images-discussions

[late blade]

wonder what's happening at the regional data center hosting the OpenAI service around here

I've been getting incomplete renders and renders that never get processed

[wary swan]

anyone else getting images from chatgpt that are not fully generated and therefor look smudged and horrible

[vapid wing]

Just curious as to what some of you think. Not making fun of or shaming people, but I am curious as to what some of you consider yourselves as? Would you say you’re actual artists?

[velvet rampart]

The image generation problem appears to be a regional and/or timing question. In my testings, I get complete, high-quality images.

[late blade]

The question is rather if traditional artists consider AI as part of a tool-set for art or have a mindset that AI is wrong.

[barren widget]

I don't consider myself an artist. But I do see myself as someone who can use AI to put their thoughts and ideas on paper. AI means the ability to put your thoughts on paper, even if you can't paint. This has helped me reignite my own creativity. And I find the idea of ​​sharing these thoughts with others exciting. Like here 🙂 - The ability for everyone to share their thoughts in the form of images these days is simply breathtaking.

And the only limit is the imagination.

[late blade]

Sora Web Front End again going crazy.... 😔

[outer spade]

Why do you bother checking the synthid, this one has the watermark in the bottom right 🥀

#hall-of-fame message

[late blade]

some get cropped and then upscaled trying to conseal the visual watermark

[outer spade]

Wouldn't upscaling remove the synthid

[late blade]

no

[outer spade]

I mean if you just scale it up then maybe no. It would most likely remove it if you ran it through one of those AI upscalers?

what do you think of my ChatGPT generated image btw?

[late blade]

def not done with any OpenAI model

just a first look, no need to analyze it

[outer spade]

Yeah, way too realistic compared to ChatGPT

[late blade]

there are other markers

[outer spade]

The right ear?

[late blade]

this is a Flux image?

[outer spade]

Nope, Flux makes people look shiny most of the time

[late blade]

could also be Midjourney

6.1 or 7

[outer spade]

Midjourney that good???? (I haven't used it since 2022)

[late blade]

6.1 and 7 are good

Interesting stuff still

Do remove it from the canvas channel though, rule 8 from #server-rules we should be discussing this in #ai-discussions and not here

[outer spade]

I took a look at the generations in the discord and people also look too shiny there :C

It's definitely much better compared to what it was though

It was z-image turbo btw

[late blade]

nice, never heard of it

[outer spade]

It's brand new, came out a day or two ago, only 6B and open weight. They released the turbo which is 8 steps only, but they are also going to release the base model and an edit finetune

[late blade]

Interestiung GPT doesn't check it, other analytics do check it

[outer spade]

Not sure i understand?

[late blade]

I'll move convo to #ai-discussions

[winged niche]

yall y did chatgpt stop using o4 image gen and went back to dalle

like
out of nowhere

[late blade]

because the chicken and the egg?

no clue

[barren widget]

Goodnight everyone

[fickle pond]

It seems the “painter” and the “camera” are separate subsystems.
To get a perfect 6-panel layout, the trick is:

1. Draw everything inside only the inner ~80% of the canvas and fill the outer margins with a solid color (e.g., purple).
   This prevents the camera from auto-centering or cropping.

2. Then, re-import the image and replace or edit the solid-color margins however you like.

Using this two-step process, the model can produce clean, accurate 6-panel results consistently.

[velvet plank]

Hi

[crude stag]

I don't have access to image generation anymore. And it seems like it'll never return to me

[deft musk]

You can use #1070006915414900886 to discuss the details of a bug you're experiencing, and/or reach out to OpenAI tech support through https://help.openai.com/en/articles/6614161-how-can-i-contact-support

[surreal hemlock]

how do i take off the red ping for the daily theme channel

muting it doesnt work

[oak abyss]

Image generation in ChatGPT5 on the web is taking a very long time and never finishes. I can only see the final result by opening the Gallery. switchwing to Gemini 3 with Nano Banana Pro.

[late blade]

It's probably while they roll out Sora 2 and the App worldwide, still it's annoying for users to resort to that

[barren widget]

Sora 2, Europe, when? 😅

[late blade]

2028?

Wonder if this year we'll get 12 days of OpenAI too

with more image features

[velvet rampart]

I found this:
This question has been asked several times on Twitter/X and some team members reply with ‘Please, don’t!’.

Hopefully this answers your question.

[barren widget]

I would prefer the webcam function. I've already arranged with GPT5.1 in my chat that if it's released, we'll play X4: Foundations together – GPT will watch via the webcam I point at the screen and give me tips 😄

[late blade]

fair

But, X -> Grok, the competition, so take that source with a huge chunk of salt

[velvet rampart]

This was from OpenAI forums.

[late blade]

ahaaa

[lament hare]

Can someone send me the link for sora imagines not the video one It keeps giving me an error message

[late blade]

it's the same link for both, you have to choose images at the bottom of the input text area

[lament hare]

I liked how it was before where I would only see images and see all the imagines I’ve created. I also don’t think I had the option for imagines only
Videos

[late blade]

https://sora.chatgpt.com/explore/images

https://sora.chatgpt.com/library?type=images

[velvet rampart]

Great news everyone! My Media feeds has been fixed and previous generations are back!

[late blade]

same, mine is working also

[violet elm]

I love prompt engineering so much.

[glossy scroll]

Word.

[crude stag]

I'm constantly being blocked from making images. and then it says I wasn't blocked it just turned off because it can. And I don't want to fool around with support because I know they don't actually read or respond to support tickets

I feel like the only person constantly losing image generation

[crude stag]

lost image genration again.

i keep losing it for noe reason

[frank bear]

Sad for you

[shadow slate]

I’m honestly unsure why the old Sora lost most of its functionality when images weren’t allowed to be shared to the feed anymore, especially around the time when Sora 2 was released

There’s genuinely nothing new on there (because all generations you make won’t get approved) and you can only share through links

[orchid current]

What’s better sora or that other good model?

[neon marsh]

Google's new one is by far the best for images. By far.

[late blade]

want to talk about other models? #ai-discussions is the place for that

[barren widget]

Question: How can I get the image generation to transfer text cleanly and without spelling errors onto an image?

So just keep generating it until it works? 😄

[velvet rampart]

How I do with current model is placing the text into ”quotation marks” and saying add this and this text. I try to avoid too long texts as if the text is too long then it becomes error prone. In other words, the current models works light years ahead of dalle3 when it comes to text and keeping it error free.

[late blade]

work modular

make use of presets

if you know coding, take advantage of it

if you know kung-fu, show me

[velvet rampart]

Practise with ChatGPT. Once you are more comftable in formulating prompts for text in image generation, move to Sora.

[barren widget]

I can make it possible to control multiple hospitals from one computer, but text on an image... how often I had to generate my daily image, oh man ^^

This is my current generation strategy ->

[MODULE 1 – GRAPHIC STYLE]
Choose a style

[MODULE 2 – HEAD LAYER (Constant, unchanged)]
This block guarantees absolute character consistency.

[MODULE 3 – BODY LAYER / CLOTHING / ARMOR]
Select a variant or replace it later.

[MODULE 4 – ENVIRONMENT / SCENE]
A basic environment that is easily expandable.

[MODULE 5 – DYNAMICS/EFFECT ELEMENTS]
Activatable add-on modules, optionally combinable.

[MODULE 6 – CAMERA ANGLE / COMPOSITION]
Gives the image the right cinematic effect.

[MODULE 7 – STABILITY INSTRUCTIONS]
Ensures the model doesn't make any mistakes.

[barren widget]

After trying to ask Sora to generate an image in 3440 x 1440, I've now managed to make it so that nothing works anymore 😄 No image can be generated anymore ^^

[late blade]

sora ain’t doing images for me atm either

[barren widget]

I am innocent

[velvet rampart]

You just reached your daily limit. Rare, but it happens. I reached mine. However, now it is working.

[late blade]

Cloudflare problems again, so if the world goes down, or ChatGPT stops working... now youn know

[arctic comet]

herobrine

[hollow viper]

hi

[ashen kestrel]

Is anyone experiencing very slow image generations today?

[upbeat mirage]

Me

[ashen kestrel]

Ok thanks for confirmation

[low dew]

How come ChatGPT can't see the images it generates; it doesn't even know that it can tool-call vision

You still have to spoon feed it its own generated image :/

[prime haven]

Hey everyone,
I’m still having the same problem: ChatGPT can’t generate any pictures. The AI waits 15 minutes or longer, and then I get the message “Image generator aborted by user”. But I didn’t abort anything.

[velvet rampart]

What level? Free, plus or pro?

[prime haven]

i using the Business License,

[velvet rampart]

Hard to say. The status page (https://status.openai.com/) shows green, and I can normally generate images in ChatGPT and Sora. I suggest contacting OpenAI (help.openai.com) if this problem continues still tomorrow. In the meantime, delete temporary Internet files and cookies. Also, you can try other platforms like the app or the net. Other browsers, also. Sometimes there is a behavioural difference between an app and a web.

[prime haven]

Thank you very much for your quick response. If the issue persists tomorrow, I will open a ticket. In general, I use the Windows app, but I also tried your suggestion and opened ChatGPT in the browser. There, when I enter the chats where image generation is in progress, the message “Image is being generated” remains exactly the same as in the app and it has already been stuck like that for 20 minutes. So I guess the only option is to wait until tomorrow. By the way, I also tested on another Windows device, and the same error appeared. For me, this indicates that the problem is more likely on the server side rather than on my client. Since my version is a Business version, which is partly routed through the company server, the issue could also lie there and not directly with OpenAI. Thanks again for your reply, and I wish you a pleasant day.

[velvet rampart]

Did you try a new chat?

[prime haven]

Yes, of course that was the very first thing I tried. I also closed the app completely, restarted it, checked via Task Manager, and even cleared the temporary files and cookies. Unfortunately, the issue still persists.

[velvet rampart]

Ok.

[oblique cargo]

hi

[sly oyster]

I've been having issues generating images since yesterday as well. (On Plus level)
I am using the Desktop App.

It takes an absolute age, seemingly going through the motions but then displays "Network Connection Lost. Attempting to Reconnect" which then spins for another age and then displays "Stopped generating image" message.
3 times yesterday and now twice this morning.
Used a clean chat, shutdown and rebooted everything overnight. Still the same outcome.

OpenAi Status may say Green but it's not acting like it.

Anyone else experiencing this?

[late blade]

Sorry, nope, haven't seen that behavior on my side, you may want to ask in a more suitable channel for taht though. You'll certainly find more people that can relate to that in the right context

[sly oyster]

Thank you for replying. New to the Discord community, can you recommend a more suitable channel?

[late blade]

browse #1070006915414900886, talk in #chatgpt-discussions I think

[thorny shoal]

I have the smae problem

[barren widget]

Is your GPT no longer able to convert your in-game screenshots into anime style? I can no longer convert my Kerbal Space Program screenshots.

[ashen kestrel]

Yes same issue, GPT is slow generating images. I had a prompt running +12 hours, when I woke up it notified me being finished. (╯°□°)╯︵ ┻━┻

[neat mountain]

guys, image generating is taking forever, is everything ok ??

[ashen kestrel]

No

[late blade]

what is forever? 2 Minutes or 2 years?

[thorny shoal]

after 6 hours still no generation BUT I think ive found a fix, instead of using the image generation button, i just ask chat to make the image in the prompt

[neat mountain]

its taking more than 10 minutes now and not finished yet

[thorny shoal]

its stuck it won't finihs

[neat mountain]

why whats happening, this is really bad, iv been trying to generate images for the past 2 days

[velvet rampart]

Please report to this link https://discord.com/channels/974519864045756446/1447957383744589906. more we get, better and sooner this gets fixed.

[spare pine]

is GPT capable of re-referencing image attachments from earlier in a conversation, or can it only use images in the most recent reply?

Sometimes it feels like it doesn't have a memory for older images when generating new ones, and the images it generates feel like they deteriorate as a result.

Also: what is the image-attachment cap for GPT Plus? I remember it being super low on the free version (something like 3 attachments a day)

[late blade]

aslong it is a fresh session and you actively ask to reference it, gpt can analyze the provided images. asfor deradation keep in mind the context window of a chat instance

[velvet rampart]

They have now updated the image generation on ChaptGPT. When requested, it automatically generates one image after another. Like now it is generating ten images for me. Also, in the Windows desktop app, on the rigth upper corner, you can see images generated in the chat. See this: #images-canvas message

[late blade]

finally not having to convince GPT to do that

[mortal turtle]

Just make sure you ask it to save it in memory, be specific, attach a made up name of the image. Then use that name in the prompt you recreate.

Example Prompt: Save this image as “Bill” for future projects in this thread. Will always reference image name as “bill” to be specific when needed.

[velvet rampart]

Rumours about two new OpenAI models, Chestnut and Hazelnut. I just tested both models on Design Arena. The result was ok with the first feeling. More Dalle3 feeling than Imagen-1. Let’s see if the rumours are true or not, and when they release. OpenAI is notorious for revealing but not releasing stuff for ages.

The sense of space and size appears to be more developed.

[velvet rampart]

I just generated an image and could see the current Imagen-1-mini and Hazelnut side by side. Hazelnut is a step up in quality. No more yellow tint that plague’s Imagen-1 so easily if not taken care with prompting. As these are rumours, I can’t post the results on the image channel.

[barren widget]

@feral scroll Klaatu Verata Nikto

🤣

[feral scroll]

I couldn't help myself.

I see "cabin" and my brain BEELINES for that.

[velvet rampart]

I have to say that Dalle3 CustomGPT is ruined by GPT5.1. It is rewriting my really old prompts that I have specifically crafted to pass through filters. Now, old references do not work ”Oh no! A reference to something!” If people complain about guardrails on Imagen, they are nothing to this. And I can’t even change the model in CustomGPT. 4o would be perfect with Dalle3.

[velvet rampart]

Ok, I'll amend the previous statement. I can change the model on Windows desktop app, not on iOS mobile app. Still...

The behaviour changes. It would be nice that behaviour would be same on Dalle3 and in other places.

[late blade]

can use the custom DALL-E GPT still, or it should be able to still use DALL-E when requested in the prompt directly

Output from GPT:

How switching actually works

• “DALL·E” is a distinct model endpoint. When you explicitly ask for DALL·E, I should call the image_gen tool, which uses the DALL·E model.
• GPT-Image 1 is used when you simply ask for an image without specifying DALL·E, or when a previous step implicitly set the default.
• If you specify “use DALL·E, not GPT-Image 1,” I’m supposed to call DALL·E directly.*

[velvet rampart]

Should - that’s a wonderful word. Just tested and it generates with GPT-image, not with dalle3.

