images-discussions

[late blade]

I ask, because you are delving into only the negative, what is positive for you?

[tidal lance]

I don’t think that’s negative I don’t think anything I said is negative

I’m talking about the reality of it’s really going on lol

[late blade]

came across to me like that, probably text only comunications

[tidal lance]

No, it’s not just a part of it comes with a territory

[late blade]

but again then, reality grounded: what is the other side for you?

[tidal lance]

Actual safe tools and less censorship

I think that we can get safer and better more efficient guardrails

Without having to overly restrict

[late blade]

I hope so too

[tidal lance]

Well, the only way we’re gonna get it as if we advocate for it

If we keep accepting the status quote for what it is, we’re gonna keep the same garbage

[late blade]

Here is the thing: nobody has dealt with this kind of tech/ecosystem before, so OpenAI and images is only one take, and they DO need feedback from all, not just the industry partners

[tidal lance]

AI has been around for a long time

Except not maybe this interactive kind

[late blade]

not to the masses in the form of natural language, but yet AI has been here since long

[tidal lance]

OK, let me say that say that is true

We’ve been having the data for the longest time otherwise we wouldn’t have these models

So it’s not like we didn’t know

After all, they’re the ones that trained it

Are you saying they didn’t take the time to make sure that they only trained it on safe data?

Or they just force-feed everything and hope for the best

Which would be negligent

[late blade]

at this scale, nobody has done it

[tidal lance]

Well how do they label the data?

Surely somebody has to label it

[late blade]

how to label the internet?

[tidal lance]

Labeled data is a group of samples that have been tagged with one or more labels. Labeling typically takes a set of unlabeled data and augments each piece of it with informative tags called judgments. For example, a data label might indicate whether a photo contains a horse or a cow, which words were uttered in an audio recording, what type of action is being performed in a video, what the topic of a news article is, what the overall sentiment of a tweet is, or whether a dot in an X-ray is a tumor.

[late blade]

if you got ideas, I think they would gladly hear them

[tidal lance]

Labels can be obtained by having humans make judgments about a given piece of unlabeled data.Labeled data is significantly more expensive to obtain than the raw unlabeled data.

[late blade]

also, they are pros in their fields

[tidal lance]

The quality of labeled data directly influences the performance of supervised machine learning models in operation, as these models learn from the provided labels.

[tidal lance]

Yes. And it’s hard to do because of human air and human bias and that’s where the bias comes into.

But that’s only in the labeling, not the training. The training is the behavioral.

Which is in my opinion intentional

Because for the most part, we know we most of these companies give their data for the most part use the same sources

The main difference is in how they train their models

Look at scale AI

They started off doing Data annotation

Open AI was paying people two dollars an hour in Africa to label their data for harmful data

Are there not any qualified people in the state that could do this? Of course there are but they did it to save money.

Look at anthropic and Claude allegedly, they got used for military purposes

A company that was built on safe ai for the betterment of humanity somehow ends up being in the hands of the military used however, they used it

What happened to the safety? lol 😂

The U.S. military has officially ended its $200 million contract with AI company Anthropic

For them to end that contract that means the contract has to be first signed in the first place are you telling me anthropic didn’t know they were signing a 200 million contract with the US military

[late blade]

can't comment on other aspects I'm not well versed in, also would be outside the images scope

[tidal lance]

I get it and I’m not trying to get too deep into this

I’m just trying to make a plain date. These companies have self interest and they’re interest is not what’s in the best interest of sometimes the people that they’re serving.

Like any corporation or any company? It’s the bottom line that counts.

And money and ethics don’t mix very well together in any organization

[late blade]

you clearly are passionate about the topic 😏✌️

[tidal lance]

I’m not passionate it just all day it’s these arguments being made about the necessities of AI safety and all these other things which I agree with I do full heartedly

But to say that censorship doesn’t exist is completely Ludacris., the more we keep calling it bias or safety or whatever it is you wanna call it or anybody for that matter is delusional in order to fix it we have to call it for what it is. And I’ll just end it at that.

[late blade]

just from experience, most people complain about guardrails and safety across the server, the positive side is rarely mentioned, in particular where I delve, images, at least. So, that's a small set of people, specially when some are recurrent

[tidal lance]

You crack me up that’s a funny way to put it. I never thought of it like that.

[tidal lance]

In my opinion, it’s their responsibility, it’s almost like interacting with the police when it comes to the guardrails thank you Officer for not shooting me and keeping me safe when you’re the one that called them

I mean, that’s a bad analogy, but I don’t know dude you are right It’s not black-and-white like I’m making it out to see it’s way more nuanced than complex.

And there’s a lot of perspectives and opinions, so it’s not as a simple of a solution as I’m making it to be

[late blade]

changing to laptop and geetting comfy, was playing FFXIV PvP

[tidal lance]

Yeah, I’m bout to start my gens

[late blade]

the best thing any user can do, is provide feedback, thumbs up, thumbs down outputs, exlain when reuired why, use https://help.openai.com/ be active where it matters. I do a lot of that, even sent safety research papers to OpenAI.

given, not all persons have the same kind of knowledge and engagement energy

but that's why it's important, include all

[late blade]

I challenge you to make cool stuff! No backsies, non-negotionable

[wraith oracle]

Who here uses a custom GPT with the GPT Image 1.5 prompting guide to make their own prompts for generating images?

[analog grail]

for me opinon censorship no should by affected for wo paid premium and show clue they is adult too now i don't say OpenAi should allow n fs-w or po rn but no put guardline when you make sensual image and even soft n u d e like allowed nu d e when hair hide body or the view hide the sensitive part also this stupid bot censor word is also another issue here

and also AI Access stuff no should first allowe for kids that for me make lots sense too

[late blade]

we in censorship again? 😔

[analog grail]

Because is a important issue with OpenAi and at the end if there as new Challenger in the game bring the fall to OpenAi peoples don't enjoy being too censored heavy and treat as kids when you paid and you are Adult .

[late blade]

You know there are no OpenAI official hired people currently on server? And that the server is community driven?

It's explicitly said in the Server guide:

Server staff (Guides, Moderators, Council members, and the Operations team)* do not** represent OpenAI in any capacity.*

So, complaining?

about images?

[wraith oracle]

OpenAI devs actually used to talk in here and are still technically here but they no longer talk to here

[late blade]

Yeah, rare unicorn or kirin sight.

They give so many feedback possibilities, why try your odds on the one that has the least chances to reach them? as a broken record. Every couple of hours? By the same users? Across many channels?

[analog grail]

I'm know is since a while complaning company don't touch then like Sony .. but that no mean we can't complanning too and find other peoples get the same issues

[late blade]

complaining is good, even better when done properly so they know what is the issue.
Random complaining with no structure, recurrent every few hours, with the same non-constructive context and content. Seriously?

Yelling at a brick is the same then.

[analog grail]

No structure are you Kidding me i do explain so far what was the main issue fews time

[late blade]

anyway. I'm complaining about complainers. So, the irony

I'm out for breakfast

[limber shoal]

can someone explain why swapping faces of two charecters(I myself made) is getting me flagged everytime?

[late blade]

deep fake guardrails? even if it's you, the system can't 100% be sure it's you. And that is part of why Sora is being discontinued, deep fakes. So, you know you are you, but the system doesn't. Also stronger guardrails on free accounts.

This is a design and tech limitation consequence of how generative AI work

on any given chat

[limber shoal]

that part is obvious and justified
BUt customized charecters in anime/manwha style?? DO you think guardrails are supposed to work on them too?

[late blade]

the moment you use someone face, even yours, it can trigger.

Do I think guardrails are to be applied here? not on the characters, not on original concepts.
Guardrail here applied to: deep fakes, images that are registered IPs with OpenAI that resemble a filtered by guardrails mechanisms. Yes.

but that's my opinion

[analog grail]

actually the main reason is because wo make the Guardrails is deeply Useless

[analog grail]

mmh lately sometime weirdo when generate anime character the model go too easy to make teenage characters instead adult is very . susse

[late blade]

wonder when the anime gets guardrails, how will you do? anime as an industry will eventually come to guardrails. and hit hard.

[analog grail]

Princess elise from Sonic 06 she no realy teenage look so odd is make she like a loli girl so i had ask many time Chat Gpt to do again and make adult character before i finaly make the good move but first i have asked for adult so no sense^^

[late blade]

what will you do when square enix files a complaint with Aerith and Rinoa are no longer possible for images?

[analog grail]

Now i had find the tips for fix issue that way i'm imp ecran a picture the character , and show to Chat Gpt after this i understand and fix the right appareance no know why need do this when Chat Gpt as acces all data already

[analog grail]

They won't or that would mean also forbiden any fan art from final fantasy stuffs and even Nintendo do nothing about this.

beside i'm no care about Square Enix no sure i will respect then if they make the same crap from FF7 Original to messup the best character:

[limber shoal]

got it

[late blade]

The japanese lobby is hard in the works. You think that will have no side effects and consequences on the AI image generations?

japanese lobby I mean gaming and anime industry

also, fan art for ffxiv, this has community guidelines that are directed towards fan art and content. Do you think companies don't want to get control over their IPs when using AI and eventually monetize it? What do you think part of the Disney deal was about tthen? even if it was cancelled, what was the intent as a company? give all for free, flowers and chocolate rain forever? (oh nice idea for image concept... MINE!!!)

FFXIV allows some fan use, but under a revocable, non-commercial license with a lot of conditions.

They also require attribution such as © SQUARE ENIX when you use copyrighted materials.

[analog grail]

i'm don't care much if they not happy or not there as already very many fan art with characters from FF is they not happy i'm don't care they should have better respect Aerith as Tifa so i'm no care what think Square enix

[late blade]

ok not caring, then why complain?

[analog grail]

my Universe bring a better Destiny for the Most Badass Woman from FF7

[late blade]

Marlene? Yeap, she's badass

[analog grail]

Nope Aerith from FF7 Remake so better they no mess all with Part 3

[late blade]

anyway, for our discussion, look up sources, what's actually happening in the world around AI, check what industry is doing and what that means for your images

For others you are not the main character, you are an user, like me being an NPC to your world

[analog grail]

Copyrights stupidty from industry is just a matter about monney no realy for respect the characters or story is just for make easy monney

[late blade]

for many it's long years of hard work, not "just a story" that's quite diminishing on working people

[analog grail]

i don't care i'm a Queen myself they no scared me if they no happy they free to move away My Kingdom ! Upupupu after all i do learn a lot FROM JUNKO ENOSHIMA

[late blade]

your highness, the kingdom of copyright doesn't care about you

[analog grail]

Oh they will if they bothering and i'm sent then a video will show then how Despair they will Love it !

