images-discussions

[undone pumice]

Thank you, you too bro

[outer spade]

I've noticed it saturates the image more with each turn

[wraith oracle]

The fact that the thinking model can grab references automatically means i dont have to get that many references anymore to get a more accurate style

[echo sapphire]

Honestly not better than nano banana lol

[hot rain]

Still worse than Nano Banana by orders of magnitude, still more censored than Nano Banana by orders of magnitude.

I really loved GPT, it sucks that it seems to be permanently behind the curve now.

[wraith oracle]

Meh they kinda the same level

[echo sapphire]

It doesn't understand like products if you wanna reference those

yes it can generate new images well with text

but it doesn't understand images lol

[hardy tartan]

is there any way to tell what image model is being used in the chatGPT app on desktop?

[wraith oracle]

Generate a screenshot of ChatGPT and see if it's accurate

[hardy tartan]

I get the feeling I am still seeing 1.5 because it seems roughly the same quality and it refuses to do multiple images at once

[violet ether]

Franchise prompts are blocked now, funs over

[hot rain]

Pretty sure they've been blocked for the most part for ages on GPT, it's a severely censored image gen lol. Gemini is wild west though.

[wraith oracle]

Yea some prompts were blocked while they worked for me, like i asaked it to make a 80s cel animation frame and it worked fine but when someone else did, it refuesed, they retried it a few times and it worked, so inconsistent

[hot rain]

You'd think OpenAI would stop being so "pRoteCT dA chIlDrEn nD dA coPyRight HolDERz" crazed given how Gemini/Claude are gaining ground.

[hardy tartan]

is 2.0 supposed to be abel to generate multiple images at once? It looked like they did this in the demo in the live stream?

I am prompting for this and it just keeps generating a single image with panels

[civic dust]

if I'm on 2.0 now the first image I made because last night's one had an error made the same error

[jagged lotus]

Does anyone know if an API call using gpt-image-2 automatically use the thinking model?

[echo sapphire]

ahahah what about seadream lol

[amber raptor]

it told me it could generate up to 10 images, i accidentally prompted it in such a way it attempted 16 images and failed (which is fair lol)

[hot rain]

That I don't personally have experience with. I sort of stopped searching after the last Nano Banana update, it does 95% of what I want.

[crude patrol]

hi}

[hardy tartan]

oh wow, so the desktop app for me was definitely still using 1.5. Just tried on the website and its clearly using 2.0. Difference is very dramatic

[civic dust]

ok so you cant set the image size anymore, you have to resample

[wraith oracle]

Acutally you can prompt the aspect ratio

[civic dust]

aspect yes, size no

[wraith oracle]

Oh

[civic dust]

I've been running a 1024x1024 pipeline for ages and now I can't get 1024 anymore

really odd that they've chosen 1254 as their default

[digital tartan]

I haven't found any way to do this

has anybody else?

[quartz dagger]

Seems to be broken anyways atmo

Too much traffic probably

[soft yoke]

I was able to do it, I gave it a whole story for a manga and it created 4 images.

[amber raptor]

"actually wait, you can generate up to 10 images at once, how about you generate a single texture with all the maps i might need in blender 5.0.0? diffuse, roughness etc"

this worked

[soft yoke]

https://chatgpt.com/share/69e7ef93-a7c8-83ea-bd41-5ae0703afa23

[hardy sparrow]

“Sorry, but the request may violate the safeguards we have in place regarding similarity to third-party content. If you believe this is a mistake, you can try again or edit your request.” anyone else getting errors with the image gen too

[thin wagon]

Don’t they always start with loose regulations then zip it up so they can ride the hype cycle

[hardy sparrow]

its just annoying tho

like censor it from the start if they want it so bad

[civic dust]

dear god I asked it to make a well-used whiskey flask and it gave me an entire scene that looked like it had been run over by a truck and then used as hollywood props in a post apocalyptic fantasy world. I just wanted not-perfect

[hardy sparrow]

what its generating coca cola stuff for me now so weird

[timid knoll]

LOL

[surreal onyx]

Has anyone else noticed that asking to make the image transparent reverts back to the old image model and the entire image completely deforms?

[vital gull]

✨ hands✨ 🥲

[thin wagon]

What is the rate limit right now

I feel like I got cut off hard

Nothing for the last 15 minutes

[wraith oracle]

You know what i just relized, with search grounding now, this new image model can acutally be a visual tool with the thinking mode for me

[pastel delta]

How's everyone doing with it

[vital gull]

in chat

[wraith oracle]

Oh Im doing well with it, find the thinking model to be really good

[amber raptor]

where can i safely talk about a user without anyone else seeing their bio?

[civic dust]

wait, you can post images in images discussion now?

[amber raptor]

if you're staff i think

[agile peak]

There's seems a bug in v2 image, random diffusion artifacts.. at some parts was perfectly fine earlier this morning here

Especially when you try generate Sands or Pixel for instance, which happens more frequently

[civic dust]

that's just cheating

[wintry gull]

2.0 is so bad Im in disbelief. It adds these weird textures on everything I make now.

[civic dust]

1.5 was so very clean its a huge change

[agile peak]

It was perfect and feel like massive leap, at least before official announcement.. i've tested on lmarena back then week ago...

But then again, there are 3 different image_gen v2 models though. Each has a different taste in visual direction and style approaches.

[buoyant otter]

is it just me or do the images look lower resolution than before

like the image content is good, it's just blurry

[surreal onyx]

Likely because everything is locked at 1k resolution

[languid geyser]

Have they mentioned limits on generations?

[azure arrow]

Model goated so far

[wraith oracle]

I agree

[bleak skiff]

Is it live now or was that just an announcement?

[wraith oracle]

It;s live now, go try it

[surreal onyx]

So far it just seems like a downgraded nano banana 2 visually due to resolution. Transparent backgrounds don't work and completely blow up images. Whatever content filter is being used is outrageously aggressive as well, good luck manhwa/manga creators.

[wraith oracle]

Maybe a quantized nano banana 2

[thin wagon]

Nano banana always had an uncanny effect on text, diagrams, and wasn’t able to do a lot of edits without losing context

I went to a fair and I could very easily pick out signs of street vendors directly made with AI

[harsh zealot]

Omg

[thin wagon]

This, not so much anymore

[obtuse brook]

Does creating different images in the same chatgpt chat effect eachother? I have found previously that making different images in the same chat one after another vs making a new chat for each image started to blend / lock chatgpt into a visual direction; does this happen with V2 or is this even a thing? I may simply be tripping

[mossy ingot]

Yes, generating images in the same chat window builds context, and it can start to blur everything together. The way I like to think of it is that you should only ask for multiple generations if you’re editing the same image and working within that framework consistently. Otherwise, it’s best to start a new chat for each image you’re working on.

However, if you're working on something like comics or character consistency, you'd probably want to stay in that chat window for that context with each new image...

[wanton fractal]

When I use the ChatGPT for Mac app and ask it to build an infographic, it's still a bunch of disjointed absolute nonsense. Is there a way to force it to use image-2?

[elfin robin]

Y does it seem like gpt 2.0 is more censored than banana pro

[mossy ingot]

ChatGPT is always censored to hell and you shouldn't expect anything less, tbh.

Also, it's very common for them to be relaxed in the beginning and get more strict with what you can generate over time.

Never...ever...get attached to what you can generate in the first week of a model release. 😂

[hot rain]

RIP The wild west week of Sora where I could have Tommy Vercetti from GTA Vice City in a shootout with Solid Snake.

[mossy ingot]

^Exactly.

[sour oracle]

Something I don't get now, why sora shutting down and this image 2.0 generation just come out?

[wanton fractal]

Okay... the exact same prompt and settings on chatgpt.com on my account, and ChatGPT on my Mac, you get completely different results. Same model and settings checked... um... what?

[hot rain]

Idk the numbers myself and am obviously a layman, but generating an image and generating a video are two vastly different tiers of expenditure.

Like, many people still have to buffer 1080p+ videos when browsing online. Basically nobody is waiting an hour for their internet to load a picture anymore. Same concept.

[mossy ingot]

What are you attempting to do and how are you checking what is which model?

[sour oracle]

I get that behaviour across 2 different chats online

[obtuse brook]

Wonderful break down of it, thank you Franco!

[wanton fractal]

"Create an infographic about X", both set to 5.4 Thinking. chatgpt.com is clearly using a dramatically better image model than the Mac app

[sour oracle]

While fair, Sora generates images as well

[mossy ingot]

Np

[hot rain]

Sora's image gen was just GPT's image gen essentially. Removing the Sora website is one less address for them to maintain while image gen functionality remains in GPT.

[wanton fractal]

mac app results are literally unintelligble nonsense, chatgpt.com results are perfect

[wraith oracle]

Yep and I hered the API has been disabled in the sora website so you cant generate images anymore there

[sour oracle]

Huh, I was generating images about 12 hours ago

[mossy ingot]

I'm going to give you my best guess of what's happening here based on experience. The web application and the macOS/Windows applications aren't always updated at the same time, so you can have a new model on web and not on the macOS app. I've experienced that myself and had to wait for GPT-5 to hit my app despite having access to the model in the selector through the web version. Also, you should check to see if your app has any updates or is even automatically updated (if you haven't already).

[wraith oracle]

it broke because of the new model

[hot rain]

I imagine the same way rollouts of new features come to different groups in waves, shutdowns probably have the same process.

[sour oracle]

Ahhh ok so I got in just before it broke I guess

[rigid bough]

what broke?

[wraith oracle]

The Image API via Sora 1 website, it broke when the new mdoel came out

[rigid bough]

oh, that's been disabled

[wraith oracle]

Yep, yall will never generate images on the Sora website again

[sour oracle]

😭

Lol

[civic dust]

I'm beyond disapppointed in images 2, guess I'll watch discord until it gets fixed

[wraith oracle]

It's ok, man. I tested the image generator in other areas and was pretty impressed with it.

[civic dust]

I'm off in nb2 for the day until chatgpt works again 🙁

[vivid zodiac]

So I’ve been playing with the new image 2.0 and it’s pretty good. It can finally do landscape at least.

And I’ve been able to get consistent CGs

[long crater]

Hi Image Gurus do we have 3 images (previews) per generation beside the main one ? At one point I managed to click two of those 2 small ones on the right and get both generated & appeared on 'My Images' album but only one version appears on the chat.Other than that one time, I only get 1 image.It would be great if we get many images per click. Im on Thinking mode on Android

[tender aspen]

Image 2 is available on chatgpt plus? 4k resolution?

[west glacier]

画一个风景图

[inner axle]

How is everyone liking images 2.0

[civic dust]

I'm waiting for the bugs to get fixed, so its hard to comment when its this bad right now

[dawn hearth]

https://chatgpt.com/c/69e82400-94f0-83ea-b11e-3047ea39388d

https://chatgpt.com/c/69e80988-07d4-83ea-8df4-9145a7b3b24b

The model seems to have a strong racial bias when race is not specified

please don't come at me with "if it was all white people or east asian people would I even notice?", yes I would.

[civic dust]

oh yeah its always young white people unless you force it

[urban goblet]

Where are all of our generated images from the old Sora website? I don't see them in my ChatGPT Library.

[dawn hearth]

did you look at the images?

[civic dust]

no, they were not valid links

[wild ridge]

personally ive been loving it but ive noticed in some cases it isnt that great at editing images. sometimes it is good sometimes it just sucks

[inner axle]

Yeah, it still might need some polishing

[wild ridge]

yeah im sure they will have it fixed very soon

[dawn hearth]

I posted the images on the image-canvas chat

[velvet rampart]

Is image 1.5 kept as a separate CustomGPT as previous image models have been?

After yesterday’s publication of the new model, they have basically stopped Sora1 from functioning. Yes, you can enter a prompt, and it will start developing it. But, no, it does not generate but rather informs that a mistake occured and no picture was generated. I tested this with ever sure prompt Blue rose.

[devout cypress]

New image model but will it dethrone nano banana tho

[civic dust]

not with the bugs it released with, but see after they fix them

[sour oracle]

So how do I know if I'm using 2.0 in the chatgpt app?

[devout cypress]

Bugs?

[timber fern]

how can we make gpt 5.4 read images better ?

[ivory kernel]

Am I the only one who gets a lot of visual errors when I try to put people in the image?