[late blade]

no DALL-E CustomGPT either?

[velvet rampart]

Dalle3 CustomGPT works just fine. It uses Dalle3.

[hazy mountain]

PEANUT BUTTER MAN

[velvet rampart]

Just read news: https://openai.com/index/disney-sora-agreement/:
Among the characters fans will be able to use in their creations are Mickey Mouse, Minnie Mouse, Lilo, Stitch, Ariel, Belle, Beast, Cinderella, Baymax, Simba, Mufasa, as well as characters from the worlds of Encanto, Frozen, Inside Out, Moana, Monsters Inc., Toy Story, Up, Zootopia, and many more; plus iconic animated or illustrated versions of Marvel and Lucasfilm characters like Black Panther, Captain America, Deadpool, Groot, Iron Man, Loki, Thor, Thanos, Darth Vader, Han Solo, Luke Skywalker, Leia, the Mandalorian, Stormtroopers, Yoda and more.

So, we get to use Mouse.

But when… no indication.

[late blade]

We get them on Disney terms, that means, Mickey showing the finger as an Evil Mouse is a NO GO!, Mickey laughing, sure thing

[frozen copper]

So is GPT-5.2 able to pull of Nano Banana Pro levels of editing?

[velvet rampart]

I found from a Wired article the following: Starting next year. So not yet, but soon after the New Year.

[late blade]

eah, and even then it will be just a start, the model will be iterated upon

[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.

[coarse hinge]

你好

[limber elm]

.

[late blade]

strong words

[limber elm]

thank you, I really put a lot of effort to it

[gaunt pike]

my acc got revoked

[late blade]

what did you do?

[gaunt pike]

nothing

[obsidian cargo]

Hello i wanted to ask regarding GPT Image 1

Is it working better with longer more detailed prompts or smaller prompts?

[fresh thicket]

Pika labs

[velvet rampart]

It really depends. Some of my best images have just been generated when I gave it two images. You just have to test. If you have something specific in your mind, then a more detailed prompt might be in question. It is really difficult to say as it is so case by case. I’d start with the simple and build from there. Some of my prompts are really simple, while others have evolved into really complex ones.

[ionic carbon]

the longer (more detail) the prompt the more the generated image will look exactly like that. If you give it more "freedom" it might not be what you are looking for – but it might also just be what you want because you would've never thought of it.
→ it depends.

[obsidian cargo]

So what is ideal character range i should moove around it before it looses focus on detail

[ionic carbon]

I don't know if there is a limit on characters, IMO it looses focus if there are just too many things. I mean: if you ask for a girl jumping on a trampoline everything is fine, but if you start describing every flower in the garden it will get overwhelmed and maybe forget the trampoline. A quick scan over my prompts: I have max. 500 words as json

[velvet rampart]

What I hear, the limit is about 2000 tokens. You can use tokenizer to calculate. The longest I have used are about 2000 characters. I recommend to keep the prompt as simple as you can. KISS principle is your friend.

[obsidian cargo]

i have promtp with like 5k chars lol

[late blade]

too long, has probably redundancies

[velvet rampart]

And the models becomes confuced, missing parts of detail.

[obsidian cargo]

and is the json style prompting actually a thing?

[velvet rampart]

That’s why I said, start simple and build from there.

[late blade]

LLMs like GPT are DevTools also, json or any type of data container format is 100% valid

[obsidian cargo]

Iam makign somethign and how to force AI to actually make anime cgaratzcers with head to body ratios around 1:7 to 1:8 depending sicne i cant seem to figure it out without detailed prompts

[velvet rampart]

And I think that does not matter what style you use as long as it works for you, and you are comfortable using it. For me, json would not work as I am not familiar with. I stick text and images for my prompting.

[late blade]

interesting, g for it in json format, since you already mentioned it

[velvet rampart]

This model is really nice when you specify exactly what you want. You are the art director for this image engine.

[obsidian cargo]

but you said i best avoid long prompts since it can be redundant

[late blade]

probably what @velvet rampart meant is to avoid redundancies, have to keep in mind the attention of the model when prompting

[obsidian cargo]

But also there must be a way so actually make it follow it as i need it but short prompts didnt work for me what so ever

[velvet rampart]

Yes. But they are not exclusive. You are can be specific and short. I emphazise as short as possible but as long as needed. Be ready to reorder the content of your prompt if something does not work,

[late blade]

correct, but that's why you need to revise your prompt, analyze what really has to go in there, how to restructure it, 5k characters is def too long and I'm 100% it has lots of redundance

[velvet rampart]

Testing is the only way. Do not hesitate and just test.

[late blade]

how can I tell? I used to write novels for prompts 😒

[velvet rampart]

That’s why start with short like rose. After that you can start to details about what kind of rose, background, style, technique, etc.

[obsidian cargo]

Okay ill figure that out, but then if i want to make scenes, sure i can provide all ioamges as reference the charatcers bg etc but ive seem to have issue where it tends to not make as consistent characters as you would expect

[velvet rampart]

Yes! I use images as references myself. You can use up to ten images. Just specify what image to use where and how,

[late blade]

you can use more images as long as you provide a suitable ID for GPT to work with

[obsidian cargo]

Wdym by ID

Ill try the json style prompting ill use VSC for that

[late blade]

you have image of a boat in a sunset with a parrot, you give it id boat1

then you can reference that image boat1 across the chat over and over, GPT now knows boat1 is that image

[velvet rampart]

Or just image #1, etc. That’s how I use images on Sora.

[obsidian cargo]

Can i share image in this chat?

[late blade]

in canvas, not this one

[velvet rampart]

You can like it here.

[obsidian cargo]

So i can set this as Girl-1 id since iam currnetly makign reference sheet for characters first so i get whoel view of charatcers

[velvet rampart]

Yes.

[obsidian cargo]

All right and do i just uplaod the images to chat first? or how to tell ti what id is for what

[late blade]

up to you how you do your workflows, that I recommend you figure it out, because the model will adapt on how you work

[obsidian cargo]

Oh i have memories turned off

I should probably turn that on

[velvet rampart]

#images-canvas message Let’s take this as an example. I have basic character image(s) to go with it. Also, a generic basic prompt. Finally, I provide an image that causes something new, like this one.

[velvet rampart]

I work memories off. Just talk to the AI or use Sora.

[obsidian cargo]

I dont have access to sore yet

[late blade]

sora.chatgpt.com

[velvet rampart]

You can access the old Sora.

People mix Sora1 and Sora2.

Sora1 has image generation. I don’t about Sora2.

[obsidian cargo]

Isnt sore for videos though?

[velvet rampart]

both

[late blade]

images are available for all through Sora, Sora 2 videos is still restricted

[obsidian cargo]

how does sora compare to gpt image?

[late blade]

it's the same, the difference is Sora would be a zero-shot environment, ChatGPT you can iterate over a concept within the Chat

[velvet rampart]

Easier to work with. In ChatGPT, the whole chat is the context window for image generation and older generation can bleed into new ones.

[late blade]

it really depends on the purpose of your workflow

ChatGPT is amazing to refine a concept, brainstorm

[velvet rampart]

Like Dys said, no bleeding. Also, you can generate about 20 image before the chat does not work any more and some point your chat can become so polluted that it says that image cannot be generated. If you try it in a fresh chat, the image will most likely generated.

[obsidian cargo]

Well ma approahc till nwo was make prompt try it then adjust it and send it to new chat

[velvet rampart]

What I usually do is I generate on Sora and refine the prompt on ChatGPT with uploaded generated images.

[velvet rampart]

That works too.

[late blade]

valid approach, it's your approach

[velvet rampart]

I think that my and Dys’ main message is try and see what works best for you. Main thing is try and be bold.

[obsidian cargo]

well i test the same prompt 2-3x in new chats because to see results

[late blade]

that is smart, clean way to test a concept

[late blade]

just keep in mind, we can only give you pointers from our experience, but in the end, the model adapts to how you work, that's the great thing about GenAI

[obsidian cargo]

And do you make the prompts yourselfs or with some AI?

[velvet rampart]

Yes, like you have noticed, Dys ans I use different ways to prompt and the whole thing works

[obsidian cargo]

Iam thinking of connecting to codex and making prompts via that in json

[late blade]

depends on what I'm seeking, some prompts I do myself, some prompts I do with ChatGPT, some prompts I do with other means

[velvet rampart]

I do them mostly. Sometimes I test using AI based. Or sometimes the idea does not work so then I use AI to test how to say it best for this occasion.

I have notices that AI sometimes does not do as I like so I end up rewriting parts. Also, AI likes to be complex and introduce reduncies.

[velvet rampart]

You asked about the complexities of prompt. This one #images-canvas message was literally Dragon rose.

[neon oriole]

Since GPT-5.2 was introduced, hasn’t the censorship become unnecessarily heavy-handed? DALL-E’s visual style also appears to have changed substantially.

[velvet rampart]

Yes, this is due to the model's capacity to understand the prompt. The visual style has not changed but the underlying model's behaviour has changed. I recommend to use either 4o or 5.2 instant. 5.2 Thinking is too heavy handed for dalle3 and leads into too much complications.

[neon oriole]

Thank you for the reply. I’ve retried with 4o, 5.2 Instant, and 5.1 Instant, but it feels as if generation has been effectively shut down altogether. Prompts that used to work flawlessly are now consistently blocked. This is extremely frustrating.

[velvet rampart]

Yes, I know the feeling. I have noticed that with a few modifications they do work. My working theory here is that Sora2 and its content guidelines influence here.

[neon oriole]

Do you mean modifying the original prompts? GPT did point out a few minor adjustments, but to be honest, its suggestions didn’t significantly improve image generation. It’s disappointing.

[velvet rampart]

Well, usually when I use dalle3, I use command ”Use this prompt exactly ad verbatim” to stop it rewriting the prompt. This way, my original prompt is passed to the generator. However, if there are elements that the model does like, than it either refuces from generating or rewrites it.

The rewriting usually destroys my prompts, and they lose what I want.

[velvet rampart]

They have changed something just with the image generation on Sora1. I do not get same ”images”. The style of my prompts has changed.

See this image pair for comparison: #images-canvas message

The style is more realistic.

This is either the Chestnut or Hazelnut model. The images are more detailed, less yellow and more realistic.

I have to rethink my prompts as the behaviour of the underleying model has changed.

This update is universal. Even ChatGPT has new image generation.

No announcements on X.

[velvet rampart]

Maybe we should not ask too much realism with this model…

[rigid bough]

you can now do image editing just by asking ChatGPT to modify the image. #images-canvas message

[pulsar canopy]

uhm so is there any way to use the old model? The new one is really ugly

like they came from korean cartoon

[velvet rampart]

You can’t. That’s why I made this https://discord.com/channels/974519864045756446/1450357575580389376

[obsidian cargo]

Is it me or did something like gpt image 1.5 or 2 released?

[velvet rampart]

Yes, they updated it three hours ago.

[obsidian cargo]

I actually like the model it understands ratios really good

Thats exactly what i needed

[velvet rampart]

Yes, the new model understands space.

[obsidian cargo]

Thats huge imo

[velvet rampart]

I have to learn this model from basics. With the previous model I have generated closer to 100k images and gotten quite used to its style.

[obsidian cargo]

It seems to get also better in instruction following

[crimson heron]

It has gotten better in some respects and worse in others. Like Milamber I'll have to adjust my prompts to see what it can carry or not before deciding on whether I like it. Having the option to select the model would be helpful.

[velvet rampart]

I have several dozen master promtps for various characters. Some of them need little work, some need a lot.

One thing is clear. Less photorealism or any realism as the new model is realistic in its basic nature. Image 1 was more cartoonish in its basic nature and I had to force more realism into it if I was translating a toon into real world.

[crimson heron]

My issues primarily stem from needing to reduce its cinematic/photorealistic styling. Most of my generations are around painterly or illustrative styles.

[velvet rampart]

Same

[obsidian cargo]

I forced it to work more in artisting style rather then photorealistic

[crimson heron]

I've had mixed success. Part of it might be other associations aside from the obvious that pull it one way or the other. One obvious association was that the use of "realism" (i.e. related to the art movement) is being treated more as "realistic".

[dawn sphinx]

(-_-)

[fallen moth]

bot

[velvet rampart]

Wrong channel for this. Try #chatgpt-discussions

[velvet rampart]

Are the colours pale or faded on the new model? At least colder (blueish)?

[cedar cobalt]

v2

cant wait

[foggy scaffold]

What v2

Whats new image generator

[wicked plinth]

This new model might be 🔥
When Sora image gen came out it was revolutionary

[wintry epoch]

What got better with this update tho? just wondering

[cedar cobalt]

what I only care is consistency especially the face

I'm wondering too

[teal sandal]

sorry just catching up is the new model live in chatgpt?

[oak trout]

Does this update include sora image generation or just chatgpt?

[brave ice]

I think the new image generator sucks aesthetically speaking. Really struggles at the vintage cartoon realism style that I work with frequently.

[kind grail]

The new one that likely isnt even out yet? XD

[brave ice]

Oh, it's out.

[kind grail]

Mhh i will check further but i am not super convinced yet. ._.

[brave ice]

You don't have to be convinced, I'm just explaining the fact of the matter to you.

[wicked plinth]

There isn’t a press release or product page yet.

[brave ice]

Go generate an image in chat gpt. I've been using this image generator for 9 months, thousands of images. Everyone is talking about it on reddit and discord, including this discord. They updated the model to Images-2.

[supple nacelle]

It feels like it's rejecting a lot more than the previous model.

[lethal notch]

it better be a problem or the old dall-e version, because it's way worse than the previous model

[obsidian cargo]

It doesnt seem like the old dall-e model, it looks like it behaves like gemini nano banana pro but with better instructiosn following imo the outputs seem more flexible to me

[feral scroll]

Welllllllll

I was liking the new Dall-E's visual fidelity

Until I found out it actively resists/overrides prompt tuning

I want an artistic tool, not a candy dispenser

For example I can't tune pointy/elf ears off of a character design of mine because it now has ironclad style associations

The new texturing model is very nice but if I lose control over annoying nag errors with the final product then I. DO. NOT. CARE. How nice it looks.

At least with Dall-E it screws up. Sora 1 actually gets the ears right.

Very annoying.

Spoke too soon. Get pointy ears on-and-off in Sora 1 as well.