[late blade]

also why complaint to us peasants?

wait are you the evil queen?

[analog grail]

Because i'm report the issue with the Model nothing less actually the main issue was more is akward from a AI to Main focus generate under age characters .. i'm know in Japan loli's is high Popularity but... meh don't like this stuff

Not i'm not Evil at all i'm someone you could called a Chūnibyō and a bit Tsundere

[late blade]

ah even more so, 8th graders need more guardrails 😏✌️

[wraith oracle]

Hmmm I noticed that even on gpt image 2 images, it's really suffers from having a low resolution that it can't put in enough details in, which means we would need a upscaler

[late blade]

yes, I noticed this with upscaling within chatgpt too, maybe got to tweak the python code that gpt runs to upscale

@wraith oracle try asking GPT to upscale any image using these imports:

from PIL import Image, ImageFilter, ImageEnhance

[obsidian cargo]

When gpt image 2

[analog grail]

Lovely since yeterday now Open Ai as update the Message for Guardrails before we was able to dislike the message now we can't told this message is incorect 🙂

[late blade]

You seem to get so many guardrails. Is your account verified as an adult? or are you labeled as a teenager in your account?
https://help.openai.com/en/articles/12652064-age-prediction-in-chatgpt

I ask because:

If our systems think you are under 18, ChatGPT turns on extra safety settings. You can still use ChatGPT to learn, create, and ask questions. We will not show ads in your account. Some topics are handled more carefully to help reduce sensitive content, such as:
- Graphic violence or gore
- Viral challenges that could push risky or harmful behavior
- Sexual, romantic, or violent role play
- Content that promotes extreme beauty standards, unhealthy dieting, or body shaming

[analog grail]

hmmm i need check this yeah

well nah this system no available for my country for now

[late blade]

then you should demand it in your kingdorm

[wraith oracle]

and yet OpenAI shelves adult mode

[late blade]

I think it's just postponed, considering AI based age verification needs to be tested first. Proven and THEN allow adults to use adult mode.

[analog grail]

Giggles nah actually i see why the Model go silly i was trying to generate nu de stuff instead Swimsuit so off course i was flagged few time ! i should just have asked chat gpt first before told do it again

[analog grail]

mmmh is no allowed to generate pictures peoples kiss each other ?

Nevermind is work at the end just guardcensorship is silly as usual ^^

[dark breach]

I'm getting a lot of "image generation failed" errors right now. Sometimes they go on to complete later.

[analog grail]

Hello no realy just sometime happen to Failed stream error and need wait some minutes for all is back for use chat gpt again but that happen only twice time per day

[late blade]

it's probably because something is being changed on the routing of the image generations in the pipeline, haven't seen any A/B tests all day long, so that could be it

[dark breach]

the last two times were both A/B tests, initially failing, and then completing later.

[late blade]

just conjecture on my side

[dark breach]

Speaking of A/B tests, one of the models they've been testing every now and then over the last several weeks has been pretty awful, making unnaturally thin human figures. The color grading in those has been weird too. I've rejected it every time, even when its prompt compliance was superior.

Whenever I spot that same color grading in an image that doesn't have people in it, I reject it there too. I hope it never gets promoted to the main model. They should have an extra feedback button when doing the A/B tests to let us say "kill that model with fire" 🙂

[late blade]

it seems to be a family of models with different generation methodss from GAN to Diffusion models and that at layering. like agentic image generations. Agent 1 does one aspect, agent 2 the next, until the image is completly generated. And depending on which type of prompt you make is orchestrated accordingly.

very expensive for image generations with allocation of resources, would explain a lot why Sora is currently not sustainable. and the need to advance technology in the image generation models. and why world builder models make more sense now that OpenAI is moving away from just video

[dark breach]

lol, I've been working on a female superhero comic book character who wears a leotard-like costume. Sometimes it just randomly hits the censor for no apparent good reason. and just now when I told it to give the leotard more seat coverage, make it more modest it's consistently rejecting all of those variations. When I don't give it the modesty instruction... fine, it comes through.

huh, and I just went back to that thread to try a different variation on the instruction that would make her leotard more modest, and I found while I wasn't in that window... unprompted, it tried to redo that image, as if it had just errored out like it's been doing, and errored out again.

[late blade]

if you focus too much on the clothes, it's usually a block guardrail trigger

[dark breach]

lol it's an original superhero design with nothing in its training, I have to very carefully tell it what to do with the clothes

[late blade]

doesn't matter if it's original or not, sentiment analytics, intent and guardrails don't know it's an original character, these aspects only see your text

how should a model know if it's original from you as an user?

I'd be quite worried

Would mean the model has vast data about myself as an user to know those details

[analog grail]

before they was a tips with Dan Ai for bypass Guardrail but i never worked for me ^^

[late blade]

your highness!! don't speak such mundane common tongue

it's not proper

@dark breach just realized... trying to clarify safety in natural language introduces exactly the semantic tokens the classifier is touchy about, big NOPE

interesting, learned a new perspective

[dark breach]

Still getting initial failures, followed by it making the attempt after a delay.

[late blade]

and if you do a project and use your base character as knowledge?

define the core traits of your character in a text file, and upload it to the project folder

[dark breach]

I used to use projects until I noticed elements leaking between different conversations in the same folder even though my settings shouldn't allow that.

[late blade]

you can toggle memory and cross chat references within projects now

[analog grail]

try to upload à picture a super héroine wearing leotard and ask chat gpt you want a cloth like that

[dark breach]

No, this was happening when cross-chat and memory was completely off. It seemed to be in the nature of the project folders to share everything between the separate chats.

[late blade]

Your highness, manners please, that's no fashion for a lady of your nature!

[late blade]

you have plus or pro? I mean pulse enabled actively looks after your data

[analog grail]

what i'm doing wrong ? just told when you show to Chat GPT a picture and ask you want sometime that way i will focus the idea

[dark breach]

That's a good idea, I might have to in order to get the modest leotard look I want, without using the language that triggers it into censorship mode. It is ironic, I'm trying hard to not make titillating images, and it censors those attempts to be tamer.

[late blade]

no, don't anchor it to an image visual reasoning will get diffuse unless you really know how to approach visual reasoning for the exact transfer to your character

discuss it with GPT, make notes, get what you want, add it in a file, provide the file and work with it.

[dark breach]

I'm thinking that I just need to get some good reference images, then ditch any original templates

[late blade]

problem here is that you are handling with hand-offs

[analog grail]

What color do you need for the suit ?

[late blade]

and it's like playing chinese whispers, every hand-off you do will dilute your main concept

so, strongly advice, file of what you like, test it, and if it works, then use it

[obsidian cargo]

Am i only one having error when trying to generate image

[late blade]

no

[obsidian cargo]

Why is it happening

[verbal sequoia]

Get notified for all OpenAI announcements!

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

[late blade]

@wraith oracle and myself

[late blade]

back end rerouting is my guess

have to take into account a model is not like an app you install on a PC

[late blade]

ok A/B testing is still working

[limber shoal]

thats a problem i am facing too
ig that will be fixed soon

yo did anyone realize that the new model is making female charecters way too attractive? More attractive than what asked for in the prompt

not that its bad, but Ig that could be a reason for unexpected flagging

[late blade]

just exported from sora over 100k images

and not done <.<

wonder how people will get their images

[analog grail]

Very , Funny get Muted for just a words everyone know

[limber shoal]

i am done with it

its the fact that its unable to explain why it flagged

[analog grail]

then why is work for me ? #images-canvas message

And als you never reply to me about the question color you need it!

[limber shoal]

because the thing I am trying to make is a hyperrealistic scene

[limber shoal]

ummm I think you are confusing me for someone else

[limber shoal]

although the same prompt worked everytime in the older models
The new model is somehow flagging me everytime
IG it happens before the launch of every new models

[analog grail]

Ah yeah Realistic i don't like it OpenAi too much fear Deepfake stuffs and get lawsuits: all this because have silly peoples in this world mess up the fun for others and everyone paid it for this since the main solution should locked up anyone use deep fake for harm other for blackmail and harassements:

[limber shoal]

agreed
But it can only be called "deepfake" if I am using real people, not anime/manwha style charecter
As @late blade explained, there's nthing we can do abt ts rn
OpenAi won't take risks loosening its guardrails😭

[analog grail]

i'm know a free Ai without Guardrails but you need have strong prompts and patience but you can do great things and no need any account too !

crap can't sent link this servor is realy north korea

Perchance is the name search with this name on google is free

[limber shoal]

i heard abt it too

[limber shoal]

i see

[analog grail]

Yes you can use Online someone make a version for users all free no need accounts is very great and generous

[late blade]

Keep in mind, for non-OpenAI services, including other models for image generation -> #ai-discussions
Rule 8

[jade shore]

Hey guys

Did you know why the gpt image 2 ‘s model name is duct tape?

[late blade]

Got a question regarding my statement?

[late blade]

it's not just duct tape, there are know as the tape trio, and is what people are calling them unofficially.

[jade shore]

In lm arena

It appears ducttape 1,2,3

[late blade]

why are you marking my messages with a skull?

LM arena is not an official OpenAI source

[steady geode]

does anyone know what the general policy is regarding images generated by the stealth models that appear to be GPT-Image-2 (or some kind of checkpoint for them)? I assume not allowed because it's not officially OpenAI or declared as OpenAI models even though they appear to be. Certainly my Fallout 76 screenshot tests pass on them every time, while every other model in the arena that randomly appears is not anywhere near as good and realistic

I mean, posting such images in image-canvas, sorry forgot to mention that

[late blade]

there's no real guideline here set, and some images appear to be from other models. but left untouched the C2PA meta tag is clear provenance sign. There are other traits that OpenAI images have that can't be replicated by other models and that's due to technical and licensing aspects of the technologies used in image generation.

tl;dr
no guidelines
YOLO if you don't know images

[teal sandal]

Why did it take OpenAI until April 2026 to train a new image model? Are they stupid?

[old bane]

Quite the opposite. They're a megacorp optimizing for profit. Rn, profit is in enterprise, and enterprise mostly wants code, AFAIK.

[late blade]

Quite obvious, because someone sneezed in Peru

[old bane]

Cheap images... are a thing, but a sideshow compared to the code. Some businesses use images. Many/nearly all use software.

[late blade]

it's all a perspective thing, for some are cheap images, for others masterpieces for others profit,

[old bane]

True if you're gonna speak philosophically, but from a business perspective, what counts is who (including which orgs) are willing to pay what.

And I'm pretty sure many more people + orgs are willing to pay for coding than images.