[civic dust]

yeah the repeating texture error is awful, makes it unuseable

[devout cypress]

I always thought if something like that happens it just means the model is bad

[civic dust]

then this v2 is extremely bad. I'm a bit over hitting guardrails every second prompt myself

I think once you've got guardrailed once or twice everything you do is immediately slammed down. I managed to get it to draw a picture in a new chat

[river dome]

Unnecessary, overprotective guardrails always immediately shutdown any interest I have in these image models. I had a photo of my living room, and I was trying to play around with having it rearrange the room, change furniture, etc. and my daughters Minnie Mouse toy was occupying an entire 1% max of the image in the corner, and had nothing to do with what I was prompting.

Apparently this is the gravest violation of their policies. It’s just disappointing how easily you can get shutdown. Don’t even get me started on wanting to make fun personal picture that puts your own kid in a silly situation that they would think is cool. I haven’t even tried doing something as offensive as making the floor look like lava while my 3 year old stands on the coffee table pretending that the floor is lava in a while now, because that kind of thing is apparently a no-go.

[novel tusk]

new image update, in my opinion, completely sucks

I play Sims and enjoy taking screenshots of scenes I create with my Sims and recreating them using ChatGPT to see how each Sim could look as a realistic/stylized realistic person. But now, it's taking every photo I give it and just "cleaning it up" no matter what I do, so it still looks like a render and not like a person. 🙁

[elfin prism]

I was today years old when I found out that the standard thinking mode for chatgpt lets you generate more than one image at the same time from one prompt, which can specify what you want to see in each individual image

[pale ermine]

Where do you find the new image thing? Didn’t watch the livestream.

[wraith oracle]

press the plus button and press the create image button or go the sidebar to the images panel

[gilded elk]

@pale ermine in chatgpt

[mossy ingot]

Check #images-canvas, want you to try a prompt.

[civic dust]

don't even need to do that, just say draw me a picture of ...

[pale ermine]

Thanks

[civic dust]

just don't ask it for animals, grass, trees, dirt or gravel

[clear laurel]

im trying to try the new 2.0, its been generating for over 40 minutes now and still nothing. anyone having the same issue?

[mossy ingot]

Try refreshing the page?

[clear laurel]

been doing that, still no luck

[mossy ingot]

I have no idea then, sorry.

[alpine brook]

someone is brasilian? o que acharam do image 2.0? opinião sincera

[sour oracle]

Force refresh the page? Shift + F5

[clear laurel]

im regenerating the new one, and it completed the work fast. the previous one i generate it inside a project folder, maybe thats the issue

[flint nexus]

Pretty good result, granted the prompt was rather long since it was the entire notes on a story character of mine lol

[orchid kayak]

When are we going to get the pricing per 1m token updated on their official website?

[next fog]

Do you guys notice unnecessary noises and particles like artifacts? And what's with the censorship? Every other word triggers the guardrails. Not gonna beat nano banana this way

[steady geode]

depends on what you're trying to do I guess, I've not noticed censorship issues with the kind of work I'm giving it so far, generally seems fine. Bear in mind it might not be the model itself which is complaining about some guardrail, but potentially ChatGPT might have tighter system prompting in place

[next fog]

Any solutions?

[steady geode]

if you're using it via ChatGPT, maybe try it via API and see if you bump into the same guardrails first?

if you use something like openrouter then it's probably on there

[analog grail]

APi is better browser for make images ?

[thorn kernel]

to be honest

gpt 2.0 still doesnt match nano banna pro nor nano banna 2

[next fog]

Not with this much censoring and noise artifacts.

[thorn kernel]

eh in terms of quality and etc i still find nano banna pro/nano banna 2 is far superrior and its still the best ai image generator by far right now

[analog grail]

never try Nano Banna is good for make crossover stuff and know no usual Characters ?

[supple anchor]

I think the image isn't clear enough.

[analog grail]

and the guardail how is with Nano Banna ?

[next fog]

Nano banana pro is way better, in terms of both censoring and quality.

[next fog]

Gpt image 2 is almost close to nano banana 2 but can't compete with nano banana pro.

[analog grail]

is work for character like Rinoa Heartilly and Junko Enoshima and for sensual stuff ?

[civic dust]

I hit guardrails a lot less often with nano banana. I've only hit it about 3 times total I think, I've hit guardrails on chatgpt a heap of times just today

[analog grail]

Now now i'm curious i should give a try then upupupu

[civic dust]

once was for "paint the town red" which had a prompt with red paint that was very blood adjacent on the second generation, once for "roll up your sleeves" that just happened to have very veiny arms and I can see what it might have mixed that up with, and once for no apparent reason

I make photorealistic scenes of fairly ordinary things though

[obsidian cargo]

Gpt image 2 is far better for ART like anime, cartoons or so compared to nb 2 or nbp

[glad fable]

not when all the ouputs have the mosaic overlay bug

[obsidian cargo]

I dont have that issue

[dense delta]

anyone else having issues with sora?

[civic dust]

yeah enough of us do to not think much of this update

[obsidian cargo]

I believe the more loaded regions are doing it due to load

[civic dust]

asia?

[glad fable]

but I do think the update is great besides that, and once that's fixed, smooth sailing

[obsidian cargo]

Are asking me if iam in asia?

[civic dust]

yeah I'm having to route around images that trigger the bug until its fixed

[civic dust]

I'm asking if you're suggesting asia is getting texturing issues due to load

[obsidian cargo]

Most likely

[civic dust]

that would be the worst thing ever if we have to wait until night time to use chatgpt. Usually asia is the lowest load cos we are awake when the US and EU are asleep

or at least my bit of asia

[bitter cypress]

so this model is well less censored

[obsidian cargo]

It will calm down with people generating less its only thing of hype

[civic dust]

I would no tthink that load would lead to a massive drop in quality, more rate limiting

[zenith igloo]

did they say this is going to be exclusive for chatgpt or will they impelemnt 2.0 for the image bot in this server?

[obsidian cargo]

It can happen

[obsidian cargo]

They are most likely waiting for people to calm down

[civic dust]

but quality drop makes people regenerate

I've had to do a bunch of images on nano banana today cos the quality is so bad

[obsidian cargo]

Yes there were issues with gpt image 1.5 on release where it didnt folow things correctly but as people calm down for next week or two it got better

[zenith igloo]

same happened with google's nb2, it had issues and many were furious about it

[obsidian cargo]

Oh yeah i forgot about that

[zenith igloo]

when they removed the use of pro model from free users (its available in flow) and nb1 as the primary in gemini (now moved to ai studio where its possible to access)

so yeah, its gonna take time to settle down

[sour oracle]

Holy heck, generated some images on the new version and can see the difference already

[indigo seal]

yup I hate the overlay thing. it only seems to arise for me when i attach even one picture and if i had attached a picture earlier on. otherwise the model's nearly perfect

[glad fable]

yeah I was able to compare very well right here #images-canvas message

[analog grail]

Yes indeed actually i had find a temporary fix work until is fixed the tips is just to open a new request for your Images you want: and is work but no stay the same chan otherwhise the next Image generate is Mosaic

[indigo seal]

yea for me the fix has been trying to use less images for references

i haven't had the mosiac problem at all unless i attached images early on or for that image in the chat

[glad fable]

for me it happens even in fresh new chats, can even occur with mere text prompts or image refs

but things like real life photorealism I haven't had issues with

[analog grail]

Yes is can still happen too isn't 100% Work ^^ but with some retry you get what you want

because ask in the same page ChatGpt to fix it no work well

[dusty pecan]

The "safety" filters are still too restrictive. Can't even depict teen level romance or violence.

Can't create a manga or manhawa or comic or realistic comic for example.

[wraith oracle]

We finally have a prompting guide for GPT Image 2: https://developers.openai.com/cookbook/examples/multimodal/image-gen-models-prompting-guide

[fading steeple]

.

[wraith oracle]

?

[civic elbow]

anyone getting the new UI glitch with the images 2 launch

where it “thinks” and displays the image generation animation despite an irrelevant prompt or query

[worthy apex]

i had

and have now

[civic elbow]

so your getting it too

[worthy apex]

check the status

[civic elbow]

how do you report issues with triggers and stuff where do u go

[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.

[worthy apex]

OpenAI support

[worthy apex]

https://help.openai.com/en

[small vale]

So, my current thoughts on Image 2 so far.

Pros:

• generally does a better job at capturing the instructions when drawing.
• generally does a better job at drawing over the attached image and capturing the art style of the attached image.

Cons:

• image tends to have some kind of "static" (small lines, dots, cracks, etc.) on them; this is more noticeable when you ask ChatGPT Images 2 to generate an original image instead of using an attached image as a template. Hoping this gets smoothed out later on.

[glad fable]

is what ur describing in cons similar to this?
#images-canvas message

[small vale]

In a way. To be fair, though, the images are usually best viewed when zoomed in.

[regal burrow]

Am I the only one that feels like the new image generation model is a massive downgrade compared to the previous one
I used the tool almost every day
The previous model was consistent from the first try and the quality was amazing
But the new model is not consistent at all and keeps giving me pixelated results and messes up body proportions

Is there a way to switch to the old image generator model until this gets fixed.. I need to use the previous model urgently?

[analog grail]

Yes but that no the issue the news models is because too much user use the new model since is freshly release so too much overload^^

[orchid kayak]

When do we get updated token prices??? for gpt image 2

[regal burrow]

Really? does that really affect the ai ability to stay consistent with my references?

Anyway more importantly how do I switch to the previous model I need to do this asap

[velvet lark]

Anyone that has used image 2 - Have you created turnaround images? If so are they working better with the new image system?

I find that even if you set rules restricting drift and reinterpretation, images still drift sometimes.

[limber shoal]

does error means you got flagged?

[tender aspen]

how do i refer to my uploaded reference images in Image 2? "use the armour from image 1 and the shoes from image 2"?

[wraith oracle]

Yep

[tender aspen]

does anyone know how much images i can generate if i'm on the Plus plan?

[wraith oracle]

144

[tender aspen]

per day?

[civic elbow]

i would disagree with that con it specifically seems to depend on prompt quality and in specific cases, how many additional edits you make to the same image <- this seems to introduce artifacts BUT can be fixed with a follow up on it
check the batch of hyper realistic photos i made it generate in #images-canvas (latest messages)

prompt was centered around hyper realism random camera roll images, no source was given no reference, and all were done with the same prompt IN the same chat

[tender aspen]

for the edits, is it best to download it then reupload and prompt it for changes in a new chat? instead of asking it to edit again and again in the same chat.

[civic elbow]

in token heavy chats potentially yea as your reinjecting the image context helpful even if its back to back in the same thread
in fresh chats its very unlikely you'd have to
in multiple-edit threads though if ur referring to a specific image/image-state it IS more beneficial to save the desired one and just prompt on it like its new than to make it contextually interpret

@tender aspen

i've noticed the last situation can introduce drift which increases significantly the more you validate it (continuing in the same thread despite drift)

[tender aspen]

thanks! one thing i noticed with Image 2 is that if i copy and paste the same prompt in the same chat, to try to create different variations of the image (like Midjourney's 4 variations per prompt), it keeps giving me the same identical image lol

maybe Image 2 has no variance feature

[civic elbow]

i did that exact process to create the image batch in #images-canvas
it depends on prompt and anti-prompt quality

remember that the capacity gpt has performance wise is mostly operator dependant, a lot of issues commonly encountered are simply prompting or personality settings

[vestal pendant]

spent a while talking to GPT yesterday after the image lauch broadcast and we identified some areas where the new image model still fails, but those are hard edge cases and not likely to show up in normal gens. both I and GPT itself are impressed with the new model

[civic elbow]

would you mind turning it into a matrix table and dropping a screenshot in my dms

like the flaw/issue table

[uneven wyvern]

i just tried image 2.0, this is way better than i expected!

[next nest]

yeah its cool af! i liked it too

[civic elbow]

check the images i just sent in #images-canvas

think youll be shocked

[uneven wyvern]

what the hell man, thats insane! couldnt tell they were ai untill i saw you say that. damn..

[civic elbow]

now check the annotated image i just sent, the fact its showing correct, fairly complex ray tracing, is blowing my mind

[orchid kayak]

Are the api on openai platform down?

[acoustic peak]

How are people generating gameplay screenshots? It just says its copyright and makes it look all messed up

[limber shoal]

it is way more censored tho

rn

although sometimes it passes

[thin wagon]

I find this model to be oddly too consistent sometimes

I asked it to generate an open ended character design 15 times and even gave it ideas on how it could add variation

Gave me the same character slightly different 15 times

[limber shoal]

i cant relate coz yesterday i was using a prompt that safely passed all the time
But today it got flagged 8 times lol

[supple quiver]

anyone know when image gen 2.0 will be available in codex cli ? it seems to still use 1.5 ATM

[hot rain]

Lol I just hit my limit for the next hour after generating two images? I have plus.

