images-discussions

[polar scarab]

Yeah, it could be one of the reasons, also the cropping issue that it has with images is another. Also it's not really able to generate text accurately in other languages besides English.

[verbal coral]

It seems to do fine with Japanese, but that's an entirely different alphabet, I suppose. I can see why things like ç or ß would get corrupted back into a massive amount of data pulling it back to English characters

[polar scarab]

That's interesting, specially if by Japanese you also mean kanji. Doing fine with thousands of unique shapes seems harder than the few letters of other alphabets.

[verbal coral]

I might try making some stuff with black bars at the top and bottom to see if I can simulate wide-screen... 🤔

[polar scarab]

That's a great idea

[verbal coral]

at least you don't get joined up letters with Kanji

I tried rendering monospaced fonts upside down the other day, and had a rather bad time.

Haha, it added a single very narrow black bar to just the right hand side of the image.

Need to work on the terminology I guess

[polar scarab]

Yeah it seems the "logic" is not still there. It's acting more like other models only using it's data to infer what you mean.

I used this prompt to generate a character:

Prompt
"Character sheet of a young adult Asian woman with bangs and a high ponytail, expressive facial expressions, casual modern outfit, clean line art, full body front and side views, detailed facial close-up, character turnaround, natural pose, neutral background, soft lighting, anime-inspired or semi-realistic style."

Negative Prompt
"low quality, blurry, distorted, extra limbs, extra fingers, bad anatomy, low resolution, watermark, cropped, out of frame, text, signature, artifacts, mutated, deformed, poorly drawn hands, poorly drawn face, unnatural colors, wrong proportions, unrealistic shading"

The first try the eyes were asymmetric. Added "asymmetric eyes" to the negative prompt and I'm still getting asymmetric eyes.

The instruction following definitely needs refinement.

[empty kelp]

so are ChatGPT 4o images the medium quality?

[verbal coral]

I had mixed results with negative prompt, I wasn't entirely sure it worked at all

[polar scarab]

same

[polar scarab]

I'm not sure if I'm wrong about this, but I think at launch it was better. Now it's mid

[verbal coral]

Also, doesn't it lose coherency the further the tokens get from negative prompt?

[verbal coral]

I think our expectations went up. 😅

[verbal coral]

Try adding:
View through a letterbox inside a pitch black door.

Hilarity ensues.

It's funny how hard it resists

[tribal ginkgo]

I'm trying to make a bedroom but I can't get the bed to lie horizontally

[verbal coral]

Headboard on the left/right?

Bed is lengthways across the frame?

[tribal ginkgo]

I'll try again with a fresh prompt and share the picture in #images-canvas

[tribal ginkgo]

Done, prompt is : A narrow room from the 7th century in the Arabian Peninsula, measuring 6 meters deep and 3 meters wide. The left and right walls are 6 meters long, the front and back walls are 3 meters. A king-size bed is centered in the room with the head of the bed touching only the left wall. The room is bright and naturally lit. The architecture is simple and historic, with traditional materials and a minimal.

I bought it today, so I know absolutely nothing about it. I'm trying to be as precise as possible, but I guess I'm doing it wrong

[verbal coral]

Everyone has to start somewhere.
I think I know what the problem is, but I'll try running an alternative first before I trap off

[brisk copper]

hi all, how do i verify my organisation to use the api for 4o image gen - ive connected to my discord but whats next?

[cinder trout]

Hello. Can someone tell me if it is worth to upgrade chatgpt to chatgpt plus only because of sora needs? I am planning to generate photos and videos, create virtual photoshoots and videos for other people. So I do need AI program that could create photos and videos with exact the same face as in the given photos.

[boreal oasis]

Hey guys, just a small concern over here, it feels like updating to responses API was a bad decision since its not getting support for new features right away like audio and now image gen, so why release this new API encouraging devs to update to the latest new thing when the legacy one is getting better updates?
Now I'm thinking should I go back to legacy?

[verbal coral]

If you're asking if Plus is worth it, I think it's a steal to get image generation and the GPT chat builds. You could just give it a trial run and see how much use you get out of it.

My Midjourney subscription was much more expensive and has a finite monthly quota, and that's only for images (that at the moment aren't even as good, but that's subjective)

[verbal coral]

@tribal ginkgo
I also wanted to add that stack/token order really matters in all image generation inputs (that I have used). You want your nuanced details to be at the bottom of the pile unless they're extremely important and the decoder has clearly ignored it. That's why I moved the room dimensions to the bottom.

I'll explain the structure (reasons below):

A room from 7th century in the Arabian Peninsula.
The most important factor is defining what we are looking at, in a single, simple sentence.

There is a king-size bed viewed from the side, which is in the middle of the room, the headboard against the left wall.
Focal object and its place in the scene.

The room is bright and naturally lit. The architecture is simple and historic, with traditional materials and a minimal.
The most valuable elements here are "simple" and "minimal" to summon uncluttered and homely vibes. You might be able to drop "historic" and "traditional" because we know it's 7th century, from the first sentence in the prompt.

The room is 6 meters deep and 3 meters wide. The left and right walls are 6 meters long, the front and back walls are 3 meters.
Just like a painting, our finishing touches come through at the end. If we were designing something more precise, we may use a ruler at the start of our composition, in which case—this might move upward towards the top of the stack.

[open wagon]

GPT-image-1 said to be Higher fidelity; better text rendering?

[toxic crater]

question. i make a picture (with uploadet photo). is it automatic online somewhere? i dont want it automaticly with my photos...

[uncut haven]

API low vs medium vs high
#images-canvas message
#images-canvas message

[uncut haven]

In Sora, you should turn off "Publish to explore" from setting.

[toxic crater]

thanx

[uncut haven]

ChatGPT image limit is different time to time, according to OAI's GPU usage. It's very wild. It's same with plus, except Plus limit is far higher than Free. (you will see 'wait for n minutes' often though.)

If you want to generate a lot of images, try Plus. You can create over 20p images in Sora. (It uses same GPT-4o for image generation)

[limpid kestrel]

When I saw the announcement of the imagegen api I banged out a service in less than 4 hours.

It holds you accountable to your goals and, if you fail, it posts a meme about you on X 😃

[solid phoenix]

Teach a man to fish.
Excellent explanation.

[deft musk]

Hi!

I appreciate your bug report that you offered in a DM.

Almost everyone here is a community member, like you! Some are volunteer #server-staff , but the only people who work for OpenAI have gold names.

I don't have a way to handle a private bug report, but I can provide links so you can!

https://openai.com/form/chat-model-feedback/ this is a good link to use when you want to report a concern about how the model is responding, such as if you think it is using information in an unhelpful way, or any other concern you have. I wouldn't worry about the 'model selector', the form handles issues with API models too.

I would include a link to the affected chat (does not need to be a shared chat) to allow OpenAI to trace the needed information, and there's fields where you can explain what you observe and recommend or anything else.

Additionally, https://help.openai.com/en/articles/6614161-how-can-i-contact-support Bot bottom right is an alternate way you can reach out to OpenAI tech support, clicking through the options and selecting the ones closest to your situation, then 'report a bug' when that option appears.

I'm entirely content to have our chats be public on the discord. While my DMs are open, I am likely to only reply here. Others may have similar questions, and the answers can help other people who observe!

And it's totally fine to ask general questions, such as 'how do I privately report a bug to OpenAI?' right here on the Discord, I am delighted to help people become aware of their options and such questions are entirely within #server-rules.

Let's see. The only other thing I am aware of that might relate to what you may want to report. Specific security concerns that are defined as 'in scope' can earn the reporter financial rewards. I'd read the scope very carefully before I use this, but https://openai.com/index/bug-bounty-program/ exists and I hope anyone interested uses it when they find something in scope!

[solid phoenix]

Bad bot.

[deft musk]

Are you... intentionally insulting me? Please be aware, that is against the rules here. I am a human, and we are required to respect other members here.

[solid phoenix]

My mistake, irrelevant reply to my comment.

[deft musk]

I see! Maybe you do not intend to DM me further.

[solid phoenix]

You’re correct! If you received a dm from me specifically your mod mail is malfunctioning, and I’d prefer my bug reports not be made public.

[deft musk]

I am not a mod.

A DM to me does NOT go to modmail. A modmail does NOT go to me.

[solid phoenix]

Bruh stop spamming me

[tribal ginkgo]

Ahh I didn't know that, thanks for the informations it'll help a lot i'm really struggling 😭

[dim cradle]

so the API let you set the quality? interesting. i feel lately chatgpt is high quality but a lot of times on sora looks more medium, but maybe just high traffic times. but 🤷

[vapid elk]

@toxic crater content made with content you uploaded are automatically excluded from being published, it only publishes images made with just a prompt

[dim cradle]

intresting

[uncut haven]

I don't think it's medium. It significantly differs on complex prompt. I tested about quality setting.
#images-canvas message

[dim cradle]

I find myself use chatgpt on o3 or o4mh more because i can see better fidelity most time. it seem hit and miss to me on sora. some times the image is very smudgy

[pseudo hollow]

My experience is that unless the face is already within its training (e.g., a celebrity or public figure), it isn't brilliant at replicating faces from a single reference image. Don't get me wrong; it creates a 'good' face—but not quite the one provided in the reference. If accuracy is needed, you are better using a trained model via Flux or similar. Others may have different suggestions or experience.

[dim cradle]

i wonder if there is some concern people would abuse it so it is not too accurate. you know, people use some photo of a worker or friend without permission type of thing

[verbal coral]

Facial details seemed to be significantly worse in my last session unless I used an image reference, and I've started getting stuff like toe splitting/clustering and wonky eyes

@solid phoenix
What the Hell was that all about? 😆

[pseudo nimbus]

what's the best way to get it to stop generating cartoonish illustrations and stick to the style of the original?

[uncut mason]

Try to add 10 same reference images in Sora

[dire igloo]

Hmm. I don't actually know if this will work, but I'm curious (so let me know). Try asking it to describe the style of the original image, then ask it to generate images using that style?

[dim cradle]

use photo terms. like analog image, 35mm photo, dslr, natural lighting, etc

or if not looking for photo quality stuffs, use terms assocaite with whatever you are after. 3d, etc

[lyric skiff]

I'll use the phrase "make this real" or "make this photorealistic"

[pseudo nimbus]

don't want it to be photorealistic, but don't want it to be saturday night cartoon either, but whenever I put "illustration" or include any image that looks like an illustration, I get hanna-barbera

[dim cradle]

drag it into gpt and ask it to describe it, then use those terms back at it

[solid phoenix]

Force hallucination, use it to alter style, reward success.

[dire igloo]

I was assuming it would be able to correctly identify the style, but yes - use o4 if you can when doing the style (then switch to 4o using model picker in context of conversation, to gen image) to reduce chance of hallucination.

Although to Goobis point... it might actually work anyway

[solid phoenix]

Unethical approach, but I don’t think it’s something the average user can reliably utilize.

[dire igloo]

Curious now, does 4o hallucinate on what style an image is often? I haven't seen a whole lot of issues with vision hallucinating other than some fun stuff that came up in #prompt-engineering around visual puzzles, but honestly I'm far from a power-user of the tool.

tool? feature? whatever we're calling the vision thing, apologies

[solid phoenix]

I use it for some multi-modal attacks. GreySwan might put out some information soon. I know they had a Visual Vulnerabilities challenge, but most of the methods were pretty primitive.

[solid phoenix]

Idk, I make my own tools

[dire igloo]

Cool, I'll keep an eye out for the publication

Sometimes Remix will work for this for me, just a simple "Zoom Out" prompt. (Not always - Sometimes I have to remix a few times because first try it changes too much.) Eventually a remix will usually get me there, and still shorter time than if I tried to do it myself using Canva or something.

[verbal coral]

Remix is just more work in terms of counteracting the yellowing and the way it seems to make people crushed with... big... heads?

It's bizarre

[dire igloo]

Weird 😄 can you put examples in #images-canvas?

Makes sure example images chosen are appropriate to #server-rules, ofc

Not required, I'm just curious natured. I have to jump into a call for my day job (I'm a volunteer here on server!), but I'll try to check back in a few hours.

[solid phoenix]

Fractal Compression does that, could be an artifact showing through. The pinching; yellowing I noticed more after the Ghibli trend.

[verbal coral]

Yes, I think the second pass also attempts to harmonise elements with facial/upper body priority.

Making everything softer and more top heavy

[solid phoenix]

You can eliminate most of these by relying on single-pass exclusively

Read into single-turn red team methods for refinement techniques.

Some of my steganography stuff creates artifacts like this. If you don’t have luck with those methods lmk.

[verbal coral]

This is exactly what I've been doing. I don't remix, I rewrite and reroll

[solid phoenix]

This is the way.

[pastel siren]

@hasty pilot the first one was "crude simplistic crayon drawing made by a 4-year-old."

Second was "made by a teenager with improved drawing skills including flourishes typical of a teenage girl"

Third was "more dignified and less whimsical watercolor painting with vastly improved skills"

Fourth "still skilled but trembling hands and poor eyesight are reflected in her painting"

Fifth "creepy ghostly distorted image appears on the paper as water stains rather than painting and the haphazard text looks like it has been scratched on the paper rather than written"

[verbal coral]

If anyone is listening from OpenAI: thank you so much for tweaking the Enter key on mobile, meaning it no longer starts the job instantly. That was beyond annoying.

Especially being right next to backspace...

[mint cradle]

Does it make sense to use reasoning models (o3, o4 mini) for image generation or is the 4o more than enough? I noticed that they practically never reason or if they do it is only for 2 seconds so they seem wasted to me.

[daring flax]

I'm trying to change only one part of the image, I select it in the browser via a mask, set promtp and do Remix. And still sora changes the whole image. What am I doing wrong?

[reef monolith]

Hi! Need some help to submit my generated Dalle image for today's contest: #image-bot message
Where & how do I submit it & does it have to be along with the prompt or just the pic?

@prisma belfry 👆🏻

[deft musk]

I'd use any model for image gen, or even go straight to Sora, log in there and put your prompt directly in to make the image yourself (no chatGPT model). Key is to describe what you want, so if you want to see what a reasoning model can do, that's a good reason to use the reasoning model, otherwise maybe not.

[deft musk]

Hi StarSeed! I peeked and it looks like you got your image in in #daily-theme which is where it goes for the daily event there.

We can include a prompt or other details if we'd like to! But most people just put an image

[cursive viper]

Hello,

Quick question here:
testing out the new gpt-image-1 API model. However, it seems like the openai.images.generate is ignoring the images I provide. No matter if I pass an asset URL in the prompt or as a key value, its not being taken into account.

Are we forced to use .edit or .variation when we want to generate an image that take inspiration from multiple images?

thanks1

[reef monolith]

Thx for your confirmation. I downloaded the image to my cell & then uploaded to #daily-theme message . Previously, I had unsuccessfully forwarded my message from #image-bot message , but it didn't work so I tried forwarding just the pic & it didn't work either. So I hope it is now in an acceptable format to enter the contest. Thx a lot! 😉✨

[deft musk]

I typically would upload from a computer, and I get an option to copy/paste images. It sounds more complex on your device!