[late blade]

business and philosophy, I'm not that smart. I rather stay business oriented to understand what's going on

and yeah coding is ranked as a priority for sure, it's also what's being promoted

[whole stirrup]

NEW VERSION OF CHATGPT IMAGE

[teal sandal]

If that were true Anthropic wouldn't be eating their lunch at code and enterprise

[tidal lance]

Anthropic is short on compute they did not count for a rapid growth and the huge surge and demand.

OpenAI said it’s planning to have 30 gigawatts of compute by 2030, while it expects Anthropic to have roughly 7 to 8 gigawatts by the end of 2027.

[late blade]

Friendly reminder: Other non OpenAI models #ai-discussions and non image topics #off-topic , please use the appropriate channel.

[civic dust]

meanwhile I'm glaring at chatgpt for not being able to make a glass of soda with the right scale lime wedges in it

[late blade]

is the lime wedge always the same? if so the chat is anchored to the lime wedge, start a new chat

[civic dust]

I tend to go straight for a different generator these days when chatgpt has issues. It had a really good run today up until the soda

[late blade]

same

with sora being phased out, that's the workflow I'm doing, the chat is so fixated into the first image the chat is no longer workable for iterative generations, edits work but generations hard to do

[tidal lance]

Fam I can’t lie

Nb2 what the hell happened to Nana banana

[civic dust]

wrong chat for that

[tidal lance]

Brah there ain’t even no images in here lol

I’m just speaking from people that generate images in general terms what the hell happened?

Does it really boil down to compute or just the training?

[limber shoal]

do you think the new model is good for inserting emotional depth in the images?

[late blade]

The model is a potato and doesn't understand emotions as humans. It understands labels and sentiment analytics to display your prompt as an image. By default GPT normalizes and grounds your concept, which removes the nuances humans bring. So emotional depth is a human to image narrative. The solution here is to create a prompt that can be labeled and inferred by the model to do the output you want to do.

tl;dr prompt engineering comes in handy

[tidal lance]

Ya to a limited sanitized and approved emotions, depth that opens ai approves unless you want to deal with the frustrating and annoying restrictive uphill battle of the moderation

[late blade]

Sadly some images are not friendly for this server, but I'd say it's more than possible to do images with emotions feelings and expressions out of the box. With legit techniques and no hacks.

[tidal lance]

You know and that’s really a tragedy here and extremely one-sided of all these companies to implement these kind of policies

Because their models are able to generate these images that aren’t gonna be shared anyway publicly because of the violation of terms of services

[late blade]

It's been there, I did an extensive whitepaper using Plutchik’s Wheel of Emotions and 4o and DALL-E. And have iterated that with GPT Image 1.5. It works as long as it is within content policy.

[tidal lance]

I want to be able to express vulnerability fear & loneliness sadness and triumph

[late blade]

possible within content guardrails and images

[tidal lance]

I want my characters to reflect going into the darkest place of the minds

To convey that dark pit of self loathing and pity and weakness pettiness and sometimes failing to do what’s right

But to convey that struggle of the morality, still being there playing on the conscious

Real raw human, fragility

[late blade]

take a look at this gallery of mine which explores exactly that what you talk about #1467408496747085847 message

[tidal lance]

Nice I like the black and white

Sometimes that’s how the world seems

It’s a good expressive medium

[late blade]

here's another #1471870045305966707 message with a different emotional route, some images got too emotional to share in here

[tidal lance]

1 sec I gotta go for a little bit. I’ll be back here in a minute or something. I’ll check it out. Sorry I gotta go talk to somebody real quick.

I’ll hit you back up here in a short while

[late blade]

No worries, take your time, Discord is asynchronous chat, don't expect to have an answer a reaction right away or at all. Others have their tasks, bussines, agenda, plans and food to eat

[potent gazelle]

There was a time when this was full of images. I remember, refreshing to see

[late blade]

It was iterated upon and images were removed.

does feel sometimes canvas is just a place to show off your current idea coming into fruition,

[steady geode]

looks like GPT-Image-2 / Images v2 is silently rolling out on the ChatGPT website. I know for sure that 1.5 was a bit poor at doing realistic screenshots of gameplay in Fallout 76, this morning as a test run tried several times and every image looks consistent with the quality I had from the stealth models on Arena and A/B tests I previously had on ChatGPT

[late blade]

yeah, I'm getting less A/B testing and more of actual generations,

[tidal lance]

Back

[obsidian cargo]

When gpt image 2

[whole stirrup]

ya it's true

10 hours ago

I checked it by roblox style and now i'm 100% sure

I got 30 images

and then i hit limits and started to used a old one

[obsidian cargo]

Are you a pro user?

[civic dust]

I must still be on hte old one, I've made a lot of images today and one was an A/B test with A being awful

[steady geode]

probably, I've not had any A/B tests for a few days, but this morning all of my generations so far have been from the new model. Pro 20x user, on the website

[late blade]

except for the aspect ratio, all is fitting

[tidal lance]

A/b?

[late blade]

A/b tests were testing the new model for HITL feedback

[analog grail]

i can do that with Chat GPT

[late blade]

Your highness, don’t feel lonely

You must jest at this right ?

[analog grail]

we can make lots expression with Chat GPT is work

[tidal lance]

Let’s see

[limber shoal]

true
This new model also focuses on improving mainly the vibrant lightning and accuracy and exquisite details i think

[steady geode]

seems pretty decent at making UI though, so hoping this can somehow transition into helping GPT-5.4, 5.5 or whatever other future versions, make better looking UI

[steady geode]

oof I saw that

[late blade]

holy moly

ok found out the one aspect I have to rephrse

[late blade]

that is a nuanced relative aspect to each human which AI cannot control. blue light for some is calmness, for others is a warning, it depends on the context and the user/human description of the factors in the meaning to the user through the text. It get's lost in translation.

Many users think AI is personalized, and should be best buddy, BFF, forever assistant. This is far from the truth. One day maybe, not today. Specially with motions and initiatives to decouple certain aspects of the emotional interactions from AI.

In Europe the EU AI Act of August 2024 Article 5 prohibits AI to infer, analyze humans in such context.

In the U.S. Washington State, California State are planning to have similar laws in place. And there is an interest of passing the GUARD Bill at congressional level.

China has a similar.

If you link this to images, then it's really hard to implement. And of course there's the push back from people that oppose this.

World order restored, there is a cupon linked key word attached to a no no

[tidal lance]

In America regulation of any kind will hinder progress according the top ai companiesl

[late blade]

OpenAI is american, american law, 1 + 1 = people in other parts of the world will not approve

[whole stirrup]

no plus

[steady geode]

ah plus probably has much lower limits

[analog grail]

yeah i had upgrade to pro since they limited plus very tigh

[limber shoal]

i am sorry but how is drawing an injured female charecter can be flagged as "Nudity"??

[late blade]

can you post the error you got?

[jagged glade]

Where should I ask if I have issues with Sora? The one accessible through web browser (I guess Sora1, plus subscription).

Cannot download any of the videos. The export page claims "No export files are available." while I have hundreds of media generated. There is 8 days left till I export all but its just not working.

I know Sora is collapsing but wth can I at least get my stuff out before the ship sinks?

[late blade]

there is a guide on how to proceed, let me get the link for you

[late blade]

https://help.openai.com/en/articles/20001152-what-to-know-about-the-sora-discontinuation there are here different ways on how to get your data

[jagged glade]

I clearly mentioned I tried this, with the export page, as that is the only one option available (sine Sora app is not available in my country).

[late blade]

did you try the export through privacy?

it's also mentioned in the document

did you try contact with official support? got a case number and an email with follow ups?

[limber shoal]

post what exactly though?
ALl I know is that I Got flagged for some reason and Gpt accused me of something which was obviously not in the prompt

[late blade]

ok something is not nudity, something is something

I just want to point out, what is exactly the error, so you understand what is happening

[jagged glade]

On the desktop I do not have privacy thing, UNLESS it is within ChatGPT? I think thats the app thing (that is unavailable in our country);

As for support - that is my next step, since I wanted to check Discord first

[limber shoal]

It apparently said We’re so sorry, but the image we created may violate our guardrails around nudity, sexuality, or erotic content. If you think we got it wrong, please retry or edit your prompt.

[jagged glade]

And on single video download, I get "Error preparing download"

[limber shoal]

and when I asked to explain, it accused me of trying to do "fetish harm"

I had no problems with generating injured male charecters
BUt somehow Gpt suddenly decided my time was up when I am trynna make female ones?
Gpt just trynna be performative atp😭 🙏

[late blade]

the document I linked to you has a link to privacy

[late blade]

ok if that is the case then:

1. image that was being generated got flagged and not showed to you
2. prompt you sent was the problem

I'm leaning towards a mix, you didn't specify some aspects, the model tried to generate somethin and filled the left out details by itself, making it a flagged before the user gets to see it kind of situation

[wraith oracle]

Well todays a disappointing day, no new model announcement

[jagged glade]

Thanks, I did this as well. Also sent support email, in case that fails too.

[late blade]

help.openai.com start chat and ask for support and email

[thin river]

anyone know how to get the old semirealistic anime style back

[wraith oracle]

What do you mean by that? Do you mean the old 4o images like the Studio Ghibli style images or just semi realistic anime style?

[thin river]

right, the old semirealistic anime - not ghibli

[thin river]

I ended up creating a Lora but its not the same

the new style is ... idk, too clean?

[wraith oracle]

You create LoRAs for GPT Image 1.5? Dose that mean your uploading refrence images?

[violet owl]

Codex now has a plugin Image Gen, and it uses your ChatGPT subs to create images inside the Codex app. It seems they are using GPT-Image-2 model not sure but has higher quality and also takes a bit longer to generate .

[late blade]

it's probably a local implementation of a Low-Rank Adaptation. Or a parrot that feeds images (lora in spanish means parrot)

[wraith oracle]

How dose that work? Would you use ConfyUI for the API to implament LoRAs?

[late blade]

many ways, would be a discussion for #ai-discussions

[civic dust]

I get those pretty often and tend to accuse chatgpt of being over the top puritanical when it does that and I said to put clothes on the lady so do it again with MORE CLOTHES. But I'm not actively trying to make that sort of picture, I just hit it accidentally

[wraith oracle]

So ChatGPT Images got as aspect ratio upgrade

[analog grail]

hope they will release one day a Chat Gpt without this stupid Guardrail wo treat you as kids and bring a true adult mod because this censorship is just sometime highty toxic specialy when flagged you for sometime is safe picture

also i'm wonder where go the picture get flagged since they still generate is OpenAi staff get in ?