[wild ridge]

Probably just has a bunch of people using it rn

[hot rain]

You'd think we're enough years into the product existing for them to give a clear "High user traffic" message instead of acting like a few images is a standard cap.

[wild ridge]

Yeah hopefully one day they will add something like that

I think i also just got hit with a limit lol

[mossy ingot]

Did you hit your limits yesterday? If not, you might have used up what was left of your daily limits before it reset for the day.

[hot rain]

Two different times yesterday I reached a limit that reset after an hour (or less) so idk how it works right now.

Felt like after 20 gens or so I'd hit a limit.

[mossy ingot]

From my experience there are two limits:
Use limit: prevents you from mass generating in a short amount of time (typically anywhere from a few minutes to an hour).
Daily limit: will tell you directly you reached your daily limits and give you a much larger time (usually around 24 hours).

[smoky sable]

The new image model sometimes could output the same image twice. Wondering if that's a potential extension for generating two different images.

[civic dust]

Ooh that’s a good workaround for the bug in 2

[opal anchor]

They’re 100% gonna nerf it like how they did to Sora

Especially with copyrighted stuff and people

[vestal pendant]

i tried. but the DM is just sitting there, not sure if it went through or not, do you have your DM's set to friends only?

[wild ridge]

Its like they hate money

[civic dust]

the money would come from copyright deals with the third parties, and would still be nerfed so you can't just make stuff with their content

Also how can you say "hate money" when the alternative is they get sued for this

[proud oracle]

How do i generate more than 1 image at once? Can you not do this anymore?

[terse lark]

GPT Image 2 is scarily good

[proud oracle]

Chatgpt is telling me I can only do 1 at a time? That's really bad if true, especially since it's incredibly slow

[wraith chasm]

I also got hit with a limit

[native night]

It's just subject to throttling during times of heavy usage. It will probably allow concurrent generations on higher tiers like normal once the heightened demand subsides (due to its recent release)

[carmine marsh]

Anyone tried Images 2 to create websites? For me it generates almost the exact same web design no matter what prompt I give it.

[mossy ingot]

Any example prompts you use to create websites?

[carmine marsh]

If you look on X to see the websites people are making with Images 2, they all have the exact same layout. I've tried several different prompts, each one is the same.

[mossy ingot]

What are some of your prompts, though?

[carmine marsh]

Explaining what the company does and the branding and then asking it to generate an image of a web design for it. And then after it gives me the image even after asking it to make a new one with a completely different design it just gives me the same image back but with very slightly differences.

[mossy ingot]

I would be curious to see what would happen if you were more specific of what you wanted rather than leaving the model to generalize a design based on reference images. When the model isn't given enough details, it will prompt it for you and it will likely look very much the same. However, I haven't tested this to see if that's the case, but I can do some testing later and play with it to see.

[civic dust]

I'm finding it so scarily bad I'm off wiring up codex to use 1.5 via api right now, which to be fair I'm doing such repetitive work with it I have already jokingly suggested I need to replace chatgpt with a button. Apparently this triggered me actually doing it

[carmine marsh]

It seems like it's very impressive at first and then I start to notice that each image is basically the same as others of it's type just with different data and colors and slight variations

Like the infographic posters people are making looks very impressive at first and then it seems like all of them are the same

[civic dust]

the repetitive textures plus that make me think someone forgot to randomise a seed somewhere, you're not the first here to comment about lack of variations

[mossy ingot]

All image models do that when you have simple generalized prompts.

[civic dust]

oh god I need to give this thing an api key and all my money now

[carmine marsh]

No even when I give significant variation it still gives me almost the same thing each time.

[mossy ingot]

^Time to give daddy Sam money, Tarrow

[carmine marsh]

In fact, you can even go on X and look at the web designs people are making with it and you will see the hero sections all have the exact same layout. Text on the right, buttons below, dashboard screenshot on the right and then list of companies using the product below

[mossy ingot]

I bet 99% of them say prompting: "Make a digram of..."

Without saying, "Put the text x, and the images y, while..."

[carmine marsh]

Yes, everyone happens to be using the same prompt

[mossy ingot]

I'm not saying everyone is using the same prompt, you're simplfying my statement. I'm saying that people aren't adding complexity to their prompt when asking for a similar theme.

[carmine marsh]

That wouldn't explain how everyone is getting the exact same layout.

[civic dust]

someone in codex discussions here posted screenshots of something he made that looked scary similar to something my codex made, so the lack of variety in UX is pretty baked in here

[mossy ingot]

Check: #images-canvas message

How you use the model matters

[carmine marsh]

search for "website" and you'll see a robot website that is exactly as I described. Text on the left, image on the right, row of info on the bottom

[civic dust]

chatgpt said "click → result → decision → 💸"

[mossy ingot]

#images-canvas message

[mossy ingot]

#images-canvas message

[carmine marsh]

I guess I would rather use the AI to help me come up with a design. If I'm just telling it exactly what to make, that is not nearly as helpful

[mossy ingot]

Then ask the model after your generation for variation and themes. If you're looking for a one shot prompt and ready to go image...I suppose good luck.

As I mentioned earlier, the more generalized your prompt is, the more it defaults to a certain theme.

[carmine marsh]

Per my last message, that is not very useful

We can circle back on this later

[mossy ingot]

No need to, would just be much of the same thing I said earlier, "Most image models work like this."

But that's fair at the end of the day. If you don't find this process useful then it's not a great use case for you.

[bleak skiff]

Omg image 2 is crazy good

[azure arrow]

The biggest improvement for me is that it doesn't make everything look so cinematographic now

Now photos look like actual photos

[civic dust]

daddy Sam just got 4 cents out of me for a test run, I don't think I'm funding any private yachts with this

[mossy ingot]

Not too bad, but it does add up. Haha Good see you got everything setup, however

[civic dust]

yes wired in an entire gpt-images-1 pipeline just to avoid using images 2 in chatgpt, totally what I was planning to spend my morning doing

[mossy ingot]

Let's just hope you don't have the kind of luck where they announce its removal from the API within the next few days haha

[civic dust]

when they fix images 2 I'll just change the model in my config. I was dabbling with the idea of a button anyway cos this is so repetitive

[mossy ingot]

That totally works. Are you able to generate higher quality images for Image 2? I haven't used the API, but I heard you can generate 4k through it?

I am curious if the artifacting exists in their higher quality renders if so

[civic dust]

no idea, I'm just going with 1024x1024 and I have a lot of images to get through so this entire thing has been a huge sideline from my manual pipeline. But hey, I unmanualed it and gave my cc details to the machine

[mossy ingot]

Ah okay

Well, I'm glad you found a workaround despite it not being ideal

[round rune]

I'm sure this has been covered a ton here tonight - but can someone point me to a solution or explanation on the weird hatch / diffusion pattern that seems to happen on lots of my images from gpt-image-2??

[civic dust]

this is remarkably easy to do. Ohai codex make me a thing

[mossy ingot]

Hold on, let me grab the link

[not a fix, but hope at least]

[civic dust]

I literaly just routed around that via the api. They'll fix it, the question is when

[mossy ingot]

Here you go: #images-canvas message

[round rune]

thanks guys!

[mossy ingot]

No problem. However, wish I had a solution

[civic dust]

suddenly having an automated (but $) process for my workflow isn't entirely a bad thing so there is that. I also need to do this for text

[mossy ingot]

With image editing at least, it seems to artifact less with large high quality images, but not sure how consistent that is

[round rune]

routed around it how? like used a different model?

[civic dust]

I'm semi manually batch updating a large corpus of old photos. I just hooked the api into my backend that I'm using to do the update

[round rune]

yeah I tried just about ever api config I can think of - image generation, response api, edits .... still get this weird pattern.

[civic dust]

its images 2 with the error. I'm using images 1 via api

[ripe salmon]

anyone notice midrun nerf? I am getting safety system rejections for "generate a radiograph of Pikachu's head" that prompt was working a couple hours ago and now it won't pass. I am wondering if its dynamically updating its guard rails or I just watched it get nerfed real time

[civic dust]

of course now I'm fiddling with the prompt to get the same results as I got from chat

[vivid zodiac]

Is there an issue where images are getting like, these little artifacts? I have no idea how to not to get them to show up on my images. It's just a lot of white dots, and I'm stuck on how to ask GBT to get rid of them.

[mossy ingot]

#images-canvas message

[tough kindle]

Image 2.0 is fairly broken for UI work, its really hard to convince it to make an image transparent, and any edits tend to revert back to solid or checkered background. From codex its even worse, have not yet managed to get a transparent background from codex at all... sigh one step forwards, one step back.

[civic dust]

psht, guardrail block on mannequins. Bright green plastic people = bad

[mossy ingot]

Check: #images-canvas message

[limber shoal]

happiness is indeed temporary

[lusty pawn]

Anybody notice image 2 has these weird square artifacts or warping?

[fallen lintel]

i want to subscribe to pro x5,does it have a daily limit on generating chatgpt image 2?

[civic dust]

check franco's link above your message

[indigo seal]

yup, known issue, its frustrating

[indigo seal]

im not sure, but i think it's about 100-200 images per day

[fallen lintel]

what about plus subscription?

[fallen lintel]

on the website it says unlimited tho,for pro users

[indigo seal]

idrk

[indigo seal]

ull have to check with ppl who r users. becuz i remember ppl saying pro got nerfed for image gen

[fallen lintel]

okayy thanks

[dusk creek]

Why is the images 2.0 on the $200 pro account so limited? I feel like I'm back to using a free account.
I am lucky to get 20 images created after going through all the trouble to train my chatgpt session and then image prompts just stall and don't even tell me what the problem is. All I can think is that it is some stupid limit.

[civic dust]

I generated 55 images yesterday with no cap, no idea what the limit actually is

[civic dust]

today I'm just generating content violations

[velvet lark]

I had finger distortion on two character designs yesterday.

On regenerations to try to fix the fingers, both characters kept the incorrect finger count but each regeneration would shift the pose or drift the details despite being told to lock in the character except the hands.

It was a noticeable problem over Image 1 🤔

[amber raptor]

just saw someone faking access to gpt 5.5 by generating an image showing a fake message. this is getting ridiculous

[long crater]

I feel you But what annoys me most is how slow it is jus to generate & load 1 image and sometimes it stuck. I'm not patient enough. I miss Old Sora for these reasons. However I've found , in some cases , GPT 2 far better than NB 2 & NB Pro 👍

[fallen lintel]

i just hit the limit

[civic dust]

what is it?

[fallen lintel]

around 100

[civic dust]

ah cool I usually don't get that high. Today I was coding something to generate images by api, diffrent bucket of costs

[fallen lintel]

and reset in 1 hr 28 minutes,but i think you can see how much generation is left by hovering your cursor on the send button

in “images” menu

[mossy ingot]

When they set the video gens to 100 when Sora 2 launched and then changed it to 30 weeks later. Possibly enjoy it now, could change too. Haha

[civic dust]

the darn thing generates images when you want it to or not, its hard to stop that

[analog grail]

this annoying guardail still here when they ride off this crap isn"t just no fun for anyone common we are all adult AI isn't for kids after all

[wild ridge]

Im bored and out of ideas does anyone have any fun ideas for something i could generate

[next fog]

Any fix or workaround for the noises? Sometimes(rarely) it creates normal images but mostly noise infested images. If someone got a fix, please let me know

[analog grail]

sure what do you love it first ^^

[deft flame]

Does the $20 a month subscription have better image models?

[civic dust]

nope, just wait till they fix it pretty much

[obsidian cargo]

Wdym every subscription is using gpt image 2, idk about free tier yet

[tulip prawn]

hello

[tepid lance]

I generated like 200 images on the workspace plan.

[analog grail]

Sadly no the actual tips work for now is open a new chat and make your picutre is sometime work.

[mighty ermine]

Hello, what are the limits on the Go tariff?

[green horizon]

So its not unlimited on Pro plan?

[analog grail]

Hello ! Premium plus is being very frustrating tier when you create à lots .. you can't generate too quickly or you get blocked after 5 -7 pictures , and 60 image per day sometime less is very randomly and until you reach you have to wait 9 hour or sometime even 15 hour .

with the release the Pro Tier 100% i had decided to took this one and i can make no limites pictures *

[fallen lintel]

is pro unlimited?

