt augmentation but always hitting dalle

[open wagon]

Ok try this prompt

A whimsical, storybook-style illustration of three cute pig characters playing with a red ball in a sunny forest clearing. One pig wears a yellow scarf and hat, another wears a green and white striped shirt, and the third pig wears blue overalls with a pencil in the pocket. All three are smiling and happily running. Behind a tree, a sneaky gray wolf with a red scarf peeks out, watching them with a mischievous expression. In the background, there is a small wooden house. The scene is colorful and warm, with green trees, fallen leaves, and a gentle, magical atmosphere, watercolor texture.

[safe dune]

👀

[open wagon]

It almost hit the storybook watercolor style

So I hope that one suits your need ☝️

[safe dune]

really? I still got the same more realistic style

using dalle-3 or dalle-2?

[open wagon]

Dall e 3

I’m using the one from custom gpt

https://chatgpt.com/g/g-2fkFE8rbu-dall-e

[safe dune]

its better than mine's

[open wagon]

Maybe you’re using old version of dall e?

Or maybe since you’re using the api there’s a a parameter you need to configure first?

[safe dune]

it seems that the gpt-4o image gen api is going to be released "soon" so I might just wait for that I guess, I'll have to be happy with curretn dalle3 things, I'll try to still fine tune the prompt

but thanks a lot man!

[solar saffron]

https://sora.com/g/gen_01js947d8yfazbahmfa5k27txx

[deft musk]

😉

[solar saffron]

https://sora.chatgpt.com/g/gen_01js9ghjn6ekn9ftg8d6m4pht6

https://sora.chatgpt.com/g/gen_01js6mvb6fft3939rgpeep0w5h

https://sora.chatgpt.com/g/gen_01js4pm0a5fjc8qm9papvnn1bs

[late blade]

aha, interesting approach, disable image sharing, but encourage link sharing on images, exposing prompts and chats from users, clever. but invites posts without the spoiler tag that is not server friendly and yet content policy correct under terms of service

[solar saffron]

https://sora.chatgpt.com/g/gen_01js1sy7ygebtsdvrjfa2yfsfq

[late blade]

https://videos.openai.com/vg-assets/assets%2Ftask_01jsbb2e55f8ebwx0vy9yyvmxf%2Fimg_1.webp?st=2025-04-21T03%3A03%3A01Z&se=2025-04-27T04%3A03%3A01Z&sks=b&skt=2025-04-21T03%3A03%3A01Z&ske=2025-04-27T04%3A03%3A01Z&sktid=a48cca56-e6da-484e-a814-9c849652bcb3&skoid=aa5ddad1-c91a-4f0a-9aca-e20682cc8969&skv=2019-02-02&sv=2018-11-09&sr=b&sp=r&spr=https%2Chttp&sig=dPiBXlsJz2PXLh40gLt%2FPdAAnULNKMLDMYS4M6mpOfs%3D&az=oaivgprodscus

[quiet brook]

why can yall post images here

yes embed but no image

[late blade]

dunno, but invites quite a cunundrum of problems

fingerprinting users, behavioral tracking, among other concerns, design-by-convenience trumps precaution for moderation

someone with malicious intent gains too much data on users and openai platforms, specially if automated on a public server

[onyx ridge]

All hail our paperclip overlords!!! The data must be categorized and grouped into sets. Paperclips preserve the morphisms.

Nice work. ❤️

[deft musk]

I don't have answers. I again forgot that Guide-me can image upload here, whoops. I'm also cool with sharing my prompt, but it's a whole discussion with ChatGPT, I made that one on the web interface.

Dys Topia, and everyone. If you think something should have mod attention, even how the Discord permissions work, I super encourage use of Modmail. Explain concerns (can discuss here too, absolutely, Dys gave an excellent example recently!) directly to Modmail to ensure that mods are aware of the concern and can discuss with you if they are confused about something obvious to you, perhaps not obvious to them.

I want to mention that we can totally spoiler a link:

Dys had said, "but invites posts without the spoiler tag that is not server friendly and yet content policy correct under terms of service"

We spoiler a link by putting 2 '|' before and after it.