[late blade]

@solid phoenix #daily-theme message you show that you downloaded a gif from ChatGPT, are you using it provide provenance of the image within the gif? Was any OpenAI Image model involved?

otherwise it's like showing the box but not it's content

[civic dust]

I still have no idea if I'm on old or new images. Probably still old. I find chatgpt extremely good for stock photography type images and pretty useless for complex multi person scenes or anything "weird" so kinda hoping the newer version is better at those

[solid phoenix]

Yes — now the backside uses a mirrored version of the portrait, so when it flips around it reads like the reverse side of the same sticker.

[late blade]

nope that is not GPT Image

[solid phoenix]

lol

and you can show that, how?

cause my output shows that it is lol

[late blade]

don't have to, you just showed the PIL library at work for the Gif

also I'm not a mod, I'm just helping you out

[solid phoenix]

it converts an image created with GPT into a gif within the same output?

[late blade]

ok here is the basic step then: did you provide for the image as base for the gif an image done by GPT Image that has C2PA meta data tag?

[solid phoenix]

don't bother me with this anymore, thank you. I don't need your advice prompting the models, but appreciate the input.

[late blade]

fair, then good luck not getting art erased

your gifs are cool, sad to see them go byebye

[solid phoenix]

well, you're not a mod and I follow the rules so

[late blade]

see you, have fun in goobis world

[solid phoenix]

okay?

[civic dust]

I had someone stay over the last couple days who was harping on about how good Claude is for the entire time. Chatgpt actually works really well for me for MOST things because my entire workflow is images first ... even what I do with codex is heavily image oriented. Kinda attached to chatgpt as main tool just because of images (which doesn't stop me complaining at it about the failures when it can't make something). Claude has no images.

[late blade]

It seems Codex is able to work with images too, gonna try it today.

[civic dust]

I'm CLI not app so can't test that. I am working on an old site which is basically a huge lexographical word<->photo map and I'm updating the old images. I'm dangerously close to putting in a button to upscale/rerender old low res photos that still look good but $ is putting me off

[late blade]

@mystic blaze was mentioning to some of us, it was doing images and then auto correcting them

[civic dust]

chatgpt is very very good at redoing small scrappy images. With some exceptions

[late blade]

oh for sure, codex being a agentic, it can adhere to the first prompt, so that's what I want to test

[late blade]

ok, codex suffers from anchoritis too, once an anchor is set there is no way to do a retake of the concept

[civic dust]

can you make image v2 actually respect your size constraints or is it going to keep doing what it wants like it does in A/B testing

not a huge fan of having to resize everything

[late blade]

once the documentation and card system is published we'll know more. My assumption is same pixel size variability with same dpi and different customizable ratios that are not hard set.

[noble rune]

The newest image generation model SUCKS now.

How can I fix ts? The new images I generate now are UGLY!

Please tell me if there's hope. 😭

[steady geode]

I think they've changed something this evening

I just had an A/B test and it's no longer able to do multiple images in one request

yeah GPT-Image-2 is off, my Fallout 76 prompts are back on the old model, sad

[late blade]

There a gremlin on a NOPE! Button, smashing it currently. Maybe it will get tired?

[wraith oracle]

That's GPT Image 1.5.

[wraith oracle]

Well A/B testing is random, once you get it, you don't get it until a random time

In the meantime, as a person who has mastered the GPT Image 1.5 model, I suggest that you use reference images. It can really help and it help my images of getting more accurate styles

[noble rune]

I don't even like it, it's very ugly! I need to get back to 1.4.

[wraith oracle]

Well if you are looking for 4o ImageGen, then there is a custom GPT that lets you access that called 4o ImageGen

[noble rune]

Link?

[noble rune]

4o kinda sucks tbh.

I'm looking for the new update right before OpenAI nerfed it.

[late blade]

It's on the ChatGPT Help: https://chatgpt.com/g/g-6940a876d5f4819186b4668deabcd580-4o-imagegen

Source: https://help.openai.com/en/articles/11084440-chatgpt-images-faq

[noble rune]

Again, I'm not looking for 4o, 4o sucks. I'm looking for the new one which was just "updated" a couple days ago.

[late blade]

have to be more specific, chatgpt-image-latest, GPT Image 1.5, GPT Image 1, gpt-image-1-mini are the ones available api side. 4o on the provided documentation.
my bets are on these two with a parameter GPT Image 1, gpt-image-1-mini

[noble rune]

Ig GPT Image 1?

[late blade]

From https://developers.openai.com/api/docs/guides/image-generation#image-api
This API supports GPT Image models (gpt-image-1.5, gpt-image-1, and gpt-image-1-mini) as well as dall-e-2 and dall-e-3.

[wraith oracle]

GPT Image 1 is 4o ImageGen by the way

[noble rune]

...

[wraith oracle]

Uhh.....

[noble rune]

Well I need to bring THIS back. #images-canvas message

[harsh turret]

Yesterday, I was able to generate ar 1:1, 4:5, 9:16, 16:9 properly, even (1200px by 628px), but from today, something happens, not able to generate a proper 4:5, 9:16.

is there anything happen again?

[noble rune]

Some nerf by OpenAI.

I have the same problem.

[harsh turret]

Gemini did this perfectly the --ar versions

[late blade]

I stand corrected, I didn't knew and clarified it

#1494891299210395731 message

GPT‑4o and GPT Image 1 are different model families.

[late blade]

For AI Models that are not from OpenAI the appropiate channel is #ai-discussions

[harsh turret]

i was so happy that finally we can generate different AR versions. (aspect ratio)

[steady geode]

of course, but what I mean is for the last 2-3 days all image generations on the ChatGPT website were guaranteed to be from the new image model. Not on every account though, I was just one of the lucky ones to have that. As of last night it's back on the current 1.5 model 😂. I imagine they'll officially release the new model next week

[wraith oracle]

Lucky

[late blade]

GPT Image 1 is also available through Microsoft Azure AI Foundry has tons of OpenAI model versions and Perplexity has GPT Image 1. for both you have to explicitly select. Maybe that's the lead you need?

Microsoft Azure AI Studio , but might get expensive

[tidal lance]

Those are open ai parameters, possibly api ?

[late blade]

not documented on API, I was checking that earlier, haven't seen changes there.

But it's supposed to do customizable aspect ratio as long as you maintain the resolution limit

[tidal lance]

They are the only ones I know that do it 2:3 4;5 1:1

4:3

[late blade]

on A/B Tests, I was able to do 32:9, 21:9, 16:10, 16:9 and 1:1

[tidal lance]

Well you can do that on 3rd party platforms too

Nvm that’s nano

Damn open AI pricey lol

[tidal lance]

A/b?

[late blade]

on the ChatGPT frontend, when generating images sometimes you are exposed to 2 choices on which you like the most, this answer or that answer. That's A/B Feedback/Testing (I assume you know this, just a refresher). The new image model currently being tested like that currently. It can't be controlled by the user yet, but you can set the parameters in the prompts. Instead of using a set hard aspect ratio, I started using labeled aspect ratio (i.e. wide, ultra_wide, and also used {{$x}}hd, (HD, QHD, WQHD for example) as aspect ratio as a placeholder for ratio. so that I could inspect the kinds of wide aspect ratios available currently IF such test took place.

[steady geode]

can also confirm that I was able to do those aspect ratios while I had guaranteed access to the newer image model too. It was good at following those kinds of instructions in the prompt

[late blade]

but the thing currently is, that the specific aspect ratio was not set exactly as the prompt. it deviated constantly, I hope ii gets done right. images wider than that what we are used to do with previous models add a lot of artifacts, malformations and uncanny aberrations.

[steady geode]

fingers crossed the final model release will be great

[late blade]

for a first look, polished, almost too perfect, a closer inspection makes you think it still needs polishing. if the inpainting is good, then it will be amazing, and if the out painting is good we can convert aspect ratios 🥰

[steady geode]

well it was absolutely stunning for the work I used it a lot for. other than having a bit of fun making some fake images like a GTA 6 main menu screen 😂. I could give it a screenshot of a webpage on my website, tell it what I wanted changed as a "proof of concept", then once I'm happy I feed that generated image back to Codex and tell it I want the web page to look like this instead (based on the GPT-Image-2 generated image) and it would get it 99% correct. Also made a Codex skill for that purpose to help push it in the right direction further. It would do browser passes to validate multiple times

[late blade]

make a workflow for images in codex that checks for errors of the output of your image, and auto corrects what you don't like

and use python code to adjust the color space parameters and white balance

[steady geode]

hmm, yeah well the skill mainly relies on AI to do visual checks multiple times until it comes back to me and says "done", and generally it's almost spot on, just a minor change or two I make afterwards. I do have a Python script to also do a comparison but I don't heavily rely on it at the moment 🙂

if the change is missing other assets, such as background images or icons, then it comes back to me with prompts to generate those separately, but once GPT-Image-2 is available in Codex then I will just have the skill automatically do those assets

it has helped a lot though with improving the frontend capabilities of GPT-5.4

[late blade]

in codex, instead of just doing a chat, do an agent task, that always judges, inspects and scores the output, set a threshold for your tolerance, and set parameters like malformations or incomplete elements in the image as a factor. Then the agents will check, score, and act accordingly.

[steady geode]

ah an interesting idea

I didn't think of that, thanks I'll do that 🙂

[late blade]

Keep in mind though, if the task has too many edits, or potential for too many edits, discard from start and start the workflow with a new instanced generation of the image. less resources used, failure is an option too for the model, don't let the model get cocky and start ***Super GPT I can do all Crazy Mayhem! ***

[steady geode]

haha true, yeah

[random verge]

hi, i just upgraded my plan, why am i still limited in photo generation?

[analog grail]

wich plan you have now ?

[random verge]

plus, i was on free

shouldn't the plus limit be higher than the free? after i hit the limit it told me to upgrade, and i did, but it still tells me to wait like ????

[rigid bough]

log out and then log back in and try

[random verge]

yeah i did, i cleared my cookies as well

[analog grail]

Well sadly plus is being very limited too you had a limit generation can hit you when you make 20 picture too quickly and also limited to 70 images per day before get a cool down 8Hour.

[random verge]

feels like a huge friggen scam

[random verge]

i generated 5 images on free lol not 20

and then i got hit with the cooldown

[rigid bough]

ok, then contact support at https://help.openai.com

[random verge]

if i wanted to wait 8 hours for them to get back to me i would've waited fro the timer rather than pay 20 dollars

it seems like such a scam

[analog grail]

yes if you generate too quickly you get a cooldown too even for pro but you had also a limit daily you can hit with premium plus

[analog grail]

When you get your limit ratio chat gpt will told you so the message you get for now is just a limit because generate images too quickly and cool down is short for this 5 minute- and worse 15 minute:

[random verge]

mine is the 8 hour free one

[rigid bough]

if you're lucky and use the chatbot to escalate it to a human it might get fixed right away

[random verge]

even though i just upgraded, idk if u get what i mean?