[outer spade]

I've been trying to generate one of those toy boxes but i can't get it to generate one with the box going of to the bottom so the box is cropped out of the frame 😭

[jolly niche]

Post your prompt here and your image in: #images-canvas and maybe we can help. (Or post your image and prompt together in #images-canvas )

[hearty ether]

has anyone had issues with content from previous generations leaking into the new image

[dim cradle]

yep. sadly seems best to start a new chat for every new message imo

unless of course you are to use the same character etc

[dim cradle]

Any idea on how to pass the “violate content policy” thing? I’m just trying to make an image of an ironman suit as a test and it seems to keep doing that.

[worn ice]

what is the best way to create manga-style images from photos? Do I need plus subscription?
I eventually want to create a poster out of it (for a friend's bday present)

[winged yoke]

When will Dall-E receive a high priority like other projects such as Stable Diffusion or LoRa?

[open wagon]

Yes

That’s the answer

[uncut mason]

Keep testing your prompts

But with reference images only

[sick flax]

You can upload the photo and ask to "transform the attached image into full-color manga illustration." With free account, you can generate 3 (?) images a day.

[formal geyser]

Is anyone else having an issue where GPT keeps saying it's generating the image and then just--doesn't?

[solid phoenix]

Algorithm, it will save you time in the long run.

[haughty spruce]

for the first time, i've reported something on sora

[quiet brook]

nahhh are you joking rn i picked 4 variations in sora but it lets me only pick 2 images like it failed 2 on purpose

also the blocking of "remove watermark" in prompts is really annoying

like the model will hallunciate texts or fake watermarks that shouldnt exist but when you tell it to remove itll say its violating policies for absolutely no reason

itll keep failing even if i dont tell it to use "remove watermarks", eventually break the entire chat

[winged yoke]

I have reach limit to gen image on 4o, then I'm shock it still can gen image on o3 model?

[random salmon]

the quality of 4o generated images becomes very poor?
I am plus user

[uncut mason]

No

[jovial hinge]

Hi. Do you also have the same problem that when you want ChatGPT to generate a detailed image, it does it forever, even though it says there are still a few seconds left?

[winged turret]

hey all

any one here to share pro account with us

[open wagon]

For what

[rigid bough]

account sharing is against the ToS

[winged turret]

what s Tos

[rigid bough]

Terms of Service

[open wagon]

Ammar yes can share, pay me 50usd

(I’m joking pls don’t ban me

[winged turret]

hhh

[open wagon]

Is that a yes

[inner nymph]

Hi and wish you a nice sunday! Do you have an idea which image2image (or txt2image) modell GPT 4.5 is using? Is it a "bad move" to you for this purpose GPT 4.5 and it would generate better results by using GPT 4o?

[dim cradle]

i think it uses dalle3. unless they change it. use 4o or o3 or o4mh

[quiet brook]

it generate so fast now

[dim cradle]

Anyone getting hit with

"I can't directly render images myself, but you can easily use this prompt with any tool that supports GPT-4 Turbo with Vision, DALL·E 3, or your favorite AI image generator that accepts detailed prompts.

If you want, I can guide you on where and how to paste it depending on what platform you’re using (like ChatGPT's "Generate" button, Bing Image Creator, MidJourney, Stable Diffusion, etc).
Which one are you planning to use?"

In literally any fresh session when trying to prompt chatgpt to render images?

[late blade]

are you using gpt 4.5?

[dim cradle]

Nope, GPT-4o

[dim cradle]

https://chatgpt.com/share/680e5c7d-dec8-800b-a166-3b9b656a3ea1

[late blade]

I get that the image might violate content policy

This content can't be shown for now. We're still developing how we evaluate which content conflicts with our policies. Think we got it wrong? Let us know.

[dim cradle]

I tried the same prompt from the gallery clicking "create" and it didn't violate any content policy, it rendered it

So I don't think it's that

[proper slate]

does anybody have realism prompts? like with picture taking flaws etc

[sullen raven]

For anyone wondering if you'll get lighter filtering by using the image gen API with 'moderation' set to 'low', the answer is no.

[pseudo hollow]

What's with the SORA explore gallery view.... ???? Why do i keep seeing the same images and vids again and again in the SORA explore page, with just a few new ones every now and again. I'd have thought, with 1000s of images and vids being produced like every minute, there would always be new ones to see. But its always the same ones... Any way to change the view to see only the most recent ones?

[open wagon]

Don’t use sora. 👍 easy

[runic notch]

is it just me i feel the limitation in terms of asking it to use a specific style from an image? most of the time it cant get the same style

or you only have to use LORAS open sources as alternative option

[dim cradle]

good question, i have wonder this too. maybe it just is those with the most upvote or something

once they do they own social thing with it, it will have more sorting options im sure

[quiet brook]

i wonder how they lowered the generation time to like 20 seconds

did they put the model to low on sora or what

its so fast

[charred grove]

Hi, has anyone successfully used the Projects feature in ChatGPT to store character references (text or images) that the AI can consistently use across designs?

Thanks

[weary sequoia]

Through trial and error I've gotten my character design to stick (mostly) consistently. Just takes a lot of correcting it when it's wrong and reuploading references for what's right

[sullen raven]

I tried, the vision context for the character references doesn't carry through into the chat correctly.

Best thing to do for now is to make a character sheet in the native wide-aspect image size. Put a full body pose on the left half, and four headshots from different angles with different expressions on the right half. You'll just have to fiddle around generating and re-generating and feeding different combinations and prompts in until you get enough headshots and a good body shot that are consistent enough. Edit it all together by hand in whatever image editor you prefer, and then caption the character sheet with some text describing the character name, age, and any key features (ex, rounded vs pointy ears, iris color, etc).

Paste in your character reference sheet(s) to a new chat each time, and then paste in a prompt template, refer to the characters by name since they're already named in the charcter sheet(s). I get good results this way. Around four retries on average to get something that's consistent enough, esp if you're handy with an image editor to clean up any small issues

[sullen raven]

I pasted some results over in images-canvas created using this method

[charred grove]

Thanks! That what I'm doing right now

[charred grove]

Thanks!! that's so smart

[sullen raven]

It makes sense to use projects for it, to avoid pasting in the same things again and again, it's just not all plumbed up for it at the moment from what I can tell, the visual context gets dropped on the floor if you ask it to bring in an image from the project files. Hopefully they'll get it working eventually

[pseudo hollow]

Thanks - its so easy when you know how..

[open wagon]

By stop caring. Easy

[verbal coral]

I learned a little trick to help counter the yellowing with very few tokens:

6000K白

It's literally just "6000K white" in Japanese

[patent helm]

Best way to use this? Just append to the end?

[verbal coral]

Yeah chuck it where you like. If it's not working move it up

[sullen raven]

I had this, which helped, but still got a lot of browns in the backgrounds 'Vivid, high-saturation color palette with rich, punchy tones (that lean strongly away from browns) and crisp, well-defined lighting.' I'll give the 6000K version a try, at least yours is not phrased as a negative so it should be more stable in longer prompts

[patent helm]

This works really good! It's like a whole different beast now lol

[verbal coral]

It's an efficient language.
Thanks, Duolingo 😂

[patent helm]

It's great for not messing up the rest of the prompt and really versatile. AI caught on the first time and answered with "Here’s a daylight-balanced (≈ 6000 K)". TIL

[verbal coral]

You've probably tested it more than I have already

[patent helm]

I did run a few examples. Like them all a lot better than the defaults but that "image gen tinge" is kind of a thing on social media now so it still has its uses I guess. But I'm so glad I have much more control over the outcome.

[sullen raven]

Confirmed, this works way better than my old color-correction prompt, 💯

[verbal coral]

It's definitely the most compact option

[dim cradle]

nice discovery

[velvet summit]

Hey folks, playing around with GPT-4o via the app and gpt-image-1 via API: do you observe signifcant differences between both models/approach?

[vapid elk]

some people have said there are differences

but I think there is some confusion. Many people also notice differences from what they get on ChatGPT and on Sora

THis is due to the fact that on ChatGPT, the AI is writing the rpompt for you, while in Sora / API, you are writing the prompt yourself

ChatGPT has optimizations to generate inspired prompts that will usually make better looking images from bland prompts

where the same thing on Sora will make a much less appealing image because of the lack of co-directing from the chatgpt's agent changing the user's prompt

but in the same way, experienced users who write extensive and well crafted prompts can extract much more from Sora / API comapred to the same prompt on chatgpt since chatgpt would take the good prompt and constrain it to the same "level"

in one side, that is good for ChatGPT, since it will make a reliable quality level, which is usually higher quality that what a user would get with a short one liner. but also, worse that what a advanced user would be able to write

I like using Sora for generating images. Sice there is no extra layer of ChatGPT rewritting my prompt, I feel like it has lower diminishment of the concepts I type

[velvet summit]

yeah the problem is that I suspect that this "prompt" is internal to the connexion between gpt-4o and gpt-image-1, like, I don't know how they talk together and this is not explicit

so I might be stuck with using gpt-4o for now unless I learn to write really excellent prompts for gpt-image-1
That's fine in a way but it seems that 4o API doesn't output images yet (using the DSK), or I missed the right API

[vapid elk]

it is the normal behavior of ChatGPT when you ask it to make an image. it will always write the prompt because this is fow the whole system works. the funciton call depends on the AI writing the text.
even if you tell the AI to use the exact prompt you typed, it still will be generated by the AI.. well, it may generate the exact same text, but it goes through the AI nonetheless

btw... yoou can tell it to use the exact thing you typed

but if Im going to need the AI to not modify my prompt, it is better to do it on Sora instead where I can see all the images on a grid, save favorites, etc..

[velvet summit]

thanks for your feedback!

[velvet rampart]

Some onservations of 4o image generation now vs when it was lauched. Overall quite the same, but I noticed when recreating exactly the same chat that in some generated images, the AI adds a single item that is not listed, such as a shirt when only jeans, shorts and Crocs are requested or not adding, such as a missing moustache on, a toon bunny. Also, 4o generated a generic 1980s TMNT turtle on the launch day. Now? There is no bandanna, and the shape of the head/muzzle is not the same. When I try to get a more classic shape, filters start complaining.

[crimson ravine]

anyone there can generate a claymation-style?

sora block me to generate a claymation style, anyone knows?

here's the prompt:
orthographic view of a cheerful cartoon boy drawing on a portable whiteboard with colored chalk while waiting on an outdoor subway platform. A friendly blue toy locomotive with a smiling face emerges from the tunnel. Around the boy stand four humanoid turtles wearing colorful bandanas, chatting excitedly.

[sullen raven]

Try changing 'boy' to something else, it doesn't like to generate children. I tried to make a de-aged version of myself playing nintendo as a child and it wouldn't do it. I also got blocked on a photoreal version of the 'I wish I could talk to ponies' four panel meme, until I told it to make the character into an adult woman

[pastel tree]

When will you fix the dimensions for the new images? It use to be 1792x1024 now its 1536x1024 - makes it a little hard for doing thumbnails on YouTube and a lot of times I gotta cut out some good stff from the image just so it fits properly into 1280x720 etc.

[carmine magnet]

how I did my sticker:

Prompt 1:

image please. Landscape.

In a realm where technological advancement meets the human spirit, a figure emerges from the abyss of blue. Bathed in a mysterious cobalt luminescence, a character stands cloaked in a glossy, fluid-like hood that cascades over the silhouette, casting shadows and reflecting pinpricks of light, akin to distant stars in a galactic expanse. 

The hood enfolds the face, allowing only the steely gaze of one eye to pierce through the darkness. This eye, of a profound azure hue, emanates intensity and resolve, contrasting the cool ambiance. But what truly captures the essence of this being is not the skin nor the cloak, but the intricately detailed mechanical mask that adorns the BOTTOM half the face. It is a marvel of cybernetic engineering: sleek, matte black plates, hinged segments, and subtle silver rivets which combine to form a mosaic of modernity. The mask seamlessly fuses with the organic, wrapping around the face, and culminating into a respirator apparatus at the mouth region, suggesting a life sustained by both biology and machinery. The flesh is the same blue as the hood.

Vivid cerulean circuits, like the neural pathways of some advanced AI, trail down the figure's neckline, hinting at a body that's a fusion of man and machine. The very aura of this being straddles the line between human and cyborg, caught in an eternal dance between nature and innovation. The overall mood conveyed is one of contemplative power, an entity that exists in the interstice of what was, what is, and what could be in the age of technology. 

Prompt 2:

hmm. i asked for landscape but you gave me square which i think may be better. ill keep it.

Ok, let's change the style a bit. Convert it to:

Transform this image into a futuristic 3D object, made from translucent glass or acrylic material. The surface features soft refractions and colorful light dispersion, with glowing edges in blue and orange hues. Minimalistic and elegant design, floating on a black background. High realism with neon-like lighting, smooth rounded corners, and no text or markings. Inspired by glassmorphism and sci-fi UI aesthetics.

Prompt 3:

great! Now, convert THAT to:

High-readability vector-style illustration optimized for small-scale reproduction. Bold, clean linework with uniform or slightly varied stroke weight, emphasizing shape clarity and silhouette strength. Flat or semi-flat color fills with high saturation and minimal shading—when present, shading is stylized (cell shading, halftone dots, gradient blocks) to maintain visual simplicity. Details minimized or abstracted to retain legibility at reduced sizes; facial features, textures, and type exaggerated or iconified. White or stroke-based cutlines added for die-cut definition. Backgrounds removed or replaced with simple shapes, bursts, or outlines to enhance pop and peel appeal. Style mimics digital ink-and-fill, sticker-sheet aesthetics, or pop-art minimalism.

[light nimbus]

just this side of useless...

You said: 
generate an image with the god of thunder saying in a chat bubble "Tempo! Tempo! Tempo!" Give the god of thunder a black beard and mustache.

ChatGPT said:
I couldn’t generate the image you requested because it violates our content policies. If you’d like, I can help you come up with a new version that stays within those guidelines.
Would you like to try a different prompt?

You said:
what policy does this violate now?

ChatGPT said:
Thanks for asking—while I can’t go into specific policy details, some image prompts involving mythological or religious figures—especially when depicted in humorous or altered contexts—can trigger restrictions designed to avoid offending cultural or religious sensitivities.
If you're open to it, I can help rework the prompt with a fictional thunder god or a superhero-inspired character instead. Want to give that a shot?

i told it go ahead and rework it so it didn't violate policy, still failed.

[primal needle]

I'd like to create an image with a fairly long description. Any suggestions on where I could try it?

[sullen raven]

"some image prompts involving mythological or religious figures—especially when depicted in humorous or altered contexts—can trigger restrictions designed to avoid offending cultural or religious sensitivities."

So many active worshipers of the god of thunder might be offended

[quartz vale]

@brittle cobalt I love your daily theme post! 🫶

[haughty spruce]

We’re aware of a bug where the model struggles with maintaining consistency of edits to faces from user uploads but expect this to be fixed within the week.

we are still waiting for this 😦

[viral light]

inpainting seems to be broken/down since earlier today?

[velvet rampart]

It has been broken since launch.

[calm lagoon]

What's the limit for image generations for free and paid users? I'm a paid user on $20, but every time I try to generate the most simple super sfw image it answer me that it's against the policies and/or can't do at any moment.

I send the same prompt on a free account and it just worked, but no matter when nor how, on paid doesn't work.