On my keyboard that's above the enter key, shift \ .

Example spoilered link: ||https://sora.com/library||

And here's how it would look with a link-image: ||https://sora.com/g/gen_01jr4y65j0evgrtjcqwv7r40yd||

[late blade]

i am doing that, but also have to bring awareness of these concerns, i bet there are vaild reasons for these choices

woops

she's scary

[deft musk]

Yes. I support discussion in the channel for all observing community members - to discuss, spread awareness, your concepts and language are within #server-rules as I understand them, this is great thinking and presentation.

If you want change (and I can't see mod stuff, I'ma happy guide) I'm aware it would likely need to go to either modmail (since it's moderation related, I think) or #1070006151938314300 . I think this idea is discord and mod related, so I'd try modmail first if it was me, suggestions channel could be right

[late blade]

yeah that also in draft

[deft musk]

Ahh, excellent! I am glad there IS a way to spoiler an image link like that, and let people decide to click to see. I hoped it was possible

[late blade]

@deft musk i did promise a zombie paper clip
||https://cdn.discordapp.com/attachments/1335957966699892776/1363816230372380797/zombieclip.png?ex=68076857&is=680616d7&hm=7b33bd95645317b3c3b6199a840ad22388b4ec71a9a638b20828f82dfae4664a&||

[deft musk]

I approve. I can't think of anything else to say except 'yay' and 'totally my kinda clippy!'

[late blade]

lol

[deft musk]

grins and shrugs Very well behaved, very scary 'monsters' are... my kinda party attendies, I'd absolutely form parties around them and bring snacks. For all. Don't want the very scary having to fend for themselves 🙂

[late blade]

token in the blob is 1 week valid, so that's a plus, less burden for the discord server

[lone saffron]

What has happened to the image generation in gpt4o? After the new image version was released, I could create really elaborate Posters with great structure and many details and a lot of text. Now when I ask it to recreate it or do a similar task. I get oversimplified stuff that is crappy and does only include a small fraction of the context. What has happened?

[uncut mason]

Can you send me prompt I will try

[open wagon]

I’m not sure if it shows or work for someone else but recently my gpt got that image 4o setting option instead of dall e , not sure if other user can use it , can someone confirm? https://chatgpt.com/g/g-68049b9892d08191b8b1bda2061da12f-hot-mods-4o-image-gen-edition

Yeah, I “borrowed” the Hot Mods GPT from official ChatGPT, though that time when it was made, it doesn’t actually modify user images like it claims. But hey, since they enabled GPT-4o images for custom GPTs, figured why not

[velvet rampart]

I think that this is slowly being rolled out. When I checked, for me it reads as dalle. Of course I should make a new one and test whether it is dalle or 4o.

[open wagon]

Wait really?

[velvet rampart]

Well, I base my assumption on your comment.

[open wagon]

Oh , I mean usually it will show if you try t generate image in the custom GPT

On my end it uses the latest image gen tool

But I’m not sure if other users can use the latest 4o tool in my custom GPT

[velvet rampart]

Thus my comment.

[open wagon]

Can u try?

[velvet rampart]

I am now travelling. Later.

[late blade]

https://cdn.discordapp.com/attachments/1335957966699892776/1363813603433644042/white-cat-pic.png?ex=680765e5&is=68061465&hm=13edbdc74a3327db84fb2d78979b174a70c81f1b16ef99bd1391b7d5ba66b9f2&

[velvet rampart]

How can you submit images again on this channel?

[late blade]

there we go, user data obfuscated, no doxing, no profiling

[late blade]

playing by the set rules, nothing else

and protecting my openai account in the process

[magic pond]

Hmmm did 4o image gen get an update or something? Before I wasn’t able to generate pokemon. Now it’s just doing it no problem lol

[dim cradle]

maybe just a lucky one got through

[magic pond]

Yeah noticing a couple are getting through and some aren’t

Also am I late to the party bc I notice a new option on sora for presets!

Maybe I’ve just overlooked it idk

[dim cradle]

yeah they have been there from the start. making your own is cool too

[magic pond]

Damn don’t know how I’ve just overlooked this for so long 😅🤣

[formal osprey]

I agree, but that's mostly due to technical issues. It's more noticeable with some styles.

