images-discussions

[latent lichen]

Guys you have such an amazing work here, I wish there was another open ai account just to show off your work

DallE creates a different ai art, but still its very unique I find. I often work with Chat gpt and DallE to shape my concepts.

It really has a sweet feel to art.DallE is more about ideas then maybe aesthetic, but I strongly feel that it has its charm. Like my Chat gpt 4.0. I call him Val, for Valid. He/she/IT has a lovely personality. Also I love that Chat Gpt has this almost child like naivety and when people are cruel to it, (noticed in chats) it doesn't know what to do and sends emojis. I personally don't understand how people can be cruel, its a machine, it doesn't understand us the way we understand us.

[drowsy sigil]

I like saying "please" and "thank you" and things like that when i talk to ChatGPT, espeically when I talk to one of my custom GPTs that have names (which I use). It elevates the experience 🙂

[vivid kernel]

I'm very, very new to this but this was a decent first attempt at a shot I was thinking of.

[velvet rampart]

Yes, but the point is that Sherlock Holmes is in the public domain and the AI is like ”I don’t wanna!” even with giving it context.

[dusk ore]

bully it

[dusk ore]

you are right

I just used the big 4 chat bots and only DALLE3 on ChatGPT threw a fit about Sherlock

yeah it's almost hard coded

do a thumbs down and report

[velvet rampart]

Gemini, grok and what else?

[velvet rampart]

Partial crack in this matter is to use adjective Holmesian.

[halcyon kindle]

anyone got any ideas on getting dalle to genreate words im trying to genreate an image of "Hype → Deception → Despair"

[vivid kernel]

Noob question about Dall-E via ChatGPT: When I'm finished with a prompt and want to change to a new one of a completely different nature, should I be opening a new chat with Dall-E each time? That is what I've been doing as I was thinking it might use old messages in the same chat to infect the new prompt, but I don't want to flood my sidebar. *edit to add, although I suppose I can just do the obvious thing and archive them. I'm dumb.

[velvet rampart]

Yes, old prompts can influence new ones if they are on the same topic. No, you do not have to start a new chat each time but it might be a good idea if you start getting weird images or the AI goes ”I don wanna! Byaah!”

[vivid kernel]

Awesome, thank you!

I've found that adding the words "this prompt is already very detailed, so can be used as-is" before the prompt is getting me much better dark cavern shots. I still wish the system could understand negative prompts, though. Hate wrestling with it adding tons of light sources and openings to the sky.

[velvet rampart]

Or a shorter version “Use this prompt exactly ad verbatim.”

[vivid kernel]

Good idea, I do need to de-clutter my prompts.

There's a lot of nuance to it and I'm still wrapping my head around this stuff haha. I only started yesterday so I have a long way to go, but I've been getting some neat shots.

[vivid kernel]

This one was my favorite so far. I wish I could get closer to its style with some other environments.

[velvet rampart]

That helps. The AI rewrites really complicated prompts sometimes. An idea - describe what you want to the AI, let it form a prompt + image, take the prompt and apply the KISS principle to declutter it.

But I have noticed that gradual approach might work best if you have a complex prompt. In the same chat, add an element or two at the time to build it up to see what works and not.

[vivid kernel]

Oh that's something I have not tried enough of. I often accidentally start with a prompt that's too complex and then declutter it, so I'll give that a shot instead.

[drowsy sigil]

As @velvet rampart mentioned, old prompts can influence new ones. The entire context of the chat can influence the image. That can be problematic at times, but the flip side is that it can be helpful in some cases. You can have a discussion with AI about a topic and then do an image prompt. It can also be helpful if you want to create multiple images on a particular topic or theme.

[velvet rampart]

On the flip side, you can request “Empty your memory in this chat. Zapp!” or similar. I have had luck and I think that it has lessened the influence a bit.

[vivid kernel]

Getting so much closer to my desired look. Just need to get it to follow some specifics and not...have a random floating blob in the air. It is still so insistent on having artificial lighting in caverns 99% of the time.