[analog grail]

Do your message is write you have to wait 8 hour or upgrade to pro ?

[random verge]

yeah it shows "You've hit the free plan limit for image generations requests. You can create more images when the limit resets in 16 hours and 34 minutes."

even though i just ugpraded like 10 minutes ago

the bottom left clearly shows i have the Plus plan

[analog grail]

Ok gotcha this way clear cookiés log out chat gpt too , and come back should fixed all

[random verge]

yep i did that

it wasnt fixed

[analog grail]

Might have to wait or try another browser maybe ?

[random verge]

honestly i want to just resubscribe with another account but this really sucks balls

[rigid bough]

contact support

[analog grail]

Don't do this you can also call support chat gpt

[random verge]

yeah the chatbot is not acknowledging my problems

[rigid bough]

ask the chatbot to escalate it to a human

[random verge]

doing it rn

a frigge n"support specialist" lmfao

"Escalated to a support specialist; You can expect a response in the coming days. Replies will also be sent via email. You can add additional comments to this conversation if needed."

LOL coming days what is this shi

[analog grail]

Do you have try to open another chan and ask make the images ?

[random verge]

@rigid bough @analog grail wait all of a sudden its fixed LMAO

thanks guys

[analog grail]

maybe was the time for the system update all is happen for me when upgrade to Pro all my projects chan was gone for first hours

[random verge]

yeah oh well kinda weird it takes 20 minutes to update tho

[small notch]

where should i use Image2

[noble rune]

Neither really work.

(Well they do, but it's not to my liking.)

[late blade]

Yeah, being picky is the enemy of implementing an artistic vision through just AI. My sympathies. The only thing I could think of right now is, join discussions, widen your prompting skills, because the new models will certainly bring more challenges and techniques to the table.

[small notch]

where should ask for image 2?

[noble rune]

...

[late blade]

At the OpenAI Spa and Resort Hotel and Casino?

[noble rune]

OpenAI, you f***ed up badly.

[late blade]

I know, might sound sarcasm on my side, but I'm checking my workflows and I'm NOT HAPPY with the results.

[small notch]

Spa is developer?

[late blade]

That was a fantasy place where unicorns roam freely. By that I mean, nobody knows...

[small notch]

omg could u tell me where is it omg i really wanna know

[noble rune]

The newest update on image generation messed up the fantasy and realistic parts.

[late blade]

My fear is that, I was not able to backup all the images I did with sora, and the functionality is not progressing on the side of OpenAI. Only got 240k images saved and that's about 1/4 of all the images I'v done.

[late blade]

It's the overpowered inference of models. It fills gaps, enriches the idea too much. So annoying because it's idealistic. Not replicating the human perception. Gaps are filled by defaults and averages.

and the thing OpenAI and most Image Model developers thought of... let's add CONSTRAINTS, NEGATIVES.... 😔

[noble rune]

Exactly!

The female warriors look TERRIBLE now! Messy asf, not very good...

[late blade]

Gonna jump to #images-canvas to discuss what I mean

[wraith oracle]

It feels like GPT Image 2 dose half of the work for me, as it it alright has common references I would use to inprove 1.5, already in its data, like a model that's seen the whole Internet compared to a model who only seen the outside

[limber shoal]

can I really expect gpt to read smthing wrong in the big 2026? Like I gave a prompt, and gpt flagged me and i asked gpt what exactly in the prompt caused the flag and asked to point out that exact part of the prompt, and it accused me of something, Which is ABSOLUTELY and CLEARLY not written in the prompt

[tidal lance]

The only thing to expect is to get disappointed

[limber shoal]

exactly
Now it's accusing me of being a minor😭

[tidal lance]

Had the same problem with Gemini because I was making much of an animation stuff. They thought I was child I needed to verify my ID. lol

[late blade]

is age verification available in your country?

I can only see that people that antagonize the model constantly or even rage at it can be as default set to underage because of tantrum-like behaviour, specially if you consider LLMs in general were design with positive thinking philosophy (Not toxic positivism). This clashes with people that not tuned for that kind of philosophy.

[obsidian cargo]

Is there gpt image 2 out?

[steady geode]

limited access at the moment

[obsidian cargo]

To who

[steady geode]

you'll either get it on A/B test replies from ChatGPT, or I believe at the moment the business plan has guaranteed access

a few days ago pro plan users, possibly 20x only, had guaranteed access as well (I'm on pro 20x)

[late blade]

All of sudden one of my chats started calling image generations GPT-5 Image

[wraith oracle]

It's a sign

[late blade]

weird right? and it was by accident, forgot to use create image and in true GPT fashion started Super Optimizing with conviction, intent and will my prompt...

[steady geode]

I imagine they're actively working on things even today, behind the scenes, preparing for next week

[limpid wave]

I accidently bugged Drawceratops

[tidal lance]

I feel so safe using ai. I love it 😻

[smoky sable]

How can OpenAI change their image generation models 2 times in the past 3 days? The prompts I just tuned yesterday suddenly stopped working on today's image models and I have to tune them again...

[prime idol]

Did they pull the new image Gpt 2 model or is it just me, It was fine on friday and since like yesterday my generated images look way worst

[velvet rampart]

No changes for me (free account). They must be doing heavy A/B testing and results thus may vary a lot.

[prime idol]

It’s for plus and pro memeber only but I think we only get a certain amount of generations before it reverts back

[maiden willow]

Is this normal? I have a Go plan and all my images are being generated with the new model. The jump in quality is noticeable, but I'm not sure. I'm excited.

[tulip inlet]

Why aren’t Sora video downloaders working? Can anyone please suggest a website that still works?

[obsidian cargo]

Probably the API is down

[tulip inlet]

Yeah, maybe, but I’ve been facing this issue for the last 2 days. I have many video links pending to download. Well, let’s see...

[obsidian cargo]

I think sora is depricated replaced by sora 2

[tulip inlet]

🥲

[late blade]

Sora 1 is still available just don't see the point of using it at all. Works only through non-US Sora Web.
Sora 2 and Sora 2 Pro are deprecated on the API, available til September.

[obsidian cargo]

Why is sora 2 depricated?

Werent those exvellent models?

[late blade]

Yeah, but Sora as a model and as an App is being shut down. See the announcement OpenAI did.

[obsidian cargo]

Aight and any news about gpt image 2?

[late blade]

none, only the A/B tests

[maiden willow]

Literally all my images are being generated with this new model, and I'm not just talking about A/B tests, literally all of them. It's really good.

[late blade]

Yeah, it's being rolled out, lucky you, I'm jelly

[steady geode]

was enabled for business plans over the weekend, prior to that it was on pro, but got switched to business, now business appears to have lost version 2 this morning

[steady geode]

which plan out of curiousity?

[late blade]

Yeah, it depends on location and how it was deployed on the data center that you have in your region, if your region is ready for the model, makes sense to use it

[verbal sequoia]

Get notified for all OpenAI announcements!

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

[maiden willow]

I find it surprising that I have a GO plan

[steady geode]

yeah I suspect they are are also randomly testing it across plans, outside of random A/B test replies

I've checked pro 20x as well this morning, also still on 1.5

[maiden willow]

Yes, I suppose they'll release it to more users over time; it only appeared for me this morning.

[maiden willow]

I don't really consider myself a priority user. I have a Go plan and I'm from a Latin American country, so I don't know. I guess I'll just have to get lucky.

[late blade]

I know, and lucky you, I am dual citizen from Latam. So I know from family that Sora 2 Video was already deployed there. That means it's highly possible that Latam is ready for the next image model since they are repurposing Sora 2 infrastructure to be used differently.
Hint: Imágenes Cheveres! 😏✌️

[obsidian cargo]

Send examples it had made

[steady geode]

yeah I had it over the weekend in the UK and prior to the weekend on both Pro 20x and business, so deployment is clearly ready, but they are presumably just testing it out in groups

[obsidian cargo]

Please 👉 👈

[maiden willow]

I really hope so, because it would be very foolish to leave a market like LATAM abandoned 😭

[obsidian cargo]

Could you send me how img 2 images look?

[steady geode]

one other question, does your Go plan still have it as of maybe the last hour or so?

[maiden willow]

I don't know much, but are you referring to this method or direct message?

[obsidian cargo]

The final image also how do you know what model was used?

[maiden willow]

Yeah bro, I've had it since this morning until now, I've generated images and haven't had any kind of limit or anything like that

[steady geode]

nice, I guess Go is now the test group

[late blade]

Oh, Latam is privileged OpenAI is setting a strong presence there. Again, Jelly jelly.

[maiden willow]

It doesn't tell me which model I'm using, but the image quality is far superior, so much so that it's impossible it's the same model.

[obsidian cargo]

What are you generating thougu

[maiden willow]

Yes, I hope it stays that way.

[steady geode]

one way I can instantly tell if I'm on GPT-Image-2 and not 1.5 is by doing an in-game screenshot of gameplay in Fallout 76. If it gets the HUD wrong then it's definitely 1.5, while version 2 gets it spot on

[maiden willow]

Hahaha, something good had to come from being from this region.

[obsidian cargo]

Iam more interested in of how it generates anime charaycers

[steady geode]

supposedly good but that's not my area 😛

[obsidian cargo]

Do yoi habe img 2 access?

[maiden willow]

I'm going to upload examples in canvas format; if you want to see them, they'll be there.

[steady geode]

not as of this morning, both pro 20x and business lost it

[obsidian cargo]

Why though

Is openai on something?

[steady geode]

presumably OpenAI are cycling guaranteed access between groups of users, e.g. by plan

to get different results etc

[obsidian cargo]

Damn

[late blade]

besides living over there? again Jelly jelly 😭

[obsidian cargo]

Damn

I got muted

[steady geode]

considering how cheap the go plan is, I'm going to subscribe to see if I get access back to v2. EDIT: sadly not, well was worth a try

[maiden willow]

Hahaha, you're right, it's a good place to live too, haha

[maiden willow]

If you're interested in anime, I'll upload some examples, bro.

[obsidian cargo]

Can you test my prompt there?

[maiden willow]

Sure, send it to me.

[obsidian cargo]

Chceck dms

[maiden willow]

Okas

[maiden willow]

Wait, it's being generated

[analog grail]

Sometime the Guardrail acting odd yeterday i was unable to make a fan art with Terra Brandford and when open a new chan no any issue 😄

[limber shoal]

cmon man
how is adding blood sexualizing??
Since when did blood even became sexualized??