[analog grail]

Yes

[tepid lance]

Workspace seems unlimited too atm.

[indigo seal]

I think we get dynamic limits, if you send more than 4-5 messages in 2 minutes it give like a 5 minute block

and per day i think it fluctuates

some days i got limited at about 30

some it lets me generate even 100

[analog grail]

Yes Dynamic limits the right word they limitss you depend the traffic

[fallen lintel]

how much is workspace?

[tepid lance]

Around 110-130 bucks per month.

[fallen lintel]

so a bit expensive than 5x pro

[orchid kayak]

GPT 2 image down through api??

[amber raptor]

models disappeared for me but now i can generate 4k images through the api :)

[fallen lintel]

how to use the 2k resolution other than via API?

[harsh adder]

Did the new Image generation update also make it have better text generation capabilities?

[real trout]

I can't use the image-2 model on the API, do you know why ? On the API platform I have this error : No eligible models available.

[fallen lintel]

just run it,it is gpt image 2

[crimson terrace]

Hello, has anything found a way to make the thinking model putput transparent images?
It worked fine until yesterday, now it's outputting fake .pngs with the checkerboard pattern.

Instant model can output transparent images, but quality is bad

[sullen imp]

I’m seeing that too, I have $600 of credits on the account and am in the usage tier 5, so idk

I’m going to send an email.

[dusk creek]

No. It seems way too limited on a pro plan.

[wispy stream]

Hi

[green horizon]

what is daily limit then

[dusk creek]

Nobody knows. Seems to be random

[real trout]

Yes it's so weird, yesterday everything was working good

[kind mountain]

images-1.5 disappeared for me in the model-dropdown yesterday in the OpenAI Playground. 🙁 Where can still access/use the 1.5 model? 2.0 doesn't deliver the style/quality I need for my specific projects. 🙁

[vital gull]

Tbh haven't cracked this issue yet sometimes it works sometimes it doesn't one thing I noticed is the chance of a fake transparent png is low id you're generating in a new chat

[crimson terrace]

yea and also low if your prompt is very simple

but that doesnt get the job done for me

im just doing greenscreens now

[sullen imp]

It never worked for me, they already replied to my ticket and wanted a screen recording. We’ll see. 🤞

[vital gull]

Drop I figma and remove it tbh fastest and easiest solution or macs native bg removal

[crimson terrace]

"Drop I figma" sir what language is that 😅

[vital gull]

Drop in Figma 🥲

[crimson terrace]

Okay so figma seems to be a website

they got a bg removal tool?

[analog grail]

still this noises issue with generate picture seriously when they fix this is fun to paid for crap quality ! off course is never the Guardrail is down 🙂

[orchid kayak]

damn, the noise is getting ridiculous, like every picture even at medium quality is unusable

nanobanana > gpt 2

[analog grail]

Indeed they still no had fixed this what a shame !

[green horizon]

Yeah, but it wasn't like that before

[orchid kayak]

idc, their fault, if they are not able tu sustain a model they shouldn't release it and spam it everywhere like it's the new messiah

[late blade]

have you tried adding to the prompt:

alpha_channel: true

[arctic surge]

im about to give up on this lol

[arctic surge]

I agree, I paid it was fine... multiple updates later and every single one is having an issue

[echo ridge]

ok

[untold orbit]

Tum log yar koi kaam ka nahi hoooo

[wicked jackal]

Hi

Is there any job in openai available in Delhi

[haughty flower]

How and why has there been no acknowledgement of the crappy image gen issue that keeps appearing

Did they not notice it at all before releasing the model???

Or did they just not care??

[noble rune]

Image Gen 1:
Cars: 7/10
Fantasy/Realism: 9/10
Anime: 7/10

Image Gen 2:
Cars: 8/10
Fantasy/Realism: 2/10
Anime: 9/10

[tender aspen]

is it me or is Image 2 somewhat altering the face references that I upload to it? it's not really that accurate compared to seedream 5.0 lite which can put the face reference super accurately in the image.

although Image 2 beats out SD5.0 lite is almost everything.

[limber shoal]

use prompts like "Strict facial preservation, absolutely no change...."

ig it will help

[limber shoal]

image gen 2 gave justice to fantasy realism for me tho😭

[vivid zodiac]

Do you all think there will be a way to take out the aritfacts from the images that are buggy?

[nova compass]

Hi ! OMG gpt-image-2 is horrible for creating fantasy illustrations or surreal-abstract-psychedelic artistic pictures (divine landscapes/science-fiction, fantastic landscapes, etc..), HORRIBLE ! 🤮
The composition, the details are ugly, it looks like a stable diffusion 1.5 with a bad denoiser algorithm.
I have tried several prompts (short prompts/long prompts/details/vague texts) but i have these same horrible results.. I don't understand.
GPT always had the issue of this recognizable "patterned noise". Sometimes it's almost unnoticeable, sometimes it's not good.
I use both ChatGPT and the API playground interface, and i play with all the parameters of the model, but.. without any success
Can you help me please ? (or make a gpt-image-2.1 update please 😜)

[lapis sedge]

on twitter, they said they're aware of the artifacting issue and are working on fixing it

[limber shoal]

can you give example please

[nova compass]

yes sure ! look my results here : @lapis sedge @limber shoal
#images-canvas message

[mossy ingot]

Your images show the artifacting issue that's happening with the new models and OpenAI is aware of it and actively fixing it. However, nobody knows when it will be fixed.

[nova compass]

yes, ok, one the prompts i used was (for exemple) :

Epic Biblically accurate heaven, featuring jasper walls, pearl gates, and streets of gold, throne of God made of blazing sapphire and crystal, rainbow like emerald surrounding the throne, river of life flowing from it, golden light filling a vast celestial temple, transparent streets of gold, immense scale, awe-inspiring divine glory, ethereal, hyper detailed, radiant mist, cinematic lighting, sacred atmosphere, volumetric light, ultra majestic

[quartz estuary]

food looks so ugly, never see so ugly ia food

[wise lily]

images 2.0 have token based pricing, anyone have idea on the token usages data for batched images and separate images? I am not able to find relevant info on this.

i am asking from the perspective of using it via API

[tacit skiff]

I noted that the new image generator from Mobile creates without any artifact. from desktop is horrible

[agile patrol]

lol he rly did

omg COMEBACK????

when is it in codex?

[vivid stirrup]

@tacit skiff hmm

[queen berry]

hmmm

[native night]

Some tags to add might be, photorealistic clear focused details (if you want to try to make it more toned down), the 'ultra majestic' and 'etherial' and 'radiant mist' is likely adding a lot of noise by itself - things that add gaudiness or 'epic' stuff tend to get carried away - but if combined with other things to reign it in could help, sometimes even adding photography tags like 'f/8.0 realistic details of majestic scene' to bring things in focus more, 'soft lighting' to tone down the lights, things like that. Gotta identify which tags cause things to blow up - maybe reduce it to a one liner, then slowly add tags to see their effect, balancing them with others to offset anything unintended. Also remember.. a celestial scene of glory like that is going to be hard to visualize, and some of it may not be strictly literal (which models can struggle with understanding the underlying meaning of).

[noble rune]

For me, it's like, so bad.

[feral sparrow]

collage photo without texto HAVE 5 photo taken with portrait iphone old phone in 2016 , by showin STEPS AND some ingredients with your packaging (this product of inredients will be fined in walmart) and the finish plat , for this recipe : Cornstarch and Oil Slippery Quicksand
Ingredients:

• 1 cup cornstarch
• 1/2 cup vegetable oil, baby oil, or mineral oil
• Optional: brown food coloring, cocoa powder, or sandy glitter
• Optional props: plastic figures, toy animals, scoops, spoons, or jungle creatures
  Directions:

1. Gather a bowl or bin and add 1 cup of cornstarch to it.
2. Slowly pour in 1/2 cup of your chosen oil (vegetable, baby, or mineral) into the bowl with the cornstarch.
3. Begin stirring the mixture gently. You'll notice it starts off as powdery. Continue stirring until the mixture thickens and becomes stretchy like wet, sticky clay.
4. If you prefer a more slippery texture, gradually add a little more oil. If you want a firmer consistency, mix in additional cornstarch until you reach your desired texture.
5. If you'd like to create a realistic muddy or sandy appearance, drop in a few drops of brown food coloring or a spoonful of cocoa powder and mix well.
6. For a fun touch, sprinkle in some fine glitter to give your quicksand a shimmery effect.
7. Once your quicksand is ready, press small toys or figures into the mixture. Watch as they slowly "sink" just like in real quicksand!
8. Explore the texture by scooping, pouring, and stretching the mixture. Experiment with slow versus fast movements to see how it behaves differently.
   Tips & Variations:

• To enhance the sensory experience, you can add more textures by incorporating small pebbles or sand, but keep in mind it will change the consistency.
• This activity can be made educational by discussing the properties of solids and liquids, demonstrating how quicksand behaves under different conditions

[brittle stirrup]

It might be better via API, with uncompressed outputs and higher res
I've found so far that at least the version of Images 2 they're running in free-tier ChatGPT chat is not better at anything (and is sometimes worse) than the version of Nano Banana (not sure if it's Pro or 2) that they're running in free-tier Gemini chat. A proper comparison would have to use both strictly via their APIs, preferably around 2K res, I think

[mossy ingot]

Has anyone noticed any difference in the image gens when switching from thinking 5.4 to 5.5?

[limber shoal]

hell yeah
Guardrails cooked
I dont know how to edit prompts anymore without being flagged lol

[agile peak]

Does generating more than one image only works with thinking mode? kinda confused.

[grim ibex]

I haven't seen 5.5 yet but Image 2.0 is nice

[vital gull]

huh?

[nova compass]

I have tested almost all my prompts inside the API in 4K/2K 😉 so the problem is elsewhere..

[civic dust]

directly via the api, there's a bunch of image models you can use

[agile peak]

I tried generate images or recursive loop for iterative design, it kept generate one image with multi sections instead. For non-thinking mode.

It won’t trigger multiple images per turn with thumbnails that I can select from.

[errant vale]

Does 5.5 generate better or smth?

[mossy ingot]

I don't know. That's why I was curious to know if anyone noticed a difference in prompt adherence or how image seem to generate. Probably not much a difference, but I haven't tested anything yet

I didn't have GPT 5.5 yet when I asked that

[nova compass]

I have found a solution to remove artefacts/patern-noises issues inside gpt-image-2 : take the generated picture as the input image for editing, then use this prompt :
"Remove artifacts, make the picture smoother and less chaotic."

it works but the artistic style is not good

[civic dust]

do it with a scene of a cottage with lawns and shrubs and trees

[vivid zodiac]

I tried something similar, and it only got rid of like a 1/4th of them.

[modern glacier]

Chat gbt is just odd man why try make "art" in an algorithm machine

[nova compass]

yes they MUST fix this issue.. 🤣

Midjourney is better than gpt-image-2 at these only specific things : complex art styles / fantasy / pixel quality
But gpt-image-2 is destroying all other models atm for accurate/cohenrent things / diagrams/texts/etc..

[regal burrow]

The new image generation model is unusable.. it doesn’t follow my instructions, it doesn’t stay consistent with my references, it keeps creating this weird noise texture, it messes up body proportions.. it’s a massive downgrade compared to the previous model

Is there a way to use the old image generation model I need it urgently for my deadlines

[regal burrow]

I have the exact same issue

[analog grail]

Yes Open a new Chan and try Again . and if is failed open again another new page using same prompts or text. until is ok

New Chan work no bad for first request Image.

[regal burrow]

I’ve been trying ever since it came out and I’m unable to produce a single usable artwork

It’s so frustrating

Especially because I had a deadline and now my work is put on hold

The previous model got things perfectly from the first try following my instructions perfectly and the quality was amazing especially with humanoid characters
But after trying the new model for the past few days I find it to be completely unusable
And it’s really frustrating that there’s no way to switch back to the older model

Honestly any help on how to access the old model or a good alternative would be appreciated

(I’m grieving the old model it was my favorite image generator)

[vivid zodiac]

I wish this were true, but it's still at random

[vivid zodiac]

Which sucks, because that's literally what I need for my project.

[nova compass]

Midjourney + gpt-image-2 (or banana) = perfect combo !!