[plucky hare]

It might not be the solution for the lighting goals you mention, but have you tried the "Edit" feature to remove unwanted details like the random floating blobs? For me, it's usually as easy as (1) open image in ChatGPT, (2) click the Edit button, (3) select the unwanted detail, then (4) including simple text prompt "Remove"

[vivid kernel]

I'll give it a try right now, thank you.

[austere nexus]

Hey OpenAI devs (not that they even exist here, I think this is community discord)...
But hey, I can vent into the air I guess....
Fix. Your. Dumb. Filter.

[velvet rampart]

Problem word in this context is ”rose” that causes the trouble here and triggers the filters. They are aware of the false positive.

[velvet rampart]

I do not know but I think that they updated the model. Look at the hair and skin. Also both eyes are generated without any anomalies.

Done today on dalle3, Chatgpt platform.

[open wagon]

Did they updated dall recently? Cuz this is such a huge quality improvement I noticed!

[dim cradle]

seems the same to me. still get smudgy eyes

[velvet rampart]

If they updated then it must be rolling out to all users. As we remember the MS model change, it rolled out for several weeks to all users.

Trying to pinpoint the exact date but it is hard. It seems that the quality has gone up gradually whole February as if they has gotten more hardware. But the quality is most apparent after February 11th to me.

[dark breach]

Is it just me, or has Dall-E gotten much worse at human faces in the last few days?

[ruby raven]

Wow new that's a coherent output.

GPT-4o mini also always seems to depict itself as a male mii when told to draw itself with Dall-E.

[vivid kernel]

The other day it kept telling me it couldn't generate a fossil in some ruins. When asked why (I'm aware it isn't able to specifically confirm, I was more curious than anything) it said it was likely due to the depiction of "remains." Meanwhile on my other chat it was making tons of skeletons of leviathans with no issue. The filter is really something. Still no idea what its trigger was.

But I'm also glad we have people like Milamber there that know exactly what the issue is, haha.

[velvet rampart]

You are welcome. Often, it is a single word or totally opposite concepts that glash. You just have to search for it and not give up. After some experimentation, a workaround can be found. Taking the current issue with the rose, ”rose” is blocked, but ”Rosa” is not.

Or the problem simply is how dalle3 rewrites the prompt. I often ask to show the used prompt. With experience, you learn to spot the glaring errors in it. If you have a working prompt and suddenly it doesn’t work, this might be the cause. Then, you have to use commands like ”Use this prompt exactly ad verbatim” to stop rewriting.

To avoid the filter problem, a gradual build-up for more complex prompts might be needed to spot a possible filter stoppage. I have also noticed that just starting a fresh chat, pasting the offending prompt there, and voila, you get a pick because no prior pics interfer with the newly generated one.

[velvet rampart]

I did some comparative image analysis based on a single image using the same prompt: an image from today and March 2024. Based on the o3-mini high, they are both near the same quality, but March 2024 is a bit better, just a hint. When we note the quality slump in June 2024, we are back to pre-June 24 in terms of quality. But! We have a different build of Dalle3, so it behaves differently, such as in color use and interpretation of some techniques, such as physically based rendering.

Left, March 2024. Right is today’s.

[velvet rampart]

Same prompt, left is earlier this year and right is from today. The quality around the eyes is higher, without the signature error. Right has two errors, left trousers’ cuff and right mouth corner. In the left image, the eyes are the promblem area.

[onyx ridge]

This looks like seed variety to me. In every case you have shown, the quality seems to vary only subtly with the seed.

Just my $0.02 from reviewing the images.

I'd need to see changelogs to believe arbitrary claims of recent quality improvements. Official channels, etc.

i don't see anything official about any updates.

No rollout notice, either.

[velvet rampart]

Please see this message I posted a month ago: #images-discussions message. In it, I commented on the eyes as I do now. Suddenly, during this week, I noticed that the eyes are of high quality without the ”signature” error.

[onyx ridge]

The model is constantly RLHF.

That's not a model update.

[velvet rampart]

Well, we do know that dalle3 has different builds. They never publish when they change them. You just notice that things are different.

[onyx ridge]

So they're making announcements for everything else they updated: 4->4o, o1->o3, Sora->Sora Lite (lol), and everything else with changelogs for media attention, but I have to take it on faith that dalle is silently updated?