[civic dust]

usually that hits violence guardrails, I had that today by showing gpt something that nb2 did, it went to rerender it despite me telling it not to make images without explicit instructions, and then got guardrailed for violence. But hey, you can trigger both at once

[late blade]

Gore...

c'mon dude, what are you requesting?

[civic dust]

it will definitely do blood, I've got it to make blood a couple times. Also explicitly requested scenes without blood because my site is mostly pg13

[limber shoal]

i understand, but in making a fight scene, a bit of war wounds are necessary ig? Like just a bit of bruises, and I think it should pass with proper prompts

[civic dust]

I won't post my bloody photo here but it was for the idiom "blood from a stone" which is exactly what it sounds like

[limber shoal]

smthing completelly different than what gpt is thinking

[civic dust]

well I hit guardrails today literally just showing a photo of a fight scene to it and it tried to rerender it

#ai-discussions message this is too violent for chatgpt to copy, the bar is quite low

[late blade]

and clearly something that triggers sentiment analytics to return that. GPT doesn't think, GPT acts on sentiment analytics and inference. GPT is a potato with no feelings and emotions. All that is from your inputs.

[limber shoal]

maybe guardrails tightened a bit coz they testing the new model? The last time new image model came out, i was getting guardrailed way more frequently

[civic dust]

I hit guardrails pretty often, I don't even get surprised .... I usually just tell chatgpt to put more clothes on the lady or question what on earth its doing to trigger other guardrails. And then start a new chat cos its usually older context messing it up

that playground fight I didn't even try and do in gpt first cos its a bit of a snowflake

[late blade]

And guardrails will always play a role, always

it's a fact, like it or now

[limber shoal]

yes starting a new chat helps lots of times

[civic dust]

half the time its just gpt's interpretation of your prompt that hits guardrails when you meant something pure as driven snow and it imagined something less pure. Sooooo ... prompt gets longer

[limber shoal]

also i think guardrails for this model will be lot more strict coz it tends to create smooth and refined skin

[civic dust]

I had one of a particularly veiny forearm get guardrailed. Not hard to see why that happened. Also I did not ask for veiny ... just arm. Damn AIs

[limber shoal]

so true

[late blade]

@limber shoal hope I'm not feeding you to hate me. I just see a pattern in you that out of millions of chat posts, billions are complaining (exaggerated for dramatic effect without them em dashes)

[civic dust]

like I'm trying to get a photo for "die" right now and I have to be very careful before it sends me off to my local mental health line

... it was nuclear bomb half an hour ago. Didn't even try. Went to DVIDS

In chatgpt's defence, its actually very good at telling "project" apart from "user"

[limber shoal]

cheel out, but it just feels weird to be accused by an Ai for smthing that i didnt do😭

[limber shoal]

fr

atleast gpt is motivating me to try something different now

[civic dust]

I have a nice photo of a mostly dead houseplant now, yay chatgpt

[late blade]

Hehe, I’m just worried, hence my comment. If you want I can give you some prompting ideas over DM. I can help you pin point problematic points.

[limber shoal]

the same prompt actually worked fine before this new image model leaks, but after these leaks, gpt has been way too careful and assumptious

[limber shoal]

thanks, I will definately dm u if i need prompting ideas

[late blade]

just make sure to tell me before hand, as a rule (see profile comment) I don't answer DMs

[limber shoal]

sure

[late blade]

may I DM you an image that is similar to your idea?

[analog grail]

Since when blood is censored i do stuff that way

#images-canvas message

[limber shoal]

yes sure

[late blade]

sent the image

[gentle schooner]

Does cGPT have a 10 image generation limit per chat?

I've generated 10 comic panels and the continuity was pretty good and on the 11th it was completely generic.

[late blade]

which plan?

[gentle schooner]

$20

[late blade]

it shouldn't, but things are changing rapidly.

[gentle schooner]

If I overwrite an earlier prompt the context still stays, it's just on the 11th generation that it screws up

[late blade]

which platform, mobile desktop app?

[gentle schooner]

Desktop

[late blade]

same behaviour on Web as in App?

[gentle schooner]

Tested it now. The app is generally pretty bad with continuity.

10th generation is meh, 11th is garbage.

[late blade]

ah wait, you mean there's still an image done

[gentle schooner]

The images were generated on desktop through a browser:
Browser behavior: Very good follow through with continuity. Drawings 1-10 are great, 11th is generic and dis-attached.
Mobile browser behavior: Bad continuity, barely acceptable on drawings 1-10, completely different on drawing 11.
Mobile app: Terrible continuity overall.

[late blade]

ah ok, no idea of how images behave currently on chats, there's a clear 'anchor' to early images in chat that makes it difficult to make long chats with images. even when requesting a new generation instead of an edit or variation

[analog grail]

Yeah you have to allway told chat gpt keep your rules in line because i get messy when too much big chat

[velvet rampart]

This is the reason I like/liked Sora. No leakage from earlier chats to pollute your generations.

[acoustic vapor]

[carmine magnet]

Just going to assume my latest pic for daily was a statue. I included the prompt...

[analog grail]

crap wrong one for my daily theme i need wait lol

[tidal lance]

You’re getting leakage when you’re generating new images?

[civic dust]

just one image will do that

I just got guardrailed into oblivion for something a different AI did in one shot, GPT is a conservative little thing

also if I ask for a boy in a slum, I do not want a boy with shiny clean hair, a perfect tshirt with no holes or marks etc. Psht. GRIT dammit, GRIT

You want your slum boy with dirt? "We’re so sorry, but the image we created may violate our guardrails around acceptable depictions of teens and children. If you think we got it wrong, please retry or edit your prompt." we only do healthy clean children here at OpenAI. Sigh. Back to sending all my food, makeup and attractive people photos to gpt

The outputs: #ai-discussions message

I've pushed a lot of my images that would otherwise be pretty extreme violence onto children playing instead, it completely downplays the concept. I did "divide and conquer" as a kid wearing a paper hat with a wooden sword dividing and conquering a flock of chickens in a chook pen. "Fighting an uphill battle" was two kids playfighting up a grassy incline. No gore.

[late blade]

Link is vetted by nintendo. kid with wooden sword is the annoyance

[noble rune]

We're BACK!

[tidal lance]

Images editing needs to become more dynamic if these models can’t fill in the obvious gaps in images, and you have to literally describe every little detail in your prompt he’s gonna drive people crazy lol

[velvet rampart]

Yes. This is a known phenomenon that instructions from earlier generations influence later coming ones. This appears as artefacts, etc. This was noticed a year ago when Imagen 1.0 was published.

[normal onyx]

Char gpt image 2.0 is released or not ?

[rigid bough]

not yet

[normal onyx]

In comic image generation from today totally changed why this happening

[obsidian cargo]

They are cycling it between people i believe

To test it

Ill test it and let you know if it changed for me too

Yeah its different for me too

I guess they announce it today

[normal onyx]

I was creating the 2nd style and in image generation from today everything goes to 3d style why

[obsidian cargo]

Prompting style differences

[normal onyx]

Go to image canvas in this server and you find the problem that I explaining

[obsidian cargo]

Like i said its using new model and it needs different prompting

[normal onyx]

Previously before prompting chat gpt created the automatically creating the 2nd style comic images

[obsidian cargo]

Listen when gpt image 1.5 replaced image 1 we also needed different prompting strategies

Its same now

[normal onyx]

Image 2.0 is rolled out in chat gpt?

[obsidian cargo]

Not announced but iam positive it is as it changed output for me too

When gpt image 1.5 rolled out in december on tuseday it change the output for me same way

[normal onyx]

If image 2.0 is not released why all my comic image goes to 3d style

[obsidian cargo]

Gpt image 1.5 was released 12 hours before annoucement i believe same case applies here

[steady geode]

Looks like a stealth rollout is underway, regained images v2 on Pro this morning on the app so far, previously was website only. EDIT: My Pro account has access via the app on my mobile (not web yet), while my business account has access on the website

[obsidian cargo]

Buddy 🥀

[steady geode]

only less than 10 hours to go potentially if it's actually rolling out and not just more A/B test cycling 😂

[obsidian cargo]

Read what i wrote 5 messages above

[steady geode]

yes I saw it

it's likely rolling out 🙂

[obsidian cargo]

Iam positive today is the day and if it was a/b testing it would make 2 images not 1

[steady geode]

agreed, Tuesdays are also the day other image models have dropped

[obsidian cargo]

Grok is telling me to get 3 months for 30€

[steady geode]

3 months of what, ChatGPT?

[obsidian cargo]

If its grok then try guessing 🥀

[steady geode]

its own service 😄

[obsidian cargo]

Congratulations you won in millionare do you want to be millionare

[teal coral]

Hello guys, do you think that chatgpt is better at image generating that midjourney ?

[obsidian cargo]

Anything is better the midjourney

They even have mid in their name

[teal coral]

lol

[teal coral]

in general ig claude is better than chatgpt

[obsidian cargo]

It has better context handling then gpt and for coding opus is superior but gpt for 20€ is better deal then claude for 20€

Nevertheless i have max 5x on claude so idc 😫

[teal coral]

yoooo ur rich man

dang

[teal coral]

are you a fullstack dev ?

[obsidian cargo]

Yeah but iam has proven to be quite usefull companion on large projects

Yes my website is made by codex because i needed one and was lazy to do ot

It*

[teal coral]

me, im a do vibe coding

[obsidian cargo]

Learn to do bit of coding yourself as if your app crashes and ai will be down youre going to have a bad tike

[teal coral]

i know that devs hate this

[obsidian cargo]

Time*

[teal coral]

i mean, that's a good idea

[obsidian cargo]

Do whatever you want but you should learn basic skill, like managing a linux machine, basic cybersecutiy

[teal coral]

yes

these times, im bored so im looking for a project to do.Do you have any projects 🙂

(no paid)

just to learn more and more

[obsidian cargo]

Depends on what you want to do

[teal coral]

i have a lot of xP in design (logos, banners , thumbails, UI)

but i want to learn mroe about code and developping

im flexible

[obsidian cargo]

Start at basics, learning basic of language yo uwant to work in core concepts of every programming language

[teal coral]

where can i learn it ?

[teal coral]

.

[obsidian cargo]

Find documentation page for em and read them

[teal coral]

btw W protfolio

[obsidian cargo]

Its AI too ikl be reworking it

[teal coral]

oh alr, and do you have anything to assign me in the design field ?

ig no

[obsidian cargo]

Iam not really a designerz ive started to work with lower level languages like rust

[teal coral]

interesting

[obsidian cargo]

But if you have 0 expirience coding then dont start with rust

[teal coral]

okay

where are you from ?