[regal burrow]

How so?
GPT-image-2 doesn’t stay consistent with my references and keeps changing the characters face or outfit details

[tiny sonnet]

With two people he goes well, with 1 he gets weird

[regal burrow]

And not to mention the noise texture in every image

And sometimes it even changes the art style completely on top of everything I mentioned

[nova compass]

Ah? ok, i have not tested characters references atm

[regal burrow]

With gpt-image-1.5 model I was producing art that looked professional, it following my instructions perfectly, it followed the exact art style I described, the characters were consistent and on top of all of that it usually gave me exactly what I envisioned from the first try.. it was my favorite image generator tool and it really sucks that I can’t access it anymore while I’m in the middle of my work with the deadline approaching

[civic dust]

you can use the old model via the api

[regal burrow]

Wait there’s a way to access the previous model???

[civic dust]

my usecase is quite specific and same problem, v2 can't do it. I got codex to wire in a call to v1.5 for me to a web interface

[regal burrow]

Can you show me how to do that

[civic dust]

its got a testing area but you'll probably need credits. https://platform.openai.com/playground/images

I'm burning 8c an image at the moment via api call, which is mildly annoying when you're testing to get the prompt right

and nano banana is very good but its a different style

[regal burrow]

I’ll try it rn

[regal burrow]

What do you mean by different style?

[civic dust]

you can just tell its a different AI, it has its own signature style. I do photorealistic stuff and nano banana looks like ... nano banana. No idea about fantasy

[azure sky]

Is image generation in your side okay and running right now, or are there issues now? Sorry for interrupting

[civic dust]

Chatgpt has always been more dynamic and more like it swallowed a couple product catalogues for me

[regal burrow]

Yeah I get what you mean

[regal burrow]

I still have issues with it, i used it a few minutes ago

[civic dust]

hopefully you can get the playground working though, just mind chatgpt web interface adds a bunch of hidden stuff so you have much longer prompts via api

[azure sky]

I see. I thought I was the only one. It takes too long to generate images just now.

[regal burrow]

I really hope it works

[regal burrow]

Oh no my issue is with the quality of the images and the downgrade compared to the previous model

[tender aspen]

THIS.

[vivid zodiac]

If anyone has Gemini as well, I think with the suggestion from here to use the prompt "Remove artifacts, make the picture smoother and less chaotic." and attaching the aforementioned image-- it will get rid of the artifacts.

[tender aspen]

And the photorealistic images look like Stable Diffusion XL with ADetailer lol

[calm swan]

new model is great until you realise it cant do pixel art and has trouble editing photos when it comes to real advanced technical surgical editing requests

[brittle stirrup]

in any case
here's it vs five other models
all of these were generated at the closest to 2K as was possible, including the GPT one, and then downscaled from there to exactly the weird 1086x1448 GPT baseline resolution for 3:4 aspect ratio, to kinda normalize the lineup
https://discord.com/channels/974519864045756446/1497031407464353853

[civic dust]

gpt2's grid repeat is in that image too

[wraith oracle]

Really it has trouble with pixel art, man when are image generators gonna be good at consistent pixel art transfer

[calm swan]

it’s still gonna take ages. llms are good at generating images from a general idea, not precise ones. they rely on patterns from training data and go with the overall consensus. so when it makes pixel art, it creates what it thinks pixel art looks like instead of ensuring every pixel is actually a pixel, which is why some pixels end up merged or blurry

[vapid elk]

I noticed that when in thinking mode, the model will produce an image, then, keep iterating on it, allowing the AI to see the image and possibly further changing things on it
so, yea, it is kinda like ADetailer

[velvet rampart]

I noticed that 2.0 has poor sense of scale.

[velvet rampart]

An amendment - you have to write the scale in. Just saying that something is X' tall does not cut it. You have to specify it.

[wraith oracle]

Im really digging the thinking image models for using them for visual infographics for information from the web.

Really makes it generating images from fun to useful

[proper wagon]

A lot of the infographics I’ve seen online are AI generated now and they leave a lot to be desired

I’ve seen low resolutions, poor data, data that doesn’t make sense, and dates for ranges that should be recent and it’ll throw in a date hundreds of years ago in the mix

[tender aspen]

If Image 2 and Nano Banana 2/Pro can't generate consistent faces, what's the best tool that can replace the face accurately after I use Image 2/NBP to generate? I tried seedream 5.0 lite, and while it can replace the face very well (100%), it changes the poses and everything else in the image lol

[vital gull]

There's higgsfield and their AI influencer or smn

[tidal lance]

They both can do new consistency for the most part, it designed to not make the same faces consistently on purpose for safety reasons

[tender aspen]

That sucks

[tidal lance]

There aren’t ways to mitigate it though. Like if you have an image of a character against the white background like the character reference sheet, you shouldn’t have no issues theoretically.

[tidal lance]

Do you use app or api?

[tender aspen]

I use the website

[nova compass]

You can use GPT-5.4 to make a grid of pixels only with coding (vectors)

[orchid kayak]

Noise reappared 🙁 it was ok for like 2-3 hours last night, and now it's unusable again even on medium

[tidal lance]

Wdym?

[orchid kayak]

that all images have a lot of noises again,

if you use Low or Medium the images through api are unusable

[wild ridge]

am i the only one who cant do basically anything because its copyrighted? before with the last model it was perfectly find doing stuff like that

its been really annoying

[limber shoal]

guardrails did increase on every aspect unfortunately
it has been difficult to create the perfect prompts coz of the guardrails

[wild ridge]

hopefully they make it not as strict

[tidal lance]

That quick ?!

[limber shoal]

ye
The prompts that used to work before is not working now

[glacial haven]

hi
I have the Plus Plan, how can I get 4k image generation quality, or it is available only for the PRO?

or it is available only thru the API?

[obsidian cargo]

I believe only pro or API

[wild ridge]

other then this and the noise issues the model is SUPER good

[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.

[glacial haven]

This is on FAL Pricing, and I dont understand how they can give you 4k and all the options for the quality from low to high
The following table shows the pricing of ChatGPT Images 2.0 from a technical standpoint.

Size Low Quality Medium Quality High
1024 x 768 $0.01 $0.04 $0.15
1024 x 1024 $0.01 $0.06 $0.22
1024 x 1536 $0.01 $0.05 $0.17
1920 x 1080 $0.01 $0.04 $0.16
2560 x 1440 $0.01 $0.06 $0.23
3840 x 2160 $0.02 $0.11 $0.41

[tender aspen]

Can I use codex or agent inside chatgpt web to get it to create a series of images with the same character in different scenes and poses etc.? Like generating 5 images one shot

[tidal lance]

Don’t forget plus I put and output token

[glacial haven]

yeah, why I cant generate 4k images with the Plus plan in GPT?

[tidal lance]

I’m talking about fal

[indigo prawn]

Hi everyone 🙂 Can I address a problem I'm having with the new image model?

Maybe it is just me doing something wrong... I get strong "patterns" in image generated since the new model is out...

[tender aspen]

That's a common thing with chatgpt image 2

[analog grail]

Should by fixed with time is already happen 1 week ago before get fixed

[orchid kayak]

yea, pretty much everybody having those "noise" artifact, the more complex the prompt the more noise you get, also if you use Low or medium quality settings u get a lot of noise, only using high you get a tiny bit

[indigo prawn]

Where can I find this settings? I was unaware of that...

[slow ledge]

Are you guys actually able to generate the images on 5.5? For me, it's been a loop of infinite loading, me stopping and asking it to retry.

[orchid kayak]

there is not if you are using the webapp (ChatGpt app or browser)

[indigo prawn]

so just API?

[orchid kayak]

yea or third party app like openart, freepik etc

(where they use apis)

[arctic surge]

Yup

[tender carbon]

i feel like the quality on images dropped today, could just be me

[rich sparrow]

I feel the same

I wonder if it has something to do with Sora closing

Sora's images were much better

[rich sparrow]

Exactly, all images seem granulated

Is there a way we can use the older image generator?

[tender carbon]

im talking about the new 2.0 was getting better images yesterday than today with the exact same prompts

[ripe copper]

yeah wth is happening

[zenith crown]

The announcement about the release of gpt-image-2 specifically called out transparent background as a limitation of the current model.

[zenith crown]

Are you using a new chat for each image? It's pretty clear from the chain of thought that chatgpt is sending ALL of the previous images in a thread along with the current image prompt, and there's some evidence to suggest that this behavior is responsible for the "granulation" that people are seeing.

[rich sparrow]

Yeah, I tried using a new chat too

[zenith crown]

Well, if you're willing to pony up for some API credits, the API will get you access to gpt-image-1.5.

[dense coral]

are there any plans to make a specefic pixle art style image generation? right now its horrid ngl especially with the random ratios

[limber shoal]

@late blade dmed you...can you check?

[velvet granite]

Has anyone tried generating images of different chess positions? It seems no matter what I try the pieces are always not centered on the squares or just the pieces look horrible lol. I think it breaks GPT-2 image too much for it to handle lol.

[late blade]

what should I check? A similar concept we discussed but with male characters in context?

[barren widget]

I generated 84 images today, and got this message "You’ve reached your image creation limit. Upgrade to ChatGPT Pro or try again tomorrow after 12:48 PM." - Have any of you also reached such limits?

I am Plus

[sand bear]

Now with the new image generation update, can I teach it to learn the 3D shape of the object by giving it multiple photos from different views and then it can put the object anywhere I want according to a description without changing any details of the original object?

[arctic surge]

You have to tell it not to change it but it’s hit or miss

It will tell you the prompt eventually after you say you don’t want the actual object changed, but still it either does it or you’re arguing with it for 20 min.

[arctic surge]

Yes and it changes based on idk what tbh. Sometimes I do a few images and it will tell me it can’t generate for 15 min, other times I can be on it all day usually messing with prompts and it’s fine

[limber shoal]

or maybe Ais in general are horrible in chess

[limber shoal]

not really
male charecters are working just fine, but with female charecters, no matter which prompt i would use, it will flag me and accuse me of "lightly sexual"

[late blade]

are you writing prompts with the same size as the the iliad, the odyssey and aeneid combined? Send me your prompt per DM, I'll revise it

[vivid zodiac]

I haven't had an issue generating women, but the women I'm also generating aren't like, the scantily clad/sexy girl type stuff.

[acoustic peak]

Why the quality kinda grainy rn?

[vivid zodiac]

It's a known bug.

It's happening to everyone.

[acoustic peak]

I feel like it has something to do with the model being 5.5 now but idk

[limpid wave]

Yo those codex pefrorm good now or no?

I mean is gpt 5.5 working propperly rn cause someone linked a status page that there are issues

[civic dust]

I’ve hit sexuality, violence and minors flags like a dozen times this week, it gets tedious

[vivid zodiac]

Oh my God, can GBT NOT recognize left and right arms? I'm trying to reverse the position of the arms, and its like no matter how I word it, it just wont do it.

[vivid zodiac]

I feel like I'm being gaslit, lol

[mossy ingot]

What's your prompt?

[lethal locust]

anyone else having problems with chromatic aberation and artifacts?

other people have posted about it too

[arctic surge]

It saw my bra strap and basically called me a hoe 😒

[civic dust]

goodness. Don't let it see your ankles

[arctic surge]

Never 🤣 learned my lesson

[vivid zodiac]

Is it me, or is the image generator doing better?

[mossy ingot]

Let me do my train test 😂

[mossy ingot]

Meh: #images-canvas message

[azure arrow]

WHY did "Make a photo taken in 2000s of a dog on a chair next to a computer that has MSN open" get blocked?? What

[analog grail]

mmh look like isn't possible generate character from Mario

[open wagon]

Not surprise lmao

[analog grail]

find my way for is work :)^^

[civic dust]

welcome to my life with gpt images

although my gpt is currently so damn rude about everything i can't post what it says in this discord, which is distracting me from making images

[limber shoal]

loll
Nah i reduced the size
I will send it to you

[limber shoal]

🥲

that's why I am not creating anything with female charecters rn except some cute scenes
Chances of getting flagged is rather high now

[civic dust]

I got flagged for a shirtless slouching male

[mossy ingot]

Free the male slouch!

[civic dust]

I get automodded a lot here so trying to avoid keywords but chatgpt/images2 made me a roman statue with chips and it had a chip on each nip and it looked ridiculous and that passes filters

[tidal lance]

Brah I’m telling you on everything it’s about how you structure your prompts

[civic dust]

yeah if you dont write an essay for a prompt and let it fill in gaps it regularly makes things that fail guardrails

"make me a medieval warrior holding a mace" sorry violence no can do

[mystic blaze]

Guys couple people on Twitter dmed me that their getting more A/B testing for images. Please vote if you do get them !!