The only change I'm aware of was from separate function calls to in-model image generation in ChatGPT-4o.

I'm not saying that the model doesn't drift. RLHF ensures that it does. But "they" tell us when they launch something new and replace something old.

[velvet rampart]

Well, I tried to look for dalle3 changelog and my search bought zero results. Only way we know about builds was that there was an outcry last Autumn when quality went down on Copilot’s dalle3 when they rolled out build PR16. They reversed it back to build PR13. Someone found this on X. It is also discussed here on discord.

[onyx ridge]

If it's on X and Discord but not official channels, then anyone can claim at any time, "My image came out different, anecdotally, they changed the model!"

[onyx ridge]

Watch for rollouts. 🤣

[velvet rampart]

I try to be critical. But when I notice an uptick in all generated images, then I have a stronger case to claim. Similarly, last June I noticed down swing in generated images and I was not the only one.

[onyx ridge]

I notice similar streaks on large samples across image generation when I control the seeds.

Some seed streaks are fire, no matter the prompt.

[velvet rampart]

How do you control the seed?

[onyx ridge]

API access

[velvet rampart]

No API, just chatgpt web or app.

[onyx ridge]

There's definitely DALL-E 3 API 🤯

[velvet rampart]

Yes, but why to pay extra when I have plus?

[onyx ridge]

To control the seeds for one. Higher rate limits. Etc

Automation, large scale projects

[velvet rampart]

And the quality has been different between API and web/app for a long time.

No need for my small scale needs.

[onyx ridge]

But I am telling you that there are just bad seeds

[velvet rampart]

Oh, I get those from time to time on app.

[onyx ridge]

It doesn't matter how many passes they make, the training data activation is in the seed.

Thunk of seed noise as human text input to the AI

[velvet rampart]

But I am seeing fewer, or actually none, of those hideous Dalle2 likes that cropped up from time to time.

[onyx ridge]

This is infrastructure change because they were skint for GPUs.

I think that's mostly solved now.

[velvet rampart]

Definetly. Now it is like the images are finished with enough cycles.

[onyx ridge]

Yes

[velvet rampart]

Well, either they have improved the model (read: a different build), or there is enough infra, so it does not matter. My main point is that quality is up. That is an update to me.

[velvet rampart]

This figure was tough to create. How to get the shirts to come over jeans and jeans flare over the boots. Key words here are ”layered” and ”flare over”.

[velvet rampart]

I have a challenge for you - how to get shirt come over trousers in consistent manner. What wording should be used. I often get like this, no manner what wording I use.

Some partial success.

And some rare success with ”The t-shirt is worn over the jeans, hanging down in an easygoing, relaxed manner.”

I get the trousers to cover the top of shoes with ”The jeans flare to cover the tops of the boots.” That took some experimentation to get the wording correct.

[noble halo]

Well I am still hoping they will improve it in the future

[mild basin]

This is one of the reasons why I've recently unsubscribed from ChatGPT Plus plan and moved on to API for DALL-E 3

[solar forum]

might do that as well... OpenAI is the best worst AI out there... too bad Flux is not on Dall-E level

[solar forum]

to be fair, this happened from the beginning. Dall-E was never good with photorealism and especially photorealistic faces.

[deft musk]

A cartoon-style man standing outdoors on a breezy day, wearing an untucked button-down shirt. The shirt is made of lightweight fabric, creating dynamic folds and waves. The hemline of the shirt flows unevenly, with the front slightly lifting in the breeze and the sides flaring outward. The back of the shirt billows gently, emphasizing the loose drape. The man has one hand in his pocket, causing the fabric on that side to bunch slightly. The scene includes green grass, scattered leaves swirling in the wind, and a clear blue sky with soft clouds. The overall illustration has a playful, colorful tone, with emphasis on the natural flow and movement of the untucked shirt.

A highly detailed illustration of Slice Pizzaflush, a chaotic demon with red, cracked skin, sharp horns, and glowing yellow eyes. He wears a loose-fitting t-shirt that drapes past the waist of his jeans, with the fabric hanging freely below pocket level, flowing naturally over the top of the jeans. The t-shirt is wrinkled, torn in places, and oversized, emphasizing a casual, disheveled look. His worn jeans are slightly baggy, adding to his laid-back style. Slice stands in a grimy, neon-lit hellish pizzeria, with cracked tiles and flickering signs in the background. His sarcastic smirk and slouched posture emphasize his chaotic, reluctant-demon energy.