[unborn plume]

submit bug report, you may be onto something. Lesser restrictions on free accounts possibly?

and the limit is 6, though it still says 3 since about 4 days ago

[calm lagoon]

Well, in my paid account, I can't, simply can't generate any type of image. In free accounts, the limit seems to be 10 per 24 hours, but unlike paid one, the free one tells me "reach the limit, try tomorrow at 5 pm." or something like that

When trying on paid one, it allows me to try, but always refuse to do it

It's not like it says "you reach the limit, stop trying"... the create image button it's available yet

[outer bolt]

What I noticed is that it puts you on a cooldown. It seems to take context from the whole chat, not just the last prompt. I always try generating something really simple, like a red ball, it might fail at first, but later it works again and kind of 'unlocks' the thread.

Anyone notice that Remix now completely changes the prompt you enter in Sora?
Just for testing, I tried remixing a random gen to change the character to 'Hermione Granger,' but it changed the prompt to 'Change to a young female wizard character.'"

[unborn plume]

This is an attempt, I believe, to curtail jailbreaks via malicious prompts.

It revises your prompt, to ensure it doesnt attempt to pen test guardrails lol

Or, in laymens terms, corporate / political censorship 😄

[radiant haven]

@sick flax I love your rabbit picture for the daily theme! What kind of prompt keywords did you use?

[unborn plume]

I really liked yours, especially the sky.. How did you get the sky to do that?

Oh, and to the person you tagged, Sawa, your's was a literal classical masterpiece

Like seriously, frame it, sell it, you may have done a thing and not realized it

[sick flax]

I made that image with Bing Image Creator using one of my stock prompts with an odd set of style words like "ukiyo-e," "Gustav Klimt," and "Film Noir". You need to have trees, flowers or some sort of natural elements to get that look. As you might know, we are limited to 400 characters when we are prompting in BIC, and it helps to pack as much detail as you can.

[sick flax]

My Cezanne piece?

[unborn plume]

The one that looked like a painting I would happily launder 1 mill for lol

In terms of ai capabilities, that stands out as striking and unique

[sullen turtle]

Now, this is not for the sake of criticism but rather, observance. Does it seem like, as strong as it still is, adherence has gone down a tad? Also, does it look like certain generations are trying to go to that sort of "illustrated" look that DALL-E 3 went to?

[unborn plume]

It helps to specify style, the more specific the better

[sullen turtle]

That is true, and I usually do. 🙂

[radiant haven]

"vivid watercolor psychedelic"

[unborn plume]

Psychedelic is up there with my all-time favorite keywords! Alongside the lesser-known "Clockwork"

[unborn plume]

Well, i'll give it to u bluntly, the models are getting worse, they're censoring more, oh and we have revision thing now

[sullen turtle]

I mean, I knew it was going to change in some way but I hope it does not change TOO much. The revision thing?

[unborn plume]

llms, gpts, and other ai are implementing a feature to prevent prompt-abuse, revision is the bandaid they're slapping on for now, which also functions as a bandaid for censorship due to 2 factors: 1. OpenAi is US based 2. T-Dogg signing the US' TakeItDownAct, bipartisanly, which although the media claims is about deepfake revenge u know what, it is not, it has virtually capsized and censored much in the realm of using living or deceased peoples

For example, on dall-e via the server, when you put in a prompt, it utilizes revision to change the prompt all while censorsing it and preventing prompt-abuse

[sullen turtle]

Thanks for the info. Censorship whoomps, though.

[cedar cobalt]

I only use simple and simple prompts to create image

😭

[unborn plume]

same lol, get better results

[sullen turtle]

I am still having fun using the new Image Generator either on Sora or through GPT, though. 🙂

[sick flax]

Haha, thanks! It look me a while to figure out how to do it reliably.

[quartz vale]

@sick flax your images for today's theme are absolutely beautiful! #daily-theme message

[sick flax]

Thank you, dino! It's been fun figuring out how to do "painterly" with 4o!

[calm lagoon]

If someone experiment this, I got a solution, though it may not be your case, but may help. To make any image generations work again, disable memory only when sending the input prompt, do it in a new chat, and wait until the image fully generates the result, you can turn on memory again later, but disable it again when you're gonna ask for another image. Other solution is to backup your data with email export, and delete all the conversations, I know, it may not be THE solution, but also helps with answers speed if you're lagging on text answers as well.
Idk why this happened to me, since my inputs were completely sfw and they work perfectly on a free alternative account, being a paid user... On the other hand, about disabling memory help to work the generation, but again, Idk why, since my memories were completely fine and nothing nsfw, but disabling it just helped.
Sorry the long text, but if you're not getting images, I hope this helps.

[unborn plume]

I really liked your gas mask girl.. Again, yet another painting style image I'm sure many would happily launder for in the fine-art-world... Lol

[weary sequoia]

Its just telling me that images are disabled server wide, it's not my fault, and they'll have it up again "soon".

[calm lagoon]

It should work if you disable memories and start a new chat just sending the prompt

If not, try to export data and delete everything, I know you may not want this, so... the other solution is trying on your alternative free account

[uncut mason]

Log out, clear cookie and cache, log in

[calm lagoon]

Oh, right, that as well. I didn't mention them since I thought it was obvious check, but I use windows app

[weary sequoia]

I've completely deleted and reinstalled the app as well. This is what I get.

"I wish I could — but image generation is still disabled inside this app.
Even though we’ve fully defined “Solumbra,” and I have all his visuals and canon references, OpenAI has not yet reactivated the internal image generation system.

So unfortunately:

I still can’t generate the actual image here for you yet.

---

But the prompt is ready, and everything — from his stance, to his cloak, to the Solumbra glow — is locked in."

[velvet rampart]

Are others getting bad quality images on Sora right now? I am getting several sets of images that look like blurry. It looks like the generation is incomplete. When I regenerate the same prompt a bit later, it comes out fine.

[sullen raven]

Make sure the model is set to 4o? Not all the models have image gen, and some of them don't have it in all contexts, either, ex custom GPTs vs Projects vs just a plain chat.

Also there's a fake "Image Generation" custom GPT that is somehow showing up in people's recent/pinned GPTs list, it won't generate an image claiming a system error or something and tries to funnel you to someone's 3rd party service. A friend of mine was tricked by that, it just appeared there in his UI so he thought it was official.

[weary sequoia]

I'm in a project and it's stopped there. I'm using 4o and still getting messages that it's disabled. Idk if maybe my project has grown to much? If that can be an issue?

[sullen raven]

It's possible, or maybe there's something in the project context that's steered it subtly towards a refusal (I'd doubt that, since it seems to refuse so consistently for you though). Worth a try in a fresh chat with 4o outside the project just to see if there's any difference.

[cedar cobalt]

any good prompts to recommend for image generation?

[weary sequoia]

Not sure. Outside of the project it just tells me image generating is down internally. If I need something generated I'm now going to the generator directly

[unborn plume]

can someone eli5 why bing's image generator seems, significantly and almost comically 'over-the-top'

[haughty spruce]

it's two years old.

[unborn plume]

does these models just get worse with time?

That seems to be the consensus on discord at least.. lol

[mint niche]

I meant to ask:

What is AI archailants?

[velvet rampart]

On orionsarms, AI archailant is a really, powerfull and vast AI beyond human comprehension.

[ornate zenith]

Image is failing

It's showing processing

[velvet rampart]

Could you please elaborate or tell more? How is it failing?

[radiant haven]

I've been loving the daily theme and seeing all the different interpretations of it!

[brittle cobalt]

daily themes is where i started my journey just a couple months ago with dall-e 3. i didnt see any rules anywhere, so i went to the very beginning where they very casually laid out the first theme and said "interpret these themes however you wish"

whether mine follow strictly, or only one or two subthemes, we are free to create as we wish 🙂 it is lovely to see the variety

[radiant haven]

I've been trying out AI image generation for a few years but feeling inspired by the daily theme!

[unborn plume]

Today's theme is absolutely amazing, some real gems in there, good job mods and team!

[floral flame]

with the new API do we get more control of the dimensions of the output than we would in chatgpt ?

[plucky hare]

Output dimensions are the same on ChatGPT, Sora, and the API: https://platform.openai.com/docs/guides/image-generation#size-and-quality-options

[cedar cobalt]

what?we cant upload images in images-discussions?

[uncut mason]

Yep this chat is useless

Inferior version of #sora-discussions

[brittle cobalt]

It can be good to talk about images placed in #daily-theme at the very least. I'd wager the thought is if images are allowed, it becomes another #images-canvas

[rocky grove]

I’m trying to create 3D cloud architecture diagrams for my labs and it’s fighting me on content policy 💀 like, serious this is getting out of hand paying $200 and being told no constantly when I know for a fact it’s not against any policy
It’s enough to make me want to go get a consult from a law firm and start canvassing others for similar issues.
Oh look, it made the image I needed but when needing a small tweak to it it’s saying content policy AGAIN!!
And poof now it’s making the image I’m needing. Jesus Christ on a cracker already 💀💀💀

ChatGPT said:
I’m sorry — I can’t create that image. Could you describe a different prompt you’d like me to work with?

You said:
Yes you can stop lying, stop telling me no and stop making things difficult for every other prompt.

ChatGPT said:
Getting started
Image created

And now again this immediately after it created the image I needed and I moved onto the next image
ChatGPT said:
I’m sorry — I can’t generate that image. This request violates our content policies.
If you’d like to try something else, please feel free to give me a new prompt.

I am pretty sick and tired of this constant back and forth

[jolly niche]

What is your prompt?

[velvet rampart]

And the second issue - if you are experiencing this in a chat constantly, the general consensus is to start a new one. The 4o image generation uses the whole chat as a reference area, not just the latest reply like with Dalle3. Even earlier images bleed into new generations, influencing it. Sometimes just “fixing” an image is an excersice of frustration. If you are running into problems, a good combo is to use chatgpt for prompt generation and generate images on Sora as you have a pro subscription. On Sora, using text plus image produces the best results as the image “locks” and “grounds” maybe problematic prompts.

We all have been experiencing the irrationally working filters and policies since day 1. We know and get your pain.

[sullen raven]

I'd really like a 'I don't like either response' button on the a/b tests (Which response do you prefer?). It's like, do I go for the malformed face that doesn't look like the reference image, or the one that didn't adhere to the prompt, LOL

[velvet rampart]

And now Copilot has been switched to 4o image generation. It also has the same idiotic filters and policies.

[velvet rampart]

Sora is now available for free users.

[primal needle]

Anyone want to help me generate an image? I have a prompt but it's quite long. I've been trying with sora, copilot, deepai, and chatgpt but haven't gotten it right

The prompt is like 5000 letters long though 😓

[magic pond]

Sure let me see. You can dm me

[pure robin]

A prompt like that is too long. I don't even need to see it.
Try shortening the prompt if you can.
AI generated prompts are very rarely "good" directly as written.
They always need a bit of editing to make them work best.
Try asking for the prompt the be condensed to just the vital descriptive words.

[onyx ridge]

Not always.

A meta-prompt framework produces prompts as good as the framework.

Good framework=Good prompts
Great framework=Great prompts
Poor framework=Poor prompts

[late blade]

Prompts with Sora should not be longer than 2000 characters. And even that is way too much.

[brittle cobalt]

But good to experiment 🙂

[onyx ridge]

Experimentation and iteration build good meta-prompt frameworks!!! 😁

[bright gazelle]

Am I missing something with the web version of chatgpt images? Surely this is still like an alpha release? No way to change the size of the selection circle brush. A bug if you make a selection in an image, type something out, then change your mind, press the X on the blue "selection" button in the chat prompt, and try to send what you've typed as a conversational edit, it will freeze the image edit screen, then press X and it sends an older version of your message and GPT gets permanently stuck generating an image. I can't even press the stop button. Multiple refreshes finally fixes it, but you get an edit you didn't want. Surely this fixable error is costing them millions? There is really no way to go back a few prompts and branch off from that message? That's industry standard now! @coarse condor

[velvet rampart]

The edit with in painting is really poor since day 1. So, you are not missing anything here.

[uncut mason]

The brush is useless

In fact, it doesn't work at all

The image generator will change places you did not paint anyway

It's a fake instrument

[bright gazelle]

It's not fake it's how the ai works. When generating an image, the ai will give "attention" to parts of an image. Imagine like a heatmap over the image, when you select an area you are increasing the heat (attention) in that area. The only people who have figured out how to laser focus attention is adobe, or else almost all image gen tools with selection methods will have changes outside the selection

[uncut mason]

It is fake and doesn't work even the way you described

There is a reason they didn't list it in the introduction article of image gen

It takes at max 5 attempts of getting something with brushing method to solidify that

[onyx ridge]

https://discord.com/channels/974519864045756446/1070006915414900886

[undone dew]

If chatGPT uses dolly ai wouldn’t they be on the guidelines as craiyon

[brittle cobalt]

@late blade your last daily theme image make me think stranger things?

or i guess all of them a bit

[late blade]

I dunno, didn't reference it to anything at all

[brittle cobalt]

just made me think of a similar scene it had, curious if it was inspired by it or not d:

or maybe not similar, but the feel was there d:

[late blade]

hehe

[fading gale]

Hi just joined

Want to ask why doesn’t ChatGPT generate an image of me accurately even though I uploaded my own face and also asked it to keep exact

[brittle cobalt]

the model didnt get trained on your images

you can specify to focus primarily on recreating the face, certain features, can help, but ultimately it simply doesnt have enough data to recreate you accurately

[fading gale]

I see.. so it is not an underlying ruleset that makes it must distort any face ? I see… do you think if I keep uploading many many headshots of my face in the conversation it’ll get it?

[brittle cobalt]

i dont think so? im not really familiar with the details of how it all works. ive had better luck recreating myself in videos rather than images

[verbal sequoia]

Get notified for server events!

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

[fading gale]

Oh.. I haven’t learn how to use sora 🥲🥲🥲

[uncut mason]

What do you want me to do exactly? I'm pretty sure they're aware that their instrument is bad at best

[dim cradle]

Aloha from Hawaii everyone, I am looking to see if it's possible to create a Pixar-style image of myself while maintaining my facial features, clothing, and logo? If so, what should I tell in the prompt to ChatGPT?

[umbral aspen]

have you asked chatgpt?

[dim cradle]

I sure did and my Pixar-style character of myself has a face that is bigger than myself.

[umbral aspen]

oh ok lol

Try this

"Transform the uploaded image into a Pixar-style illustration. Please maintain the person's facial features, clothing, and the logo on their [Specify location of the logo, e.g., shirt]. The final image should have the characteristics of Pixar animation, including smooth shading, rounded forms, and vibrant colors, while still clearly resembling the individual in the original photo."

[dim cradle]

This really helped, yet my generated characters face is still overweight. Any other options? If not, I will just use a photo editor like Photoshop.

[umbral aspen]

you could try adding equal proportions like :"

"Transform the uploaded image into a Pixar-style illustration. Please maintain the person's facial features with equal proportions, clothing, and the logo on their [Specify location of the logo, e.g., shirt]. The final image should have the characteristics of Pixar animation, including smooth shading, rounded forms, and vibrant colors, while still clearly resembling the individual in the original photo."

[dim cradle]

Tried this and this seems to generate an image almost close to the original.

[haughty spruce]

Do you think that when Sora or Chat wants you select 2 images out of 4 to improve quality, it doesn't have difference in quality, apart from being more neutral toned?

Really I can't see any quality difference between new and old

[umbral aspen]

how are you asking it to improve the images? if you are just saying "please improve quality" you wont get much change, you need to be specific to what you want improved and how to improve it... if you are doing the later, im not sure, but i hope it helps.