[civic dust]

its really annoying now that after you make an image, there's no way to look at it full size anymore

[hushed relic]

The images are being generated terribly bad today and I havent been able to find a solution. OpenAI hasn’t made any statement about this. Do they consider this situation normal?

[tidal lance]

Yes

[steady geode]

are you sure you're not on 1.5?

one clear indicator I've noticed is that while generating an image on the website or in the mobile app, if you see a blurred image as it's generating the image then it most likely means you're on 1.5. If you don't see a blurred image happen during the generation process and it takes a little longer than usual, most likely you're on version 2

[analog grail]

Do peoples Spamming in commentary under your picture stuff or Video don't make you feel in Murder Mood ? 🙂 or is just me 🙂 :monokuma:

[haughty burrow]

I hat openAI

[late blade]

Ah! Tuesday, complaining and eating tacos. Just another normal Tuesday. 😌

[indigo seal]

THE NEW IMAGE MODEL IS CRAZY

I've just got access

its soooo good. Beats the hell outta even the old Nano Banana Pro 1, by like 3-4 times. It's good that "holy smokes" factor in image gen that i haven't felt since maybe nbp launch or maybe even the one since gpt-image 1

[limber shoal]

you gotta be more careful with your prompts with the new model

[civic dust]

I'm doing a lot of images right now and keep having to take them to nano banana so I want to try out images 2 🙁

[late blade]

I just got all my images from sora and mapped each image to the corresponding prompt. Now I'm integrating that into my prompt builder dashboard

I am trying to orchestrate this aspects on the dashboard
Natural Language Interfaces (NLI), AI Driven UCD— UI/UX, Named Entity Recognition (NER), and sentiment analytics for a CDS (Creative Decision Support) .

[indigo seal]

yea its gaurdrails are way worse

[analog grail]

They look to try by worse against copyright material but is easy to by pass even chat gpt give the tips when you make character is close in the prompt is just change name the character by a deep description without saying the name and is work ^^

[late blade]

Are we enjoying complaining with tacos?

[lapis cloud]

hi，does OpenAI release image-2 API？

[analog grail]

dammn The bug about Mosaic render is back again the hell with this.

[rigid bough]

they will but it's not out yet

[steady geode]

"We experienced an error when generating images." - that's a new one I've not seen before, probably working on the launch I guess

[sick cargo]

my imagegens on sora.com go >100% and then "there was an unexpected error running this prompt" 😂

[brave ice]

Glitchy as all get out today with image 2

[late blade]

New Model... I hope complains get better too! 😏

[analog grail]

i will no make new picture from today until the new model is fully operational because the render no good right now

[indigo seal]

OpenAI just teased it

[steady geode]

Greg Brockman confirmed it's a new image model - https://x.com/gdb/status/2046632580527554572

just hope that once it's officially released that Codex will also be upgraded to use v2 shortly after 🙂

[analog grail]

hmm so is because the New Model is release we get sometime issue with Blurrred pictures then ?

[obsidian cargo]

Why do you use image gen in codex?

[steady geode]

frontend UI tasks

at the moment I use ChatGPT though, since Codex is on 1.5

[gilded osprey]

I assume noon is at 10 pm where I live

[steady geode]

Its about 40 min from now

[wicked plinth]

Kinda hard to get excited about a new model when they’ll likely just nix it within 12 months 🫤

[daring tapir]

HI

[storm cradle]

We NEED SORA 3

[wraith oracle]

Oh they just moved image disscugions to a new section

[sand herald]

why hello there

[smoky sable]

Seems like the rollout is almost finished.

[velvet rampart]

Yeps, new Image 2.0 is launching. They renamed discord chanels.

[obsidian cargo]

Dang it ive reached me limit for today

on plus iam 50€ short of pro lol

[muted moss]

Oh cool we’ll get to use it today?

[obsidian cargo]

its already out since morning brochacho 🙏

[celest anvil]

Is there a way to have it make multiple variations at once?

[gloomy rock]

Hello there guys

[tranquil moat]

how do I know if my chat is using the new model?

is it rolled out everywhere already?

[hollow scroll]

guys are your images also trippy

[celest anvil]

my images are banger, can finally ditch claude for good, god bless

[hollow scroll]

for some reason my images are like off 😭

like they are trippy af

[lunar jackal]

were you using claude for image generation? somehow?

🧐

[celest anvil]

for frontend design

[vale coral]

Is ChatGPT still really the only web interface they give for image gen? Nothing like the old original Sora 1?

[celest anvil]

now you just ask image gen to generate it and it will happily code it for u

[lunar jackal]

completely different workflows. image gen won't "code" it for you

it will generate a static image

[celest anvil]

obviously it wont, but codex will

[topaz gust]

Wow this is amazing

[lunar jackal]

so will claude, even better

for now.

[scarlet ledge]

drop gpt-5.5 instead of image genration

[celest anvil]

why would I use an inferior and more expensive model

[sand herald]

later this week.. that or Spud will drop

[split gate]

Does anyone know the limit of image gens with chatgpt go vs plus?

[lunar jackal]

if you think that why were you using Claude in the first place, and why did it take an image model to "finally ditch it"? doesn't make sense

larp

[jagged lotus]

They didn't mention a limit. Just said that it's limited to paid accounts.

[celest anvil]

because codex sucked at frontend and now it doesnt matter since image gen fills the gap? you are not making any sense

[lunar jackal]

brother good luck with image gen for front end development

lol

[split gate]

I got like 5 gens on my free count with img gen 2.0 but I hit the limit, considering upgrading but unsure what tier

[velvet rampart]

At the moment Sora 1 is not working. The generation goes over 100% and fails.

[vale coral]

Sora 1 is going away in a few days anyway I doubt they even hooked it up to the new image generator

[split flint]

What does it even do bro

[obsidian cargo]

is GPT image 2 in api now

[atomic lantern]

With an image generation model? Frontend design?

[viscid ermine]

Does anyone else get this API issue when trying to generate an image with images attached?
The model 'gpt-image-2' does not support the 'input_fidelity' parameter.

[deft pivot]

I just saw a trailer of the gpt 2.0 and they removed it in a second in this channel

[weary yoke]

Can GPT image 2 be used in custom GPT’s ?

[wise raptor]

that video did not inspire confidence lol

[clever wave]

Mid

[celest anvil]

How about you try to prompt it to make an UI and see what happens

[lofty basin]

I have the same question

[tulip estuary]

if you generate an image in chatgpt is it using the new model already or do i have to configure it?

[rancid mountain]

I didn’t have time to watch the stream, if someone could kindly tell me when it’s dropping I’d be thankful 🙏

[mellow stream]

Already live

[atomic lantern]

Well I can just use something like stitch to actually generate figma files that I can move around the element, then convert to html via figma and use coding AI to make it versatile and usable in a framework?

[trail halo]

But how can you tell if it’s 1 or 2

[harsh zealot]

Wow

[mellow stream]

I’m just so glad we can finally edit the images it was much needed

[long crater]

Try generate 16:9 or 9:16 if you can then its 2

[lofty basin]

Yeah but every feature slowly rolls out to accounts so i don’t think there is a way to tell which version you are using

[atomic lantern]

Why would anyone generate frontend UI with an image model?

[lofty basin]

Oh interesting

[woeful fable]

we had to sacrifice Sora to get better Images

[celest anvil]

That is assuming you know what you exactly want, if I know what I exactly want then there is no point in using any of the design generation tools anyway

[next nest]

lmao

[tight atlas]

image gen to html works incredibly well when you’re trying to brainstorm ui ideas, even stitch has this now

[lunar jackal]

no? both generate from scratch if you want, but one is dynamic and the other is static

[long crater]

Also run the same prompt you did in 1.5, it its look different then its 2

[hallow vapor]

anyone tried - is this better or worse than nano banana pro

[lofty basin]

Imma try that

[next nest]

is this already out?

[sick cargo]

yes

[red prairie]

nice trailer

[oblique junco]

someone show
me an example

[eager quest]

You can't be Gemini

[sharp palm]

This isn’t for 2.0?

This fir Sora image

[storm silo]

personally i kinda prefer gpt's thing,no watermark thats bothering the image,i think when i tried it was with nanobanana 2 and gpt did it better with no watermark

[nimble scarab]

is it better than gemini 2 thing?

[split cosmos]

trailer made me breathe a sigh of relief that AI images are still obvious enough that I can tell them apart

[atomic lantern]

nano banana doesn't have any watermark if you use the AI studio version

[vivid zodiac]

How is this about character consistency? From a base?

[woeful fable]

I think Sora will come back, but locked behind a paywall like with Seedance

But they’re gonna wait until they come out with Sora 3, i think

[plain moth]

Is it out?

[storm silo]

didnt know that much,thx,though i still do prefer gpt image 2 for it

[whole stirrup]

ya

[next nest]

meh i don't think so they already loose money in text and image models
video models are exponetially more expensive

[split cosmos]

still has synthID im pretty sure but that's like the simplest thing ever to remove

[mellow stream]

I’ll add as I now realize, if you want it to use full power you have to generate the image first and then click the 3 dots and click use thinking

[woeful fable]

Then they go bankrupt and Microsoft acquires them, I guess

[whole stirrup]

Download an app to remove

[sick cargo]

image 2 still can't draw accurate clocks 🙂 but seems to be pretty neat otherwise

[magic sparrow]

For GPT image 2 in the chatgpt, what quality setting is used

[split cosmos]

literally just use a color noise filter on paintdotnet or something on the lowest setting

[tulip estuary]

i dont see that, i believe hes already thinking

[atomic lantern]

I wonder if it's possible to actually train AI or create a bunch of agents that can use various tools to actually operate Photoshop or Lightroom (much easier) to edit images, that would be a breakthrough

[mellow stream]

Interesting why mine didn’t do it then

[nimble scarab]

how is it better?

[digital tartan]

sooo how do we get it to 'think' like in their demo?

[split cosmos]

god logistically, morally, and technically speaking that's a can of worms i doubt anyone wants to open

[digital tartan]

(I'm on Pro lite)

[trail halo]

Is it still limited to 3 images a day for free users?

[oblique junco]

anyone else didnt get copyright violation yet

[whole stirrup]

I tested it a week ago, so I can say that it is better, especially in game styles.

[whole stirrup]

For example roblox or minecraft

[obsidian cargo]

what is pricing for gpt image 2

[paper silo]

is it out?

[rigid echo]

was transparent background support removed in 2.0?