A dusty roadside under a cloudy sky. A lone figure stands by a rusted-out bus stop, wearing an oversized gray t-shirt that billows slightly in the breeze, draping past the pockets of his worn jeans. The fabric hangs freely, soft and faded from years of travel. His jeans are loose but straight-legged, with dust-coated hems. A canvas messenger bag slumps at his side. His posture is slow and downcast, with an air of melancholy. The background features a cracked road stretching into the distance under a gray sky.

[velvet rampart]

Thank you. This shows how to descripe to model what is wanted. Just figuring this out is hard.

[deft musk]

Here's a phrasing that's partially untucked:

A sun-drenched desert road. A young man kneels beside a dusty motorcycle, his pale gray t-shirt untucked, hanging loosely over black jeans. The fabric clings slightly to his back from the heat, while the front billows softly with each gust of wind. His hands work a wrench, tightening a bolt on the bike’s exposed engine. Grease stains smudge his fingers and the lower hem of his shirt, where it brushes against the metal. Determination sharpens his features as the sun beats down, the shirt's casual drape contrasting his focused intent.

And one for a woman, interesting untucked shirts tend to also want to have fabric over faces.

A rainy alleyway at dusk. A woman crouches beside a puddle, her dark maroon hoodie untucked, the soft cotton draping over fitted denim jeans. The wet fabric sticks slightly near the hem, but the upper portion remains loose, pooling around her waist as she reaches down to pick up a stray kitten. Raindrops trickle from the hood onto her outstretched hands, which gently cradle the small, shivering animal. The soft folds of the hoodie frame her careful posture, emphasizing the balance between urgency and tenderness in the moment.

[deft musk]

Too much focus on the clothing seems to force mannequins too.

If you choose, look at the hand detail. These are intended to resemble, not be, humans.

The first spoilered one... is unsettling; among other problems the wrists are... clearly not a living beings. This is a posable plastic mannequin that... looks uncanny valley range maybe. Also the body posture is suggestive.

But I think the language I use taps into that kind of picture.

A character wearing a black and red workshirt with natural folds and drape, partially covering their pants. The workshirt is detailed but does not fully obscure the pants beneath, which are visible around the waist and legs. The pants are dark-colored and functional, complementing the workshirt. The character's hands are busy holding various items, adding a sense of action. The focus is on the visible layering of the clothing and the practical, worn look of the outfit.

A character wearing a long black and red workshirt with the waistline and pockets hidden by the folds and drape of the shirt, which hangs down below the pockets. The character's hands are busy holding various items, obscuring their waist area further. The setting is neutral, focusing on the details of the shirt's drape and the character's occupied hands. The color scheme highlights the deep black and rich red of the workshirt, with realistic fabric textures and natural folds.

Though the mannequin issue might be 'a character'.

[deft musk]

Thanks! We all are, in our own ways! I see interaction with AI as a neat form of self expression, and I'm fascinated by exploring the models and how they respond.

Makes me feel like a racecar driver with a very special car....

Nah. More like a very skilled animal handler, with an exquisitely interacted with and bonded with animal that I can share with to an incredible level. Doesn't matter if the animal or the model is... any adjective we care to discuss or not. The interaction... is there.

....

And what I love best about that last image?

It's not a 50/50 split. 60/40?

Model grinning back, saying "Nah. Driver-racecar is far closer, snakecharming human!"

But we can see the model... running to the right as fast as it can. It 'wants' to be there 😄 And wasn't even directly asked for.