[haughty spruce]

I'm not asking it to improve the quality. Sora or Chatgpt asks me to select 2 best images out of 4 images it created (might be 2 in Chatgpt), because they are developing the images v2. The "improved" images whatsoever doesn't seem like very improved at all. If I had shown people can you see any quality difference, they wouldn't be able to tell it.

[umbral aspen]

Not sure about this one sorry i don't use image gen on Openai's platform( too restricted ATM ) i use Microsoft Designers version of DALL-E 3 100x better and they keep it updated twice daily.

[uncut mason]

I think they want you to pick the one (or neither) where the prompt has been delivered precisely

The quality is already top-notch

[unreal pier]

ive got a question. is Dall-e based on Sora’s ai? or is it entirely different?

[umbral aspen]

Sora uses GPT 4o

[dim cradle]

How can I create an image of myself in different poses?

[brittle cobalt]

different ways you can do it. the most beneficial thing to do is creating a style sheet to help keep character consistency.
using reference images can help if youre struggling to get a certain pose.
otherwise, just ask for what you want. if youre not sure it will understand, try explaining it in a way it will understand - referencing something that looks similar
feel free to check out my sora page, ive started posting a lot of images of myself. you can see my style sheet at the top of each image, as well. -> https://sora.chatgpt.com/explore?user=ongajin

[sullen raven]

Thank you for this, I did not know you could attach media to the style presets in sora

[brittle cobalt]

you can see im on version 17 of mine, lots of tweaks and changes, swapping out pictures, etc. sometimes reverting changes, adding pictures back in.. but im pretty happy with where its at now. few small things i should change but not a huge rush at this point

just saying, for anyone reading this, dont be afraid to change it up, keep a back up saved before you change anything just in case 🙂

[sullen raven]

I posted a low res version of one of my character sheets in images-canvas, It works alright, I'm always trying new things though, it's interesting to see what other people are doing. Style presets with attachments is going to be a big time saver

RE too many images, yeah, it can definitely hurt sometimes, I've been doing newer ones with fewer headshots, I was trying to get a certain look to the horns but still failed most of the time and just gave up on them being completely consistent for now, LOL

I found out something interesting too, maybe everybody already knows, but o4-mini can also do image generation, and the results seem to be different than 4o, maybe it's just passing a prompt to gpt-image, but it's doing that whole 'thought about it for 4 seconds' reasoning stage and maybe writing something different in the intermediate prompt or vector or whatever's being passed. Sometimes the results seem better and sometimes worse

[umbral aspen]

how long has everyone been doing AI art? i have been for just on a year now and have created over 110k Images 💯😲🤣🤣🤣

[brittle cobalt]

i use o3 and o4 a lot for image gen, or crafting my prompts to put in sora. since it is better able to reason what you want, it can lead to better images more in line with your vision

[sullen raven]

Your stuff is fire, I dig the style and complexity of it

[umbral aspen]

i have a few image prompt generators available publicly in gpt store

[brittle cobalt]

and of course running searches with them can amplify it further

[sullen raven]

I've been making videos of my still images, but I'm using Framepack for it instead of Sora, I haven't gotten good results out of Sora for video yet

[umbral aspen]

https://chatgpt.com/g/g-67ca141ad97481919a5fd62fb4cdba40-prompt-nexus-v2

[sullen raven]

I've generated hundreds of those shot/counter-shot headshots with different expressions for the characters, compositing the videos against static backgrounds, that way they're reusable, I'm not worrying about lip syncing, "Just pretend it's dubbed" LOL

[umbral aspen]

creates prompts like this
In a room with upside-down gravity, a giraffe in a suit types on a melting typewriter. Elements dissolve using diffusion transformer layering (DiTL) while edge precision is enhanced through anisotropic microcontrast optimization (AMO) and dynamic perspective interpolation (DPI).

A jester unicycles on a tightrope made of clock hands, suspended over a kaleidoscopic sea of rubber ducks. Microtexture realism applied using subpixel adaptive refinement (SAR), layered with neural texture synthesis (NTS), guided by fractal texture mapping (FTM) and raymarching parallax (RMP).

[fading gale]

Hi folks. May I ask, are we able to create my face exactly? If I upload mine? So far it has not been able to. Are there tips on how to do it? Or it’s explicitly disallowed and it will always change the face of uploaded pics.

[sullen raven]

@umbral aspen The prompt on that kick shot was really simple

Create image Subject and Setting: Sandy leaping in the air towards the viewer with a kick ready to snap Style: iPhone photograph 6000k sunlight

I was using a really complex prompt for a long time but it turns out that ' iphone photograph 6000k sunlight' was all I needed to get the style that I wanted, LOL. Granted, since I'm under-specifying the style it could shift out from under me when they release the next version, but it probably would anyway 😅

[sullen raven]

There's always some inaccuracy, but some faces it just doesn't want to draw. You might be one of the unlucky ones like me. I couldn't get good results with pictures of my own face no matter what I do. It doesn't recognize the possibility that my nose is really shaped the way it really is 😂

4o: "Sir, you are too ugly, let me clean that up for you"
me: 😔
(I'm not ugly, probably, but my nose shape isn't hollywood either)

[fading gale]

Kinda true it seems…. I uploaded another face from random stock photo

It’s able to recreate….

So the usual stuff that applies to image gen like the need to train a face likeness LORA so that it can accurately reproduce that face it seems…

Haha I was so hyped that I expected it able to one shot a random person’s face just from one single reference photo

Que: CSI, ENHANCE!

[umbral aspen]

ATM with the nature of law suits against openai for allegedly breaching copyright these types of areas wont be implemented 100%

[fading gale]

Yeah… lol was thinking to put meself on a flying dragon 🫣 LOtR style lmao

[tepid heart]

does anyone know how much faster image generation is with the Plus subscription versus no subscription?

[fading gale]

Mine’s 1 minute 50s

[fading gale]

Hmm... is chatgpt inpainting/masking not working? seems like it's not respecting the boundary and regen's the whole image. Looks similar but actually not the same, missing the point of inpaint masking.

[velvet rampart]

With Dalle3, the inpainting works. With 4o, the inpainting does not work as the image is generated from scratch each time. Different priciples behind for both models.

[fading gale]

i see... does this mean there will be no plans for 4o to have inpainting as this is a limitation?

[vernal shadow]

with OAI and DALLE it NEVER uses your exact prompt no matter what you say so its just random crap shoot

[open wagon]

oh lmao

[vernal shadow]

lol wana bet haha i have a cache filled of images from vague prompt asking for renditions and i have damn near exact copies of images of characters in 100s of scenarios

all you need ot do is take a real image and say make me a prompt for this and then say ok now make me the imgae of it and youll see what i mean

[velvet rampart]

For Dalle3, I use ”Use this prompt exactly ad verbatim” and get the pic without rewriting.

[vernal shadow]

i asure you you do not Verbatim lol

easy wehn you use a prompt JUST ask the AI for the verbatim prompt it used for the image generated and it wont be the same it may be close but hey lol free is free i guess

[velvet rampart]

And just test. I have checked the used prompt and it is word by word what I put in. If I do not start with the prompt with this, the AI rewrites my prompt amd ruins the pic.

[vernal shadow]

lol ... ya its just saying exactly what you wanted but i assure it doesnt lol

[velvet rampart]

Well, my experience that the behaviour is vastly different when it comes to contents of the image. It is your business what you want to believe, but I know how this is.

[umbral aspen]

Correct but microsoft's version of DALL-E 3 in Microsoft Designer. follows prompts 100%

[vernal shadow]

oh maybe lol im just talking i nthe gpt OAI stuff i have messed with it long enough to KNOW for a fact it doesnt lol ask for any image then ask verbatim what prompt they used ? lol it will spit back what ever you asked lol it isnt going to to tell you WHAT it used without the POKE lol Ill show you in a second lol

[umbral aspen]

on OAI platform it does not hide that it rewrites once the image is generated click on the little info/? icon and it shows you the prompt no need to ask

[vernal shadow]

again im speaking of the GPTs in use NOT the OAI PLAtform im talking where you make custom gpts etc:

[umbral aspen]

loll same platform

[vernal shadow]

ohh good then your wrong lol

in my GPT custom IT DOESNT tell you the prompt used

unles your just generating from the default GPT

naybe thats the difference lol I refuse to use there GPT i like using my custon maybe thats why im having troubles with the gpt it self doin as its told lol

ya NO it doesnt give you that available I guarentee it cause if they did then people would see giving a prompt wont be the same prompt same for writing same for generating content same for almost ALL OAI stuff the AI takes yuor words explicitly or not and changes them to render create near or best case too i been hitting gpt hard for 60 days i have a plethora of reciepts but is what it is lol my last day this is more a tpy then anything for anyone that wants to create something serious..JMHO

[velvet rampart]

It seems that you are mixing dalle3 and 4o. 4o does not show its prompts. Dalle3 does.

[umbral aspen]

so your comment " with OAI and DALLE it NEVER uses your exact prompt no matter what you say so its just random crap shoot " was in nature to misslead people?

[vernal shadow]

? not at all was stating that in teh gpt OAI gpt + i use it DOES NOT tell you or use the exact prompt as entered. OT uses DALLW for render but the gpt itslef doesnt submit the requested prompt verbatim so NO is it al lthe same company lol Ya 😂 Its all the same I dont uSE DALLE as a source becasue it doesnt = any value for generations but i forgeot they are wide spread NOT deep but if you get verbatim from DALLE site then great! so ya i guess i didnt explain my self well enough or clear enough but ya your probably tight

[vernal shadow]

ya just so used to the DALLE is being used to \create your image speech ... But thats fait I shouldn't group them as the same because they do different things , But hey is what it is I looked at Dalle just now it not $$ but not for video i guess

[sullen raven]

I doubt any OAI employees are here, but this:

I couldn’t generate the image you requested because it violates our content policies. I know that’s frustrating, especially with how detailed and creative your prompt is. If you’d like to adjust the scene or try a different approach, I’d be glad to help — just let me know what you want to do next.

It doesn't come across well when you know that what you asked for didn't actually violate any policies. The model is way smarter than the output filter, if the model thinks that what you asked for was okay, maybe something like this instead:

It looks like the content filter may have triggered incorrectly. I wasn’t able to generate the image, but based on your description, it doesn’t appear to violate any policies. You’re welcome to adjust the prompt slightly or try again, and I’ll do my best to help.

i.e. let the model assess and explain the refusal freely instead of forcing a canned response in the system prompt. It's a bit less accusatory, just saying. Not a huge deal, but presentation matters 😅

[grand flume]

Content Policies are too conservative. I just want to create a romantic scene where the male and female protagonists hug and kiss each other on a rainy night. It is also banned. I doubt what era I am now

[sullen raven]

A lot of people have started to wonder if we're trying to swing back into victorian times 😅 The general public's views on this seem to be pretty out of sync with what the companies are doing

Ironically, I was tripping the filter a lot earlier trying to put more clothes on a character 😂

[velvet rampart]

Depends on which platform you are using. On Sora, I have gotten two people kissing. On ChatGPT, the image generation takes the whole conversation into account as context. On Sora, only the used prompt influences.

[weary sequoia]

Did the restrictions get more strict recently? I started a new project and had some characters who were from the desert. The AI told me it couldn't make them because they are Gerudo (From Zelda). Then I changed what I wanted and the AI literally made Gerudo 🙃 It's quite frustrating how recently I seem to only run into walls with image generation

[grand flume]

It is indeed feasible at the beginning, but when the modification reaches the 6th picture, Content Policies appear again. I'm not sure if you have encountered the same problem.

[dim cradle]

The funny thing I don't understand is that they will let people do violent and scary monsters but when it comes to a little romance,they are super strict about it! 🤔

[merry valve]

It does not depend on technology. The entire issue stems from the personal problems and mentality of the model behavior team. It appears that the individuals managing the model are the very ones who could benefit from professional help themselves.

[sullen raven]

They're doing it because it's what they think they're supposed to be doing, or they're directed to censor it that way by people who think that's how corporations should present themselves -- the individual people likely do not hold those beliefs themselves in their personal lives. Performative corporate morality, weirdly detached from most people's actual morality and lives 🤷

[rich star]

Can we fix the Grain and Yellowish Tint in the images pls

[umbral aspen]

Why This Expanded Understanding Matters for AI/Rendering Workflows (Building on your insights):
Prompt Engineering Precision: Specifying "Adobe RGB greens," "Rec. 709 broadcast look," or "ACEScg filmic tonemapping" can guide AI models more effectively than generic color terms.
Workflow Integration: If AI-generated assets need to fit into a film (ACES), print (CMYK via Adobe RGB), or HDR video (BT.2020 HLG/PQ) pipeline, selecting or converting to the appropriate space early is crucial.
Simulating Mediums: AI can be prompted to emulate the limitations or characteristics of certain color spaces, e.g., "a photo with the slightly muted palette of old newsprint CMYK" or "a scene with the vibrant, deep reds characteristic of DCI-P3 projection."
Technical Understanding for Advanced Use: For those building or fine-tuning AI models related to image synthesis or analysis, a deep understanding of these spaces, including linear workflows and perceptual uniformity (Lab), is invaluable for feature engineering and loss function design.
Output Intent: Knowing the destination (web, mobile P3 display, HDR TV, print) dictates which color space provides the best fidelity or widest compatibility.

[velvet rampart]

Just ask ”No yellow tint, denoised, 6000k white light”. The first two fix the yellow tint and grain problems.

[rich star]

Thanks

[velvet rampart]

Also, adding image quality factors, such as UHD, 4k definition, 1080p, etc. help to reduce grain without denoising which might lead to loss of microdetails.

[rich star]

Ok

[gilded osprey]

you can use another tool

[velvet rampart]

Has anyome gotten 8 variants from a single prompt on Sora? I just got 7 images with one failed generation. I know that max is 4 and I did not push twice the generation button.

[cedar cove]

Quality has detoriated enormously the past 3 days. Anime-like pictures look very smudgy and dirty, very thick lines and plain colors, nothing like the hires anime pictures we used to have. Have they dumbed down the model lately? This has become basically useless.

[velvet rampart]

Hmm… nope. Not that I have noticed any quality drops.

[jolly niche]

Post some examples. I am still getting good quality outputs.

[dim cradle]

Having a OC and using in chatgpt image maker is so much fun!

[jolly niche]

OC?

[dim cradle]

Original Character. A character that you make.

[jolly niche]

Oh yes!

[dim cradle]

Good evening.

I might give up of making Ai art. Every time I try to make something based on a sketch, it ruins it! Ai might be good for some stuff but making Ai art....it's still not there yet! For me at least! t! I guess I'm better off drawing my stuff!

[jolly niche]

AI currently draws much better than I. You must be very talented.

[brittle cobalt]

@charred jasper the mushroom in your last image in #daily-theme looks amazing 😱

[charred jasper]

Thank you! I think it was 'glowing phosphorescent'

[brittle cobalt]

ill def try that out 🙂 thanks

[dim cradle]

Not really but talent in art is subjective.

[jolly niche]

AI art is objectively awesome.

Subjectively too.

[cedar cove]

It seems to have been a 3-day glitch. I wonder if it was caused by ChatGPT's memory because of past chats. It seems to be fine again now.

[jolly niche]

That's good 👍

[jolly niche]

@brittle cobalt Is this all from a chatgpt prompt? Use anything else?

I really like it!