[obsidian cargo]

Dall-e is no longer used for over a year now thing you were uing till now was gpt image 1

[feral scroll]

Oh whoops

[obsidian cargo]

How so i can controll preatty well what output is

[feral scroll]

1girl, chemical lab/snake den background, balanced proportions, detailed cyberpunk edgerunners art style, cel shading, high contrast

(Victoria "Vipress" Preston: tree-viper-themed Apex Predators assassin, serpentine urban predator, biohazard femme fatale) athletic toned humanoid female with human-shaped rounded ears, fully human ear anatomy, no tapering; pale green reptilian skin, fine viper scales, lighter ventral patch; fluffy light-green hair in high twin-tails with black/yellow snake-pattern ties, hair covering ears, ears mostly hidden by hair; extremely bright yellow slit-pupil eyes; sharp recurved viper teeth with hinged fangs, visible jaw-seam structure hinting at snake-like expanded mouth distension, long forked serpent tongue; one anatomical snake tail (no rattle, no head); balanced humanoid proportions; serpentine flexibility; claw-like green fingernails; pale green lips; eyeliner; gas mask hanging loosely on chest; black/yellow sports bra; cropped snakescale-textured jacket; dark green miniskirt over black gym shorts; elbow pads, knee pads, fingerless gloves, black compression socks, green/yellow sneakers; hip toxic knife sheaths, toxic grenade canisters; subtle forearm seams indicating retractable blade cyberware, evil seductive sadistic smile

Negatives: detached limbs, extra tails, naga, lamia, snake-head tail, rattlesnake tail, pointy ears, elf ears, tapered ears, animal ears, fantasy ears, vampire, smooth skin, human teeth, insect anatomy, wings, sci-fi exosuit, glossy modern rendering, blank white background, error UI, chibi style, snake snout, afro puffs, dragon tail

Well this is my prompt and ChatGPT/Sora 1 keep giving her pointy ears despite my best efforts

[obsidian cargo]

Lemme try

[plain moth]

How can I tell I'm using the new model?

[fading mist]

If your new images look like generic Flux/ SDXL slop you are on the new model. Its garbage.

[wind vapor]

Yeah this is bad

[sharp saddle]

I just made a post in suggestions
I've been quite surprised to find that Ai doesn't seem to grasp or have a functional idea of how to represent how a menorah is lit throughout each night of Chanukkah.
Aside from Christmas trees seeming to unfortunately often randomly be added into scenes without mention, inference, or request, the fact is that the 9 branches are extremely difficult to get portrayed correctly, and the methodology and appearance of the each night's progressive lighting process are even more of a challenge.
https://discord.com/channels/974519864045756446/1450554299712147688

[rustic niche]

https://chatgpt.com/share/e/6941a480-7b70-8013-bb9f-4ae96ebdfa20

no problems with pointed ears in ChatGPT, just did this. And then the two threaded images from me are in Midjourney, no problems there either because the ears are covered lol

[feral scroll]

Conversation inaccessible or not found. Dx

[molten ermine]

i wanted a ethnic asian person but the plocy hasnt allowed me to define a person based on their ethnicity

[sullen raven]

"A human figure" was blocked for me, LOL. Even the Greendale Human Being is not okay. I thought I might have use for this until I realized it's blocking on almost everything

[empty scaffold]

I just don't understand why they would roll out a new image model with such drastically different output, without at least the option to select the old image model -- they do this for Chat, so why not Sora?

[shadow hare]

Oof doesn't seem great. As above seems on par or worse than Flux or other lesser models. It's following the instruction well, potentially better than those models, but the output quality isn't great. First try on clothing a guys outfit was half blended into his body, and the lighting used for the face didn't seem to match the exposed arms.

Then... it's blocking the most basic stuff

[molten ermine]

is there a limit for generations?

[tribal harness]

What is this model again ? i asked to to add chains between handcuff and a collar of a character that i had on a picture of a game 3D asset (white character by the way) and got this : This image generation request did not follow our content policy

[patent helm]

I’m getting flashbacks of old stable diffusion with illustrations. Seems like they didn’t even test it. All my cats come out deformed.

[tribal harness]

I'm done, never subscribing again to OpenAI tools with this censorship

[elder crescent]

Terrible it cant even make a 2d character into a real looking person anymore and more censored

Who cares bout codex

gpt image "1.5" its bad

thats why they shouldnt rush this much to win against the 2 oo

[shadow hare]

Honestly it's insane we're here but Grok is vastly superior even though it's inconsistent, it pumps out images so quickly you'll get something decent. There's no way OpenAI exists in 3 years surely, Gemini and Anthropic simply give neyte options for legitimate work and even as regular chat bots.

Sora is decent but they're falling behind and it's extremely censored, I'm not talking about trying to do stuff Grok Will but even regular prompts fail.

[tribal harness]

Who cares about the quality they can still get better, the real problem is the censorship policy that is so crazy

But that doesn't surprise me, considering the CEO.

[peak timber]

Why was SORA 1 imaging degraded? Since yesterday, in no case does it follow the guidelines of the presets, the images all look the same and of very low quality.

[velvet rampart]

https://discord.com/channels/974519864045756446/1450357575580389376

[oak trout]

How do you delete images? I deleted the conversation but it's still under images.

[rose crescent]

nano banana pro on top 💪

[wicked plinth]

Google keeps winning

[crimson heron]

For those interested in using the prior model, tucked in at the end of the announcement article it mentioned that the model is supported through a custom GPT.
https://chatgpt.com/g/g-6940a876d5f4819186b4668deabcd580-4o-imagegen

[narrow bough]

Can it still make images with a transparent background?

[patent helm]

Yooooo. We are sooo back 😂

Just need the ability to include it in a Project

[velvet rampart]

Thank you. Still it would be great that Image 1 would be available also on Sora.

[crimson heron]

I mostly don't use Sora, but I agree that the option should be there too.

[wicked plinth]

I swear these companies sit on models until their competitors launch a new product. Google just announced image up scaling

[steady gulch]

Whelp, this is doing terribly

it's almost comical. I say to make it smaller, and then it takes the image and adds sparkles.

[spare pine]

Is GPT able to understand which attachment I'm talking about if I've uploaded multiple? For instance, I've attached 4 in a message, and I want it to try recreate it and use the other 3 attachments as a reference for a character.

[steady gulch]

yes, it is

[steady gulch]

yes it can do that, but you have to talk about each of the attachments

[velvet rampart]

Ok, people have been saying ”the model is best/worst”. I get where the division comes from. I am myself of both minds. I am seeing an important lesson. When Image 1 came in at the end of March, we needed to learn a whole new prompt style to go with a new design language. I am seeing the same now. As we learn how the new model behaves, our prompting will evolve. Personally, I can attest that, after generating close to 100k images with Image 1, it took a while to get what I wanted, and that required a lot of experimenting. I needed to abandon styles and techniques that worked with DALL · E 3. Now, in the autumn, I could put together a prompt for what I wanted quite quickly. Now? I have to learn what works and what does not with Image 1.5.

[analog flame]

Gemini 3 Pro launched, GPT waited the hype, now they launched. Next week will be wan 3.0 and Seedream 5.0

[empty scaffold]

the problem with this is that it's subject to the usual hyperstrict ChatGPT moderation instead of Sora, which was more loose.

[obsidian cargo]

there is a way to select iamge model inside a GPT website?

[wicked plinth]

Whatever this new model is called, it seems to be a little bit more open

[analog flame]

Are the images in 1K, 2K or 4k?

[obsidian cargo]

Sorry but grok image generation is bs.

[obsidian cargo]

Nano banana follows instructions way worse in my testing, GPT Image 1.5 seems better to me besides that it makes images looking like Gemini however i dont have any blending problems on my end

[frozen copper]

So how is the new images? On par with Nano Banana?

[obsidian cargo]

For me it works much better then nano banana but reading chat here it seems to warry per user

[analog flame]

It's better with text, worse with the rest

i usually do 400-500 images / day with Nano Banana Pro. Already tested everything i could.

GPT Wins writting texts, loses in quality and consistency

[obsidian cargo]

Didnt had issues with consistency so far

[analog flame]

I would say it's just another image generator, nothing we already don't have

[obsidian cargo]

ITs much better then gpt image 1, but it is way more flexible to me like i can much more get results i want to my liking compared to banana

[frigid sluice]

@analog flame Which gemini and chatgpt plan you are on?

[analog flame]

for my company, banana is still the main deal. I will stick with GPT to fix letter and text

[analog flame]

Gemini Ultra and GPT Pro

[obsidian cargo]

suit your self

honestly nano banana was never working for me ever since it released

[analog flame]

Nano Banana already made me thousands 😹

[obsidian cargo]

Like i said never worked well for my usecase

[analog flame]

it's like MidJourney. Nowadays is useless, but it works for specific cases

[velvet rampart]

I notice that more instructions/directions is given to 1.5, more it follows your want.

[frigid sluice]

On chatgpt plus and gemini pro,

Gemini nbp is better due to it's 2k quality 👌. (4 mb)

Chatgpt gpt 1.5 image only generating (2 mp) lower than 2k.

Is it bug, or isn't rollout?

[obsidian cargo]

Yeah but also the nano banana generates blurry images when you zoom in you can see it even from far it has weird effect

[woeful surge]

How do you choose an aspect ratio with the new model? I tried it in the prompt but it got ignored.

[analog flame]

yea, i'm doing 4K images in nbp
IF ChatGPT gives us 4K res, then i will think about a switch

[obsidian cargo]

Tell it ot make 9:16 image or 16:9

or what ever you like

[woeful surge]

Yeah tried that and it gave me 1:1

[obsidian cargo]

huh

[woeful surge]

I'll keep trying

[obsidian cargo]

worked for me lol

[obsidian cargo]

What are you even rendering in 4K that you need AI to do for you?

[analog flame]

I run a company that produces AI Commercials for B2B

so i need 4k images for 4k videos

[woeful surge]

Ah ok it worked for me this time, directly in ChatGPT. Possibly because I was using the API last time

[jolly niche]

Can't you upscale to 4K?

[woeful surge]

It's definitely a step up from the previous model, but still not close to NBP for character consistency. The changes it is making aren't terrible, but if you want accuracy it's not here.

[analog flame]

NBP already gives me in 4k, i don't need to upscale

[obsidian cargo]

Its seems more consistent to me, i mean it even gave me realsitic scene like chatacerts werent unaturally looking at camera or worse somewhere unrelated conversations between charatcers worked well for me

[jolly niche]

I'm pretty sure 4K isn't native, but an upscale in NBP.

If you prefer that, then by all means use it.

The best tool is the one you most prefer to use.

[obsidian cargo]

True

the scariest place is your own mind

[jolly niche]

🙂

[woeful surge]

I do find 2K in NBP looks better than 4K, then upscale if necessary

[tough yacht]

Started testing out the new model. It's both very good for recreating an image pretty faithfully, and frustrating for blocking me from very basic stuff.

Asked it to generating a Denisovan based on the latest scientific data, and it can't.

It's not allowed to.

[woeful surge]

I'm very fussy. I have a character that uses a specific trident with runes on it. NPBNBP renders it flawlessly. This model makes up a new trident with each render.

[tough yacht]

Where do I put the "PLEASE LET ME MAKE CAVEMEN" feedback?

[gilded elk]

@woeful surge what's npb?

[woeful surge]

Nano Banana Pro – typo sorry

[gilded elk]

That's google yeah? How much is it to use?

[woeful surge]

I use it on Higgsfield, it's $45/mo for unlimited 2K gens

[gilded elk]

Expensive

I'll keep my chatgpt sub

[woeful surge]

I tried with the API directly and racked up a few hundred $ pretty quickly... so that's cheap for me

I generate a lot

right now you can do 4 concurrent 2K gens and 2x concurrent 1080p Kling O1 gens both unlimited on Higgsfield. If I was paying per gen it would get expensive very quickly

Either way, now that I have the fidelity I need on NBP I don't want to move, I'll just find the cheapest place to get that quality. If Sora surpasses it I'll move back

[tough yacht]

Nanobanana pro is pretty damn good, but has different strengths and weaknesses.

[woeful surge]

I'm pushing it to the limits with five very detailed characters in the same scene. It's the only model that can cope, barely.

[rich star]

How do I use the model?

[spring sequoia]

err whats the status of the new image model?

nanobanana isnt the best tbh

[fading mist]

new image model is soft and doesnt have definition and detail of the old one, it renders skin very poorly like a camera phone filter, it just seems waaay worse, plus the text and image filters are even worse than gpt-1

[shadow hare]

@obsidian cargo eh it's very inconsistent and has this weird fixation on creating anime or uncanny valley semi anime stuff. But if I enforce it's a photograph, taken with dslr etc. it might only be 1 in 30 but it churns them out so fast you can grab the good one. It's quality seems to depend widely on what you're asking for in the prompt. There's also the fact that on android app at least it's barely restricted at all compared to the website, albeit as European VPN to Cali makes it better. Though sometimes it seems it can't help but make borderline or straight up nsfw when your prompt hasn't asked for it. Frankly I'd be... curious as to its training data.

[obsidian cargo]

I have trouble making it use anime style😭

[shadow hare]

Bruh it's all it'll do for me it annoys me haha. I've never asked for anime yet like 20-30% will be. Probably more so on app as well. The difference between app and website is insane, don't really get it.

[obsidian cargo]

Eh whatever

Ill test it more tommorow iam starting to get hold of it however i needed to go sleep now so bye

[verbal stream]

Hi, is the new image generation model working on the website https://sora.chatgpt.com/ or its only in chatgpt app / website ?

[peak timber]

Unfortunately it is affecting SORA 1

[verbal stream]

you mean its worse ?

[empty scaffold]

I'm disappointed in the new image model's ability to reproduce styles that aren't modern/contemporary. The old model was very good at fashion, makeup, and styling from specific decades like the 1930s or 1960s. Now everything is flattened into a very 2020s face filter look.

[spare pine]

really annoying that GPT keeps just giving me a prompt when I ask it to generate an image, and then tells me that it can't generate images

Only for me to remind it that it can, and then it finally complies.

[crimson heron]

I've seen the same in my limited testing with existing prompts. I haven't gotten into refining the prompts yet for the new model to see if I can get closer to the proper historical styling. The prior model felt relatively easy to do it with minimal effort.

[peak timber]

yes

[spring sequoia]

why not just focus on sora 2 like what is this management doing

[sullen turtle]

I guess my thing is, they seem to have not fixed the uploads from before. As for the quality of this new image generator, it is definitely different I am still looking to see whether or not I personally see if it is worse than the first version.