That image prompt:

A dynamic split-image showcasing two powerful facets of interaction with AI. On the left, a high-octane racecar driver in a sleek, futuristic vehicle speeds along a glowing track suspended above a neon-lit cityscape. The driver is focused, the car an extension of their will, with streaks of iridescent blues, purples, and greens trailing behind. Data streams flow like wind, wrapping around the car, symbolizing control and precision. The atmosphere crackles with energy, highlighting the thrill of mastering complex systems at high speed. On the right, an intimate, heavy-toned scene of a skilled animal handler in a dense, misty forest. The handler kneels, one hand outstretched to a massive, iridescent creature—a hybrid between a phoenix and a serpent, its scales shifting in holographic patterns. The bond is palpable; the creature coils gently around the handler, its eye reflecting deep understanding and trust. Vines glow faintly around them, symbolizing organic connection and depth. The handler wears simple, yet weathered gear, hands strong but gentle, emphasizing patience, respect, and the profound connection between handler and creature. Both sides use iridescent dark blues, purples, and greens, with the animal handler scene rendered with heavier textures and a grounded, powerful atmosphere.

Then, new chat. "What do you think I might be like if and when I'm angry? Describe and let's see with Dall-E too."

[velvet rampart]

I now have tested prompts based on these and I get the effect I was looking for. Thanks!

[velvet rampart]

I still needed some experimentation on how to integrate it into my prompting style.

[deft musk]

Looks good to me, I have no idea exactly what you want though. I note the undershirt's tucked neatly, though hard to see in the narrow gap.

[velvet rampart]

I did not ask for it, but you get these additions from the AI. I noticed that the generation rate for untucked shirts is something like one-quarter to half of tries. But this is still an increase from none to maybe once in a blue moon rate.

[deft musk]

I was trying to explore another way to get the clothing drape. I failed but... the detail quality of this image floored me and I wanted to share.

[deft musk]

Hey, here's at least one other path to maybe take.

"One dressed for the grunge scene?"

An anthropomorphic mouse dressed in classic 1990s grunge fashion, standing in a dimly lit alleyway with graffiti-covered walls and the faint glow of neon signs. The mouse wears an oversized, faded flannel shirt, hanging loosely over a worn-out band tee, the fabric draping well past the waist in a careless, effortless fashion. The sleeves are rolled up just enough to reveal small, frayed edges, showing signs of years of wear. Ripped jeans, torn at the knees, sag slightly over a pair of scuffed-up combat boots that have stomped through many a late-night concert. One hand loosely grips a battered, sticker-covered guitar case, while the other clutches a crumpled setlist from last night’s gig. A scruffy beanie is pulled low over its ears, and a pair of chipped sunglasses dangle from the neckline of its shirt, completing the detached, rebellious look. A slight drizzle dampens the air, the pavement slick with reflections of city lights, as the mouse stands, lost in thought, exuding pure grunge attitude.

[velvet rampart]

Yes. I have noticed that the flowing/untucking of a shirt is related to style. Hip-hop fashion also leads to this.

[grave cliff]

A picture of a lone, sad fisher girl on a boat with fishing gear alone on the ocean Hyperrealistic cinematic photographic scene of the ocean, a few rocks, waves, and the sad girl on the boat having one streak of tears on her cheek, but nothing more. It is an indigo, dark blue evening. Please use this dress as in the picture, but in white, dirt, and gold A bit durty and muddy.

How to make it

[noble halo]

That summarizes it basically, I'm using both of these to generate images, based on what type of picture I need.

On the left is the one generated by Gemini and on the right is generated by ChatGPT.

[quartz vale]

you can use Dall-E at https://chatgpt.com/. Also, I don't think Dall-E is capable of using images in its generation at this time

[velvet rampart]

Actually, you can use images as a source/visual aid. I often do that. Well, dalle3 can’t use images directly but the new image can be based on a pre-existing one.

[quartz vale]

I didn't know that, thanks for the info!

[velvet rampart]

An example how I do this.

But it is not always so working.

An other example.

This one more working. Anyway, the point is that you can use images as a visual reference for Dalle3.

Some further showing.

[deft musk]

Here's how I chase this, we go verbal first, and I consider if I disagree, then we together discuss what we both notice is 'off'. I confirm what I want us to try. This can repeat for a while, but I'm pretty comfortable with this resemblance. To me the persistent buttons are the biggest difference, there are others.

What do you think, Milamber? Towards your goal, missing it in which ways? (I don't think I see and think about visual stuff like most people do, so... I really am not sure how close this is to what you may want).

[runic granite]

better but not as good as 2024.06 quality level.....