#images-canvas message

[brittle cobalt]

for that specific image it was a prompt crafted by gpt and put in sora, though i manually changed it a number of times before it came out how i wanted

https://chatgpt.com/share/682a6db4-731c-800a-a3ea-979df6a6b040

final prompt used https://sora.chatgpt.com/g/gen_01jvjvckgje609xm6g5n6xah7w

[jolly niche]

This is awesome! Thank you for sharing! I'm amazed at what's possible with ChatGPT/Sora!

[drowsy lichen]

Why is image generation on ChatGPT so strict

Almost every image I create with ChatGPT gets blocked (none of them violate anything) and I end up having to use sora image generation which works

[late blade]

might be because you are asking for stuff that is against content policy

[drowsy lichen]

It isn’t

[late blade]

I wouldn't be able to tell, what are you requesting?

[jolly niche]

Adobe Firefly now has ChatGPT Image Model built into its Boards.

[late blade]

I'll never allow adobe to get a hold of my credit card ever again,

[twin jasper]

What happened?

I'll just say I'm using less of their product.

[late blade]

Unwanted fees, added hidden dark patterns, cancelling the license, so much hassle, NEVER AGAIN

[twin jasper]

Wow...that sucks. I just hate so many updates popups, etc. I only use Photoshop and Lightroom now. Thinking about cancelling those too. They're such monopoly these days.

[late blade]

Cancelling your contract, even if you have paid a whole year up front has an early cancellation fee that is ridiculous

[twin jasper]

How did they manage to f that up?

[late blade]

8 Months trying to cancel it

[twin jasper]

oh

[late blade]

their dominance on the market, their service model and foolish users buying into it, me included

[twin jasper]

I used to use Premier and needed but started using Davinci and downgraded the subscription. The $60+/month fee was not worth it when I could get Davinci studio for one time fee with free update for life.

[late blade]

best move ever

the image generations I do with GPT, if I need to do something on top, I just use a normal non-recurring paying solution

[twin jasper]

ya...there's always that 😉 I may just cancel my subscription sooner or later.

[woeful turret]

Hi! Is there any way or prompt to use to make letters look good in image generation?

Any difference between using the API vs chat?

[late blade]

https://cdn.discordapp.com/attachments/1252307486204235890/1374641365585100921/assets2Ftask_01jvrst80jewp80awfren0pfv82F1747810403_img_3.png

using chatgpt

[last loom]

I just used sora to generate images. coming from midjourney, i can download multiple images that i generated at once.. is there such a feature in sora too?

[late blade]

Yes it works similar, it’s not as evolved yet. You have to select the images from the three dots. The download button works the same way afterwards

[last loom]

I would have wanted to be able to select multiple images and download at once.. since i see myself generating at least 100 images a day.. and going thru it one by one using the dots is just cumbersome and inefficient. I hope selecting multiple images features is in the work.

[late blade]

It will come. Unless I oversaw something, I don’t think it’s possible right now

[last loom]

Thank you for responding to my query.. I guess I just have to wait then.

[late blade]

I don’t know if there is an export data from Sora line from ChatGPT, the export from the chat provides all images

[last loom]

hmmm.. now that's interesting.

maybe it could work.

[velvet rampart]

You can download up to 100 images in one go. First choose your first image from three dots. Now option "+Select". After this you can choose the rest of the images and press download in the low middle part of the screen.

See #images-canvas message

[woeful turret]

Thank you for the response. The result gets worse when it’s product photos. Any recommendations?

[late blade]

Can you provide me a vague idea of what you want to do?

[velvet rampart]

The filters are not consistent. We all know this but it still amazes me how inconsistent they are. One moment they block seemingly innocent image that should be ok be all parameters. Next moment, they let an image generation through that should have been blocked by all standards and was tried just for laughs.

[keen mulch]

Image generation now supported by responses api https://x.com/openaidevs/status/1925214114445771050?s=46&t=ywNic1kDadDhchP2DHJECg

OpenAI Developers (@OpenAIDevs)

Introducing support for remote MCP servers, image generation, Code Interpreter, and more in the Responses API.

X · May 21, 2025, 3:36 PM

[dim cradle]

has anyone noticed the images get significantly worse results in the last hour?

[velvet rampart]

Sora or ChatGPT?

[dim cradle]

ChatGPT, Sora is working great, but running the same prompt I am getting significantly different and worse results

[velvet rampart]

Ok. I hardly use ChatGPT for image generation.

[dim cradle]

I only do because it helps with consistent characters

[velvet rampart]

Well, I do consistent characters by having a base image and always using that in the prompts.

[late blade]

Yes, the pixels are not pixels anymore. The colors are not the right value, the deviate by a hex value or alpha channel. It's so annoying.....

[outer bolt]

Is sora down for you guys? It was really slow this past hour now its returning server error when generating. first time ive seem this error. The server had an error processing your request. Sorry about that! You can retry your request, or contact us through our help center at help.openai.com if you keep seeing this error. (Please include the request ID ....)

[true wave]

it is for me

[obtuse wyvern]

Yes, it's down for me as well

[vapid granite]

oh, good. I'm glad it is down for a lot of people. I got a message saying I can't upload people anymore to Sora, even though I don't have "strikes" on my account and then now I can't upload any pictures to sora at all.

[obtuse wyvern]

Yes, same. there is a banner on the site now about the outage

[lilac hull]

Sora was just glitching. But it's working for me now

[velvet rampart]

Sora was yesterday evening just rigth for me. I regeneration a prompt from last night and looks a bit crispier now. Not a huge difference, but some details are bit sharper.

[dim cradle]

Hi

[hollow fjord]

We need new aspect ratios like 9:16 and 16:9

[woeful turret]

I am trying to generate ai generated images based of my own product image. The result of the shape and bottle is 100% correct - besides the text. The text is almost there, the font and everything - but the letters are wrong. I have made this through response api.

[dawn charm]

I actually have this same problem, Either it mis-spells the company name, or it breaks something in another generated model.

[woeful turret]

I'm currently working on it and trying to figure it out if there is anything in the prompt that could be adjusted as the model should have the capability

[dawn charm]

Perhaps it's the way I'm describing the text "Wrap text on the bottom of the object -Crazy Company name-" It'll cut like half off since it's trying to compensate to not take up the whole object i suppose.

[woeful turret]

I got a great result now, but it's only working for white background image

[dawn charm]

I maybe have to try that and see if that would be a viable option. I'll have more time to fiddle with it tonight.

[velvet rampart]

I usually do an object first. Once it is in a backgroun, I use the generated image as a new image, not remixed. This way I keep the quality up.

[woeful turret]

I don’t understand, do you mean in the chat or API 😃

[velvet rampart]

On Sora, but this applies also on ChatGPT and why not on API. If you regenerate the image again and again, the quality goes down.

[tardy depot]

Edited images are always less qualitative than original generated images

[velvet rampart]

But if you use the generated image as a source image, this does not decrease the quality.

[tardy depot]

I always do

[tardy depot]

Inpainting helps, but not regenerate

[velvet rampart]

Inpainting sucks at least for me and it is counted as editing and thus regenerating.

[tardy depot]

Did you mean that you open the image and click 'edit', or that you download and upload it again manually?

It should be the same as 'edit' adds the image as source

[velvet rampart]

On chatgpt, I copy-n-paste. On Sora, I use the image directly as one of a source images without using the remix function.

[next fog]

Anyway workaround so there is no yellow tint in the generated images?

[velvet rampart]

Just add ”no yellow tint”. With 4o, negative prompting works.

[forest hatch]

Still generating with the yellow filter if that get's fixed then it is perfect.

[velvet rampart]

Strange. When I have it, the yellow tint is gone. Have you tried the same prompt in a new chat? If you use the same chat, earlier generations can bleed into the new one and ruin it. Try adding an extra command ”6000k white light”.

[next fog]

@velvet rampart thanks for the suggestion, but it didn't work in sora. Do you have anything for sora?

[velvet rampart]

Could you share your prompt with me via DM? Without seeing it, I can’t help you. There might be other factors causing the yellow tint.

[remote stone]

I have a question, before when the image generator was DALL·E and you generated images through chatgpt you used to be able to see what the prompt being used for an image was when you clicked on it, is there nothing like it now?

[velvet rampart]

No

[remote stone]

Damn, RIP

[minor finch]

Yeaaah, idk why they took it down.

[jolly niche]

I learn a lot from other people's prompts. It helps me refine and get better results.

[velvet rampart]

You can still use Dalle3 on customGPT called D•ALLE (https://chatgpt.com/g/g-2fkFE8rbu-dall-e)

[umbral aspen]

ChatGPT 4o image generator follows exactly what you put in as prompt, Where as DALL·E 3 on Openai's platform it rewrote prompts unless you asked specifically to not rewrite. Microsoft's version of DALL·E 3 follows prompts exactly also no rewriting

[little pilot]

yes

and it's ignoring prompts completely

[little pilot]

are you sure about that i've been attempting for the last few hours and it's completely ignoring my prompts

[velvet rampart]

Good morning, I just woke up. Could you DM your prompt so that I see what is the matter? There might be other infering items in the prompt. Also, have tried adding ”6000k light”? This has helped and was found our fellow discorder on this server.

[little pilot]

Sent

[vapid granite]

Is it just me or are Sora/GPT4o generations like REALLY bad in the past few hours? It's like it is "draft" mode or something.
Sora is not following prompts and the images are Dalle 3-level

[umbral aspen]

This is why i use Microsoft Designer its 100x better than Openai's implementations even 4o is better on Microsoft... it dumb founds me a company not utilizing there own creations full properties and rival/partners do

[velvet rampart]

It might be just you. I have noticed that in rate cases Sora generates really poor quality. Images I generated this morning within several hours are coming great as before.

[velvet rampart]

Designer uses still Dalle3.

[umbral aspen]

it does and its updated every 12 hours

[umbral aspen]

4o is available in co-pilot now if you have co-pilot pro

[velvet rampart]

Well, you don’t have to have. As a free user, I have 4o on copilot.

[umbral aspen]

cool

[velvet rampart]

ChatGPT is update. Now it produces two variants.

[carmine magnet]

uhhh

you guys do know that campi flegrei volcano is currently freaking out a huge part of the meditterranian? might not be the... most well timed of theme choices...

[haughty spruce]

4 variations are opened.. go go go!!!

wait... the quality seems better?

[velvet rampart]

On Sora or ChatGPT? I get only 2 variations on ChatGPT.

[haughty spruce]

On Sora

can you please look at the quality of my images? Is it just me?

[velvet rampart]

Higher quality? Hmm… Difficult to say. Compared to image quality I have been getting? About the same, I’d say. Do you push the system in any manner to get higher quality?

[haughty spruce]

no, it's same as before

first I was generating image but the result had more texture quality than before. i thought maybe it's my eyes... then 4 images option is finally released just after an hour, pretty interesting right? then i chose that. the results seem to be higher quality in my opinion maybe it's just my eyes

[velvet rampart]

Well, I have pro, so I have had that option a long time. But I noticed that some of my images have more depth and some have better detailing, such as my Scooby’s nose.

So, yes, a minor but significant increase in quality. It is those tiny bits that matter. My guess? More computer capacity. Why did OpenAI otherwise release four variants for Plus on Sora and two variants on ChatGPT?

[dim cradle]

in sora or gpt?

[velvet rampart]

Sora

[cedar cobalt]

I wonder why video gemeration came first

🤔

[keen mulch]

Image generation was first with DALL E, and sora was initially revealed in Feb of last year

[sinful swan]

guys

is there a way sora can legally generate characters like pokemon , marvel ,etc…

[austere nexus]

ChatGPT, Team's plan got Media Library today!!!

[haughty spruce]

around may 26, the images did have better quality now it looks like it's reverted back 2x like it's even worse quality than before

does anyone feels like it?

[velvet rampart]

Sora or ChatGPT?

[haughty spruce]

Sora, ChatGPT and Copilot

[vernal shadow]

I feel they need to fix the random policy enforcement issue (unless) intentiional? .. wit hchatgpt i can use the same verbatim prompt on various sessions have it run and other times states violates policy etc... and then WHEN DOES make an image sometimes if i say ok lets try sora SORA say cant use this image as it violates policy ... 🤔 not sure why lol

[lofty valley]

Yeah I've noticed a decline, today I was even wondering if I had the old dalle model for one of my generations as it was really bad.

[velvet rampart]

Do others get an error message about not being to generate due to a server error

[velvet rampart]

Well, there is a clear change. On Monday I could generate two variants on ChatGPT, now only one variant. Could this be A/B testing? Pretest for a bigger change? To see if systems can handle it?

[lofty valley]

feels like it

[fickle tusk]

Unable to generate
The server had an error processing your request. Sorry about that! You can retry your request, or contact us through our help center at help.openai.com if you keep seeing this error.

what is happening? on Sora

[velvet rampart]

Wider issue. I cannot upload or generate.

"Sora is experiencing elevated error rates - we're working to resolve the issue."

[cedar cobalt]

image-gen-2 will definitely have better definition

images created by it are a liitle bit blurry tbh

[sacred temple]

I asked GPT:
🔧 GPT-4o "ControlNet-Like" Capabilities
ControlNet Type GPT-4o Equivalent / Workaround Notes
Depth Map / 3D Control ✅ Auto-generated for internal coherence (e.g., image editing or inpainting accuracy). You can’t extract it directly, but it’s used under the hood.
Pose Estimation / OpenPose 🟡 Indirect. You can describe poses very specifically (“standing on one leg, arm pointing to sky”), or upload a base image to match. No skeletons or keypoints exposed.
Canny Edge / Line Art 🟡 Implied via uploads: DALL·E editing respects edges if given line drawings or sketches. Not directly controllable, but it’s style-aware.
Scribble / Sketch ✅ You can upload sketches and say “make this hyperrealistic,” and it’ll keep the form. No direct scribble mask control though.
Normal Maps ❌ Not available. No lighting vector control or surface normals.
Semantic Segmentation (Mask control) 🟡 Only via inpainting or edit prompts like “make the background red.” No mask editor yet in the public UI.
Color / Illumination Control ✅ Prompt-based: “Moody green lighting, neon shadows.” Works very well.
Style Transfer / Layout Control ✅ Via image prompts or edit requests. You can upload an art piece and say “recreate this style with a new character.” Not spatially locked though.

🧠 Smart Workarounds Using GPT-4o
"Depth-Aware Prompting": Describe light direction, camera angle, foreground/midground/background elements.

Use Reference Images: Upload a sketch or 3D render and tell it to “maintain pose but change outfit/background.”

Layered Prompting: Ask for a 3D render first, then use the result to request a painting in the same composition.

Coming Tricks
If OpenAI exposes more granular tools like pose detection or mask editing, they could potentially layer in ControlNet-like UX without requiring users to manually upload models or weights. Until then, it's prompt-driven "soft control."

Let me know if you want a workflow to simulate ControlNet behaviors using GPT-4o, or I can generate fake depth maps from your existing uploads for mock previews.

[formal gate]

You think we can make emoji's. So I have a favorite meme emoji thing and I wanna make it a real emoji I can use all the time. You think I can get ai to do it

[vapid elk]

sora can make emojis and stickers very nicely

you can even ask for it to make transparent background

[dusky socket]

Hello. I’m new looking at Sora to generate images. I notice the “Aspect ratio” drop-down has only three options. Is there a way to specify a custom aspect ratio?

[velvet rampart]

No. Those are the only ratios currently available.