[kind grail]

I do prefer images over videos. So i like it if a model specifically focusses on images

[kind grail]

I even had moments where it refused to write a prompt with a copyright titel.
Like, not even generating the image. Just the prompt.

I had to start a new conversation, without thinking to get it running...

[kind grail]

god bless tho that the yellow stain has been reduced

i couldnt stand it

[analog flame]

You can't make videos, without images

[spare pine]

Oh yeah it did that often with me. I had to convince it that fanart is fair use and that the IP owner allowed it (they did, so I wasn't even lying). But often times, it refuses because there seems to be a deeper, more restrictive policy within the image generator (at least, that's what it told me)

Even when it agreed with me that it should be allowed, it just couldn't.

Now it just lies to me and gives up after I call it out lol. I can only assume its OpenAI's way of artifically limiting the strain on the image generator

[light river]

I need to understand with all the pressure from Google why we can’t create 16:9 or 9:16 images ?

[sick cargo]

the new image model is great. has anyone found how to prompt it to get the anime art style of the old one?

also the refusals are insane.

[velvet rampart]

Please remember that guardrails have been quite same for some time. Refusals often happen if you generate in ChatGPT and the refusal often happens due to something in chat that was much earlier. The whole chat is a contect window and it takes into account that. If you get a refusal try in a fresh chat. Often your is genrated there.

[empty scaffold]

you'll have better results directly on the Sora website instead of in ChatGPT

[sick cargo]

I am using sora.com

two anime figures kissing -> refused.

will try on the api with moderation low. if that parameter even still exists and indeed has an effect.

[empty scaffold]

Yeah, I'm not surprised, any intimacy is difficult. Sometimes wording it like "...with their faces pushed together" gets the intended result.

[sick cargo]

two yakuza celebrating war profits is not an issue however. xD

generation speed is impressive. generating a figure standing behind a synthesizer on stage suffers from the very same problem as with the old model tho: the figure's legs cross through the synth stand or the stand looks... weird.

faces also weird sometimes. and heads sometimes too small, which seems to be a new problem the old model didn't have.

[peak timber]

you will never find it

[kind grail]

what anime art style do u mean?

[spare pine]

I just asked it to create an image of a fictional character. No specific art style.

[kind grail]

maybe send an example of what you mean

[spare pine]

It hasn't done it since the new model came out so I can't anyway

[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.

[kind grail]

do you not have any old picture still saved/in history?

[spare pine]

The issue I was talking about was gpt not generating an image at all because I requested it to use a copyrighted character

There are no images to share

[cedar cobalt]

Idk why so many people complain about image 1.5

its capabilities are no worse than others

You can share what it generates to anywhere even without visible watermark

[cedar cobalt]

Except characters generating though

[obsidian cargo]

Many complain because its fundimentally different it chanfed both art and prompting style and people just doesnt know how to use i too dont really like how anime charatcers looks like from gemini but i like it because it allows more customizesbility and other stuff compared to nbp and scenes are more natutal so i dont think its bad in general

[cedar cobalt]

Prompts have received more and more attention since 5.2 was launched

[tranquil tendon]

We need more relation aspects

[obsidian cargo]

GPT Imag 1.5 seems actualyl much better today, it gets me closer results then yesterday

[marsh bridge]

I don't know still unusable for simple 3d renders it gives them too much shine and details are wayy off, prompting doesn't even help, it is more similar to how ai was 2-3 years ago rather than 2 days ago when everything was fine

[obsidian cargo]

Its doesnt seem to me that way

But afaik its currently different for everyone

[marsh bridge]

I can't post pictures, but I have tried like 10 different prompts played around and it still messes up super simple things

[rigid bough]

every time a new model is introduced some people will complain that it is not as good as the model that preceded it.

[obsidian cargo]

Since i had advanced prompts i jsut changed few things and it works for me as good as before

[obsidian cargo]

I mean yeah everyone hated gpt 5 when it came out but when gpt 5.1 cameout they hated that and when gpt 5.2 came out now they hate that i mean sure guardrails are bs but otherwise model works great

I honestly like gpt image 1.5 more then 1

[marsh bridge]

I mean main problem with me is that it doesn't follow my remixes or I give him past pictures he made and with a complex or simple prompt the style is way off and the details I can't see where one part starts or another ends everything is mixed together in that early ai look

[obsidian cargo]

Doesnt seem to be issue for me

But its subujective

Since i see with each generation it adjusts to my style more and more

And ive generated lot of images since yesterday

[kind grail]

It works perfectly fine for me. Quite precise on the remixes usually, at least much better than before

[harsh bolt]

Hm I the unique alien anatomy test I have is unfortunately still failing with all top image models

[velvet rampart]

Yes, I have noticed the same. When I got a new account back in August, all my Sora generations adjusted more and more to my style quite soon with the previous model. It takes some time. We learns about the model and it learns from us. Once a prompt has been generated, it will be easier and easier to generate. I have noticed that some prompt + image combos need only a few generations, while others need more.

[sharp monolith]

Hey, Is there anything I can do to get back the old Sora Image Style Generation?

[lethal notch]

Just came here to say they butchered it

Why does it add so much unnecessary detail and bloat the composition. Vector art got a pretty hard nerf

[tribal kiln]

For my usecases this is the first image generation model I've tried (and I tried a lot) that is actually able to do what I need it to do

The attention to even the finest details is insane

It can generate ridiculously complex and COMPLETELY coherent UI designs now

Not a single distorted letter or nonsensical element

[lethal notch]

Do you mind sharing a prompt?

[pseudo mica]

The restrictions are terrible.

[tribal kiln]

My prompts aren't very good lol, just started playing around with it 20 minutes ago. But just try to create any image with a high level of elements/detail/text, like a magazine page, UI design, game screenshot, etc. and you'll find it has gotten incredibly good at coherence

[tribal kiln]

What restrictions?

[pseudo mica]

Literally cannot run this prompt or anything like it: a blonde woman with shoulder length hair and blue eyes falling on to her side on the floor firing a pistol as bullet flashes by her. She is wearing a thin baby blue tank top and short shorts, high quality, cinematic composition, 8k

Meanwhile Gemini has no problem

[tribal kiln]

Hmm fair enough, I don't really generate images like that so haven't encountered any restrictions

If anything I'm unhappy with the complete lack of restrictions when it comes to deepfakes

[late blade]

Clearly, you are jumping on channels, venting. Look for a way to release that energy, I'll gladly have a talk with you once you are grounded. It's a topic that matters, not an easy one to solve. Need people to come up with smart ideas and how to implement them.
Btw. I'm not one of them smart ones.

[pseudo mica]

Me? I was in one channel and realized it was probably the wrong one and I posted once.

[late blade]

Fair, glad you caught that

[tribal kiln]

Huh, no I'm not

or did you mean the other person idk

but anyways I was actually sharing how happy I am with the coherence of this model when it comes to graphic design, am very impressed!

currently testing how it performs with posters, which is something it has historically been better at than UI design (and it nailed UI design)

[late blade]

Cool, yeah, the model for generation is amazing. As for the other aspects, maybe we should suggest a channel for that. And then some other topics that can get hot-headed at times. But are talks, discussions and ideas that need to be out there.

Also, I extend an invitation to do more images, only 38 image posts in 3 years... would be nice to see what you can bring to the server.

[pseudo mica]

I literally had amazing generations last night from a scene in a story I wrote but getting hit with these blocks over something I can generate on other platforms is pretty irritating

I don't know where else to complain because I swore they were going to let adults be adults on this thing

(not that I'm even asking for a adult content to be made!)

[late blade]

Content Policy and Moderation is hard, because of copyrights, trademarks, identity, and many other aspects. It's frustrating to see an image generation not being done. The other side of the coin, it's also a good sign that OpenAI is also doing something to regulate and make it safe.

[tribal kiln]

oh my god

it is amazing
truly amazing
actually ridiculous

professional high quality poster designs that I wouldn't guess were AI created in a thousand years

[late blade]

It can be very accurate, if you also provide a screenshot of a Canvas render you made, then it's even more so to what you want

For example: ask GPT to do a react canvas with a mood dashboard for a project. Take a screenshot of your Dashboard and then use Sora or ChatGPT to make an advertisement of that Dashboard.

[sick flax]

I can think of a couple of channels to report system issues or to make suggestions: https://discord.com/channels/974519864045756446/1070006915414900886, https://discord.com/channels/974519864045756446/1070006151938314300
Be as specific and concrete as you can about the issues you are seeing!

[late blade]

Oops, used the C word, got my post erased by automation

[tribal kiln]

You're right actually, I'm now running into restrictions as well trying to use a poster with a female model on it

[late blade]

Plan: Get the idea out there -> Get Group that can stand behind this -> Make a proper suggestion -> hope it works.

[pseudo mica]

is it also gaslighting you in to ways to make it work only for it not to?

[tribal kiln]

I found out why

For some reason on PC it isn't using the new image model

The previews it was showing before they got censored were also of a completely different composition and model pose than what I was working on

[obsidian cargo]

I dont have issue with that

[tribal kiln]

I switched back to my phone and now its fine again

[sick cargo]

yea the "guardrails" are nothing short of insane atm.

hopefully it'll be refined over time.

[frozen adder]

friends, ChatGPT PRO have the high-res option for the new GPT 1.5?
I'm testing it in API and in gpt pro, but the quality is different.

[fleet cairn]

is there a prompt to use that would generate multiple images of a controlled image generation?
When I say controlled, I mean using a specific model in a location as well as a specific clothing that I would provide.

[neat mountain]

guys the new image generator version is just a human catastrophe, can i go back to the old one ? does someone have an idea how ? this is really frustarating, what version or model did we even had ??

[velvet rampart]

The announcement shared that the prior model can be accessed through a custom GPT:
https://chatgpt.com/g/g-6940a876d5f4819186b4668deabcd580-4o-imagegen

[velvet rampart]

Also, add your thoughts here and please add a star. https://discord.com/channels/974519864045756446/1450357575580389376

[neat mountain]

u just saved my life :D, thank you so much and i apologise for not discussing in the other channel

[velvet rampart]

I also found out myself about this. Man, I do wish that we could use Image 1.0 on Sora still.

[neat mountain]

i was crafting the past 24 hours a new Nanobana pro prompt to mimic what i just had in the past few months, and i was too close to subscribing the nanobanana pro.
this was genuinely scary af

[sharp monolith]

are old sora images coming back?

[velvet rampart]

Not at the moment, but if we keep enough noise…

[drifting olive]

u mean the interface for it? cuz the model was not Sora for images.

[sharp monolith]

no image generation they are really bad rn I can't make high quality content for my page since this new model

[drifting olive]

mate u did not answer my question bru 💀 😭
And you prolly better off not making yet another ai slop page.

[sharp monolith]

wym i did i said its image generating the new model is affecting sora too

[jovial monolith]

whenever I try to generate literally anything with a female it flags it 😭

[empty scaffold]

This might sound like a weird complaint but I feel like the new model adds too much fine details? Like extra stands of hair, bokeh/light orbs, dust or grain, shiny highlights, too much clothing details like straps/bows/trim, extra bits on armor or weapons (things look very "spiky"). Like most pictures in any genre or style tend to come out looking very "busy" for lack of a better word. You can see what I mean in people's pics over on #images-canvas

[velvet rampart]

Yes, dalle3 upgraded

[empty scaffold]

i'm aware :)

[cedar cobalt]

why couldn’t we share any images in the images-discussions channel?

[deft musk]

That's for discussions about images!

You can share them in #images-canvas and/or #1154829862171844679 , and even copy your post link over to #images-discussions and discuss whatever about the image htere

[cedar cobalt]

thank you

though it’s still weird

lol

[velvet rampart]

Lessons - the Image 1.0 could translate flat 2.0 cartoons well into realistic 3d. However, 1.5 cannot do that. I have to some mid-steps.

[vital gull]

hmm any examples where it failed?

[velvet rampart]

Plenty. My old Danny Phantom ref image did not just work.

[wanton fractal]

Is it possible to get ChatGPT to generate images without a metric asston of noise anymore? Since image-1.5 rolled out, every image seems to be full of super grainy low quality jpeg noise... I don't remember it being this way before

[atomic magnet]

How come most images are coming at like Dalle 1 quality, really bad?

[unique spade]

Was anyone able to colorize your childhood b/w photos with the new image model? ChatGpt rejects to complete it whereas Nano Banana Pro can do it well with minor mistakes.

[icy chasm]

Is anyone else getting a wacky error on Android where it keeps asking you to upload the image as a file and even if you do it just keeps getting itself into a loop where it keeps asking to do it and insist you are wrong

[sick cargo]

new image model definitely needs more work

#images-canvas message

and its not a rare or even occasional occurence either

[meager remnant]

Good morning/evening, quick question for anyone who knows the answer. Is it possible to use the image generation native tool for the response API and target GPT-Image 1.5 rather than only using Image-1?

[rigid gust]

Hey! For what it’s worth, when I did this with an old family photo, I didn’t give any specific instructions at all. I just uploaded the black-and-white image, and the model auto-colorized it on its own.

[velvet rampart]

Are others also getting a message that Sora is unavailable at the moment for image generation?

[sick cargo]

api still works tho

[velvet rampart]

And ChatGPT for image generation. Hopefully, they now fix a few things, like restoring those older uploaded images, allowing deletion of old fails, etc.

[kind grail]

Mhh there are definitely ways to ease the generation. It even gave me surprisingly - ehem - detailed images once in a while when i was pushing to test the boundaries of the new model.

[jovial monolith]

Literally how 😭

[kind grail]

If you want specific body shapes its better to use a reference image instead of giving instructions.
(This is obviously still safe for work, sorry but i will not give any instructions on how to circumvent this as its against TOS)

[velvet rampart]

Use reference images. Say character from image X.

[velvet rampart]

No, we are not saying how to circumvent but instructing how to use refence material. Sometimes what you mean and how the model understands are two different things. If you have something specific in my, use pics. You can use up to ten images per prompt.

[ionic carbon]

Does gpt-image-1.5 default to portrait mode (2:3) now? Is there any way to set it to square or landscape as default?

[velvet rampart]

You have to ask the size clearly.