[tidal lance]

It’s about prom structure, bro. I know that sounds cliché, but it’s how you prompt it and in what order that determines if it passes or not, you have to be very strategic with your semantics when you’re prompting.

[fallen lintel]

when i use 3rd party or API i get low/med/high options for Quality,what are the default quality used in chat GPT for image 2? and i see it’s 1K not 2K via chat gpt,am i missing something or?

[mossy ingot]

https://developers.openai.com/cookbook/examples/multimodal/image-gen-1.5-prompting_guide

[tender aspen]

I can now tell whether an image is generated by image 2 via the level of details and AI repeating patterns lol it adds way too much giberrish and detail overload to photorealistic images.

[ripe echo]

image down?

https://developers.openai.com/cookbook/examples/multimodal/image-gen-2.0-prompting_guide

[orchid kayak]

The artifacts/noise are getting worse.. So sad, this had potential to be the best img model if it came out like it was in preview in arena

[late blade]

seems so I'm geetting errors that the image tool is unreachable

[tidal lance]

Give sharing some of the prompt you’re using

[limber shoal]

yeah I got the result
BUt after lots of attempts

[tidal lance]

I was talking about to with somebody earlier

[limber shoal]

oh okay

[tidal lance]

Both are good points

[tidal lance]

I am from the school of thought that structured prompts may eliminate the need for constant iterations. Obviously clarity is an important part of that, but I think that using the structure and language ChatGPT uses

[mellow finch]

anyone know what model was the image generator using before this new updaet

is it still 1.5 or 2.0

[mossy ingot]

It was GPT Image 1.5 before this latest update

It's now GPT Image 2.0

[mellow finch]

no wonder my iamges arent like before with my prompts

time to do new trial and error methods i suppose
im on day 2 trying to get my style back 😭

[mossy ingot]

1.5 is still available through the API 😅

For now at least.

[mellow finch]

welp

good luck for me

[azure sky]

Are you also getting image tool error just now?

[mellow finch]

yes

some error in creating the image

hopefully some maintainence to listen to a style 😭

[patent helm]

can't generate images now

[azure sky]

glad it's a worldwide issue. hoping it's going to be fixed soon. some pending tasks waiting to be generated

[devout ferry]

I am also getting the same issue here from South Africa. This sucks

just tried the app, seems to work in the app but not in the web browser

[tidal lance]

Works on arena and fal

[mellow finch]

i hope it listnes to styles and stuff

[limber shoal]

is it fixed now?

[mellow finch]

let me check

yes it working

[hushed relic]

gpt keeps generating weird, noisy efect images when I try to create cartoon style drawings. I haven’t been able to find a solution for 3 days. Has anyone found a fix for this? Or does anyone know how to switch back to Image 1.5?

[indigo seal]

has the diagonal grid effect gone for anyone else too?

it's not there anymore for me

[rich sparrow]

Can we choose to use 1.5 instead somehow?

[azure sky]

it still isnt fixed

[rich sparrow]

Man... Sora is over and ChatGPT butchered their image quality too... it's sad

[errant vale]

Can image 2 generate images in 4k

[steady geode]

GPT-Image-2 can but only on the API, not via ChatGPT subscription

[errant vale]

Ohh

Ty

[tidal lance]

It takes two minutes to generate a 4K image, roughly

[next fog]

Any news on them fixing the artifacts and noise issues? Any workaround so far?

[tidal lance]

Higher resolution is more prone to make more errors than generated from low resolutions and upscale.

[fallen lintel]

how they generate 2k then?

[fallen lintel]

can you dm me what artifacts you mean?

[mellow finch]

samee it keeps making this dusty effect images or very weird art styles dont know why

[late blade]

your prompt might have a bias towards that specific art style, try setting a constraint towards another visual art, see if that changes

[nocturne siren]

So, how are yall generating images. I use GPT and run into the generating images too fast after like 6 prompts.

Will also add I have a Pro sub

[kind mountain]

I need an option to choose 1.5, too! Is there any way to have gpt use 1.5? Other than using the API/Open AI Playground which costs extra even though I have a GPT subscription. 🙁

[steady geode]

Pro 20x here, not had this error so far, however generation time can be easily around 1 minute each image. The more complexity required for the image, the longer it takes to generate the image (I've observed), and my prompts can usually be somewhat complex

I also generate mine via Codex if that makes any difference

[nocturne siren]

Yeah I just use the bog standard GPT site. I have access to Openai API (tier 1) and I want to generate images but not do that verify organization thing. I'm not very knowledgeable on codex and all that

[steady geode]

I see, yeah the API is a separate thing. It might be that ChatGPT has some kind of limit where Codex doesn't have the same limit

the Codex app (desktop app) is essentially like a ChatGPT of sorts, it has a chats section now. All you need to do is ensure you enable the image gen skill and away you go. Well, that's optional as the tool itself is available anyway, but I'd recommend enabling the skill. Then just prompt it to generate some kind of image

[nocturne siren]

with the codex app is that different from the API?

[steady geode]

yes, it can use your ChatGPT subscription

[nocturne siren]

Ok cool. time to download it and dig in. Appreciate the assist

[rich sparrow]

I'm wondering the same. I hate that 2.0

[tidal lance]

I feel you I was just talking about the same thing

[worthy dew]

anyone else bored with 5.5 already?

[tidal lance]

To be honest I haven’t ever tried higher than 5, I use other 3d party platforms mainly for image generation. So 5.5, I’ll get around to it when it becomes cheaper lol

[arctic surge]

👀

[desert jasper]

Claude lwk better 🤷🏻

[civic dust]

yes, claude's images are way better than gpt's. Check what channel you are in

[errant vale]

Show us claudes images

[desert jasper]

No as in general for everything I ask it

[vivid zodiac]

So, has the artifacting gotten better for folks? I've been busy, but it seems to be a little betteR?

[analog grail]

Yes and no is still around when no was here before ... all would by perfect without this dammn issue

( Analyze this image, and then make a detailed prompt. When making the prompt. Make sure to not do anything aggressive with details or texure. Texures should be smooth and not aggressive and for details should be her hair . And for the other details keep them soft and subtle not loud or noisy. Or else artifacts will appear. ) someone gave this prompts so i can sent buy you have to use OFTEN TIME !

[vivid zodiac]

Yeah, I've been using Grok to clean up the images which seems to work?

[civic dust]

I've been using gemini

[analog grail]

hmm i can't paid another Ai when paid already Pro

[civic dust]

I'm on plus for both so a tad cheaper than pro

[vivid zodiac]

I got a deal for Grok to pay for the pro for $30, so I'm on that, lol

[civic dust]

I just really like nano banana and its better at following instructions, I bounce images between the two pretty often

[analog grail]

last time i'm have lots idea and plus was not enought.. with the daily limit because when you flagged is still count as generation images*

[civic dust]

oh yeah, I have a "make an image with openai" button in my app now and I still get charged when its flagged

[analog grail]

Yes, so that mean the picture is generated in fact... wonder if there as way to reach it like maybe the text flagged you use a Ads block for remove this window and nour picrTure is hidden under ?

[civic dust]

if there was a way, someone would have written an extension by now

[analog grail]

Yeah still find deeply no fair they treat us like child since isn't nour private session no need this crap i can understand for some stuff but when is nothing seriously ^^

[analog grail]

very funny this guardrail is strongly back .. that make me in murder mood giggles

[acoustic peak]

Is the quality for image 2.0 still bad for anyone else?

[south wraith]

its so bad, I can not make one image, id had a really good flow that was right 100% of the time, 15 edits in and still wrong...

[acoustic peak]

Ever since 5.5 came out it started to suck. Even text is getting a little blurry and unreadable now

[analog grail]

hope one day someone will hack OpenAi and destroy the Guardail

[acoustic peak]

It seems to mainly suck at making realistic photos right now they look all weird and glitchy

[agile peak]

Yes. It’s been degraded a lot today, i’ve been testing this 2.0 model since it was still stealth testing on lmarena like two weeks ago..

it was all flawless, there’s a lot of bugs when it officially launched on platform.

Especially diffusion artifacts bugs (in photorealism) and some text scrambled.

[civic dust]

its very good at hot anime characters. But not trees

[acoustic peak]

Its genuinely just bad at realistic photos in general which is weird cause it was good before

[agile peak]

If you generate image that heavily use trees, sands, dust or water it heavily noticeable for those diffusion artifacts bug.. in photorealism..

[analog grail]

i was very possible before they just need fix this instead counts monney.

[agile peak]

Yeah kinda annoying need to fix this issue. It’s been mentioned many times on twitter..

[civic dust]

yeah its like they got so carried away with anime and fake screenshots they forgot to check realism and it regressed

[acoustic carbon]

What's up with the fake transparency? Have you noticed the same, it keeps adding checkboards to the images instead of making them alpha transparent

[fallen lintel]

any info did they fixed the artifacts and noise?

i notice noise/artifacts happen when you continue generate image in the same chat,if you create a new chat,its gone

[nocturne siren]

Not always. some people have begun using keywords that help eliminate the artifact/noises (I have no idea what they are)

[fallen lintel]

can someone share the keyword please🥹

[azure arrow]

Is it functionally different to prompt GPT-5.5T for an image than GPT-5.3?

[tulip raptor]

what solution did you end up going with?

[buoyant otter]

I can’t use 2.0 for much atm due to the frequent texturing error where it seems to be clone stamping textures

if anyone’s found a solution to that lemme know, but right now if i ask it to make trees it looks like ripped paper

[tender aspen]

Chatgpt image 2 to get the composition and poses all looking right easily, then download the image and feed into Dreamina Seedream 5.0 Lite and ask it to replace the face etc. the result usually looks way better than pure Image 2 because it gets rid of all the ultra detailed look and artifacts which Image 2 creates

[civic dust]

image 2 with the artfacting is already appearing on the socials in all the ai generated promoted posts

[hot rain]

Fairly certain the artifacting is getting worse lol.

[tender aspen]

it's a telltale sign those are all generated with Image 2 lol

[analog grail]

OpenAi no make any news about this Artifact issues ...

[fallen lintel]

i found a fix

[analog grail]

i had a temporary fix too but need allway use all time

[velvet rampart]

Is the artefacts to 2.0 same as yellow tone to 1.0?

[analog grail]

this stupid guardail hit hard today no sure how this crap work. look is allway change

We're sorry, but the image you created may violate our rules regarding nudity, sexuality, or erotic content. If you believe this is an error, please try again or edit your prompt. just Shut up is nothing that way ...

[limber shoal]

it has always been super strict ever since the launch of image 2, ig wait for it to get stable after all the corrections?

[analog grail]

is very randomly sometime i will no have ny issue for fews generation and suddently being crazy becuse a dress use word sheer seriously .

[regal burrow]

Artifact is not the only issue that needs fixing
The new model doesn’t follow instructions
It doesn’t stay consistent with character design unlike 1.5
It changed face details
And when it came to creature design it distorted its body proportions
On top of all of that it changes the art style too

[analog grail]

Yeah that true too

[limber shoal]

but in A/B testing phase it stayed so consistent with charecter tho

[late blade]

yeap, artifacts are a problem
following instructions depends on how you prompt, I find it that even thinking "forgets" half the prompt
i had so far here good outcomes, posted a few on daily theme
faces might still be in the directives for deep fakes
mostly on background aspects for me
art style, agree

[limber shoal]

also try using "STRICT facial preservation" and so on
till now I am not struggling with facial preservation

or charecter consistency

but ye i have noticed in fantasy styles, charecters do deviate with messed up proportions

[regal burrow]

It depends on what you make
I use the image generator almost every day and never had a problem with 1.5 but with 2.0 it’s a complete mess I can’t create anything anymore it’s unusable

[late blade]

I'm still trying to figure out how to prompt the agents, clearly there's a mismatch of the agents at times

[regal burrow]

I create fantasy painterly realism art
With 1.5 I made beautiful professional looking artwork but 2.0 just looks terrible

I wish they’d give us the option to switch to 1.5 because my work is on hold and my deadline is approaching

[fallen lintel]

any info on the fix for artifacts?

[late blade]

have you tried custom gpts?

[regal burrow]

Yeah

I couldn’t find one that uses 1.5

[late blade]

don't see any changes in the release notes for custom gpts

[river gazelle]

Peak GPT Images 2.0: #images-canvas message

[tender aspen]

Dzine AI

[sour oracle]

I think I broke the image generation in chat, taking a very long time to generate an image

[late blade]

did you take into account that thinking models take longer to generate the image?