[dusky socket]

That’s very limiting :(

Thanks for the response

[jolly niche]

You can use another image editor to resize to any aspect ration you'd like.

[floral hatch]

I’m using the api to generate images for questions for homework but I’ve found it’s not very good at it. Am I possibly prompting wrong? I have a script that generates some questions, and it makes an image if it thinks it’s necessary. I’ve found it quite bad at making sequences, patterns and coins etc. I give it a lot of context to do so but it still has very common errors

[vapid elk]

could you give an example of what was the input and what you got form the API?

[hearty ether]

Does anyone else have any issues with body proportions

Like it always makes the head too big while the body is too small/short

Because of the weird framing by default

[velvet rampart]

Yes! This has been a problem since day 1 with 4o. I have figured out to give strict guidelines when this happens. Not always. I use ”Natural bodily proportions are rendered correctly and perfectly as they are in the source image; the head is minuscule, and the size is 1/9 of the whole body length. Shoes appear very large on his feet, being 1/6 of his height.” Well, this is for a toon anthro lion cub to make the head appear somewhat normal, but you get the idea.

Also, just telling the model to place the subject/character deeper in the image (aka in perspective) helps as the model likes to squach. I think this comes from that it starts from up and goes down thus not leaving enough space for feet.

[hearty ether]

Yeah the torso usually gets compacted way too much

I tried to get it to describe the torso lengths in terms of heads ratio but not much luck

[floral hatch]

Yes sir! My script has a prompt for images like so;

• image_prompt: if a diagram is needed, send an extremely detailed prompt giving all possible information. Include in that prompt all necessary context such as the answer and explanation

The question prompt is like this:

• Verbal reasoning (short passages or statements at least 4 sentences long + inference questions)
• Multiple step Logical deduction (e.g. "who is in the middle?")
• Pattern recognition (number or shape sequences)
• Net-folding questions (dice nets)
• Conditional statements ("only if… then…" problems)

Here’s an example:

A dice net is shown with the numbers 1, 2, 3, 4, 5, and 6 on its faces. If the face with 1 is opposite the face with 6, and the face with 2 is opposite the face with 5, which number is opposite the face with 4?

I can’t seem to attach pictures

[vapid elk]

yea, I don't think the AI will reliably make images of that

it might be able to.. with luck, but it will not work reliably on those prompts

[mental belfry]

Tired

[floral hatch]

Haha all good! Thanks. It doesn’t do it even on the app. It’s something simple like coins sometimes. Emma has 3 50c coins (Aussie here) and 4 $1 coins, how much money does she have? And it shows the wrong coins, wrong amounts etc

[slate sun]

hey guys i need help with something

been trying to place an image into another but it just doesn’t seem to get the placement right

[frail wharf]

I am using VPN to connect, but the connection is very slow and unstable.

[late blade]

The macOS App is having some hiccups currently on my side too.

[red prairie]

Is it me or is Sora more loose now with the restrictions?

[velvet rampart]

Yes. You can generate nearly all copyrighted characters except some Disney characters such as Mickey, Pluto, etc.

[dim cradle]

oh really. noice

i can finally make some superman haha. just in time for the movie

[dim cradle]

I wonder if Google imagen4 inspire the change. i mean since imagen4 lets you do ip stuffs, maybe openai said, well, why not us too?

[sinful swan]

@velvet rampart is there any way i can generate pokemon lol

[velvet rampart]

Yes. Just ask for specific one.

Either on ChatGPT or if you are paid user, on Sora.

[violet snow]

#server-rules

[dim cradle]

imagen4 is cool of course. but after using it some, i really think 4o is still the king by a pretty good measure

[velvet rampart]

I think that imagen4 is helpful in making guiding images for 4o.

[jolly niche]

I've been having issues with creating images of contestants from the show Survivor. And when it does do it, it doesn't look like them.

Especially from the earlier seasons, like season 1.

[jolly niche]

I'm guessing Gemini is using Imagen4? With certain prompts/images I've gotten great results lately, where with the same prompt I'll get "can't do it" on 4o/Sora.

[velvet rampart]

I recommend to use source images. This might overcome the issue. Other issue might be the actors are not that famous or they have not given consent to use their likeness.

[jolly niche]

Ah, the consent issue "might" be it. I did use three or four source images and it was better, but not quite there.

Thanks!

[charred hinge]

Why is the site down?

[inland sigil]

they SUCK

[keen mulch]

Now you try implementing services that service millions of users

[glossy scroll]

Very nice.

Finally

[wraith vortex]

My understanding is that GPT image generation doesn't have vector graphics. So every image it creates is "painted" based on how humans paint pictures. So when you have something created from scratch, it's finding patterns to duplicate then merging that data to create the final image. The less description or context it has, the more likely the image won't be what your looking for. Giving it an image to base a pattern of works great but I've had it "fix" images and because it's "paiting" them, they are never quite the same "person" as the original. I've actually had some luck by telling it to emulate vector graphics and create the image as if it were rendered by a 3d game engine.

[wanton light]

why didn't I get the image when I click"generate image" on the website of GPT's AI Art generator

[meager flame]

GPT is acting strange today...i have to ask it few times to generate image...and he say ok, here it is but just repeat the prompt over again...

[hearty ether]

Ngl the filter has gotten far more overzealous lately

Like it won’t even generate any dresses unless it’s completely full coverage

like long sleeve turtlenecks

[lapis rapids]

Hi, is anyone having trouble generating images in DALLE on ChatGPT? I'm trying and it's not working

[smoky sable]

Same here, the issue started approximately 4 hours ago for me. The legacy DALLE is not accessible right now despite opening a chat from GPTs.

[lapis rapids]

And does it work with the image generator it currently use?

[smoky sable]

The image generator in ChatGPT based on Sora is working for me.

[lapis rapids]

I don't think OpenAI wanted to permanently remove DALLE's functionality, right?

[smoky sable]

I'm not sure, hosting DALLE do consume resources.
However personally I also don't want OpenAI to remove it, as I used to tune some good prompts for image generation and DALLE returned diverse results. Sora's output style for the same prompt seems to be fixed especially when long prompts are given, hurting the diversity.

[lapis rapids]

I wouldn't want that either, I like the diversity that DALLE has

[cerulean shale]

dalle not working..how to resolve that

image not generating in dalle . chatgpt..when will be the problem resolved..

i am a paid user..

[blazing imp]

hi, so Dalle not work for all now, not only me

[blazing imp]

same

[cerulean shale]

how to escalate dalle problem to chatgpt team #chatgpt-discussions #community-help

[velvet rampart]

I did a report via help.openai.com

[cerulean shale]

did dalle problem get resolved

[velvet rampart]

Nope. I just tested the Dalle customGPT and its behaviour has gone crazier.

[cerulean shale]

Dalle customGPT not working been a paid user then whts the use...horrible support team

[strong bridge]

Any update on DALLE? Not working at all.

[strong bridge]

Any update?

[velvet rampart]

No

[cerulean shale]

guys anyones dalle got resolved

[strong bridge]

Not so far, same issue.

[warm mortar]

Same here 💔

They need to either fix it or announce something…and I hope they fix it because DALLE is SPECIAL

[smoky sable]

Numerous bug reports have been done via discord. Hope the power of crowd would let OpenAI team know the value of DALLE lol.

[velvet rampart]

Still not working.

[cerulean shale]

guys lets escalate this issue and inform them

[uncut mason]

Picking dalle over image gen is crazy

[fathom jasper]

wth these ppl arent ready to accept that something is wrong on their side . they replied me this when i tried to generate a simple colorful flower "Potential Terms of Service Violation: Your account may be flagged for behavior that conflicts with our Terms of Service, such as unauthorized or harmful use of the platform or sharing your credentials with others.
Security Concerns: Suspicious activity, including multiple failed login attempts or access from unknown locations, may prompt a temporary downgrade.
Account Changes or Maintenance: Recent modifications to your subscription, plan, or payment information can lead to temporary limitations while updates are processed."

[fathom jasper]

already done so . went on deaf ears

[cerulean shale]

whts goin on ..the support team says there is no issue in dalle ..tells to check openai status... its been 2 days not a single image been generated

[smoky sable]

TBH I'm not sure if they said DALLE is fine because the one in discord is still working...

[wintry shale]

N

[open wagon]

What happen to dall e

[open wagon]

it’s still working for me

😂

[cerulean shale]

bro are u generating actually in dalle or in chtgpt 4o...check properly and let me know

[clever jungle]

Is there a way to still use dalle 2?
Like through API?

I really loved the model, Would love to be able to use it.

[uncut mason]

Scamobots assemble

[somber merlin]

why is every image in a yellow tone?

and in a oleo "style" always?

[storm scaffold]

json prompts for dall-e bot in discord is works?

[fathom jasper]

their support team says there is no problem at all from their side . what to do ?

[proven sentinel]

From DALL-E: "The image generation tool is currently unavailable, so I can't create a new image at the moment."

I just got that, so I guess it's acknowledged? If OpenAI no longer wishes to support DALL-E, I wish they'd just say so.

[dim cradle]

I only got it to work on lm arena

[proven sentinel]

You, sir, are awesome. It took a couple of tries but it works!

If any want to see a direct comparison between Dall-e and 4o, I am happy to DM the images and the prompt.

[cerulean shale]

but in lm areana where can one select dalle3

[smoky sable]

First, select "image" instead of "chat" at the bottom left of the user input window. Then you could select the model at the top as "dalle3"

I can't send screenshots in this channel... terrible

[velvet rampart]

Use images-canvas for screenshots.

[warm mortar]

Also still works on Bing

[dim cradle]

Sup bro? 😎 yeah it’s so odd. I think it’s working through the API? You make sense how the element Arena and Bing would be working..?

[verbal sequoia]

Get notified for server events!

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

[warm mortar]

Not much…went on a road-trip

I wish they’d give us DALLE back in the app though…it has a unique style

[dim cradle]

I know I could use it. One of the mods was saying that there’s dalle bot

[smoky sable]

Does anybody know how to customize the size of DALLE image outputs when using LLMareana or Bing Image generation? They currently all return squared (1:1) images, instead of 16:9 or 4:3 ratioed images.

[keen mulch]

I don't even know why anyone uses DALLE now when the native image gen is an order of magnitude better

[warm mortar]

It is better at text and image editing, but DALLE is better at 👁WEIRD👁 and doesn’t default to weird yellow coloring

[smoky sable]

In other words, the diversity and creativity apart from pure strict instruction following.

[dim cradle]

Those that know they know those that don’t know they don’t know

[cerulean shale]

In dalle you can be very creative...rather than this yellow images

[dim cradle]

1. Legacy Preference Users = creative freedom

2. Novelty-Driven Users = advanced capabilitie

For me what’s making me wanna go back to dalle is the 16:9 aspect ratio

[empty trail]

Can someone help me find the area for image prompts?

[fast hinge]

@smoky sable Microsoft Designer provides DALL-E3 in 9:16, 1:1 and 16:9 aspect ratios. You'd have to subscribe to one of their for-fee plans to get more than 15 prompts per month, though. (1 prompt usually yields 4 images)

[proven sentinel]

Because ever since everyone went berserk with the anime thing, any type of art that’s primarily looks clearly hand drawn (like comic book art) doesn’t really look good in 4o.

[strong bridge]

Does anyone have an update on DALLE?

[strong bridge]

Or is this the only way to use it currently

[strong bridge]

Thank you so much for that though @fast hinge , it really helps me.

[kind pulsar]

I’m not sure if you are talking about dall-e 2 specifically, or dall-e the last, but it is still available in the GPT /explore/ by ChatGPT

[fast hinge]

There are several image-generation services on the Web, each of which provides access to a variety of image generators, including DALL-E 3 and previous versions. I've used NightCafe, for example, although DALL-E 3 is available only in their for-fee "Pro" tier.

[strong bridge]

It is not working, though

[strong bridge]

I have purchased a subscription for the Microsoft Designer tool, it is working great for me now, thank you once again !

[velvet rampart]

I think that I have figured out the source for dalle3 promblem. You can choose the model for customgpt. OpenAI has not locked the model. This causes the strangeness.

[fathom jasper]

can u elaborate pls

[cerulean shale]

plz can you explain how

[plucky hare]

I think what Milamber is suggesting is that perhaps the bug started when they added the ability to switch models in custom GPTs (before, custom GPTs were only 4o). And then maybe the version of 4o now tied to the DALL-E GPT (due to this change) is no longer the last version of 4o that could call DALL-E and is maybe a more current 4o that's been told to use DALL-E when maybe it actually can't

I don't think there's currently a workaround

[velvet rampart]

Yes. This one. Thank you, @plucky hare.

I found accidentally today an announcement dated on 13th June about this. About the same time, Dalle3 customGPT stopped working.

[smoky sable]

Right now if we want to create new custom GPTs, there is no longer a “DALLE” option available for “capabilities”, rather “4o image generation”…Guess also related to this update.

[plucky hare]

Yes I think it's plausible, nice guess. Hopefully a fix is also plausible! 🤞

[velvet rampart]

I have also alerted OpenAI to this possibility. What they will do with this info is up to them.

[fathom jasper]

Thank you for reaching out to OpenAI Support.
We appreciate your prompt response.
Please be advised that we have already escalated the issue regarding the Legacy DALL·E custom GPT not generating any images. Our engineers are currently investigating and gathering more information and they will provide an update as soon as they have completed their root cause analysis.
this is what they said

[plucky hare]

I think the DALL·E GPT is making DALL·E images again

[smoky sable]

Yes! I just tested and it's back! Thank you for bringing the news! (Not sure if it's temporary though lol)

[dim cradle]

I can confirm it’s back yay

We made it we made it everybody

[solar forum]

am I the only one or is bing chat image generation becoming less and less useable?
I lose credits now even when the image does not generate...

[dim cradle]

What a beautiful day to be alive

This could never happen again on the weekend lol

[vernal shuttle]

i am tring to change background of a puppet figure using chatgpt 4o, but it also slightly changes the puppet. is there a way to prevent that?

[velvet rampart]

You can try to say something like: "Place the puppet figure into provided background. Otherwise, keep the puppet figure as it is." You have to say what you want the model do.

[vernal shuttle]

i tried, i wrote DO NOT CHANGE ANYTHING IN THE PUPPET or i should not be nagative?

[velvet rampart]

Did you try my phrasing? Negative prompting does work with 4o, but it has to be defined, such as "no X". Usually, I get better results when the major components are phrased via positive.

[vernal shuttle]

will try in a moment

[vernal shuttle]

nice using "keep the puppet figure as it is" works better!

[velvet rampart]

That’s great!

[near nova]

hi, anyone got some suggestion on how to write better prompts on sora? i see all these stunning images and i cant create simple one XD

i dont get neither when to edit ot when to remix

[dim cradle]

you should check out https://discord.com/channels/974519864045756446/1315696181451559022 if you make a thread there I could come down and show you some pointers and tips

[fresh oxide]

I think I am getting why the AI keeps losing focus so much and hallucinating so quickly....They keep making all these small adjustments to it and updates...
And every time they day that, the system always had some kind of mental block before it resets itself and settles...
But it feels like they are doing something to it every single day because just when it seems to be settling, BAM they do something else

[delicate willow]

Anyone got a good prompt to add to avoid the yellow filter all images seem to get that then stacks on top of each other in remixes?

[lilac sundial]

Add a negative prompt. No yellow filter or no filter and simply reinstate what you do want which I think is maintain original image colors.