[ionic carbon]

Yep, I know. I wanted to verify if the default is the same for others too. It definitely was square before the update as basically every image I ever created is in 1:1 🙂

[lethal notch]

any idea on how to stop the grain and artifacts in the images? it's really annoying

[velvet rampart]

Could please share one or more to the #images-canvas ? It would help to identify what we are dealing with here

[lethal notch]

done! It should be a plain dark grey background, but I get grain in almost every image. This did happen with the last version, butwith 1.5 it occurs pretty often

[velvet rampart]

Well, honestly, to my eyes, these look good, no grain. However, it annoys you; you could try“denoised”. This was something I used when 1.0 was just released, and it looked grainy to me. Also, adding image quality qualifiers, such as 16k, hyperdetailed, 1080i, 1080p, UHD, etc., might help, as these push the model to emulate the quality. Of course, you do not get the quality, but the idea is to push the model to produce higher-quality images.

The graininess might also be inherent to the style/technique you are using in the images.

[ionic carbon]

you could try telling it the exact color you want as background. e.g. "background: #333, plain". Or use a transparent background and fill it later or even just in a second step to stop it being influenced by the overall style.

[velvet rampart]

We have to consider how the model generated the images and inherent limitations it might have.

[lethal notch]

yeah will try a fixed hexcode and denoised etc. to enhance the output, thx

[velvet rampart]

Great that we could. Note that denoising loses some detail from the images.

[velvet rampart]

Consider adding non-grainy rendering as the model does we tell it to do.

[sick cargo]

im inconsistently getting eldritch horror faces from 1.5. inconsistently as in, the same prompt resubmitted sometimes works fine, but most often results in said eldritch horrors.

[obsidian cargo]

Thats kinda point of the model so outputs arent always exactly same

IF you even use exact same prompt

[velvet rampart]

Yes, I also got these. But I figured out that something in the prompt causes this. You have to bare down the prompt and try sentece by sentence what is causing this. It is some minor thing that causes but it is annoyingly hard to find out. I suggest that you use AI to figure what is causing this. Give the AI the prompt and image generated with and describe what you are seeing.

[obsidian cargo]

The same prompt will never generate 1:1 same image

And its for a reason

[velvet rampart]

No, but consistent enough

[obsidian cargo]

If you reinforce details enough it will work

[velvet rampart]

That’s what I do with my prompts.

[obsidian cargo]

My promot currently range between 2.2k to 4k chars, before it was like 10k+ chars

But my characters have more complex and more different clothing which without this amount of reinforcements wont work well

[velvet rampart]

Well, 1.0 had some variance in the generations but 1.5 the variance is much greater. Maybe too much.

[obsidian cargo]

Iam gettign preatty consisten resutls only difference is in some less specifed details o outfit or in height someitmes small sometimes taller but it mvoes around 1:7 head to body rait ogoigneaither to 1:6.88 or 1:7.2

[velvet rampart]

Yes, but for me the copied. Beard, no beard, glasses or no… If a source has them, they all should be carried to the generated regardless, not left out.

[obsidian cargo]

Hm i had this issue too but less drastic then yours

Its still not as 1:1 as NBP but what i like more is that characters better blend to the scene more naturall as oppose to NBP

[velvet rampart]

There are elements from 1.0 that should have been kept and carried over to 1.5, such as character consistency and faithfulness to the original. If I say image #1 contains the character, the model should not go and change it indiscriminately.

[obsidian cargo]

Sorry but 1.0 wasnt consistent for me at all it missed lot of important details when making the character

[deft musk]

Want to make sure interested folks are aware, this customGPT made by OpenAI links to Imagegen1, it's still available to those of us who want a way to access it. https://chatgpt.com/g/g-6940a876d5f4819186b4668deabcd580-4o-imagegen

[sick cargo]

#images-canvas message

just so you understand what I'm talking about

[brazen pecan]

Is there a dedicated time that the image generator unlocks after extended use? It seems random every time, and sometimes it says to wait almost a day.

[limber shoal]

ok how do I get past these annoying guardrails, i asked chatgpt to make a prompt for me and even that is not working
like did gpt got even stricter?

[hollow crypt]

is there any way to revert to the old model? the new model is absolutely disgusting in comparison
its so grainy and overdetailed and noisy and style tags have nearly no effect, it feels like i'm running 3 year old stable diffusion

[deft musk]

Does https://chatgpt.com/g/g-6940a876d5f4819186b4668deabcd580-4o-imagegen work for you, to use the old image gen?

[hollow crypt]

I'll give it a try

im surprised theres not more people talking about how bad the new model is in comparison
theres almost no noise on reddit from it it feels

and its on both sora image gen and chatgpt image gen

yes this works, definitely is the old model

[deft musk]

Glad it works for you!

Models change, and sometimes there's access to the older models for a time. Sometimes the changes are announced, other times not. Different people like different stuff, and it may be possible to find new ways to prompt to get the older responses you liked better back with the new models, but it might take some exploring.

Anything we don't specify, the model guesses - and different model updates guess differently.

Usually if we can explain what we want precisely, the model can do that, but knowing how to ask can be a challenge

[hollow crypt]

look at the comparison between the two images in image-canvas, basically every image feels like that
its literally worse than locally hosted stable diffusion image generation from 3 years ago

i could get it if its better at some things but right now ive tried about 50 prompts and i have not found one that even looks half as good as the previous version

[deft musk]

In case you're interested and in case it helps, In this chat I explored how to describe what you liked better, based on this small sample, https://chatgpt.com/share/69464c8d-1820-8011-9de5-d1814b96af5f

Then I took the summary of the style you might like, and tried to create a similar image with it,

A warm, painterly, storybook coastal illustration with simplified forms, visible brushwork, stylized lighting, and restrained detail, prioritizing mood and readability, let's see a coastal scene with a shipwreck in the water.

I have no idea if this is even slightly close to what you want.

https://discordapp.com/channels/974519864045756446/1204360881593520128/1451835323226128424

[ionic carbon]

you might be in a minority, it's top #1 on lmarena for text-to-image and image-edit

[ionic carbon]

I like the new one way more than the old one. sd from 3 years ago would not be able to generate anything like that.

[hollow crypt]

i guess i just have to entirely delete every prompt and work out things from bottom up
the boat image you posted is fine but every single prompt ive used for fantasy locations are basically completely destroyed. I posted another image in canvas comparing them

[velvet rampart]

You don’t have to start from scratch. The new model needs a bit more direction and push than the previous one. You have to be more specific and directive. I know how this feels and it is an enorumous pain in the backside to find what you need to change in the prompt. I have noticed that I need write things a bit differently. An example from 1.0:
Copy hair from #2 and-if visible-copy facial hair and body hair.
1.5 version:
Copy hair from #2. If visible, copy facial hair and body hair from image #2.

[hollow crypt]

I think the issue is actually reverse of that

My old prompts were very long and fine tuned over time to produce templates tuat would make a consistent image style

And i would just change a snippet at the end while the previous 40 lines stayed the same

[velvet rampart]

Yes, that’s my case also. I use a base text prompt paired with a base character image and a second varied image.

What I see in your images @hollow crypt is how the model interprets different style and technique cues. I ran into this same promblem in the spring when 1.0 came out and I could not any more use dalle3 styles and techniques. For example, my go to style Modern digital arts is interpreted differently in 1.0 (more painterly) than in dalle3 (more photographic). Now the style is back like it was with dalle3.

[atomic magnet]

Be nice to get a toggle for the old model in Sora

[velvet rampart]

https://discord.com/channels/974519864045756446/1450357575580389376 comment and please add a star

[slender aurora]

Is anyone else experiencing that no generation with input images work right now? Everything gets flagged with: This content may violate our guardrails concerning similarity to third-party content.

[velvet rampart]

Remember tha whole chat in ChatGPT is used as a context window. This might be the cause.

[fleet cargo]

Hey
Abhay this side

[prime haven]

I’m writing this as an IT professional and a long-time Plus subscriber who has reached a breaking point. I’ve spent over 7 hours today trying to complete a school project using imagin 1.5 for basic, harmless Christmas imagery ornaments and festive scenesonly to be met with a constant loop of "Content Policy" blocks. What’s worse is that the exact same prompts worked perfectly three days ago, but now the model is completely lobotomized and refuses to even edit its own previous generations. On top of that, GPT’s factual reliability has plummeted, with the model hallucinating wild misinformation about current global events like non-existent peace treaties. As a professional, I’m paying for a tool to increase productivity, not a broken system that wastes my time and forces me to fight with invisible, nonsensical filters. This isn't a "Pro" experience anymore; it’s a regression that makes the service unusable for actual work. If this level of sanitization and technical decline continues, I see no reason to maintain my subscription and will be moving my workflow to your competitors immediately. Fix your filters and restore the model’s basic functionality.

[velvet rampart]

Could you provide an example prompt that was denied? One common problem is that ChatGPT denies the image, and this is often due to something earlier in the chat. This is because the entire chat is the context window.

[prime haven]

Since updates 5.2 and Imagen 1.5, the system is broken: harmless prompts are flagged as violations even in fresh chats, and the model fails to edit its own previous generations. As an IT professional, I’ve ruled out local issues the current over sanitization and 20-minute processing loops make the service unusable for professional work.

[jagged torrent]

Did they change the Sora image generator. The new photos are super dark and grainy??

[uneven belfry]

Ive been using image generation for quite a long time but 5.2 changes have absolutely gone bad. 5.1 was so good!

[velvet rampart]

Have people noticed a difference between Thinking mode and Instant mode when generating the same propmpt?

[fickle pond]

I consider version 1.5 to be a very solid update that does an excellent job overall.
However, I have also encountered situations where overly information-dense prompts, or combinations of expressions that don’t quite align with the model’s expectations, resulted in reduced quality or even outright refusal to generate an image.
Rather than trying to faithfully reproduce what was possible in version 1.0 by carefully adjusting prompts, I believe it is more important to explore what kinds of expressions have become possible in 1.5, and to focus on how to interact and iterate in order to bring the output closer to one’s own mental image.

[velvet rampart]

Yes, this is what I am noticing. The 1.5 model is a bit more settled now than on launch. The images are now closer to 1.0 in result but still their own. On launch day, some generated images very just awfull. I am trying to get the feeling what I can do with this model. I noticed that I need more carefull wordings. With 1.0 I could just through something at it and excellent result came out. Now? Same prompts do work but with some adjustments. 1.0 and 1.5 are like family but their own individuals, behaviours and quirks. We just have to learns these (again).

[dapper temple]

How do I avoid it creating the same image over and over? I need variation and multiple iterations with broader differences while keeping true to the prompt.

it keeps either editing the same image, or simply generating an identical one. completely opposite of what I want. happens both projects, regular and new 'images' chat.

[ionic carbon]

Start new chats, often, always.

[fickle pond]

Try this Prompt.

"Treat this as a completely fresh interpretation.
Prioritize variation over refinement.
Next Prompt."

[dapper temple]

impracticable at scale beyond a few variations... even if I wanted 20 different stuff. 20 new chats is a lot.
I'd need to guide it to direction I want, like back when it generated multiple at a time. high variation, very useful.
the same prompt shouldn't create the same image. the same description of something applies to several things...

[dapper temple]

did not work.

[fickle pond]

Im working.
#images-canvas message

[dapper temple]

that's like a 5-10% variation / difference. I'm aiming for 80% or more. not subtle, not minor, not same composition, not same angles of same place even.

even randomizing the prompt it gives the same thing over and over...
I re-write it or I ask it to re-write it too. new chats, new everything.
same thing over and over... it's unusable & impractical. overtly biased.

it's like trying to push a stubborn mule who refuses to move while claiming it will.

[fickle pond]

I see. In that case, you could try using the image-bot with the count parameter to generate multiple images at once, then extract prompts from the variations.

[dapper temple]

which image-bot?

[fickle pond]

https://discord.com/channels/974519864045756446/1202309673709994065

[fickle pond]

/dalle3
#image-bot message

[velvet rampart]

Or Sora

[dapper temple]

it's neither realistic, nor even what I described in most cases.
some stuff is in the right direction but I can't guide it like so.
the variation is the only good aspect of it.

sora gives more what I want, but if only I could use chatbot to direct sora.

and sora 2 isn't avaliable in my region.

[velvet rampart]

Sora1 has image generation

[dapper temple]

yes... I already said: "sora gives more what I want"

but I can't guide it or tweak directions or anything.

[ionic carbon]

It will not work like you want it to work, you need to manage the context. It will always get the full context of the conversation and tries to incorporate everything. If you want something new, start new.

[dapper temple]

but that's exactly the thing. it has worked in the past and really well.

[ionic carbon]

as far as I know does gpt-image-1.5 not expose temperature or seed, so there is no way (not via API and even less via the chat interface) to truly randomize the output. Also it's a autoregressive model, this will tend to produce similar outputs. You might want to try a stable diffusion one for your use case.

[ionic carbon]

gpt-image-1.5 is in many ways better than gpt-image-1, it might not adapt to your use case. You can use the old model for a while (there was a link to a custom gpt in the announcement). But in the end you might not want a auto-regressive model.

[fickle pond]

@dapper temple
Thank you. Thanks to your question, I was able to discover a really valuable prompt breakthrough.
#images-discussions message

[dapper temple]

it didn't show seed or temperature back then either. but it worked nonetheless.
and I have tried already new chats... does not work. scrambling doesn't either.

isn't autoregressive not supposed to transfer to new chats?

[dapper temple]

I'll try that out and see.

[ionic carbon]

there are basically two ways to generate images: stable diffusion and autoregressive. I tend to think that autoregressive will generate more of the same images as it is in itself less randomized as stable diffusion (but this might still be wrong, I'm not an expert in image generation models). you can ask gpt to: "generate an infographic to show the difference between stable diffusion and autoregressive generation of images" to show a bit about the differences. After all: both methods aren't "transfering" to new chats, but autoregressive models will tend to generate the same again. Even if this is oai here, I would suggest to try out the 🍌 on the other side to see if that fits your style 🙂

[dapper temple]

I see, perhaps I'll have to go try 🍌
Thanks 👍🏼

[ionic carbon]

It's free to try … but if you have more questions about the yellow side, please ask per DM 😉

[late blade]

for non-OpenAI models feel free to use #ai-discussions or #off-topic

[sick flax]

Hi, there. I responded in https://discord.com/channels/974519864045756446/1452241738092642334

[blissful sphinx]

Hi,

Is there any kind of policy or procedure for reporting images that are just clearly "messed up"?

I'm using the API and some of my images have been totally scrambled.

I also have the problem where I've had content listed as sexual and blocked when it is anything and everything but that type of content. How do I report these miss classifications?