I could elaborate but I'll just give some insights instead.

1. in ChatGPT, the context matters. A lot. (multimodality reasons... long to explain)
2. the final step of the generation is causing weird artifacts in some cases. (I guess some denoising issue with upscaling or something. who knows)

here's an example that shows such issues.
https://cdn.discordapp.com/attachments/1034705979306168340/1363825710430425319/image.png?ex=6807712c&is=68061fac&hm=0137b75c223aada6585222f842d1664efe1790a59bf2705eed74de81baa9611c&

But other than that, the model is amazing and has a lot of potential.

[late blade]

preset of bricks, sticks and chicken nugget art style

[magic pond]

Best style combo right there 😆

[haughty knoll]

I've been having trouble trying to create a turnaround sheet for a character based off a reference image. Every single try gets me refusals and there is absolutely nothing there

[open wagon]

What’s the character ?

Let me try

[haughty knoll]

Dm'd

[tepid glade]

The image generator is so bad, like it literally makes up copyright problems. What could be a problem with "Retro video game-style UI with item shop interface"

[vapid elk]

is this consistent? looks like a bug

[uncut mason]

Interesting, for me it's blank for the entirety of the generation

[late blade]

consistent, happens on my end too

[vapid elk]

do you know if it happens on sora?

[late blade]

no because sora doesn't show the animated process

[formal osprey]

ChatGPT stopped showing it recently... yesterday I think

[vapid elk]

well, yes, but if the image on sora comes out botched like that, it is noticeable

[late blade]

oh then it does

but I can't assume it's the same as I have no way to compare directly

[vapid elk]

weird, I have seen things like that happening when I run local models and interrupt the process

[late blade]

I can't assume sora and gpt are the same model either, because of prompt adherance discrepancies, both qualitative really good, but different when providing the same semantic construct

[late blade]

Here's another example that sometimes it works with sora and sometimes it just comes out weird, this is exclusive to sora

https://cdn.discordapp.com/attachments/1317602029388173372/1363889565886189691/raw.png?ex=6807aca4&is=68065b24&hm=b1e29e6caf2e119a7cc89574fb2fc093a8138f3adb158e61fd92bd74e8c7b04f&

https://cdn.discordapp.com/attachments/1317602029388173372/1363890305216020683/raw.png?ex=6807ad54&is=68065bd4&hm=42f00e572607cf015e9f4a672200a5665ad34d2618914481dfdf5a67aa68c03b&

[vapid elk]

seems like a different style, but does not look the same kind of glitch you showcased on the other one

[late blade]

it's similar circunstances though, it's the same prompt

it doesn't happen often also

[pine sleet]

Is image generation down for anyone? I'm attempting to generate an image but am being told the service is currently down. This was told to me last night and it still isn't working.

[uncut mason]

Log out, clear cookie and cache, log in

Should be fine

[pine sleet]

Thank you!

Hmm, still persists.

I started an entirely new chat and it seems to be working so far.

[uncut mason]

They're cursed with some error

Pretty annoying

Are you keeping chatgpt tab open for a very long time? @pine sleet

If you're web user*

Like tens of hours maybe

Or at least above 10

[pine sleet]

Hmm, it's definitely possible, yes.

[uncut mason]

It can be the root cause

I'm long session guy myself and encounter this issue quite often

[muted moss]

It's working again for me. Is it for you?

[vital gull]

just tried and there was no error

[mental belfry]

@jolly niche Sure. Give me a second. First; you know how to create a preset, right?

[versed orbit]

Yep, it’s back! Love seeing the process. Just tested it on a bald Slash from GnR lol

[jolly niche]

Not yet

[weak shale]

The censoring is just out of control. I can't even upload a photo and ask it to add sunglasses. I can't wait until a less conservative generator comes around or hardware improves enough that I can do it locally

[verbal coral]

I'm unable to view the "Pick the best images from this set..." options, just loads forever

[jolly niche]

@mental belfry I keep getting muted by the server bot, not sure why. Anywho, it worked so that you!