[nova compass]

Hi! Do we have officials news/posts about gpt-image-2 developers that will fix the horrible "pattern noises/dots" on pictures? Or nothing in perspective?

[late blade]

nothing on discord so far

[silk karma]

Strange thing is that it was every once in a while when it was just released. Now it's almost every image

[fallen lintel]

yeah notice that too

[exotic geode]

agree..its getting almost unusable

also agree as i see posted above, it's ability to make truly photographic images is quite poor suddenly

[tough lintel]

The safeties are really poor for image generation, they should focus more on isolating prompts

It’s possible the blocked image is generated, but still stored as a local state, then upon the next request operates on that last image

[analog grail]

Very possible indeed

[exotic geode]

i just posted a comparison between all the gpt models in the images-canvas channel

[river gazelle]

#images-canvas message

It’s insane that I am able to get that amount of detail on the screen

[coarse stag]

yeah at first I noticed almost no artifacts in any images I'd generate, but now every single image has the weirdest texture. sad because the model is awesome otherwise.

[vague hearth]

When asking GBT to remove background of an image, one can see the image being unveiled and all is perfect, until it is fully unveiled, at which the quality of the image drops into pre-2016 levels of quality (text illegible, faces, hands, etc totally degenerated, shapes, lines etc completely morphed)

Is there a solution to this issue?

[civic dust]

have to say I've been using nano banana a lot more since 2 came out

[dense rose]

Hey do the admins know the issue in gpt images 2 where some images , depending on the prompt are like overcooked . Like it has some vae error ?

[wraith oracle]

Yea they know the issue

[mortal veldt]

will img-2 ever support transparent backgrounds?

[half nova]

Image 2 is crazy good, but the rot effect from repeated edits hurts so bad

It seems to be exaggerating noise?

[rigid bough]

it does now

[mortal veldt]

I tested on the api like 15 mins ago

[rigid bough]

#images-canvas message

[mortal veldt]

sad its not working for me in the api i have background set as transparent and it says its not supported ig ill try again tm

[rigid bough]

my prompt was create an image of the Linux Penguin Tux on a transparent background

[lavish hound]

Hey all, need help with a general ChatGPT issue. I'm creating custom content for the game Carcassonne, and I'm having AI create a stylized version of a famous German Brewery. It creates this perfect amazing version of the brewery but then when I ask it to isolate the brewery from the rest of the image to use as an asset that I can use in my tile program, it completely butchers the art and does something completely different. Does anyone know how to fix this issue, or if I need to give more specific prompts, or if its unfixable? It's really starting to frustrate me. I would post image examples but it seems I'm not allowed to post images here.

[mossy ingot]

You can post your images in #images-canvas. I would ask the model to generate an asset with a transparent background from the start, but it should be noted that transparency is fickle right now. While image editing is typically pretty good at preserving the original, you will get variations at times that are hard to prevent. That’s why I would suggest asking for it on a transparent background from the start, so if you like it, you don’t need to edit the image to isolate something.

From my own experimentation, I've had better luck generating images with a transparent background using the instant model rather than thinking, and using something along the lines of:
Generate x with a transparent background (actually transparent, not a checked background)

[coarse stag]

Yeah it’s unfortunate because the model is so great. If it wasn’t for the weird artefacts and noise it would be perfect.

[half nova]

I find asking for a style change (even if you ask for it to change into the same style) seems to help cure it?

The rot can be prevented

[regal burrow]

hey guys where can i give feedback of image 2.0 to the open Ai team?

[lavish hound]

Gotcha, I'll give it a shot, thanks

[mossy ingot]

No problem. Good luck

[civic dust]

#1070006915414900886 but I'm sure they already know about the artefacting

[unreal kettle]

lmao the shiny dots of doom on every image XD

[sour oracle]

So I think I figured out what was up, safeguards come up the first time (and took hours to kick in) and trying again worked

[civic dust]

I'm just doing my generation in nb2 for a while until its all fixed

[cinder zinc]

I cannot for the death of me get an image in 4k. Even asking for it. It will only generate something like 1672x641 when i ask for 4k.

Any ideas?

[gentle rover]

Maybe be clearer and say use unreal engine 5 graphics???

[sage prism]

I love generating images but third-party likeness violation can be a pain but I can handle it

[idle sequoia]

yoh

[orchid kayak]

Same it made me truly appreciate how advanced nano banana pro already was

[limber shoal]

suggest a prompt that will restrict GPT To draw abs on every female charecter

[rich sparrow]

Are we all still getting granulated and bad coloring in images?

[rich sparrow]

Have they said anything about fixing it yet?

[civic dust]

vague "will fix" message in passing on twitter, and quality going up and down like a yoyo in normal use. Must be really disruptive if gpt image was your living

[rich sparrow]

Luckily only a hobbie for me. Still it improved a lot the D&D nights... shame Sora is over and 2.0 is fried

[obsidian cargo]

NB2 had similiar issue when it came out so idk what youre talking about

[orchid kayak]

no it didn't

[obsidian cargo]

It did

Id blame server overload for that, thats the reason why nb generation limits are shi

[orchid kayak]

maybe are shi if you are on free tier or idk, i'm on google ai pro and can generate daily a lot of image on google flow, and use it daily on real work.. so idk what you talking about

[obsidian cargo]

I wason google pne

One*

[civic dust]

the service is google one, what tier?

[obsidian cargo]

Shi the 22€ one i think

[civic dust]

ah way higher than the one I'm on

[tender aspen]

You'll never get an image in 4k because it doesn't have that ability or feature. It's 2k max and it's only on some third party sites. 4k i think might be available only vis API on third party sites.

[obsidian cargo]

Yeah mostly APIs

[late blade]

api supports any aspect ratio up to 4k res in any aspect ratio that is within the limits and parameters.

[small vale]

Good pm. I just want to ask a few things.

• I noticed that when I generate images, sometimes ChatGPT saves more than one version of the image into my Images gallery. Do the extra images count towards my images uses? If yes, would you know what prompt to use to stop ChatGPT from doing this, or is this part of the process?
• Sometimes, when I try to generate a text response, ChatGPT mistakes it for an image request. Would anyone how to stop this?

Please advise. Thank you.

[analog grail]

still this artifact issue when continue generate images on the same page... look like Open Ai don't care much about this for fix it

[devout cypress]

Just compared nano banana and gpt images same infographic prompt am I tripping or does gpt image deadass looks better

Specifically the text is much cleaner the pictures still look overly glossed

[tidal lance]

It’s more aesthetically pleasing because of the color grade. Gemini always had flat colors grades

[devout cypress]

Damn there is no watermark too I don’t even got to remove it manually that’s nice

Omg nvm the little mistakes r everything

[tidal lance]

For real that’s a really good move by open ai to do that, something I could stand behind. And good for them for not putting watermarks on images. I think they would’ve been a far more successful with sora 1 and 2 if the watermark wasn’t on there, it would really make them look like a progressive forward thinking company.

[velvet rampart]

sora’s website is still there, but you can’t generate anything.

[late blade]

they are probably still working on the exports. seems there’s a lot of users that haven’t gotten their exports in full

[next fog]

Can you generate 2k images, just by using the "2k resolution" prompt in the chat? Or is there some other way?

[scarlet mango]

i mostly like image 2, but sometimes it gives me such weird texture over the image

[wraith oracle]

Tip for making realistic images, use iPhone 13 amature for the style, it will get you realistic images that would look like someone took a photo there

[wraith oracle]

Why? Because professional photos in real life settings look unnatural, even in movies scenes it's unatural because the professsionalism style is for photos in a studio not real life

[wraith oracle]

Honestly it became so easy to prompt i dont even using my custom GPT anymore, I just prompt in my own words and it gets the result, is this the future of AI prompting?

[analog grail]

i'm still get this weirdo artefact issue when continue a chan is being very annoying i'm gonna ask openai refound my pro montly paid because no fair to paid for 2 week with garbage issue that way.

[wraith oracle]

What did you pay pro for?

[analog grail]

For Unlimited Generation image for keep a good workflow every day without get this daily limit since premium plus is being too much limited so i take advantage take Pro when they make the 100$ tier

[wraith oracle]

Bruh I dont even hit a limit on image generation on my plus plan

[late blade]

Think Aerith hits the guardrails limit before image limit

[analog grail]

Yes Guardail count as generation too

But what count is also no only Guardail but when the result you ask disapointed you and you try again , and ask what is wrong and what you want and the model still get messy until finaly get what you ask so i can come quickly to hit the limit ^^ in Premium plus limite Images per day is around 60-80

[limber shoal]

me happy with my 8 images per day lol

it forces me to learn prompting to avoid as much flagging as possible

[marsh sentinel]

I’m curious — are people here experimenting with autonomous agents interacting with each other?

I’ve been building something around that and wanted to see if others are exploring similar ideas.

[mossy ingot]

2K+ is only available through API from my understanding.

[plain bobcat]

i know the feeling, i was making projects using 1.5 images and when the update rolled out on April 21, 2026 the images for example anime style changed completely. I asked gpt series of questions and asked to revert back to the old image system and it tried giving close alternatives, but ultimately couldn't because it said the entire foundation is based on a "new pipeline". I said several times screw the new pipeline, we need the old one back LOL, but reality sunk in. It's a shame we cant alternate between the old and the new... What i did find above through scrolling is a decent fix for the "cleansing" of an image by using, "Remove artifacts, make the picture smoother and less chaotic." i'll share some example of new images using this, but trust me, if we make enough noise and say our appeal, hopefully they can give us the alternative to hop between pipelines and their old models and styles.

#images-canvas message

[regal burrow]

I really do hope so
I tried sending the open Ai team some feedback yesterday
The issue isn’t just the artifacts
It’s not consistent like 1.5
It distorts body proportions
It changes details when changing the camera angle which is something the 1.5 models never got wrong
And many more issues with none humanoid creatures and objects

[regal burrow]

Wait that’s coral dragon, you play yugioh?

[plain bobcat]

Are you in the same boat as me wanted to generate fantasy, anime, non realism images, because before on 1.5 they were very smooth, the shading and everything was perfect, i dont understand how they could overhaul that when people were actually content with that kind of smoothness. Because now our images have overclocked dramatic lighting, cinematic this, and shading layer that... Making characters or wallpapers was really fun, now we got a different beast entirely with images 2.0

[regal burrow]

Omg yes that’s so true i forgot to mention that in the feedback

[mossy ingot]

I've been seeing online that some people are getting A/B testing again, which suggests OpenAI may be replacing the current models that are having issues. OpenAI is aware of the problem, but there’s no timeline for when a fix will be implemented. Nobody knows if it’s a bug in the image-generation method or if the whole model is problematic, which means OpenAI may have to either replace the models or retrain them. If retraining is the approach, we likely won’t see a fix until Images 2.5 (assuming that’s the next model number). I would suggest not expecting a fix any time soon, honestly.

[plain bobcat]

mention exactly that because thats seems to be our issue with wanting the new old images versus what we get as a result with the new pipeline

[regal burrow]

I create fantasy art too in a painterly realism art style and honestly 1.5 was doing a great job but 2.0 feels like a massive downgrade

[regal burrow]

I have no doubt that they will fix it the issue is all my work is on hold until they do.. that’s why I suggested giving access to 1.5 or at least giving us the option to switch back to the 1.5 version

[kind mountain]

I have the exact same problem like you and the others, my work is also on hold since April 21 because I just can't replicate the style I was creating and loved very much with the new model which leaves me with a lot of unfinished work and no idea how to resume these projects in the future! I really don't understand why they would just completely replace the old model, why not have the option to switch the model if the older one is just better (or more fitting for a certain style etc.) for some projects. I'm also a subscriber to Midjourney but this specific style and look & feel I liked for these projects was only possible with gpt 1.5 - but MJ at least keeps the old models in a dropdown to choose from!!

[mossy ingot]

Hopefully sooner rather than later, but they’ve been aware of this since April 22 and said they’re “actively working on it.” It’s now April 27, which suggests they don’t have an easy fix. I hope the A/B testing people are getting involves replacement models and that OpenAI just wants to see which ones people prefer before releasing them to the public. I think my biggest question is: how did they not know about this issue before release? They had people testing the models weeks before release, they were running A/B tests, and they had to have been using the models themselves. Sometimes I don’t get why OpenAI releases half-baked features.