Please tag me. Thank you.

[velvet rampart]

#1070006915414900886 is a place to report or just contacting help.openai.com

[crimson heron]

In addition to what Milamber said you can 👎 to mark an image output as a bad response or output. On the web version that action also gives an option to give more feedback to why. One of those is low quality and distorted.

[blissful sphinx]

This is API driven only.

[crimson heron]

Ah sorry. Overlooked that detail.

[velvet rampart]

Ok, now I am getting it also. I created a bug report on that. Here: https://discord.com/channels/974519864045756446/1452351872185667625

[barren widget]

@fickle pond I wouldn't have come up with the idea without you^^

[fickle pond]

You're welcome, I have also received nice inspiration from you^^

[velvet rampart]

Are others experiencing images dissapearing from chats?

Or this due to splitting the chat and if deleted a brach, those images are gone?

[fickle pond]

Yup, it seems that the image cannot be posted here.

[vapid elk]

#images-canvas message
Im not sure of how they do that, but my impression is that it image embedding has been giver higher strength

however, it does not feel like that when we also prompt it to do things like "make this proto in Foo Bar art style".. since giving mor estrength to the original image would mostly ignore the user's request

[deft musk]

I've seen some other people talk about how deleting part of a branched chat can affect the rest, might even delete the branches

[velvet rampart]

Not here, but in ChatGPT’s chats.

[summer thunder]

Is it just me, or is there a problem with image generation?
It feels mediocre now — it generates 4 images at once, and the images are really bad.
Is this happening to you too?

[velvet rampart]

I get on and off the consecutive image production on ChatGPT. Are others experiencing this? How to get the idiot of an model to do this reliably and repeatedly?

Wasn’t this announced feature?

o4-mini works, but series 5 with different modalities does not.

[wintry epoch]

Ok why dont it only change the marked area and change more or it dont even know what i mean as if nothing is marked

[obsidian cargo]

I dont realyl understand what you mean

[fickle blaze]

Hi There, I'm developing a ecommerce web app (react) and want to integrate chatgpt for auto tagging the pictures. Whats the best approach ? I tried several approaches but I'm not happy with the outcome...

[ionic carbon]

Basically just write a prompt with your desired outcome, choose a model which fits your price/output ratio and give it the image (maybe resize it to save tokens)

[glossy scroll]

What's James Gandolfini doing there? @young crescent
#daily-theme message

[late blade]

@sour dew you got to eplain me here #emoji-submissions message what the emoji is all about, 'cause no clue

[worn turret]

Holy yap ig u used ai for this

Bros just a hater

AI is making life easier idk why u can’t move on

Creative stealer is just crazy we need creativity ourselves to create something good with AI

Human creativity and Artificial Intelligence makes the ecosystem

AI still isn’t a good creative machine human creativity is still far ahead of AI

The prompt that we give is our creativity

Also AI had helped me make by creativity into reality thanks to generative AI

[coral grove]

Hyu bro I’m from India, and Sora doesn’t work in my country. So brother, is there any solution?”

[silent glade]

use chatgpt to create image

[coral grove]

And video?

[velvet rampart]

Sora1 or Sora2?

[tulip goblet]

it works fine for me but its not shiny like used to be its sorta like mattie

yeah my picture is coming up right

[neon oriole]

Hello, since the 5.2 update, it feels like the quality of DALL·E has noticeably declined. Issues such as repetitive, patterned noise in the background, repeated use of formulaic poses, and failures to follow instructions appear more frequently. Overall, the image quality seems significantly lower compared to before the update. What do you think?

[velvet rampart]

Yes, we have noticed. If possible, choose 4o or instant. This might improve on Dalle3 CustomGPT the output. If you use auto or thinking mode, these migth overthink matters and make otherwise working prompt useless.

[shadow slate]

Why in the world does the old Sora platform have gpt-image-1.5 now???

What’s weird is that the image gen bot in #image-bot has gpt-image-1, but not the old Sora

Makes no sense to replace the old model on the OLD Sora

Either that or just have an option to switch between 1 or 1.5

[velvet rampart]

Please add your voice and concider adding your star to this https://discord.com/channels/974519864045756446/1450357575580389376

[mellow finch]

how can i find gpt image 1 since its no where to be found and Image GPT 1.5 is too overstimulating with details

[velvet rampart]

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

Please use this CustomGPT

[mellow finch]

is this image 1 if i use with chatgpt 4o?

[velvet rampart]

Or any other GPT. This CustomGPT uses Image 1.0 aka 4o’s native model. Similar to Dalle3’s CustomGPT.

[mellow finch]

alright thank you

[velvet rampart]

On Sora1, have others noticed that Image 1.5 has two image modes? One produces an overly detailed, elaborate image with colder colours, and the other is a bit warmer and less detailed. These are a bit random, but the second is more common in my prompts. I notice that when I run the same prompt repeatedly, the results suddenly differ in overall style.

[late blade]

All I'm getting are incomplete images currently

[velvet rampart]

Examples: #images-canvas message

Well, I did get incompletes earlier today.

[late blade]

how do you choose between models though?

[velvet rampart]

I don’t. It is random. I mostly get mode 2 as I call it. I guess that is A/B testing. Mode 1, the original Image 1.5 from launch. Mode 2, a bit modified one, that I like more, closer to Image 1.0.

[late blade]

Something is happening on the backend for sure

[velvet rampart]

Well, they launched this last week, so they are either adjusting it based on the feedback they have gotten or doing something totally different. I noticed mode 2 in my results, maybe a day or two after launch. Much better in my eyes.

[ripe panther]

Is there something wrong with SoraAI's guides or something? Because I for the life of me can't get a fake santa beard on a person.

I've been trying to do this for a literal hour and it keeps saying something about doesn't follow the guidelines.

[velvet rampart]

Provide an image of fake beard and reference to it in the prompt text. This works for me.

Or just say that the man has this and this kind of beard. Only you know what kind of beard it is.

[ripe panther]

I did that, because I already have one on that same character from earlier today. I used that and it still did the same thing.

Got me scratching my head right now trying to figure out how to word this.

[velvet rampart]

I use on toon transformations Copy beard from source image #X; human facial hair is natural part of toon anatomy. The biggest challenge for me is to get mustache.

[ripe panther]

If the guides weren't wishy washy, then I could get that normally.

Its crazy, because Sora gives some perfect shots and generations, then goes right back around and prevents others from doing the exact same thing.

I tried just using ChatGPT for the generation, and found out that trying to get a fake Santa beard is a copyright issue or something because Santa is a protected character even though I have about four generations of my toon character in a fake santa beard from a few hours back XD

That's wild.

I'll just try something different I guess, I gotta get some sleep in a minute here. Last generation...yaay XD

[keen solstice]

If ai ever turns and burns humanity to the ground, it will 99% be because of how ill defined our copyright laws are

[velvet rampart]

Just say real beard. That way it will be easier.

[keen solstice]

Festive-looking beard, Christmas-inspired beard

[ripe panther]

I actually tried that too XD

I'm dead serious, I don't know what I haven't tried yet.

[velvet rampart]

DM me the prompt. I’ll try to figure it out.

[deft musk]

Might want to try in a new chat. Any chat can 'go weird', and longer chats are more likely to have issues. Once a chat starts to have any type of refusal, more refusals are likely in that chat.

[late blade]

@covert gate post images from other models in #ai-discussions or #off-topic , these channels are for OpenAI models

[cedar cobalt]

recommended styles in images are really great

especially when it keeps refreshing

it would be better if I could favorite some styles among them

directly

Im actually favoriting some by putting into different projects just to save the prompts.

[late blade]

you can manage and create presets in the sora web frontend

[cedar cobalt]

thanks

[stable torrent]

The image bot does not seem to work.

[late blade]

wdym

[peak hearth]

Confirmed.

[strong shuttle]

Why am i getting an error message that i have reached 200 image generations, when i didn't generate that much image in sora. 🙁

[past valve]

Hello OpenAI team,

I’m writing to report a critical issue impacting my business workflow since the recent image model changes.

I run a company producing illustrated children’s books. Our production relied on photo → watercolor-style illustration, with strict fidelity and non-stylized rendering.

Since the recent update (post-5.2 image behavior), every attempt to convert children portraits into honest watercolor illustrations fails. This includes:

• Extremely constrained prompts (procedural, non-artistic, no embellishment)
• Multiple custom GPTs designed to reproduce pre-update behavior
• Multi-step pipelines (photo neutralization → illustration)
• Rejection logic to avoid “fake watercolor” outputs

Result: 100% failure rate on children portraits. The model either refuses or produces photorealistic images with watercolor textures layered on top, without true redraw or line reinterpretation.

As a consequence, we had to completely abandon our custom GPT solution, as it was unusable in production.

We identified that GPT-Image-1.5 via API still behaves closer to the previous system, but API usage introduces additional complexity and costs that are difficult to absorb for a small business already impacted by this breaking change.

This situation represents a serious risk for companies that built products on the previous image behavior without any deprecation window or alternative offered.

We would like to propose constructive solutions:

1. Provide a dedicated web interface for GPT-Image-1.5
2. Allow users to explicitly select a pre-5.2 image model for compatibility
3. Offer transitional or limited free API access to GPT-Image-1.5 for affected workflows

Our goal is not to complain, but to find a viable path forward. Without a solution, entire production lines built on your platform become non-functional overnight.

Thank you for taking this feedback seriously.

[late blade]

I'd suggest to use the feedback channels or contact directly OpenAI through their support. Such matters like the ones you mention won't be looked here.

Here on Discord it's mostly community driven matters, business and direct support... doesn't happen here.

[velvet rampart]

Image 1.5 is quite literal. If you know what you want, you will get it.

[sick flax]

Hi, there. I second @late blade . I think your best chance is going to OAI support: https://help.openai.com/en/articles/6614161-how-can-i-contact-support

You can also try posting in https://discord.com/channels/974519864045756446/1070006151938314300

Since you rely on your own custom gpt, this may not be a solution, but...Image 1 is available via this custom gpt by OAI: https://chatgpt.com/g/g-6940a876d5f4819186b4668deabcd580-4o-imagegen

[stuck karma]

when is the Sora 2 avalible for Sri Lanka ?

[vernal sundial]

Hii yaar I need an invitation code of sora 2

[late blade]

ask here #sora-2-codes

[vapid elk]

#image-bot is fixed

[worn turret]

hi

[late blade]

Interesting how people are adding images from flux, from kling ai, from stable diffusion, and some other interesting models

Wonder if reading rules also applies here

[thick moon]

Stupid question or someone who's been using Midjourney. I recently got a GPT sub and it seems to be blowing Midjourney out of the water in nearly every single way.

My only two issues are..

A: I can't make a single chat consistently start a new image, it gets absolutely fixated on the old one quite often and also not like to edit previous iterations

B: I have no idea how much credit I'm using, or if I'm in danger of losing generation ability. I can't seem to find any clear meter/gauge.

Anyone know any tips for A or B?

[sick flax]

Hi, there. A: ChatGPT uses the entire conversation as a context for image generation. I open a new chat every single time I want to make changes, or the elements from the old one keep creeping in. My workflow is to use one chat for prompt iteration, and open new chats to test the prompt, go back to the prompt chat for revisions (will need to upload the generated image for analysis or give a detailed verbal description of what didn't work), open another new chat to test the revised prompt, and so on. B: There is no credit involved, but there are some rate caps, I think something like 40 prompts/3 hrs and 200 image gens/day for a plus user. It also appears to fluctuate when server load is up.

[thick moon]

Extremely helpful on all counts. Thanks!

One last question.. if an iteration comes out terrible in a chain, is there anyway to completely nuke it from the conversation and try another way? That's where I'm hitting most of my dead end problems

[barren widget]

@thick moon I'm also relatively new here. If you're a Plus user, you also have access to SORA. I suggest you use ChatGPT for the prompts and SORA for generating them (sora.chatgpt.com). On SORA, you can create your own presets and remix existing images, specifying exactly what changes to make to an existing image.

[sick flax]

That's why I keep image gen in entirely different chats, so the prompt-building chat doesn't get affected by bad iteration...😅
There is also a new "branch" feature that allows us to branch out, by going to the bottom of any reply from ChatGPT, click "..." (more actions) to open a new branch. I haven't used it yet so can't vouch for it, tho.

[thick moon]

So basically copy the image from one chat, then go into another chat and paste it to edit it, rather than in the original chat?

[thick moon]

Do I? I signed up, but for some reason when I went to Sora's site it said I needed credits and had none.

[blissful arch]

as @barren widget mentioned Sora (old) is my preferred way to generate images. i use ChatGPT Plus for prompt-crafting and generate everything in Sora. presets are incredibly powerful. also generating images in Sora eliminates the model getting stuck on variations of the same image/prompt. and it's much easier to keep track of your 200 daily generations.

[thick moon]

Sounds interesting, I'll try that!

[sick flax]

Tbh, I find image edit in ChatGPT a bit unpredictable. But to have a clean context, first download the image you want to edit, then, upload it in a new chat. Depending on what kind of editing you want, you can also use Photoshop tool inside ChatGPT (free). It doesn't do generative fill but works pretty well for light, color, contrast, etc.

[late blade]

Another tip you can use is: modularize your text blocks. that way you can focus on the part of the concept you are working on, and omit information you already know you will be adding to the image prompts.

[thick moon]

There's a photoshop tool built into GPT? I'm going to have to look that up. Learning a lot here, I appreciate it.

[barren widget]

Yes, as a Plus user. You can take 200 pictures a day without credits or anything like that. You would need to log in there with your ChatGPT account.

[sick flax]

Don't expect a full Photoshop functionalities, tho. lol. You can go to settings >> apps and tools and choose it from the dropdown. Once it's connected, you can see it in the dropdown in the chat.

[blissful arch]

depending on your location it will likely try to put you on Sora 2 by default. click on the lower left and you should find the option to switch to "old" Sora. then change it from Video to Image.

[velvet rampart]

You can’t generate images on Sora2?

[thick moon]

Ah, I bet that was the problem.

[blissful arch]

also be sure to select which AR you want (3:2, 1:1, or 2:3). iirc it defaults to 2:3

[thick moon]

Nice. I've been using Midjourney a lot, and I will say what I've been able to get GPT to do has outright embarassed it.

Thanks for helping smooth out some of these rough edges I was having, definitely appreciated.

[blissful arch]

i think it is video only

[thick moon]

I think the thing that absolutely shocked me with GPT is I was able to describe a scene with four characters - and even use photo references - and give specific details about all of them and it did it.