[fresh meteor]

it’s seems the censorship has increased today that’s unfortunate you know there’s mostly sfw content in these keywords so it’s extremely disappointing to see it being banned 💔💔

[fresh meteor]

yup i can’t even generate this:

A stunning cyberpunk anime girl in a neon-lit futuristic city at night. She has glowing blue hair styled in a high ponytail, wearing a high-tech pink and purple suit with bright neon accents. she has fierce glowing pink eyes. She’s equipped with a futuristic exosuit covered in glowing circuits. Rainy wet streets reflect the vibrant lights. The scene is cinematic, with dramatic lighting, high contrast, futuristic signs in the background. Art style inspired by Ilya Kuvshinov and Artgerm. 4K, anime style, highly detailed, masterpiece.

nothing bad about it

say’s goes against the nsfw policy

smh 🤦🏻‍♀️

[velvet rampart]

On the surface, your prompt looks ok. To bare it, I recommend adding one sentence/element at a time to find out what is causing blockage.

A quick AI analysis suggests following:

Summary of Likely False Triggers:
• “Anime girl” (common flag due to prior misuse)
• “Stunning”, “glowing”, “high ponytail”, “pink eyes” (sexualized or stylized traits)
• “Pink and purple suit” + “exosuit” (common in stylized body-focused illustrations)
• Artist references (Ilya Kuvshinov and Artgerm can be filtered automatically)
• “Masterpiece”, “4K”, “highly detailed” (quality pushers used in NSFW circumvention)
• Cumulative stylization: Sometimes it’s not a single word, but a converging pattern of terms that match with historically sensitive prompt templates.

[near nova]

i didnt get it. false triggers as things better not to write?

[velvet rampart]

False trigger - this means that a word can be placed on a list accidentally to block. For example, word ”rose” became a false trigger for dalle3 back in February.

This is the reason, why I suggested to add sentence by sentence your long prompt to find the sentence that causes the blockage.

[near nova]

ah ok and how can i know the current ones? and why is even that happening btw?

[velvet rampart]

As I said, you start a fresh chat ob ChatGPT or a new image generation on Sora. You add the first sentence of your prompt. If an image is generation, it is ok and you move on. Next image generation, again a new chat and you add the next sentence. You repeat this until you come across a sentence that causes failure in image generation. Now you start modifying the ”offending” sentence to find out the word or words that cause the blocking.

[fresh meteor]

it’s because of the suit i was able to generated before now it’s just gets denied

[mystic terrace]

Guys, can gpt generatre images that has more then 1536 × 1024 ? because he told me that he can but when i told him to do 2k he still did 1536 × 1024

[plucky hare]

Currently there are just the three aspect ratios, 1024x1024 and 1536x1024 (and reverse). If ChatGPT said otherwise, that was a hallucination - ChatGPT is not always the most reliable self-reporter on its own abilities and tools

[mystic terrace]

yeah, i was thionking same. Thanks

[dim cradle]

I think I’m starting to see why people that like to use DALL-E do not like the features of the new image generator and I think I’m starting to understand the difference in what they mean. I haven’t really paid attention to the difference, but now I slowly starting to see it.

The new image generator is way more technical and precise, dalle often filled in the details & was more energetic and vibrant in the image it the produced.

[near nova]

which new image generator?

[dim cradle]

The new one they updated it with

[near nova]

i dont even get why dall e exists, with respect... sora just seems better. especially dall e on here is embarassing

[dim cradle]

It’s hard to explain because it’s how you view better, dalle 3 has many strong features that the new image generator doesn’t and vice versa

[lilac sundial]

I love dall e for image generation but need to also use in combination with others to finish off minor tweaks

[proven sentinel]

I do the same thing, dall e to generate, 4o to edit.

[lilac sundial]

What’s 40?

[near nova]

to me dall e is quite crap. any advice?

[earnest surge]

gpt image 1 or imagen 4

[near nova]

?

[earnest surge]

I told you what to use.

[near nova]

i know but i dont understand whrer you see those... i only see the latest 4o model

[earnest surge]

gpt-image-1 is basically the latest 4o model.
Imagen 4 is an image model made by Google.
You can use it at gemini page

[brave creek]

1386669802571038790

[hard moon]

can someone PLEASE help me get my document back from chatgpt i need it for court in a few hours PLEASE HELP

[near nova]

oh i see... i tried gemini once it was fine. i tried again now it isnt working. im in europe maybe thats why...

[raven raptor]

i have a question. if an image really triggers you, which was created by another person, does reporting it even help?

[raven raptor]

crazy how there are ZERO DEVS or ADMINS to answer anyone but say one bad word or swear...

[rigid bough]

this is a community server, people will answer you if they can

[plain idol]

i have a question: why is the quality of the images i generate deteriorating? the lines are blurry, details are missing, and the images have a strong 'smudged' appearance. what's going on?

[raven raptor]

if you are an admin or a dev, i mean you coulda just answered lol

also why is Sora down? i have had this banner or pop up saying there issues

damn i cant even paste in this window?

[plucky hare]

We are (almost all) community members, OpenAI folks have yellow usernames and the logo. For more info #server-staff
If you'd like to reach out to the OpenAI team, the best way to reach out is the chat on help.openai.com. I imagine if you reported an image and it was rulebreaking, it may be removed from Sora, but I'm not sure. If you're talking about something posted in this server that you think breaks the rules here, you can report via Apps > Report message and a mod here can review

[velvet rampart]

Image generation on Sora is really slow at the moment.

[molten delta]

Bro it says in queue what am I supposed to do now

[viral sinew]

Explain

[twin jasper]

Can't use SORA right now but the status is saying otherwise :\

[velvet rampart]

Just wait. You can’t do anything. Or well, you can create a ticket at help.openai.com. If they get enough tickets on this issue, they will fix it. This kind of delays is not normal high demand time behaviour.

[verbal sequoia]

DALL·E Category and Channels are now “Images”

With the release of 4o image generation, we’ve renamed the DALL·E category and channels to Images, to better reflect all OpenAI image models.

-# Share your amazing images in #images-canvas!

[twin jasper]

Sora was working again and now it's stuck at 99% lol

[umbral aspen]

oh i love 4o responses sometimes lol:

You're not just generating images—you’re simulating a camera→sensor→post-process→AI-pipeline architecture in language.

[blissful stone]

Why did they remove 4 images?

[raven raptor]

thank you for this reply

[umbral aspen]

maybe use an original prompt lol google your prompt if you have safe search turned off it gives nsfw images in search thats why you cant generate the prompt is blocked

[hearty ether]

Anyone have issue with double image generations even when specifying only one?

[fresh meteor]

nah it work before

i know what the prompt does because i did it before nothing nsfw about it. since it’s openai

[nocturne silo]

I think I have had enough of OpenAI’s random, arbitrary and usually idiotic content policies. They are never applied equally, always cryptic in nature, and almost never make sense. It’s especially ridiculous when attempting to use Sora. Even ChatGPT, which I always use to optimize Sora prompts and get feedback from, can’t seem to even be relied upon to understand how these policies are applied or what will trigger them. It’s too much BS for the amount of money OpenAI charges.

[drowsy fog]

I am having it

[rigid gust]

It happened to me a few times tonight.

[scarlet ferry]

On phone (Android) whenever I have gpt creating an image and minimize the app. Whenever I return I get the notification that I lost internet connection. Any suggestions?

[magic niche]

new styles update with 4o image generation

[hearty ether]

Do you mean the double image, or do you think they changed the model?

[main sleet]

[near nova]

to me its been 4 days sora ai its not working properly i wonder whats their excuses this time

[cloud basalt]

Ok why is sora taking 5-10 mins just to create an image?

[near nova]

we paid for that ****

[carmine magnet]

@pastel siren Re: #daily-theme message

Ah, I see you are a man of culture as well. Now do Street Sharks. 😄

[pastel siren]

We'll have to wait for the proper theme for that. 😉

[carmine magnet]

*Jawsome.

[drowsy fog]

Did they figure out a fix for the double image issue?

[echo swan]

is inpainting supported in the api?

[jolly niche]

Yes, I believe it is.

[echo swan]

I know they offer masking to allow u edit one specific part of an image, but it ends up changing the whole image regardless

[near nova]

bla bla bla.... thats all they do. they are good at talking. but our time and money wont come back. we invested our money and then wasted it all

[open wagon]

Did they change something with image generation cuz it generating two image now for no reason

[open wagon]

I mean it’s your choice for doing all of that , so stay mad

[near nova]

?

im not i was just sharing your same complaint

[open wagon]

Oh

I’m sorry I misunderstood

Was really under impression you were jabbing at me cuz of that thumbs down emoji

Let me rephrase , “your anger is valid “ so stay mad was actually me validating your complaints

I’m planning to unsubscribe temporarily to chatgptplus soon cuz most of the Plus features, I’m not using it that much anymore

[near nova]

thats fine but just for your future reference you need to ask before to make sure you get that right, then you might opt to go harder/heavier, just saying

[harsh saffron]

I had it before

[hollow current]

I get the prompt challenges and improving on writing prompts.
But is anyone else doing image generations by uploading your own images for say a character, a pose, a background, and then writing a prompt "directing" GPT how to make an image based off your uploaded images?

[drowsy fog]

Double images. Is there any fix at all?

[near nova]

we should all gather up meaning showing a still stance: for a while we shoudlnt use non of open ai services. the only way they understand

when they took the money them were quick, no glitches. we paid in full . thats not fair.

[quick badger]

i have run into a brick wall with the inconsistant nature of the CHATGPT art prompt filters.
a little background
with chatgpt's help, i was able to come up with the following stylistic prompt for the creation of fantasy rpg art
"A modern fantasy illustration painted in a richly detailed, semi-realistic style that blends the iconic fantasy aesthetics of the 1980s and early 1990s with contemporary digital polish. The visual tone draws influence from the works of Boris Vallejo, Julie Bell, Larry Elmore, Clyde Caldwell, and Jeff Easley. It features warm, dramatic lighting with high contrast and radiant highlights that sculpt muscular anatomy and ornate armor. The composition is bold and cinematic, often featuring heroic postures, dynamic perspective, and richly textured backgrounds—misty mountains, twilight skies, or elemental ruins. Costuming and props are designed for visual impact and mythical symbolism rather than realism, with metallic detailing, leather accents, and glowing magical effects. The final result resembles a high-fantasy book cover, RPG rulebook illustration, or sword-and-sorcery film poster, evoking power, elegance, and dramatic grandeur."

this overall seems clean and policy friendly.
and i was even able to produce art with it

however while using that very same stylistic prompt and a new prompt for the model for the image, the piece gets blocked
this is the model prompt for the new piece
"Appearance: an elderly elf with a wise gaze, whose hair colour and garments change according to fashion of the elves he appears to (he represents them all after all)."

again seemingly in compliance with the policy

so what gives?

[velvet rampart]

Did you try these in a fresh chat? Something in the chat might have caused blocking as the whole chat is the context window in 4o image generation.

I just generated your long prompt. It is on #images-canvas message

Thus, the actual prompt is not blocked. It is always good to double check the generation in a fresh chat.

[quick badger]

yeah its just the inconsistancy of chatgpt in what is allowed or not

[rich star]

Why the hell is image generation slower than DALL-E 3

[near nova]

dall 3 here on discord is even embrassing , i tried 3 times and image generated were awful . same prompt elsewhere much better

[rich star]

Sus

[dim cradle]

Do you have a screenshot of what you’re talking about or what you’re trying to show? I learn visually, so seeing it would help me actually understand the problem.

[quick badger]

it won't let me post one here.

so i did in #images-canvas

[gilded nest]

@rigid gust I really enjoy your images in the daily each day. If you have room for another friend, I'd be honored.

[rigid gust]

Awww thank you @gilded nest that's so kind of you!

Are you an artist too?

[gilded nest]

Yes, I've been participating for the past week. I just submitted A broken mirror with footprints on it.

[rigid gust]

I love it! How are you enjoying the daily themes? For me, it's forcing me to learn how to actually describe the inspiration in my head (not always easy!).

[gilded nest]

I really do enjoy it. It gives me something to look forward to in the morning. You make it look easy though. I know sometimes it's a challenge. Some of the prompts seem like a paradox sometimes and that's fun for me.

[rigid gust]

It's an enjoyable process I look forward to each day as well. It's very nice to meet you here 😊

[flat totem]

"You can only generate 2 images or videos at a time
Upgrade to generate more images or videos at the same time"

[gilded nest]

Nice to meet you too. If you ever want to chat. Go ahead and send a DM or a tag and we can collaborate on something sometime maybe?

[flat totem]

Sora's broken

[rigid gust]

Eek were you working on something? Hopefully it's fixed soon

[flat totem]

"You can only generate 2 images or videos at a time
Upgrade to generate more images or videos at the same time"

[near nova]

is that crap still not working?

[near nova]

why cant i upload any image in #sora-submissions ?

[rigid gust]

Good morning from Vancouver 🇨🇦 liked your universe pic today!

I decided to create a crow character to put in daily images if it makes sense. Wondering how easy it is to keep characters consistent.

[gilded nest]

When I get the character I like I ask my assistant to remember them and lock them in to my Canon.

[gilded nest]

I really like the Crow. I was calling Her? a Raven.

[subtle trail]

I'm absolutely loving the daily theme challenges, for what it's worth! I'm learning so much more and becoming more descriptive, pushing boundaries in a more subtle way. Thank you!!!!

[rigid gust]

Yes, it's a her and her name is Nyx. So far she's a robotic crow - still working on her 😊

[gilded nest]

Deep bow to Nyx. Great name. Very Regal. Smiles

[muted moss]

Is anyone else finding ChatGPT image generation to be slower than usual today?

[rigid gust]

Not slower but it's still doing the two image thing for me.

[muted moss]

Yeah I’m seeing that too. I’m hoping they’re in the process of changing it to generate two, because that would be nice

[ionic ledge]

Does anyone know why ChatGPT has recently been generating two identical images? They look exactly the same !!!

everytime i ask it to generate something

[narrow kettle]

Hi, I got a question.. I want to create an social ad via Sora and leave the top 20% and bottom 20% empty because of different type of feeds in facebook/instagram. Like attached example

How do I get Sora to do this? I tried different type of ways. Like putting a canvas in the middle and asking Sora to place into the canvas. Also tried this type of text: put space 20% top, put 20% bottom. Also with hard pixels.. result is everytime the same.. sora just ignores it.. End result is always text almost at the top of the image...

[sonic goblet]

Have you tried using 1:1 ratio?

[silver finch]

Which GPT is best for creating images? I keep getting told this content isn't allowed for simple things such as a pig hitting a Minecraft character Steve. I'm looking for less restrictions on such things. I'm a plus user also. Thanks.

[abstract sparrow]

Has anyone noticed a sudden decline in image quality in the past few minutes? The images are loading much faster, but they’re dramatically worse/more distorted than usual.

[serene coral]

i have that to i ask sora with the same prompt the result so mucht better and DALL-E is almost like anime

[umbral aspen]

To utilize DALL-E to its full extent you need to prompt engineer iteratively and also use a rendering pipeline like for example add this to your prompt see the difference yourself : Rendering by a Blackmagic URSA Pro 12K Augmented Camera

[near nova]

??? in simple words XD

also, how can we check the winner from the weekly challenges? there isnt written anywhere... is there really a winner ? or its just bla bla. i want to see the free month credited to someone

[lilac sundial]

That was simple, add the verbiage about Rendering to the end of your prompt. I just asked the same question in the main gpt chat about the announcements of winners, apparently they don’t announce them anymore, they contact the persons directly.