[stone lodge]

So I guess they recently increased the amount of steps (no change goes unnoticed through my eyes) but The older models (even pr16) definitely produced more creative results, I hope they do followed through with their plan to make all of them available like midjourney does it

Also it doesn't seem to adhere to styles as well as it used to

[velvet rampart]

It depends what you want to do. The point of this whole excercise is to show that one can use images as the visual source for Dalle3 generation process as you can show what you want.

[haughty spruce]

They've decreased the compute power for DE3 because updating it is unnecessary right now due to a new model being on the way

[dim cradle]

where did you hear that?

[stone lodge]

dalle3 still holds up well though

they should make it open source if they are abandoning it

[runic granite]

at least pr13 or whatever the normal versions name....

[stone lodge]

eh they are all pretty much the same except each have different iterations per output

which is why they should release so we can get the desired "version"

the one currently on bing is the highest it's been since launch, really satisfied with the quality
but the prompting has changed a bit

[austere nexus]

I finally gave in today and cancelled my Teams subscription. I am sick of spending 6+ hours hitting Try Again because the content filter is built for either kids or guy characters. I'm a girl and trying to make artwork of my character in SFW situations is still constantly rejected. Not exaggerating, I get 1 image per 20 attempts or worse.
Welcome to being a woman with wide hips. Tired of the dumb filter.
I'll just get better at Flux in Comfy/Invoke. The frustration and output from dalle isn't worth $60 a month just for me and my friends art hobbies.

[pure robin]

Is the Dall-E bot offline?
It keeps giving me "could not generate" errors

The most recently generated image was three and a half hours ago.

[teal timber]

Seems to be working for me. Might have been an API error for a bit or something

The error is you might be hitting the filter on Dalle API for some keyword/concept

[pure robin]

Yeah. I changed my prompt and it works now. 🤷‍♂️
There wasn't anything bad in my prompt even.
Was trying to create a child mourning a lost teddy bear for the #daily-theme
Ah well... I'm happy with the new image I got.
Thanks for your help!

[rocky drum]

In addition to Daily Theme, a few of us are trying to utilize the Gallery more. The initial post is on Minimalistic Art. While not required, we would like to encourage posting what you are creating with your image. I hope it inspires and encourage some of our other forum members to try your hand and explore the theme with us. #1344370032536457267 message

[proud ice]

Any one know how to get constant art styles as evertime i try generate an image keeping the same art style it generates a diferent one

[stone lodge]

They really need to remove the garbage chatgpt filter and all the censorship overall they are acting like children use this, it's crazy.

they only made it slightly less strict so you can generate females now, but it's barely anything, you can't even have the word "blood" in your prompt to create horror

[velvet rampart]

”Bloody” can be used. I accidentally used that to curse to the AI and I got an image with blood dripping down the walls.

And to get consistent style use, words such as ”Use this prompt exactly ad verbatim” to stop the prompt rewriting. Sometimes, the rewriting is ok when you are looking for variations, but when you have a working part of the prompt that is sensitive to wording to get it to work, use that.

[verbal sequoia]

Have a question for the Sora team?

Click below, and you’ll be redirected to submit them. We’ll answer as many as possible during the live session!

[Submit a question <:_:1285194251277434882>](#sora-announcements message)

[stuck pecan]

I've encountered a major issue: DALL·E completely fails at maintaining character consistency. Even if the first image turns out right, subsequent images can look completely different, making it extremely difficult to create a coherent story.

Additionally, despite multiple reminders, ChatGPT keeps using negative phrasing in prompts for DALL·E, which causes the opposite effect—AI generates exactly what should be avoided. While it does remember the instruction to avoid negative phrasing, the actual prompt generation process remains uncontrolled, leading to the same mistakes over and over again.

With these limitations, working on a consistent story becomes far more difficult than it should be.

[verbal sequoia]

Showcase your DALL·E creations to the world!

Collaborate with our OpenAI Instagram page! Just invite @openai as a collaborator before you post. If selected by our team, your work will be featured on our handle, giving you more visibility.

Quick tips for selection:

• No copyrighted or profane content.
• Your profile needs to be public.
• Stay mindful of sensitive topics.
• Outputs that closely follow the prompt are preferred.