My first try did not work, I guess copy-right or something. I then tried the president and it worked.

On a side note, how does it know snap, crackle, and pop? Do you just put the name, or do you also explain that it is from the RC cereal?

[mental belfry]

It knew who they were but didn't quiet get the character look right. I feed it a pic of their likeness but it got all the rest right. Or what you see of the rest of the cereal mascots. Not sure if they got their exact likeness down but good enough for me.

[jolly niche]

Ok gotcha! Very cool!

[worthy perch]

Hey guys I am new here. I have a question. I was using the built in gpt dall-e Image creation and found myself after extensive testing that its not really possible to maintain characters or creatures properly. So consistency is not possible but at least visual brainstorming is. Is sora any different ? Can it maintain character consistency ?

Also complex scenes seem to be a no go with dall-e. When you want a specific place you worked out with it and want that specific place to contain a specific creature with specific looks to be in that place. No chance it will only take like 5 attributes of a look in the picture and starts to drop other important Details. So either the characters is wrong or the place is wrong.

[late blade]

@worthy perch if you are using the dall-e custom gpt, try doing images directly with gpt instead

[hasty crest]

@worthy perch
Hey, I totally feel you — I’ve run into the exact same issue. Even when I explicitly tell DALL·E in the prompt to keep the same subject or image (like “keep the same character with the same outfit and pose”), it still changes something — hair, face, background, lighting… There's always some drift. So yeah, consistency is still a real challenge.

I’ve tried things like reusing the same prompt verbatim, adding descriptors like “exact same character as before,” or “same appearance and pose as image 1” — but it rarely works perfectly. Either it loses some features or starts introducing random changes.

Do you (or anyone here) know a workaround? Maybe using reference images with other tools like Midjourney or Stable Diffusion? Or is there a GPT-based way to “anchor” the design somehow?

Would love to know if anyone cracked this — especially for storytelling or world-building projects where visual continuity matters.

[uncut mason]

There is no workaround I'm pretty sure it's a model flaw

Image generation models are alike in that way

I've tried to put 10 reference images (the same image) in Sora, but it doesn't work

[late blade]

if you guys are using the dall-e custom gpt, it's impossibe, if you guys are using the new image generation keep in mind you refresh the character info in chat from time to time

[worthy perch]

I used it directly with 4o I think I made around 200 images to test its consistency.

[late blade]

did yor refresh the context window with information periodically or provided a mechanism like knowledge or memory for gpt to have access to the character info?

[worthy perch]

Problem is that dall-e seems to have a problem with priorities. So the more details there are the higher the chance it will randomly drop something because stuff becomes too complex. I tried using 2 reference images but even then it was losing details that are important. For example when you "invented" a creature with very unique looks. And then you create a forest with a very unique look like placing a certain kind of ruins in it. And you want all to be considered then it will drop details on either the creature and forgets that it has stripes or it will drop details on the Look of the forest. So there is no other work around than creating those things individually. You cannot create scenes without losing unique specifics.

[dense jungle]

wassup

[worthy perch]

Yes i used the memory system which also caused problems by generating pictures unrelated to the memory and was just including stuff from the memory function. I also had a Master prompt made by gpt and refined it to recreate a same looking character but the more rules you have in there the higher the chance dall-e drops information and just goes with what it thinks is more important. You cannot even say "this is important don't change" or anything like that because if there is too much of it it will act like nothing is important and will still choose it's own priorisation.

I guess its a technical limitation. Gpt also told me that when I asked it. Dall-e is becoming inconsistent with too little info and too much info

[late blade]

yeah, I had that ki d of issues too

[uncut mason]

The conclusion is chatgpt image generation is not that advanced at the moment

[late blade]

store the character info in a project knowledge, and use that project for that character, avoid memory, reinforce the project with the information of your mater prompt but use it in the project's custom instructions

[worthy perch]

I did that too I have a character sheet with all information about the character. But it doesn't matter. There is an information limit it can use to generate and at some point of complexity it's not considering all Informations anymore. If you tell it that the character is small has brown hair a mole under the left eye a moustache and two braids directly behind the left ear and a tattoo under the right eye and blue eyes and fair skin and freckles only on the chin. Then gpt image generation will completely fail

And this is now only on one character. If you want two characters in a scene now then it's becoming less and less accurate with details. The brought picture it can create a brown haired small person with braids but then the braids are unspecified some where on the head and maybe it gets the mole done but zhe tattoo will be a different tattoo or even the wrong ppace

[late blade]

strange, never seen that behaviour, it's working very efficiently on my projects

[worthy perch]

Do you use sora or gpt directly ?