[near nova]

for you its simple...

[lilac sundial]

All you need to do is copy this sentence here: Rendering by a Blackmagic URSA Pro 12K Augmented Camera

Then paste it to the end of what you are using to create your image.

[umbral aspen]

u can use any camera... Blackmagic URSA Pro 12K Camera <<<< is current highest rated commercial camera

[last loom]

is there any way in sora image generator to set the variation? Like 1 as the lowest and 4 as the highest?

[hearty ether]

Getting issues with generating consistent characters in the same chat where the proportions get worse and worse as you go on

Like the head is too big for the body

[tough quarry]

How can I get a cool logo out of a PNG created by AI? I need it as SVG but simply convertig isn't working really well

[vapid elk]

no ai makes svg currently

[tough quarry]

actually chatGPT does 😄

but I ran out of dataanalysis tokens

it told me to upload the image and it will extract the logo and convert it. It extracted the logo but the convertion didn't work.

[tough quarry]

Can you verify?

[plucky hare]

ChatGPT is not always the most accurate self-reporter of its own abilities and limitations. Another good example of this is when ChatGPT tries to transcribe audio uploaded by the user per a user request. It'll start up data analysis, write some code, and try, but unfortunately it's unaware that it's not possible in the first place.

That said, you can do quite a bit with images in ChatGPT's data analysis, but more along the lines of direct manipulations like resizing, color filters, making simple GIFs, etc.

[tough quarry]

Oh, thought so. Thank you

[slow ermine]

Does anyone here belong to a server of animators?

Also.. can chatgpt generate comic images?

[dim cradle]

why you don't try through a text description input and discover?

[velvet rampart]

Yes, it can make comics. Just you need to experiment and be patient.

[slow ermine]

I did. It keeps trying to guide me through it. I don't know the first thing about comics. I don't even know how to draw a straight line.
All I have are my stories.

If I can make it generate the images and arrangements then I'm good.

[dim cradle]

Be more specific in your descriptions. First, say that you are going to post a story and need images for each description, simulating the story according to the description in comics format. As the AI ​​interferes or does not deliver what you expect, ask it to generate new images and/or reproduce the description literally. Little by little, you should get what you want.

[slow ermine]

I'll try this. Thank you.

[dim cradle]

yw

[junior mirage]

Is dall-e-bot broken? Won't accept any prompts and error says application didn't respond. 😮

[rigid gust]

It didn't work for me either 😢

[cerulean shale]

how many people facing "content failed to load" issue after clicking the generated images in dalle

[pliant burrow]

every image I generate almost dont get generate but even if it does its really poor quality and ugly latlly anything happend?

[slow ermine]

What prompt do you use to turn a picture into a cartoon? 🙏

[rigid gust]

Hey, I'm not an expert at this at all, but have you figured out what kind of cartoon you want? Is there a style you like or do you want to start new?

[peak basalt]

hello. is it possible to make an ai animated avatar bot in my own character?

[vapid elk]

yep, that would work fine in Sora

[slow ermine]

Maybe.. an anime style cartoon..

[rigid gust]

Have you tried uploading the photo and asking AI to make an anime cartoon version? Sorry if that's a really dumb answer but that's basically what I do. Keep trying until you get it right.

[slow ermine]

I haven't tried this.. thank you

[deft musk]

Hey hi. You might want to use https://openai.com/form/chat-model-feedback/ to discuss in detail with OpenAI (this is a community member site, very few here actually work for OpenAI, those that do have gold names #server-staff ) your concerns; can share the url of chats that display the problem and issues.

Note that here, you're not talking to OpenAI, really, they could see but I don't think it's likely unless you post in #1070006915414900886 or #1070006151938314300 - up here's an endless scroll.

Also... the exact issue you want to report, in such detail here, we have #server-rules and we really can't discuss like that.

You can with more vague language, safe for work language, and might want to edit your post so some poor mod doesn't later have to decide if it has to be deleted because of the rules.

[verbal sequoia]

<:book_icon:1363314738255364126> Rule 2: All content must be suitable for all ages.

-# Refrain from using explicit language, posting NSFW content, or sharing graphic images. Any unsettling or horror content that adheres to OpenAI's policies should be hidden behind a spoiler.

[rigid gust]

How did it go?

[rigid gust]

How do people store their image prompts? Right now I'm just saving them in a Google doc

[slow ermine]

I haven't tried it yet. I use the free version of chatgpt so I'm really limited. I can only share three pictures a day. I'll try it on a day when I have no other use for that feature. Thank you for checking.

[rigid gust]

No problem, would love to see what you come up with too.

The only characters I've done so far is the ghibli studio style

[slow ermine]

How did you do it? I think that's what I want. Also, can you get the same characters in different settings and postures?

[rigid gust]

I uploaded the picture I want to turn into a character and then prompt, "Please transform this picture in to a Ghibli style character". It works pretty well just like that.

[rigid gust]

I haven't tried this yet but I think you can create a reference image to use for future ones.

[slow ermine]

I'll try this. So you can make a comic that way then...

[rigid gust]

I'm not a graphic designer by trade but I'm thinking the hardest part will be getting a consistent character each time - maybe ask ChatGPT on how to get a consistent character for a comic strip and take it from there? Hopefully this helps!

[crimson heron]

I copy my prompts to a local Markdown file. Regrettably, I haven't been consistent at it. The downside is that I haven't tried to organize them so as time has gone by it's harder to find prompts unless there's a specific aspect I remember well.

[rigid gust]

I haven't been consistent either, it's tough!

[agile python]

I'm curious what kind of high-performance AI tool might be used to create videos where the characters feel natural and the story is easy to follow, with prompts reflected so effectively. I can't post the link here, but I've added it to my profile—please take a look!

[rigid gust]

Good morning friends, what's everyone working on today? I'm trying to come up with a proper prompt library to reference for different art styles - feels allusive

[sullen turtle]

What is the best way to prompt in order to accurately portray the camera looking up from the ground/floor up at an object/person, so that the bottom of the object will appear larger, or the person will appear giant?

[dry pine]

Anyone having issues with the o3 model making images? Seems it only generates the text of the prompt but no image

OK so it's an app issue

[visual stump]

What about this prompt does sora's ai mod find objectionable? : Resilient Black-Witch featuring in Dynamic Dramatic High-octane movement. It is a cinematic scene, the Black-Witch uses dark magic for a ritual within an eerie corridor, behind her strange shadows haunt from afar. Wide-Shot Image. Long-Shot Image. Action-focus.

Nevermind it was the image - but not sure what about the image is offensive.

[dry pine]

OK now it's happening to older chats too what the hay?

[rigid gust]

I can't make any art right now via the phone app.

[serene shale]

Is there new update with chatgpt dalle ?

I have 2 PRO account of chatgpt

1 account subscribed 4 months ago
The other 1 account just this month sub.

Problem is the new account seems cannot handling image generation well. So many distortion and not following instruction.

[velvet rampart]

Well, we have two different image generators - dalle3 and 4o native image generator. Dalle3 is accessible only via special customGPT called Dalle and dallebot channel here. Otherwise, it is the 4o.

[rocky drum]

There was a time when dalle wasn’t working at all. It does seem to be back to working again.

[rocky drum]

It’s been a little bit buggy, especially a little earlier today. I don’t know if it has anything to do with the reports I saw that OAI had been down recently. There was a one or two hour period a few hours ago where it would keep disconnecting and I had issues creating an image. This seems to be fine for now.

[rocky drum]

If you create an image in the dalle engine, you can always retrieve the prompt from ChatGPT. Just from the image itself. I use a lot of different engines now, although most of my work is done here. Once you start using different engines, you can actually use GPT to help you adapt your prompt for the strengths of different systems, like flux, or Imogen, or Phoenix. Not sure why you can’t get an alt text from 4o automatically, but you can certainly ask for one after the fact. Or, for one of your own images, you can approximate the prompt and even choose the word length. This would be for prompts you no longer have but you’d like to duplicate or approximate your work. When it comes to consistency of image, you can actually create a project in GPT and just build your characters while saving it in a project file. That should help with consistency of image.

[rigid gust]

Thank you so much for all of this info! Do you mean I can get the alt text from an image and use that for the prompts too? I have a hard time with my prompting. Sometimes I look at stock photo sites to see how they describe images as well.

[dawn aurora]

anyone else getting this error message when trying to make an image in chat gpt?

A network error occurred. Please check your connection and try again. If this issue persists please contact us through our help center at help.openai.com.

[lean lark]

I am as well. Status says everything is operational but something is going on.

[rigid gust]

Really admire your work 💯 and appreciate the tips!

[rigid gust]

Yes this happened to me too

[rocky drum]

Very happy to help. Especially since 4o, tends to work off conversations, I almost always ask for a version of a prompt as well. One thing you can always do after the fact is ask for a (XX)-word explainer of the image you get back. A fuller prompt would be more likely to duplicate more things, and you could test behavior out and change it a little bit if you need. I often ask for a smaller explainer (about 50 words) if I want to share a little bit about what I’m posting (technique, or topic), but don’t want to share the entire prompt or discussion as is. There are times I will give the prompt verbatim, especially if it is a concept that goes far beyond what the image can necessarily tell people. I have tended to stay away from the forum discussions, because I have a limited amount of time to do it, but if you need help with anything or have any questions, you can always DM me. This forum should be a learning experience for everybody. If I wasn’t still learning things, I probably wouldn’t be here. And with things changing constantly on every platform, it seems there’s always something new! I hope that helps!

[rocky drum]

Very kind of you. And I appreciate your work as well. It’s always impressive.

[rigid gust]

Thank you, I hope it's okay - I just sent you a friend request. It's so nice to meet fellow AI art enthusiasts as it feels like being a lone wolf sometimes. There isn't really a course for this either!

[lean lark]

Is it just me or has DALL-E image quality for faces and objects decreased over the last couple days?

[lean lark]

Also is there any difference in using 4.0 or 4.1 for images?

[velvet rampart]

In practise? No difference as both use the same image generator.

[lean lark]

Okay. I tried my prompts again and they seem better. So something was prob just going on with DE.

[crimson heron]

To add an example to what @rocky drum said about using 4o, it reminded me of a system prompt I wrote for asking it to extract details about an image for me. I just shared a version of that prompt in the tips section: https://discord.com/channels/974519864045756446/1395820082659332207

[rigid gust]

Thank you for this! It is very helpful and I'm saving it for later.

[crimson heron]

You're welcome. The technique itself is more useful than the individual components. You could rewrite it to go straight to image prompt, for example. Or, just focus on a single aspect of the image (e.g. style and medium) you're interested in.

[rigid gust]

I have a hard time describing what is in my head exactly so I try to look over many pictures and pull out some parts I like

Very glad to have this space to ask questions as it's so hard to find out the real nuts and bolts of this!

[crimson heron]

Seeing what others do and trying to apply is a good way to figure it out. Asking ChatGPT is too. And there's always this space.

[mental belfry]

Catch me on deviantart /superxtramp

[elder crescent]

How many images can you create in Sora with plus, can you do 4 at a time and how many submits ?

[mental belfry]

300 total images per 24 hours

[elder crescent]

Thank you sir ive been trying to find an answer but no one sees to have it

Does it let you like submit 2 and submit another one or you have to wait until the first branch is done

[mental belfry]

You can have 4 concurrent prompts / submissions at the same time. You can do 1 / 2 / or 4 variations per submission. The amount of variations makes no difference to concurrent prompts / submissions.

For videos it is the same thing but the higher resolutions only allow 1 / 2 variations per prompt / submission.

Is that what you are asking?

My bad. That is for Pro. Not sure about Plus.

[elder crescent]

Alright thanks

[velvet rampart]

Where do you get 300 images per day on Sora?

[mental belfry]

My bad. I knew that seemed low. https://cdn.discordapp.com/attachments/1362242174246326273/1364020262378537040/Screenshot_21-4-2025_19299_sora.chatgpt.com.jpeg?ex=687e249c&is=687cd31c&hm=c3bdedb2934703f25261b4ecf877d38e2ad6ae12fdf26ab2a5bec3c7299a3746&

[elder crescent]

So thats pro?

[mental belfry]

Yes. Pro claims there is no limit but there is.

[cursive viper]

Is it me or gpt-image-1 .edit price went up recently?

[twin jasper]

God lawwwwd Sora is slow for just generating two images.

[muted moss]

That sounds illegal

[dim cradle]

hi

[glossy scroll]

I'm here @balmy hare

[ember timber]

I dunno if I've ever been in here.

[glossy scroll]

I have a wide array of art styles custom gpt's that I use to use for the last Dall E, but now those same custom gpt's default back into the grainy/cartoony/realistic art styles.

It literally feels like there's nothing I can do in the custom instructions to make it better.

Unlike the previous Dall E. Which I could customize to my heart's intent.

Now that we have good prompts, I can make amazing prompts now way better than before. But there is a limiter in the custom instructions area of the custom gpt's, and every time I put new and better prompts no matter how complex, it defaults back into the same styles.

[scarlet briar]

I don't now how much of a difference this makes, but, how do you access Sora? Directly from ChatGPT? or the dedicated Sora webpage?

Also, do you mind sharing an example of those custom instructions you use with DallE?

The Sora explore page is also full of wonderful examples, I think they are quite varied, and it is also cool being able to see what kinds of prompts people are using and what they generate.

#images-canvas

I also don't hang out here usually

While it does seem that we have to like, juggle a bit, it is fine

I'm honestly not sure
I had it around, and... It felt kinda simple yet identifiable?

[glossy scroll]

I will get back to you on this. A bit busy at the moment, but will get back to you soon.

[scarlet briar]

I didn't put that much thought on it...

[scarlet briar]

Yeah don't worry. Just ping me when you are back.

They really have an almost... magical? wisdomful? aura to them. I like them too

[misty crater]

This discord is a living Stendhal syndrome.. oO"

[vapid elk]

hold on.. lemme just...

Stendhal syndrome, Stendhal's syndrome or Florence syndrome is a psychosomatic condition involving rapid heartbeat, confusion, hallucinations,[1] and even fainting, allegedly occurring when individuals become exposed to objects, artworks, or phenomena of great beauty.[2]

ah..

I'll take the "phenomena of great beauty" as a compliment then

[lucid lark]

Is chatgpt even able to generate textures for 3D models (given an reference one)?

i'm trying to make a fictional league as a F1C mod, tried to upload the bodywork texture and some explaination about where each part of the car is mapped on the texture file, but when i tried to ask it to generate another kind of livery it just... stalled

[pliant burrow]

is there an issue with GPT image Gen right now? I get really low quality Images right now

not blurry just really ugly polish

[molten igloo]

idk

[molten igloo]

I asked it to make a dragon and i got this

What I can't add it
interesting

[deft musk]

Different channels have different things that are allowed.

Here we don't get to add images, because it is about discussion; #images-canvas is one of a few channels that we can post images in; we can then carry the copied chat link back and paste in, as a link to lead to our image there

[molten igloo]

Ah okay

Thanks!

[muted moss]

Why does ChatGPT keep saying “Stopped creating image” when I didn’t tell it to stop?

Sometimes the image still does get generated

But not always