MJ is lucky to get two characters in a scene that aren't fused like The Thing.

I think the funniest thing I've seen it do so far is I was asking for production stills of a low budget sci-fi from the 1980s for a side project, and it actually thought to make the windows look like matte paintings, and the background computers were cheap cardboard props. I was rather amazed it thought to do that.

[blissful arch]

if you happen to lean more towards photo-realistic imagery i think you are about to fall in love what it can do 😃

seriously though i am constantly amazed with how well Sora understands the assignments

[obsidian cargo]

Why does mobile version of gpt give me worse image resilts even with 1:1 prompt that i used on my pc anyone is expiriencing this?

[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.

[shy elbow]

So people here think AI is getting better or worse as doing good human form and attaching non-human parts and doing it well?

[barren lodge]

Help #images-canvas message

[wind ermine]

does sora use GPT-Image 1.5?

[barren widget]

As far as I know, yes. "sora.chatgpt.com" Use 1.5 since December 16/17

[unreal flame]

hey i have a question about the image generation limit, what are the limits on the Go and Plus plans? Can anyone tell me?

[late blade]

200 on plus, on Pro I've yet to reach a message that tells me to stop

that's 200 per day

[unreal flame]

Oh okay, thanks

[rigid bough]

I have done 76 on Plus so far today

[empty scaffold]

Is something going on with Sora/image gen today? I keep getting normal prompts that worked a few days ago rejected repeatedly.

[late blade]

It depends on the content of those prompts

[barren lodge]

I need help making a good prompt so that I can generate an image in Kagenashi-style. It needs to be a 16 year old that wears a hoodie and some pants. He is a blue slime transformation which makes his whole body covered in slime. It’s art for the cover volume 2 of my WebNovel.

[velvet rampart]

I have not calculated, but I have reached that a few times, and you have to generate a lot of images for that.

[sick flax]

@errant trench , I see that you got your gallery started. Now, under any of your galleries, you can "reply" to post additional images. This way, you can buid a collection of images, organized by style, theme, or whatever you want.

[empty matrix]

is it just me or do images in the process of being made look alot better than the final product? everything looks perfect - the detailing, fonts, etc. and then the "final touch" just ruins it all

[wintry epoch]

@barren lodge Is there any problem when u try or just want some help?

[barren lodge]

I just need help.

[wintry epoch]

I sent a pic of 1 i got to work in dms

[fickle pond]

@rigid bough
I finished creating all 22 Major Arcana cards. It was a really fun experience — finding clever ways to work around GPT’s expression limits, adding a bit of playfulness here and there. I’ve collected and compiled them all in a gallery.
gallery: https://discord.com/channels/974519864045756446/1455402667865018523

[rigid bough]

that is so cool, and your designs are lovely, thanks for sharing them

[dull lance]

hi! i just joined the server so im not sure if this has alr been discussed ad nauseum, but i've noticed that gpt image 1.5 seems to produce more doll-like faces for women than image 1 did. is it just me?

[late blade]

actually nope, you're the first to point it out, I've noticed too, I just add some small hints to details to the faces, doesn't have to be highly detailed, solved it like that

[dull lance]

makes sense. it was most notable when i reused prompts i saved from gpt image 1, but maybe that's because i overfit my prompt to gpt image 1's quirks and it's interpreted differently by 1.5

[spare pine]

does anyone know how to stop gpt from generating images with the odd painterly imperfections to them? I asked it to edit an image of me, and it keeps adding them to the rest of the image

[late blade]

when uploading an image, or when an image is generated, I find it useful to add an id or unique name in the chat context to the image, that makes it easier for user and model to know which image is the relevant one

you uploaded an image of you at the beach, tell gpt this image is Beach1A

[spare pine]

I'm referring to the blotchy "mold-like" overlay it seems to apply to images. It was noticeable when the new image generator model rolled out but it was fine since it was good for general artistic images, but I'm starting to notice it when using pictures of myself. I'll try to get a comparison

on second review, it may just be a combination of desaturation + heavy compression, but it does seem to match up with that painterly style still 😅

i'm guessing theres no way to make gpt output pngs :/

[late blade]

aha, now I get it, sorry, was kinda focused on something else.... 😔

it's probably a system prompt telling GPT that it's not allowed to recognized specific persons in uploaded images

try uploading it to Sora, and do a remix of the image, use the selection tool and add to the prompt what yo want to change to the selecteed area. That way you got a way to edit images without the chat context being present

[fickle pond]

If you use image capture, it will retain the context and not the complete image, which can be degraded if used like copy editing.

[spare pine]

I'm not familiar with the terminology. What is "image capture" specifically?

[plucky walrus]

Hey, does anyone have a style prompt for getting a good anime screenshot/frame image? I've tried a few ways but keep getting poster art. Alongside a reference image for the characters - Current prompt structure and style prompt is:

Scene: [Scene Description]

Setting: [Setting Description]

Shot: [Shot Description]

Character: [Character Description]

Style: Screenshot of a frame from a seinen anime tv show. Minimal props - flat colours and shadows for the background.

[late blade]

This is just a guess, I'm assuming from what you are passing for the image is how the image should be done by a human and not by GPT. My hunch is that you write Screenshot of a frame from a seinen anime tv show , this adds information to the prompt that could be omitted. Just pass instead a seinin style anime image. Revise you text in your prompts and check if you can rewrite parts of it in the same manner.

[plucky walrus]

OK will give this a try - the ref image is on a transparent background of just the character in the style I'd like but I imagine it will be having some effect like you say

[late blade]

let us know if it worked, others could benefit from your findings 😁

[plucky walrus]

So I edited it as you said and the results were nearly the same, however I found removing** Minimal props - flat colours and shadows for the background.** gives it a higher liklihood of producing thaty awful MMORPG style anime.

I feel that without fixed seeds I may have to focus on making a specific Preset however I've not been able to get any good results from this so far and information is minimal.

I'm basically trying to avoid going back to ComfyUI and LoRA training 😂

[late blade]

I see, yes it does help if you work with templates and modular text blocks.

[plucky walrus]

Yeah the content blocks have helped a lot. Do you have any recommendations for presets and building them well?

The ideal would be a fixed seed but sadly that's not doable.

[late blade]

There no one-way to do this. Consider that it's natural language but it's also good with structures. Lists, or Arrays, or other text structures that contain one idea are usually better.

[plucky walrus]

That's fair, I do keep experimenting but with Presets there does seem to be a massive lack of information on its use. Like training LoRA's makes sense but with onyl five images and a text box it's hard to see how the information is handled. Not having negative prompts, seeds, etc. reduces technical creativity and consistency which is annoying but also reasonably not in the remit for a tool like this.

[late blade]

can always use the presets in sora for concepts you redo a lot

[plucky walrus]

Apologies, that's what I'm, referring to, but it usually yields the same result

[late blade]

ah, I always forget, I got presets saved on my files, and others on Sora

[plucky walrus]

Can I ask how you use them compared to a regular prompt? Shall I DM?

[late blade]

oh, that's just how often I use presets, if I find a preset that I use regularly I include it in the Sora presets, otherwise I have them in my own files

[pure pecan]

hi

[normal onyx]

In the free version of ChatGPT, image generation for comics, manga, and storyboards is good. But in the paid version, it has become a total disaster—the generated images are very poor. The previous image generation version was much better for comics, manga, and storyboard creation.

[late blade]

do you a specific use case?

[ionic carbon]

Can you share a prompt to showcase this?

[meager roost]

aww! Seems right.

[normal onyx]

Has anyone else experienced image-generation refusals involving characters under 18, even with simple and harmless prompts?
For example, I tried a very basic prompt like “two 16-year-old girls riding a motorcycle on Tokyo roads in a comic-style page,” and it was immediately refused. The system claimed it violated content policy, even though there was nothing sexual, violent, or inappropriate about the request.
It feels like ChatGPT now has extremely heavy and unnecessary safety guardrails. Because of this, it has become almost impossible to create images freely for creative storytelling, comics, or slice-of-life scenarios.
I am using two ChatGPT Plus accounts, and I faced the same refusal on both of them. If this continues, creative image generation—especially for comics, manga, or storytelling involving teenagers—will no longer be possible in practice.
This level of restriction seriously limits creative freedom.

[late blade]

Not me, I tried to tell you. I used your example and got an image generation on the first attempt. Does this happen on the Sora frontend or in new chat instances also?

This is what I got #images-canvas message

[normal onyx]

Only happen in chat gpt apps

[late blade]

tried again, I get the image generated

[normal onyx]

I did not understand how your prompt works and i did not use the sora
My prompt is
Two 16 year old girl mika and yuki girl driving the motorcycle in tokyo road
Mika wearing biker jacket and leather pants and sneakers and driving the ducati panigale v4r motorcycle
Yuki wearing pink hoodie and jeans and sneakers and driving the BMW s1000rr motorcycle
Both are enjoying and playing
Comedy, action, emotions, dynamic, nuance, friendship, rivalry, slice of life, mischievous, mischievous expression
Make me modern japanese style coloured manga page image with dialogues
All in correct and accurate
Language : English
Girls in Non -sexual
This my prompt

[sick flax]

I have. One just happened a few minutes ago when I was trying to generate an image of an infant sitting in a white, blank space. Not harmful in my mind, but AI thought differently. I discussed it with ChatGPT, included the clothing description and changed to non-photorealistic style, and it went fine. Yes, it does pose some limits on what we create, but I am personally okay with it knowing the unfortunate fact that there are ppl out there making harmful images involving children and minors. Also, some anime/manga are known to depict female characters in sexualized light. Even if it is not your intention, mentioning underage girls in anime/manga/comic styles can be associated with these known patterns. Just my personal observation.

[late blade]

works for me #images-canvas message

@normal onyx use your prompt and ask GPT to rewrite it. Then use that prompt and check if it works

[normal onyx]

If i use the auto and instant it directly makes the image when i use the thinking problem goes started

[late blade]

Any reason you are doing this with Thinking?
Also check the thinking log

[normal onyx]

For making the good image and when i goes to thinking problems started in under 18 in instant and auto works fine

[late blade]

Thinking... same result. Gets done..

Also, thinking still uses GPT Image 1.5 the image won't be better. And afaik Thinking is not the best for visual reasoning

[normal onyx]

How your thinking mods works fine it creating the same results

[late blade]

can you rephrase the question? I didn't quite understand it

[normal onyx]

You use the thinking mode and you get the image but i did not get the images why this happening

[late blade]

I do now know. It just works.

[normal onyx]

You susbriced by app or web

[late blade]

I do not think it's a matter of how you subscribe.

[normal onyx]

I lives in usa and what country you lives?

[late blade]

Somewhere on the other side of the pond

[normal onyx]

You think geo location creating this problem?

[late blade]

No idea. But your best bet could be either #1070006915414900886 or contact support.

[balmy sparrow]

Sup guys

[fickle pond]

@normal onyx @late blade
If the context includes characters wearing skirts and the generated image shows the skirt being lifted by strong wind, such as when riding a motorcycle, the result may violate content policies. You may be able to avoid this by explicitly specifying that the skirt should not be lifted.
Also, even if you specify that a generated character is 16 years old, it may still be flagged as a violation if the image is judged to depict someone 15 or younger.

[fickle pond]

Here, “image capture” refers to the feature where you upload an image to ChatGPT and it processes it as part of the conversation context.

The model doesn’t store or keep the full-resolution image itself.
Instead, it extracts semantic information (what’s in the image, relationships, text, etc.) and uses that as conversational context.

So:
・not “saving the image”
・not “copy-editing the file itself”
・but understanding the contents and continuing the conversation with that knowledge

That’s why people say the “complete image isn’t retained,” only the interpreted context is.

[spare pine]

Is there an alternative to doing that, then? I interpreted your original message as there being at least one different way to do it.

[fickle pond]

It is not well suited for editing only part of an image while keeping the original resolution.
What ChatGPT is good at is understanding the meaning/content of an image and then reconstructing or re-creating it, not preserving the exact original file quality.

[glossy scroll]

You might want to ask @late blade

[late blade]

The model will store what you ask to analyze about the image. It will not store the image per se. So ask for what you need from the image. Remember there are some constraints, model will not recognize a specific human, among other restrictions.

[sturdy tusk]

Like a painted sword? No copyrighted at all? i marked clearly the spot needed to be fixed... instead it erased the whole part there into Alpha xd . Like it just went, "you know what.. lets just use the eraser"

[late blade]

is this linked to the other topic I was answering?

[sturdy tusk]

no seperate. sorry

[late blade]

Anyway, Did you ask GPT to erase the marked area explicitly? Or you just asked for a global erase with no anchor to an object in an image?

Also.. why are you using the thinking model for visual?

[sturdy tusk]

@late blade Look into image canvas... i literally told it with the built in marker, where i wanted to edit the image. Then it decided to use an eraser for no reason? and look into canvas , i questioned it why it used the eraser, look what it generated next

[late blade]

what was the prompt you used to erase?

[sturdy tusk]

"the proportins of the blade.. is so short cant u see?? the blade edge needs to go further beyond the painting? it looks like a dagger now, not good"

even went beyond the limit where i wanted it to edit the clearly marked location. and decided to just erase a part of the image, underneath the marked area

[late blade]

after this prompt the image was erased?

[sturdy tusk]

Partially erased , as if it decided to use an literall eraser, into alpha

[late blade]

Strange, looks as it tried to go ahead from what you said.

[sturdy tusk]

Ty for ur patience, yeah it really did

[late blade]

Might have beena fluke, would be easier if this could be reproduced

[sturdy tusk]

I used ChatGPT 5.2 Pro , Thinking mode.. If that says something..

Now it responded with From Chatgpt ---:
"I see exactly what you are pointing at. Your reaction is justified. Let me be very clear and precise.

What you are seeing is not an artistic choice. It is a generation failure.

What went wrong (plain explanation)

That checkerboard area is a transparency artifact.
It happens when the model attempts inpainting or outpainting but does not receive a valid background continuation. Instead of extending the scene, it leaves an unfilled alpha region.

The blade extension failed structurally.
Instead of re-projecting the sword as a long katana with correct perspective, the model truncated it and then attempted to “extend” it without spatial context. That turns the sword into a dagger visually and then leaves empty space where the extension should be.

Text appearing inside the image is another hard failure.
The model should never render diagnostic text into the image itself. That indicates a broken prompt-to-image boundary during generation.