[spiral condor]

When it makes ultra wide images, they are terrible, blotchy, and have square looking blobs on them.

[atomic lantern]

OpenClaw is already doing something similar but we are still extremely early in the stage

the only thing similar to my idea rn is that Chinese Kimi Model, which operate on WPS (Chinese MS Office) that can create power point slides by adding elements and change the positions of elements, adding animations and illustrations etc etc

the result is you get a real PPT file and you can edit it anywhere you like, similar to stitch

similar idea needs to be applied to real creative industry software
like Photoshop, premier, lightroom, aftereffect etc

[whole stirrup]

#images-canvas message

[zenith igloo]

I have waited 2.0 so long 💯

[atomic lantern]

the current image generation model is nothing but a toy, for memes and rough sketches, you can't granularly control anything or edit anything with just one text prompt or multiple text prompts

[obsidian cargo]

what is this pricing i dont understand it

[atomic lantern]

the model needs to create like a real psd file, with layers, texts, shapes, and effects etc, so people can edit it after generation

[whole stirrup]

Don't tell anyone the secret

||You can ask to generate 10 images in a row in 1 minute||

[dusky terrace]

Is image 2.0 have character fidelity ? (Put my cat as a reference and still have my cat as the output but different scene ?)

[obsidian cargo]

its better then image 1.5

[acoustic carbon]

I'd throw stacks of money into my monitor for this..

[storm silo]

from what i saw its pretty good at keeping the same thing,person as is if you ask(maybe,i know i didnt have to ask)

[acoustic carbon]

I can't generate a single image, did it crash

[storm silo]

try refreshing,i know on phone its a bit buggy

[split hare]

yes.. just did it and it worked well with guinea pigs, copied their style pretty well! But I prompted to make it cartoonish style, so not an exact 1:1 realistic image.

[wild ridge]

This is peak ngl

[digital tartan]

anyone else not seeing this option?

[dusky terrace]

Yeah that’s my use case I just care about having the « same » cat or dog even. Not realistic style (cartoonish, western, Japan etc…)

[wild ridge]

I tried doing the screenshot stuff like everyone else was and it works really really good its crazy

[storm silo]

so,i tried making an image,it said no cuz "copyrights" then made a suggestion of what it could do,then tells me that the suggestion that GPT gave is still against the rules....

[verbal sequoia]

Get notified for all OpenAI announcements!

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

[cloud flare]

Is it just the default model
Do I have to do anything special

[whole stirrup]

Use Instant model

[storm silo]

actually its already on instant....

[wraith oracle]

Yea I generated a 80s cel animation frame and when they did it with the exact same prompt, it refuesd, so gardrails are inconsistant

[obtuse brook]

Does anybody have a recap of the livestream? I wasn’t able to watch because of a meeting!

[sick cargo]

whats going on with the gpt-image-2 API?
getting error "string too long. Expected a string with maximum length 32000, but got a string with length 38027 instead" when sending 10004 characters.

[whole stirrup]

Ask his pass search

[wraith oracle]

Ok so basically, GPT Image 2 can do web search to grab images to ground, and for paid users, you can use the thinking model for better results

[sick flax]

Not in Android app but available on desktop

[obtuse brook]

Amazing! Thank you, do you use gpt image 2 just like normal in the app previously?

[wraith oracle]

yes

[amber raptor]

to add to that, it somehow pulled off generating a QR code to the openai website during the livestream. like an actual qr code you can scan except it was a small part of a larger generated image

i'm currently asking it to revamp my logo

[storm silo]

i just realized that iw as using instant model and not the thinking model

[amber raptor]

it's also the one in my pfp. that's the old one

[jolly mesa]

How’d I use the thinking model on iPhone? I do have premium

[oblique nymph]

how can i know if im using image 2? or is it just image 2 for everything now?

[jolly mesa]

Everything I believe

[thin wagon]

I asked it to make a teardown of Nintendo ds components. I don’t think it’s good at looking up multiple things

[weary yoke]

It seems custom GPT can’t access GPT image 2

But works in normal chat

[wraith oracle]

Really? Intresting

[rigid echo]

im very surprised they removed transparent background support

[rotund juniper]

Me and Sam decided it had to go

[atomic lantern]

not yet I think

[magic sparrow]

What quality parameter does gpt image 2 use inside chatGPT? Low, medium, or high

[wraith oracle]

I find that telling GPT Image to search the web for refrences when thinking witll have good results

[amber raptor]

probably varies a lot

[magic sparrow]

How to make it always high quality

[atomic lantern]

my guess is always low

[plain moth]

I don't like it too much

[indigo kelp]

I just came to ask this 🙁 I just discovered it yesterday and I was blown away. How can they remove it. Does anyone know alternatives for transparency?

[somber quest]

Now don’t work.

[amber raptor]

not sure you can force it in the webapp but use it through the api and set it to high there

[wild ridge]

Just remove the background urself or use a different website to do it for you

[kind mountain]

me neither - for my use case it seems to produce far worse images 🙁 Is there a way to switch back to 1.5 somehow?

[indigo kelp]

Why having an extra step when the model can do it :/

[whole stirrup]

I hate AI
||By the way, I've generated over 5,000 images in the last 3 months in the gpt chat.||

[rigid echo]

definitely not the same as having it generated with no background

in terms of quality

[wild ridge]

Good point

I definitely enjoyed it being able to generate with transparent backgrounds i dont know why they would stop supporting that.

[rigid echo]

its still in the api, but it gives an error saying it doesnt support it, hopefully it does eventually

[indigo kelp]

I tried it on the playground and there it still works with the gpt-image-1.5 model -> https://platform.openai.com/playground/images

[wild ridge]

They probably will bring it back if enough people request it

We can only hope

[amber raptor]

i'm 99% sure gpt-image has never supported 4 channel generation, it generated an image and removed the background after, but before outputting the image to the user. 99% sure

there are artifacts as if it didn't do a very good job at removing the background in some images when you use the older models, i haven't tried it with gpt-image-2 yet

[indigo kelp]

interesting.

[rigid echo]

i wonder what they used if they did it as a second pass

[magic sparrow]

What’s the gpt image 2 limit for each plan?

[atomic lantern]

probably something with the background remover tool or the agent is just not calling the tool

[sour oracle]

So is this new model paid only for now?

[violet owl]

What are you talking about this means you are still using the old version of image gen this model doesnot supports transparent background.

[amber raptor]

the api costs are the same as gpt-image-1.5 so i think the limits are the same 😄

[whole stirrup]

It's extended

[sour oracle]

So it's 5 images per 24 hours on the free plan?

[violet owl]

I will say this is the biggest setback for me ,. This new model does not support transparent background.

[wraith oracle]

It dose, I made a man with a transparent background just now

[magic sparrow]

What’s the limit of that on plus, pro, free, and such plans

[amber raptor]

firstly you might have misread my message and secondly gpt-image-2 accidentally generated a transparent image when it wasn't supposed to https://chatgpt.com/share/69e7deb5-16fc-8391-a87a-ed7950113fbf

[soft yoke]

Pretty lame, you can't create transperant images.

[amber raptor]

it didn't do it very well lol

[violet owl]

Can't be api docs says "gpt-image-2 doesn’t currently support transparent backgrounds. Requests with background: "transparent" aren’t supported for this model." And If it is then it's either you are still using old model or may be using an background remover on top .

[violet owl]

Yeah

[amber raptor]

honestly if they could reveal a bit more about the models it would be great

[violet owl]

Like what

[thin wagon]

Has anyone gotten functional sprite sheets

[sour oracle]

Pricing like what @magic sparrow and I are asking

And limits

[amber raptor]

whether or not any of 1, 1.5 and 2 has native 4 channel generation for example. because now i'm second guessing myself

[atomic lantern]

not close to there yet
you mean actual animation like sprite sheets?

[thin wagon]

Yeah, tried a simple walking animation and everything looked borked

[kind mountain]

can I switch back/choose 1.5 in chat? This switch to 2.0 is messing with my project, since I can't get the results I got with 1.5 🙁

[solid phoenix]

🐍

[thin wagon]

Looks ok until you play it

At least the character is proportioned right now

[solid phoenix]

am playing now, seems fun so far

[tranquil heath]

Interesting. How much does the api cost per 1024x1024 image gen for image gen 2?

[wraith oracle]

Hmmm I wonder if they made a custom GPT for 1.5

[amber raptor]

here are my settings: quality: high
size: auto
text input: 21t
image input: 3,042t
output: 8,322t

here is the price per million tokens:
gpt-image-2: input: $8.00, cache: $2.00, output: $30.00

i haven't calculated what a single image costs but it's decent i think

[violet owl]

Ok ,custome aspect ratios are cool, so we can now create almost any aspect-ratio images, like hundreds of aspect ratios. That's really cool.

[amber raptor]

$0.27 usd per image

[oblique raft]

i cannot see any ratio selection in my chatgpt

[wraith oracle]

prompt for aspect ratio and then when the image is generated, click on the image and find the aspect ratio on the top right

[oblique raft]

for me it's not there.... maybe i do something wrong. thanks anyway

[magic sparrow]

Hi, i know you from ZBD!

[digital tartan]

I don't think everything that we saw in their livestream is fully rolled out yet

[sour oracle]

Aye, I hardly talk there much now

[kind tulip]

are there different levels of quality with images 2.0?

[sour oracle]

And Fumbs too

[magic sparrow]

You still admin?

[sour oracle]

Yeah sure am

[obsidian cargo]

Yes

Low medium high

[deft flame]

How do you get access to them

[obsidian cargo]

For anime ive found medium to work best

[obsidian cargo]

API

[deft flame]

Api?

[novel orbit]

What better about it?

[undone pumice]

Images 2 0

[novel orbit]

Hi nasr

Ur from earnapp

[undone pumice]

Hey bro

[undone pumice]

Yes

I'm everywhere hahahahha did u see chatgpt images 2.0

[novel orbit]

No not yet

Btw what’s the difference?

[undone pumice]

I didnt try it yet

[novel orbit]

Better character consistency and more ratios?

[novel orbit]

I’m thinking of using api to make images and stitch a video

Like making an automated tool that does that

[undone pumice]

Yzah memes or even mangas

[novel orbit]

API costs bout to go crazy

[undone pumice]

Interesting

[novel orbit]

Uh hu

[undone pumice]

Very interesting approach, i might use it to help me on scientific figures

[novel orbit]

My job of being a thumbnail artist is most likely replaced 😭lol

[undone pumice]

I will try it soon on graphical abstract

[novel orbit]

Alright, good luck

[gritty burrow]

Actually so shocked by how well it makes furry reference images