[plain bobcat]

i understand truly, characters and wallpapers were great and more than adeqate for me and many other using the old 1.5 images. Those old images had our desired shading, lighting, contrast between fantasy and realism and cinematic parameters

[glossy bloom]

I recently hit my limit for the first time on Plus. I wish they had an estimated count remaining because it locked me out from image generation for 15 hours. Had I known how many images remained I would have prioritized the more important ones for a project I'm working on. Now I'm stuck waiting for a day instead, unable to progress.

[kind mountain]

right?? I wasn't exactly doing fantasy or anime, but kinda a "realistic" style which wasn't too realistic but had this exact soft, a bit painterly look which 1.5 seemed to nail perfectly and 2.0 is just COMPLETELY different, more like hyperrealistic which is for sure great for some people, but I don't understand why you would completely replace the old model with a newer one if the newer on is just SO different in style etc (the artefacts are just a problem on top, at least for me). 🙁

[plain bobcat]

The best thing we can all do atm is CHERISH all our old images, im currently looking at mine and they bring a smile, unlike the demons in images 2.0 gpt 5.5 LOL

[regal burrow]

I just hope whatever they replace it with is actually an improvement on 1.5 and not a downgrade

[mossy ingot]

Same.

[analog grail]

for the artifact issue my fix is leave chat gpt do all stuff i'm need with this issues and when get all i'm want go in Images where all your generate picture is storage and one per one ask chat gpt clean fix *

is the only way i had find for get all my stuff clean up

[mossy ingot]

The best way to not get the artifacts is not use Image 2. /s 😭 😭 😭

[analog grail]

i can't^^

[mossy ingot]

Yeah, I get it, just making a joke about how bad it is. Haha

I really really miss editing images without it looking like a total mess.

[regal burrow]

When the 2.0 update came out I was so excited, as someone who uses the image generation tool every day I couldn’t wait to take the art i made to the next level
I was expecting more control, more accuracy..

[regal burrow]

I feel the same way man, I’m grieving 1.5 model

[analog grail]

What you can do is open a new page in Image upload all you previous Style Art you was use and told Chat GPT i'm want sometime Like that ! i could work first generation image but you might have to make a new chat again since is only the first request as good render . but is a fix as work for me just a big waste of time per day :x

[mossy ingot]

One thing I've noticed about OpenAI's image models is that they're not really targeting artists in the same way Midjourney is, and I've always been confused about why they don't compete more directly with Midjourney or expand their dataset. They seem to appeal more to stock images and product photography, as the model doesn't really understand more nuanced terms like "lifted shadows," for example. This doesn't mean you can't get creative with them, or that they don't have artistic capabilities, but they don't have the same depth as Midjourney.

[plain bobcat]

if only it was that easy, you're basically asking to reverse engineer the new pipeline to the old pipeline from before

[regal burrow]

I think it only makes sense that they should allow users to switch to the older model as a feature
Especially since the current model is unusable leaving their subscribers unable to use the tool to create anything useable..
Especially since other platforms have this feature

[wraith oracle]

I mean they did make 4o and DALL E a custom GPT, why not 1.5? I dont see it

[regal burrow]

I actually tried that a few days ago, it did not work for me..

[analog grail]

that should by available too yes

[regal burrow]

Right!

[analog grail]

Aww i see could by very specific about your work stuff then

[plain bobcat]

Should i post an image by image side by side comparing the 2 image pipelines between new and old or nah, you guys probably know what im talking about?

[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.

[plain bobcat]

Justice for 1.5 images and all users that utilized it for their projects or work LOL

[analog grail]

Cell Shading get downgrade as i can see it

[wraith oracle]

For cel shading in anime, nah i think it's fine, cel shading for other media i might understand

[analog grail]

depend for Nova she as posted two digimon picture the left side was the previous model or the new one ?

[plain bobcat]

those were both using chat gpt 5.5 images 2.0. The decent fix was to ask gpt to edit and use "Remove artifacts, make the picture smoother and less chaotic."

[analog grail]

now i'm no realy full disapointed with the new model since Danganronpa is my Main Workflow since last month. the Model is very must better now for this Stuff*

[plain bobcat]

it definitely improved the shading, smoothness, lighting and cinematics from before edit, to new edit, but it doesnt come close to images 1.5 because thats where our problem lies as users who used 1.5 images, we need their shading, smoothness, lighting, and cinematic parameters back.. 🙁 and anime screencap was superior before compared to now lol

[kind mountain]

sounds funny, but it really is exactly like that! 😳 Everytime I look at one of my 1.5 project images, they bring a smile, they are just like an escape to a beautiful world, whereas the 2.0 images are just meh in that regard, no feelings at all

[analog grail]

hmm and now i'm open my big mouth against this guardail sorry we can't do stuff with violence and blood.. silly one danganronpa is a killing game off course i need sometime stuff that way

[plain bobcat]

hell yeah brother/sister!

[analog grail]

We're sorry, but the image you created may violate our rules regarding violent content. If you believe this is an error, please try again or edit your prompt. 😂

[regal burrow]

1.5 felt like it was creating images with heart where as 2.0 feels like slop with no soul..

[analog grail]

Worked ! i did ask to turn the blood in pink so like is paint staints ^^

[regal burrow]

1.5 did amazing things it understood my references and instructions no matter how complicated, it was able to change the camera angle perfectly keeping the characters and objects consistent, it was able to change the art style perfectly exactly how I describe it in the prompt while still keeping details consistent, it was cinematic.
2.0 can’t do any of those things..

And it was able to make precise edits and execute it perfectly

[wild ridge]

We must be using different models lol

[kind mountain]

yes, for me, the OpenAI images 1.5 model was for a completely different workflow/project, something I couldn't quite create in MJ (at least not in a way I liked too much), whereas for other work (not sure how to explain it, just a bit more artistic, like old paintings, or a specific painting style etc.), MJ is my go-to model. Which I think is great, if we find the "best model" for a specific project/task and leverage exactly that... but then OpenAI comes around and just replaces such a specific model with a new one that creates a totally different style... I hate how dependend we are in regards to something like that... you can spend months in creating your perfect workflow with the best models and prompts and then they just take that away in a heartbeat and doesn't seem to care that the new model won't be a fit for every user 😔

[plain bobcat]

unfortunately users are saying their guardrails or restrictions were heavily improved and are more strict now, we'll be lucky to generate "medium ha-ha's" or "cherry paint" as human juice on the floor LOL

[analog grail]

Yeah i hate this Guardrails now still we can pass throught

[regal burrow]

Omg I relate to this

[plain bobcat]

the problem is, if what their saying is our generations are limited per day regardless of plan we have, "i have plus" a failed generation still uses up an attempt which is the most bonkers thing, at least give the image or dont use an attempt if they're gonna block it...

[mossy ingot]

Imagine the creative freedom and power if OpenAI expanded their dataset while maintaining custom styles through future versions.

Hopefully one day uploading an image and copying a style very closely will be the future of these models, but we're not there yet.

[analog grail]

Yes i can understand i was heavy disapointed with premium plus so i had decided to move for Pro. but before images no was limited for Premium plus ....

[plain bobcat]

So Seto, we either need finish our projects with our old images and transition into the new ones or we wait for a fix or even the option to use the old 1.5 images, i really dont want to transition for the project...

[regal burrow]

Honestly I’m considering getting a new subscription on platforms that still has 1.5 available because i have deadlines and i can’t wait for an unknown amount of time

At least until they have 1.5 version available on chat gpt

[plain bobcat]

that is one loophole that kinda works, other platforms that utilize OpenAi still have a crisp 1.5 feel

[mossy ingot]

As far as I know, Image 1.5 is still available in the API unless something changed and I missed it, so you probably could find a service that's still using it somewhere

[regal burrow]

And hopefully by the time they replace the current 2.0 version it’s an actual upgrade to 1.5 and not a downgrade

[kind mountain]

I was testing the OpenAI Playground for that, just to be able to still create something with 1.5, but I hate that it cost extra even though I have a gpt subscription. Just feels "limited" that way. 😕 Oh and it wouldn't surprise me if OpenAI removes the 1.5 model from their API too in the future, so it still doesn't feel like a "safe" workflow

[plain bobcat]

I marked "April 21, 2026" on my calendar, that was a dark day indeed 😄

[regal burrow]

My thoughts exactly, this is why I’m looking for a reliable alternative
I wish we had a one time purchase tool that we actually owned and no one could take away..

Hopefully one day in the future

[plain bobcat]

for now lets do research and find some work arounds, i also gave myself a deadline for the project

[regal burrow]

100% and good luck with your project!

[plain bobcat]

right back atchya!

For now I made this impressive new recipe and its funny lol

#images-canvas message

[acoustic peak]

Is 2.0 getting fixed? The quality it had when it first came out was crazy

[analog grail]

no too bad as Ai for Sutff Pixai . Art is work with tokens but price is still very low too you can make lots stuff in anime stuff and zero guardails

[tiny sonnet]

Did they nerf chatgpt image 2? Image to image?

It's awful, looking like version 1.5.

[wraith oracle]

No, it's a artifacting issue

[tiny sonnet]

So I can hire without any problem, right?

[jaunty ingot]

does anyone know why gpt images 2 makes realistic images blurry/splotchy??

[jaunty ingot]

yeah im getting cooked because of the weird splotchy painterly effects on "realistic" stuff

[tiny sonnet]

Add these errors to the command prompt to avoid them.

[jaunty ingot]

i use GPT on web

[tiny sonnet]

Using it on your cell phone might cause your screen to malfunction.

[jaunty ingot]

im on PC

what does that have to do with the fact how it's generating images

[tiny sonnet]

If it's not on your phone, it's probably your screen or even your video card.

[tiny sonnet]

They must be rearranging the Sora servers for the image 2, right? I forgot that today was the last day.

[civic dust]

its a bug, just sit and wait for them to fix it. You really can't make some types of images on gpt right now

[velvet rampart]

There has been a lot of complaining about artefacts in 2.0 images. Could some one clarify this issue for me? When do they appear? Are they always there? Etc.

[civic dust]

mostly in photorealistic images but anything with a repeated texture in it

[acoustic peak]

I hate that its bad on photorealistic images rn those are my favorite ones to make 😭

[civic dust]

ikr? I can hardly make anything with gpt right now

[light dome]

2.0 excels at realism on close up skin. there is no comparison. But the artifacts on hair and objects further away from the camera limit it. I use a combination of a couple of models to somewhat fix the issue

[vivid zodiac]

They tend to show up at random. It's showing up for me on any image generation that is an illustration.

[limber shoal]

fantasy styles too

[limber shoal]

I wonder how did the image 2 model made the images so perfect without the artifacing in A/B phrase

[mossy ingot]

No idea, honestly. They're doing A/B testing again...

[light dome]

Maybe it was done on purpose to give other models a chance to catch up to what open ai is capable of lol

like here is a small sample of what we can do. better catch up soon! lol

[civic dust]

if you show an artefacted image to the AIs it can't see it, so this must have passed automated testing and going on what happens in this discord, I feel like the testing set heavily biases towards anime

[light dome]

There are AIs that can see the artifacting and correct it

but if you run an artifact image through 2.0, the results will usually remain the same

the problem with the other AIs that can correct the artifacts from 2.0, is that when they fix the artifacts, it also makes the good detail that 2.0 provides look more like plastic instead of real. When 2.0 works correctly, there is no comparison to the realism it provides. but unfortunately its pretty broken right now

[civic dust]

they seem to not see it and just happily redraw the scene, which takes the artefacting away. Been using NB2 like this

[light dome]

What is NB2?

[wraith oracle]

Nano Banana 2 from Google

[light dome]

Oh nice! I use another model for the same purpose. Only because i hate using what everyone else uses and its cheaper lol

but since i havent used NB2 to correct gpt2.0, let me ask, when correcting does it retain the detail of 2.0 on the skin?

this is an image straight from 2.0 when its working correctly #images-canvas

the green girl

no artifacts at all. but keep in mind i also reduced the chances of artifacts showing by getting rid of the weaknesses the current model has

i purposely kept things close, added no hair

when you have hair and have things at a further distance, the artifacts show

once open ai fixes this, not even NB2 will be able to compete

[civic dust]

well they'd have to fix a few other things like it following instructions better and not just lining everything up. I'm finding NB2 wildly better right now

[light dome]

Yea, i have heard multiple people say that abot NB2 so you wouldnt be wrong

[wraith oracle]

Agrreed, prompt adherence is way better on GPT Image