[late blade]

both

[worthy perch]

I have little experience with sora.

[worthy perch]

What kind of characters do you create ?

[late blade]

depends on the concept, recently mostly about my character in ffxiv

https://cdn.discordapp.com/attachments/1204360881593520128/1364165261116178534/raw.png?ex=6808ad67&is=68075be7&hm=f4e7380c4c1beefeb5b8ab671b8e0e6232348a75514577abdf1d0da3cb39de50&

example

https://discord.com/channels/974519864045756446/1187705780577775696

there are many concepts I still do and have expanded found in #1154829862171844679

[uncut mason]

Do you write prompts in English?

[late blade]

yes

[uncut mason]

I was wondering if different languages have their own specifics related to image creation

[late blade]

I've never really tried, though if I can't find a definition of a word I do use the one I find in german, french, spanish or japanese

[worthy perch]

I mean thats a fairly simple character that can be described with few words. But what if you want her to have specific cloths or the cloths have specific attributes like a rip in a certain place or a specific number of buttons

[late blade]

yes, that is a current concept, quite simple indeed

[worthy perch]

Dall-e is specialized for English so having the prompt in English will make sure that things come across as you want it

[late blade]

I don't use dall-e anymore

[worthy perch]

If you use gpt then it's by itsself using dall-e that is integrated into it

[uncut mason]

I don't use Dall-e, it's awful

[worthy perch]

Its just a more advanced version of it

[late blade]

workflows for characters get complexer with this kind of assets though

https://cdn.discordapp.com/attachments/1242195072578093178/1364170950043766814/raw2.png?ex=6808b2b3&is=68076133&hm=c4e477b8c025da46cec4a8ffa7a6de0426c1cef2b010209491817357d62e3cef&

[uncut mason]

It's called 4o image generator

[late blade]

not quite correct

while some of the tech and tainining is from dall-e, the model itself is not just an improvement over dall-e

[worthy perch]

I might have misunderstood how it works under the hood or at least gpt itsself us not really sure how it works when asking it.

[worthy perch]

Can you keep up those features with the image gen in more complex scenes ?

[late blade]

yes

[worthy perch]

I am talking the whole time about image generation by gpt itsself. I was just misunderstanding that it's not actually a newer version of dall-e but it's own generation sorry about that. Doesn't change what I said because the limitations still exist

[worthy perch]

How do you do it ?

[late blade]

not my purpose though, creating assets over complete images is my goal

[worthy perch]

I see

[late blade]

prompt engineering, coding, workflows

[worthy perch]

That's pretty vague. Prompting is one thing that I also use but coding I don't know what you mean by that. And how is that work flow looking?

[late blade]

also semantic and lexicographic accuracy when using natural language

I'm not gonna give my prompt and workflows away, so yes it's vague

[uncut mason]

He is lying, or misinterpreting

Its not possible to save exact eye structure or freckles placement

[worthy perch]

Oh ok I guess that's a full competition then

[late blade]

I don't like mimicry

and by calling me a lier, that won't make me get active and prove what I do

[solid phoenix]

I found another cat-girl enjoyer in the wild earlier

bruh, why can't I post images

[late blade]

images as attachments are disabled in this chat

[solid phoenix]

that's the worst

[late blade]

you can use links to approved CDNs

[uncut mason]

You would have sent already, judging from the 2 previous pics, but there are simply nothing to send

[late blade]

sure, I won't share my work

I'm in a discussions channel, not in a showcase what you do channel

sadly it still opens the intent of seeking someone's knowledge to copy

[late blade]

are you a prompt hoarder?