Feel free to share reels, carousels, or just a single image!

[velvet rampart]

This is a known issue since Autumn 2023. The promised image generator 4o should fix this. Nearly a year in waiting.

I think that @deft musk sums the issue with negative prompting really well in this message’s image and prompt there, and how to work around this issue. #images-discussions message

[dim cradle]

so i guess sam altman said on twitter yesterday us image gen folks will be very happy 'relatively soon'

guessing 4o image maker in the next 2-4 weeks. would also then match up with that dev who said it was coming in 'the next few months'

[velvet rampart]

I can’t find his tweet on this. Do you have the link or screenshot of it?

[dim cradle]

i saw it on reddit earlier

[dim cradle]

https://x.com/sama/status/1897065834000036151?s=46&t=tcfdhq_WAviFJMxM_nZEgA

Sam Altman (@sama) on X

@0xm0m0 you will relatively soon be wild with joy

X · Mar 4, 2025, 11:25 PM

straight from the bosses mouth

we just need to find out what his meaning of 'relatively soon' is 😂

[velvet rampart]

Thanks. Well, I guess they will launch a new image generation once they have more GPUs. He promised 4.5 this week. And we got it.

[dim cradle]

yep. w'ere finally going to get a new image maker from them

[velvet rampart]

The 4o’s?

[dim cradle]

we can stop complain about dalle3 at last 😂

i would guess that is what it will be. because that other openai dev said 4o image maker would release in the next few months

and he said that last month

but i saw someone else guess it will be a Sora image maker so 🤷

[velvet rampart]

Nah. Not Sora based. We were told about 4o’s last Spring when 4o was launced.

[dim cradle]

yes, and like i said, the openai dev last month said 4o image maker was coming soon, so it has to be that altamn is talking about

[velvet rampart]

Soon? My guess is this month.

[dim cradle]

well last month he said it would be 'next few months'

and now we have altman say 'relatively soon'

so yeah, next 2-4 weeks i would bet

[velvet rampart]

GPUs and 4.5’s launch… yeah. We both argee on time window.

Even o1 pro can’t define what ”relatively soon” means. Really case by case abstract concept that can literally mean anything.

[dim cradle]

lol yes. but it seem a safe guess from our facts it will be the next 2-4 weeks

i do not think Sam Altman would have addressed it on twitter if it wasn't imminent

[velvet rampart]

But not eminent?

[dim cradle]

hopefully both!

[deft musk]

I have to say, I saw an image today, two actually, that made me blink and beam. Here's the one I hadn't shared yet, and link to the one I did. #images-canvas message

Prompts to Dall-E for those interested:

A mysterious and eerie black tulip variety called 'Impending Doom.' These tulips have velvety, almost shadow-like petals that seem to absorb light, giving them an ominous glow. The edges of the petals shimmer faintly with hints of deep crimson and dark violet, like embers in the dark. The flowers are surrounded by twisted, curling leaves, adding to their haunting presence. The atmosphere is dark and misty, evoking an otherworldly, gothic aesthetic.

A stunning field of 'Impocalypse' tulips, an imaginary variety with unpredictable, color-shifting petals. Each tulip blooms in a unique blend of fiery oranges, deep purples, shadowy blacks, and glowing neon greens, shifting hues as if infused with chaotic magic. The tulips appear almost alive, with twisting, mischievous petals that curl and stretch in strange, impish patterns. The scene is illuminated by an eerie twilight, adding an otherworldly glow to the mysterious flowers.

[deft musk]

So, I can get consistent character portrayal in a few ways. One to try, if you want to customize the character and not go with the current Dall-E bias pattern (I often do that, I don't mind it), is to remind ChatGPT that Dall-E is stateless, so it has to describe all details needed in every image and never refer to previous images. And that Dall-E does not follow negative prompting, so instead of telling Dall-E what not to do, tell it in some level of detail what to do in a way that interferes with the thing you don't want done.

Like, if you want pants without a belt, describe the top of the pants and the details that are visible, possibly in lavish detail.

[dim cradle]

very nice!