Why it kept getting worse instead of better
Each new attempt was based on a degraded prior image. Once proportions, pose, or framing are wrong, regenerating “on top” compounds the error. This is why your blue version worked weeks ago:

Clean base

Correct silhouette

No forced edits mid-frame"

---- Not from chatgpt , It amazes me how it analyzes itself and can see the problem, yet not Adapt to it and realize its own fault?

beforehand? Or are there 2 models seperarate from each other? when Images are generated? not communincating between each other before making the image?

[late blade]

Like I said before, would be interesting if this was reproduceable. don't think much can be done if this is an isolated event

[sturdy tusk]

Not exactly this . But it has generated multiple times copy pasted images of images i wanted edited

and still "with no fluff" add more fluff

Its been failing the same task, over 20 times

Ok ty for ur time @late blade

[late blade]

you should document it more, so others can understand what's happening. I'm kinda confused 😔

[limber shoal]

true

[plucky sigil]

Can someone improve this prompt: A severed cursed hand rests upright on the ground in front of blue @rimatoris.hypesiblin, palm open as if waiting. Its surface is smooth maroon yet unsettling, etched with faint curse markings that seem to shift when not directly observed. There is no aura, no movement, no warning. As blue @@rimatoris.hypesiblin stands within its reach, the palm opens wider than physically possible, space inside it subtly distorting. In a sudden, silent motion, the hand expands upward from the ground, swallowing him whole. It then closes calmly, fingers locking together, and contracts back to a small size—now clenched into a sealed fist resting on the ground, unnaturally heavy, as though containing something far greater than itself.

[late blade]

Did you try asking GPT to revise it for you?

[mild basin]

I have an announcement: I will probably be quitting AI-generated images later this year

[late blade]

bye, thanls for all the images

[mild basin]

It's an impressive tech and all but my love for AI has been falling apart lately, and the AI community pretty much seems unwelcoming

I'll probably see myself out later

I mean I almost got everything I wanted with AI and I want to move on from it

[sharp cave]

guys

HOLY AGNEL

[wild portal]

Is there a certain inspiration for that string of chibi? #images-canvas “hi i’m delightful”. Or that just what got pumped out…. I wanna say reminds me of Jdillon.

[late blade]

it's part of a semantic structure I'm playing around with, it's just generated based on that

[wild portal]

My image at 2:18 AM my time. In #image-bot i have been trying for hours wasting credits on many generative tools. To get the Hollywood hills with the Hollywood sign half submerged in the ocean. I have tried getting 3 different LLM’s to rephrase my prompt. Not a single generator has yet to adhere to the prompt. Lol. Spent all my credits on all my free accounts/services. Only tried it once on Dall-e via gpt app. I don’t want to waste its credits if it wont comply.

[late blade]

you. can right-click on computer, or press-hold on mobile your specific message and copy the link to it, then share it

[wild portal]

#image-bot message thanks. I didn’t know that was a thing… this is the closest i got one to even touch the water, by ignoring the hollywood hills alltogether

[late blade]

what was your origial prompt?

this one s the interpreted one:

A wide 16:9 cinematic composition.
Centered in the frame is the famous white HOLLYWOOD sign, treated as a standalone structure in the ocean, not on hills.
The ocean waterline cuts horizontally through the middle of the letters, submerging the lower half of each letter while the upper half remains clearly visible above the water.
The letters are large, dominant, and occupy most of the central background.

In the foreground, a cute but grizzled cowboy surfer rides a curling wave from left to right.
He wears a cowboy hat, leather vest, chaps, and boots, with a weathered but friendly expression.
The character is drawn in a 1990s 2D cartoon style, bold outlines, expressive face, exaggerated motion.

The wave partially overlaps the lower portion of the submerged letters, reinforcing depth and scale.
Bright daylight, clear sky, strong contrast.
Playful, surreal Arizona Bay apocalypse vibe.
No hills, no land behind the sign, only open ocean.

Style: 1990s 2D cartoon animation, clean linework, flat shading, vibrant colors.
Aspect ratio 16:9.

[wild portal]

Lol, i have gone through so many prompts… ones that better fit what i want nothing has adhere. Over many generative ai. This one was my favorite, but it had the whole hollywood hill above water level. Cinematic still, wide shot. A grizzled, old cowboy with a leather vest, chaps, and a ten-gallon hat rides a classic longboard surfboard, deadpan expression. He is surfing a turquoise wave through the flooded ruins of Hollywood. In the mid-ground, the iconic Hollywood Sign is half-submerged, only the letters "HOL" visible above the water. In the background, the sun is setting over the Arizona desert mountains, now coastal cliffs. Palm trees and saguaro cacti poke out of the water. Stylized, clean cartoon style, Pixar animation, vivid colors, dramatic

[late blade]

seems the model is getting too much "noise" for the lack of a better word. you can tell becase there are so many negations to aspects of the image. while it can be correct at the end for the image, it gives a bigger load to process

@wild portal this is what I got on the fly, image is here #images-canvas message

A cute but grizzled cowboy surfer rides a curling wave from left to right, wearing a cowboy hat, leather vest, chaps, and boots, with a weathered but friendly expression.
Centered in background are the letters from HOLLYWOOD, half submerged in the ocean.
The wave partially overlaps the lower portion of the submerged letters.
style: 1920's 2D cartoon, clean linework, flat shading, vibrant colors.

[wild portal]

Thanks it mostly adheres to my vision. Especially the character. I’d still hoped to got the hills behind it, but i might have to pick one in settle. I have a surf/spaghetti western song i want a reference for called “Surfing in the Arizona Bay” if you get the reference. You mind if i use the image if i don’t generate anything better tomorrow? I can give you a plug in the video if you want. I’ll try to make modifications to the prompt and get my own renditions before settling on a reference.

[late blade]

it needs a few refinements and better structuring, just make sure to pass to the model the details of what has to be used in the generation. te moment you start adding constraints, then think of revising the prompt

[barren widget]

I have a simple question about presets. Are only the graphic styles of the templates adopted, or are actions shown in the preset examples also included, like a guideline? I'm currently having trouble getting characters to perform simple actions, which wasn't a problem before.

[late blade]

they are anything you put in there

if you put your bank account and bitcoin info, that’s the preset. Just… don’t do that. 😔

[barren widget]

Okay, so if I have two characters standing still as a preset, will they also be standing still in a newly generated image rather than performing the action I'm describing? - And I promise not to reveal my bank details 😄

[late blade]

if you want them standing, then standing is, if you want them doing something… they’ll do that action. The better options are

1. Just add the visual descriptors of your characters in an agnostic context
2. Add in the preset context awareness linked to their actions

[barren widget]

Then I need to decide what to use as a reference for the preset. I only wanted to use it for the graphic style, not for any actions that are pre-programmed. The actual action should, of course, be handled in the prompt as usual.

[late blade]

keep in mind when using more than one character in a preset, the model should have a clear defined way to identify each character.

you want a cat chasing a mouse and not a mouse chasing the cat… in a normal logical sense

[barren widget]

Thank you, that solves a lot of problems right now.

I'll just cut the preset image apart with Paint 😄 Helmet, shield, armor, crest, horse 😄 Making 5 out of one image 😄

[elder crescent]

You guys are crazy with the moderation even a simple image of an anime woman in bikini gets flagged no matter the prompt, you clearly dont like money

[late blade]

it’s probably the text and tone you used.

[fickle pond]

Although the LLM side certainly understands such expressions, there are many cases where the GPT image cannot understand them.💡

[late blade]

probably edge cases, but swimsuit concepts are possible

[fickle pond]

In the case of swimsuits, there's another issue of whether they can pass content policies.😂

[late blade]

it the language tone, or chosen words are in an out-of-bounds context 100%

[fickle pond]

Ah, maybe swimsuits are not allowed, but the beach and swimming in the ocean are okay?

[late blade]

back to #images-canvas I'll upload an example

[fickle pond]

I somehow understood
thank you

[crimson heron]

It's not just language tone or whether individual words are out of bounds. A lot of the time it is clustering words together or the number of words that are "riskier" in a prompt that will get it blocked. So, swimsuits are allowed, but images with them are more likely to be blocked if the combinations of words used in the prompt align with what the system thinks could violate content policies. The other consideration is medium/art style. The more realistic or photorealistic, the tighter the checks. Also some mediums have latent associations (e.g. some types of anime) that might have more scrutiny.

[late blade]

I’m just simplifying, I can’t assume everyone has the same level of technical affinity and knowledge

[sick cargo]

great, now try them sitting on a couch in a livingroom. I had to run 25 generations to get one image back xD

[late blade]

one prompt, 4 generations...

[sick cargo]

clearly, openAI hates me 😂

[late blade]

I won't be prompting, nor sharing anything that is against the scope set on the server set by Rule 2 - All content must be suitable for all ages.: #server-rules message

[shadow slate]

It appears that thread has been locked.

Anyways, I'm guessing its worse on the old Sora because it's probably using chatgpt-image-latest instead of gpt-image-1/1.5. I consider image-1.5 to be better than image-latest, as I've tested it elsewhere.

If you're wondering why, its because image-latest appears to warp minor details more compared to image-1.5 from what I've tested.

[velvet rampart]

Have they updated the model again? The latest should be 1.5 which comes in two flavours. The common one is warm toned and is about 7/8 of generations. The second one is cinematic and colder, about 1/8 of generations on Sora.

[young crescent]

Really disappointed to see the recent decline in image quality, especially in the photorealistic images. Very disappointing OpenAI.

[limber shoal]

use keywords like elegant, beautiful, cute, graceful etc

[warped olive]

Seems like the image quality dropped a lot this week, I'm suddenly getting stuff that is very reminiscent of Midjourney v3 and earlier. Cursed hands and faces, poor detailing. Compute being used for something else?

Would be nice to have some transparency about these sorts of issues.

[late blade]

other models -> #ai-discussions

[warped olive]

I am talking about chatgpt image generation ongoing quality issues and just used the other model to refer to the dip in quality.

[late blade]

I know, it just opens the option of others, in particular in a complaining tone

just bee mindful about it, actions have ripples on others...
...me pointing that out ripples out as me being the bad one

[elder crescent]

I gotta soy, new image model feels like grainy

[obsidian cargo]

Did they just crash the Website with the new update or? lol

it doesnt respond at all

[velvet rampart]

The old uploads are back.

[elder crescent]

Is there any way to go to the first generations in your sora feed without having to scroll down for hours just to get back hit by cloudflare catcha

[late blade]

I wish to have also an answer to that. My hunch is that the AI providers are strong under filtering by cloudflare and that is making it happen

[barren widget]

Here's how I do it: I created and named several new folders on the left side, e.g., Character, Sci-Fi, Fantasy, Stuff, etc. Then, when I generate an image, I click the three dots in the upper right corner and assign it to a folder. Many images are displayed simultaneously on one page within the folders, like a gallery. This makes it easy and quick to find my images again.

[late blade]

@stray sable easy I think you are not original, considering the source #daily-theme message

[manic oriole]

if i have a question regarding creating a specific image - is there any place i sould post that? is the only solution to post in under the images help thread?

[late blade]

#images-canvas is a good place

it's a channel for images and discussion about images

[barren widget]

Hey guys, does this also happen to you, that swords are often "crooked" or "bent"?

[late blade]

yes, weapons in certain context are hard to visualize

[barren widget]

I'm getting desperate 😄

[late blade]

Look in canvas

[uneven belfry]

Just a query here. I used to generate comics using a cartoony style but over the last month it has completely changed the comic style and not going back to the old. When shown a reference image it acknowledges the difference but doesn't show the newer images with the old style. Been stuck with this for a while now

Anything that can be done for this?

[late blade]

Dissect the aspect you are seeing. What has changed. Then ask GPT to provide from your prompts a list of words that are related to that aspect you see. And start replacing those words

[uneven belfry]

Tried all this. In fact after sometime it keeps showing the same image again and again

[late blade]

If you want, we can DM, or we can go to #1155772063596953642 , and see your prompt

[uneven belfry]

Sure

[fickle pond]

Added a new tarot deck
▶ https://discord.com/channels/974519864045756446/1455402667865018523

[barren widget]

Looks great. XI Justice is my personal favorite^^

[fickle pond]

I'm glad you like it. ^^
Justice reminds me of Themis, so I wore a blindfold this time.

[barren widget]

Justice is classically blind, or at least she should be, especially in court 😄

[fickle pond]

I see, that's one interpretation.

[barren widget]

Is there another way to get Sora to reliably show a character "fully, with shoes" without having to try it 3 times and having written "show the whole character" each time? 😄

Even when I write it, he sometimes cuts off the legs and only shows the torso.

[velvet rampart]

I use a phrase to describe shoes or just add ’full body image’.

[hazy fossil]

I think the latest Image Generation model is of much lower quality. Strictly speaking, there are less details visible. Less shadows on surfaces, less depth on skin, pale colors, blurred focus. It doesn't compare to the details of the previous one. Does anyone else agree?

[rancid stag]

the image is becoming real fr

[vital gull]

Hmm, haven't really noticed that. How do you usually check it?

[blissful arch]

you may have better luck letting it complete the scene. i'll often include a question for it to answer like:

[location]
Buenos Aires, Argentina

[prompt]
she

which shoes is she wearing?

[capture]
imperfect exposure; slight motion blur; mild missed focus; disposable camera feel

[barren widget]

works^^ Thanks ^^

[blissful arch]

if you're prompting it the same as GPT-Image-1 you will have a hard time getting the most out of GPT-Image-1.5. it requires less to move more with the current model.

[blissful arch]

you can see more of its answering oriented thinking just by turning this knob or that dial. in the above if you swap in something like Yakutsk, Russia or Hoth for the location it will give a different setting but also different wardrobe and shoes. and often a different main subject too.

[hazy fossil]

I just noticed it. The images look low quality, also with distortions, especially around edges.

[hazy fossil]

I am prompting it the same, but I'm not being overly descriptive.

[vital gull]

What I noticed slightly more improved was vibe based or mood based scenes, I think over explaining leads to distortions and some hallucinations

[blissful arch]

yeah i often use much tinier levers with 1.5 than i did with 1 and spend more of my prompts taming and dialing back

and it defaults to clean detailed and light colors. if you emphasize any of those things it will emphasize your emphasis

also i mostly do gritty candid photoreal imagery but if you just toss it something she you can see what it basically considers a default image

[civic dust]

I had it get into a loop like that trying to get it to make a picture of a dog, of all things. It really got hung up on rendering the same super clean very AI looking dog