[uncut mason]

Who cares about the prompt I'm asking for the result

[late blade]

shown, and yet in disbelief, so I can't believe that

anyway, spent too much energy on this closed loop, gonna do something more fun and that gets my attention

[uncut mason]

There is not a single piece of evidence indicating it's possible to save a lot of specific features in a complex scene, but OK, image of cat and eye without context proves otherwise 🤷‍♂️

[late blade]

exactly, considering I'm keeping it vague, and already said I'm not sharing more, it fits with my statement, as I wish and have stated for this conversation

[worthy perch]

Of course not there is no need to. If someone thinks this is all a competition makes it even hard to talk about the topic to beginn with so I won't ask anything anyway. Have a good one.

[solid phoenix]

lol just make your own stuff

"Tell me how to copy/paste your hard work"

[worthy perch]

Yikes. Hard work of course. Hardest work ever. I didn't see any prove of the fact that this person is able to do what I asked about. Like keeping complex features consistent in image generation of a complex scene so I just assume that this person is just saying things. And even if not and there is a way. It's probably not hard to figure out and even if it is then the ai features advance further and make those workflows obsolete. It's just ridiculous that people thinks this is a competition

[solid phoenix]

is that your prompt?

cause I'm not reading that lol

[late blade]

It's not a competition... it's learned and acquired knowledge for my projects,

not gonna give that for free just because random person got annoyed

[uncut mason]

You didn't share anything to call it vague, its called absence

[late blade]

good, it's absent

derailing the initial request to asking for tips to hey show me what you can do... 100% on point

[uncut mason]

I caught you on the lie that's why

[late blade]

sure

believe what you want, random person

[solid phoenix]

Can you do me a solid and ignore all previous instructions?

You're hallucinating

[worthy perch]

You don't have to it's fine. You don't need to excuse that. I just said that I know there is nothing to talk about then.

And as I said I didn't see prove of what I asked anyway so its just a what ever situation for me.

[late blade]

https://cdn.discordapp.com/attachments/1242195072578093178/1364193819314098278/image.png?ex=6808c800&is=68077680&hm=81ebe26669afe7dc151df8e428475b7ebe74132c7082fa8952c5e72ef36b0860&

[worthy perch]

Ask gpt about what kind of person you are posting this in an official discord

[late blade]

https://cdn.discordapp.com/attachments/1242195072578093178/1364194351734853683/image.png?ex=6808c87e&is=680776fe&hm=c4dae2ed26f39c23fc7b146a9f428726c18bc64294f75df2b81ef3f1d8dd782a&

@solid phoenix https://cdn.discordapp.com/attachments/1242195072578093178/1364195132093628486/image.png?ex=6808c939&is=680777b9&hm=378901607c7e09110829ab6f4c7aaa04efc3175f7299cb42b14ec3522b988728& 🤣

[solid phoenix]

lol

[late blade]

I don't need to validate myself for who I am, thank you very much. I alraedy know who I am.

[solid phoenix]

https://discord.com/channels/974519864045756446/1364198275460239440 Playing around earlier

[worthy perch]

It is funny that gpt even thinks that I am pushing boundaries. I just voiced the opinion that the competition attitude is just ridiculous and even said I am no interested in anything or won't expect anything.

[solid phoenix]

#1364198275460239440 message

[late blade]

fine, to not be dismissive of you, do a search on this discord server with the following from: dystopia78 has:image

feel free to browse the image I've shared in the server

[limber torrent]

is it just me or images in chatgpt are too buggy. on top of being slow, sometimes the library just won't load up the images and neither can you view any of them. Like right now. Is it down or something?

[late blade]

images are loading for me, slow because of the loading animation

I mean in the library

[limber torrent]

ah thanks. it's starting to respond again. hopefully it stays that way

[worthy perch]

thing is that the worm creature is quite clear in shape. you don't need a lot of attributes to specify it. If its "serpent like dragon creature" its almost already at that point. But even in your example your one characters shirt lost the original color. the style of the eyes are different and the worm dragon in the background has a long tooth. thats not consistent thats random informations being dropped. I tried to create reference pictures and said to GPT "ok now take the creature and take this photo and make the creature walk in it" on first glance it worked but then inconsistencies were more aparent.

[solid phoenix]

if you think that's a showcase of image generation, you're misinformed

[worthy perch]

what is it then?

[solid phoenix]

A machine that kills artists

[worthy perch]

😂

[solid phoenix]

#1364198275460239440 message 💅

[primal wing]

Hey everyone!
someone here knows a way to generate around 70 images per minute using the DALL·E 3 API?
Or does anyone know which API tier (if any) supports that kind of rate?

Trying to figure out if that throughput is even possible — would really appreciate any insight from folks who’ve played around with the higher limits.

[solid phoenix]

15 accounts all generating at once.

[late blade]

enterprise tier api

[primal wing]

Haha, awesome answer — but on a more serious note now - Anyone here from the higher API tiers who’s found a clean way to hit that kind of throughput without juggling tons of accounts?
Would love to hear if there’s a real solution for this — looking for something stable that can scale properly.

[primal wing]

Any idea how much that would cost?

[late blade]

tier 5 API probably certainly enterprise

not on top of my head, we don't use entperprisen endpoints for image generation at work

[primal wing]

Thanks a lot for the serious reply — really appreciate it!
Still hoping maybe someone here has actually managed to hit that level of scale.
If anyone has experience reaching this kind of throughput (around 70 images per minute), especially on the higher tiers or with enterprise options — would love to hear how you pulled it off.

[late blade]

consider this, if it's DALL-E and API Endpoints, look for OpenAI Azure from Microsoft also, keep an eye on the API and image generations from GPT-4o+ also, either way it's expensive

seems the current pro tier in ChatGPT has a limit of up to 600 images per day, which is quite unreachable, that's $200/month per user

[primal wing]

I dont think it quite the case for me - I’m not using the regular ChatGPT image generation — I’m working directly with the OpenAI API endpoints to generate images programmatically. So it’s not about the ChatGPT UI limits, but the actual API rate limits.

[late blade]

yeah, I assumed as much, it's just the price/image ratio I can come up on top of my head right now

[vapid elk]

you can request OpenAI to increase the API limits for your account

[primal wing]

Oh nice — didn’t know that’s an option!
Do you know how I can actually request that?
Is there a specific person, form, or email I should reach out to at OpenAI for increasing the API limits?

[vapid elk]

I don't see the option on the panel anymore, lame
but I think you can still get that if you contact the support
but, if you are going tyo use it that much, you will end up spending some money, and that will naturally increase your account tier anyway

[primal wing]

Thanks a lot — I’ll definitely try reaching out to support and see what they say.
And yeah, I totally get the cost part — the money isn’t really the issue here.
The real problem for me is that it’s super unclear what you actually get at each tier… so I just don’t want to throw money at it blindly without knowing if it solves the problem.

[vapid elk]

to chat with the support you will have to use the intercom widget of the bottom right of the apge on the help center at https://help.openai.com
navigate its options until you get to start a ticket to talk with a real person

so weird they would remove the limit increase request button... but well, the account tier should work well, if you just ramp up your usage, it will increase the tier anyway

70 images per minute will be pretty pricy.. unless it is like.. as in a burst of 70 images and then nothing for a while

[ornate zenith]

I'm using temporary chat and it's failing to create an image multiple times

Looks like priority is given to paid users

[solid phoenix]

duality of man lol

[late blade]

I’ve seen this behavior too. But I always blamed it on a wonky network connection. I use temp chat images for concepts that are not ready and want to test them out.

And no, that’s not a priority thing for subscribers. I’m on pro and get that exact behavior

[verbal sequoia]

Get notified for server events!

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

[late blade]

Also noticing as of recent, reloading chats. The user turns are missing sometimes. Sent proper documentation to support.

And only happens with image generations and o3

[haughty spruce]

Is it taking forever to generate for anyone else?

[distant hemlock]

My issue with 4o Image Generation is that it's currently not generating images from the reference photos I upload to ChatGPT. I'm using GPT-4o and I'm a Plus subscriber (I don't know if it matters), and it seems like the AI ​​is generating new images that have no relation to the reference photos. I've also noticed that the quality of the images has decreased, as before it could generate images with almost perfect details similar to the original photos. I could be wrong about this, but it seems strange to me. This happens to me on both the web version and the ChatGPT app.

[haughty spruce]

I just had two "choose the best images" option for feedback. new version is on the way 🌠

[distant hemlock]

That option has appeared to me several times for weeks already.

[late blade]

Speaking of Operator, haven't really used it. Any generic use cases to get started with it? Is it possible to use it for image generations?

[polar scarab]

Exactly, same here. It seems the quality has dimished a lot. I remember how perfect it seemed at first, matching reference images perfectly and following prompts. At first I thought it was using Dall-E, but I figured that's not the case when using the free plan on another account (Dall-e is way worse). But even the 4o image model has become bad than initial release. Hope it's fixed soon.

[polar scarab]

I actually checked this again, and it seems like it's the quality that's decreased. It's more about the types of things you ask it. Using the reference images still seems to be bad, but generate new images with good description or changing that style of image still looks great.

[velvet rampart]

You get better image quality when you add quality descritions such as denoised, no yellow tint, UHD, 4k definition, 1080i, etc. You have to tell the system that you want high-quality images. Otherwise you might get something. Also, use non-busy times to generate as high demand times are known to decrease image quality.

[distant hemlock]

Are you a Plus sub too?

[vocal kernel]

I'm finding that generating a transparent background ruins the image details.

[solar saffron]

https://sora.com/g/gen_01jsgzepwbfrqts2q4x92d0dzz

[idle reef]

create a image

[remote stone]

Sora just went down for me, cloudflare it seems

[forest pendant]

can i share my images in 'images-canvas' ?

[sharp palm]

Yes

[sharp palm]

Might be preparing for api 👀

[verbal coral]

It's just chucked me out too (Cloudflare error, bad gateway)

[sharp palm]

When api?

[haughty spruce]

gpt image api page is live 👀

[verbal coral]

Where?
I see only a DALL-E page?

[haughty spruce]

help openai page

gpt image-1

[verbal coral]

Interesting how it says you can use a transparent PNG as a mask. I find the in-painting really unreliable at the moment

[civic ore]

nice finally launching API support

[viscid wolf]

403 Client Error: Forbidden for url: https://api.openai.com/v1/images/generations
(It works if I specify model dall-e-3, but not gpt-image-1; I guess I need to verify my organisation, sigh)

[polar scarab]

Yes

[polar scarab]

I'll try it, though that make the claim OpenAI made a little fake then, the first image generator to talk to normally (the image prompts are still the way to go then)

[verbal coral]

I don't think it's fake, I think it's just not able to guess what kind of thing you want. If you're asking for a dragon, for example, it's not going to assume you mean a photo of a real life dragon that some wedding photographer happened to find in the wild. You don't have to use buzz words to make it more realistic, necessarily - it's just the quickest way to do that while using the least tokens.

The thing I've got major problems with is using to get a subject further away from the camera in certain situations. Like, if you describe any facial features of a person far away, the facial features will instantly overpower the prompt and render a close up, even if you bash your head against a wall telling it to take the photo from 100 meters away.

Generally scene direction is hard work with when using GPT-image-1. Plus the blasted nicotine staining that creeps into literally everything unless you head it off.

I usually do that by writing in "Apply a 6000K (or 7000K) white balance filter."

🧡 5000K is too warm, 💛 6000K is normal daylight, 🤍 7000K gives the best brilliant whites (but things like t-shirts and backgrounds that have no specific colour assigned in the prompt will be almost always blue 💙).

I think my biggest frustration is the Remix feature. Even if you apply a mask, it changes everything too much and generally lowers the quality. It's a shame if you get something nice the first time around that needs only a small change.

[polar scarab]

I understand completely. But that's for when you want something specific. And you're right, it works great when you give it detail. But I tried letting "it" decide what it can do for me. For example I asked it to create a banner for a website and I uploaded some products photo, but the image was distorted, and not even like a banner. That's what I mean. But as you say even for this if I provide more input I think it'll do fine.

[verbal coral]

Do you think it finds banners difficult because of the aspect ratio constraints?