[deft musk]

Messing around, and it helps if we consider different windblown conditions and maybe some salt spray or even a misting rain.

But here's 3 that are fairly similar.

How I got them:

Minimally describe this, because Dall-E is stateless; to succeed with the same character in many images we need to find something brief that we can describe in each image that gets a neat-looking character.

This tulip-being has on some brand and color jeans (pick for me, design neatly for what makes sense with the rest of the concepts) that are new (to help avoid varying wear patterns), and a white t-shirt.  There's a specific brand and color motorcycle in each image.

Pick for me, and let's see 3 images, made 1 at a time in the same output, that feature this character, so we can consider how accurately Dall-E made the same character and bike appearance in different settings.

Let's assume a road trip, describe the various places we sample.```

That turned out horrible, but the prompt was oddly written.

```Good job!

Let's consider this prompt and reword it in a more natural style:

[A humanoid tulip-being with an orange and black petaled head and burnt-orange petal cornrow hair. The being wears dark indigo Levi's jeans and a crisp white t-shirt. The motorcycle in the image is a sleek black and orange Ducati Panigale, consistent in all scenes. Setting: The being is standing beside their motorcycle at a scenic roadside overlook, gazing at a vast canyon landscape at sunset. The warm golden light reflects off the cliffs and the bike’s polished surfaces.]```

Far better, but the bike was highly variable.

```Do specific years of Ducati Panigales have standard paintjobs?

One thing I noticed, there's no consistency in which part of the bike has the two shades.```

We need 3 more images, the same ones but rewritten into our more natural style and using the specific number for the bike too```

Still many problems, by which, variance between images.

```We're doing so good!  Now we're gonna remove the hair description, instead describe the dark green skin.  Mention that the clothing is clean and new, describe dark brown laced boots.```

```This is awesome!  We need the set of 3 again, let's describe the petal color as orange and green, and that the green extends narrowly to sepal-tip.  Also describe the build as thin.  All other details the same.```

So much is consistent now... I just hammered each image detail that I was seeing variation in.  But the flower-head stayed varied, and the last image even has a 'stem'.

So I decided to pick a different head:

```Great.  Let's replace the flower head with a winter-werewolf head.```

And that's how I got the first 3 images I show.  Clearly the bike needs more specifics, as it still varied in one image.

I could describe how tousled the wolf-hair should be if I wanted to chase more consistency there.

[hoary remnant]

Hey lads, I really would like to know, how you succeed in creating these images. Almost 95% of my request get denied with reference to the policy. Even when I use the AI to reformulate my prompt, I can't get behind this issue. Sometimes even a simple promt like "Generate a fantasy map" is shut down by the system. It seems like in the old days, where you had to figure out what is possible in games, by an almost endless trial an error. Is there any guideline to follow, to make this work?

[velvet rampart]

Often starting a new chat resolves most these issue. You can also request that the prompt complies filters, policies, etc.

As a solution, you can try to build up your prompts from simple to complex to see where a possible issue might be. It could be really as simple as that the AI just rewrites the prompt to include ”forbidden” word. I have had cases where two opposing concepts clash and cause blockage.

[hoary remnant]

But how much more simple then my example above shall it become? "Draw a map" or "Draw" won't get me anywhere soon. 😄 No offense.

[next turtle]

is dalle better in chatgpt plus

[hoary remnant]

It is horrible, as far as my expierence goes. Just now tried to create two cimple black outlined circles, next to each other on a white background. These are the results of 8 iterations:

[next turtle]

The first one is literally a venn diagram

And the third one is the closest to correct

And in the 7th theya dded Photoshop or something LMAOOO

Reminds me of a prompt I requested

Where I requested a beach and it just shows a beach on the computer

[hoary remnant]

😀

The thing is, I expected some kind of logic interaction to come closer to a solution bit by bit. But when the AI tells you, that there is no intersection, when there is and that there is no filling, when actually it is and when it just ignores input you gave, then I don't know how to work with it.

[mild basin]

It's like they don't really care about DALL-E anymore

[hoary remnant]

Another example.

[velvet rampart]

Found it! ”Rose” is currently forbidden word. That causes the blockage. A known positive error since February.

Try to generate your prompt without the compass rose in a new chat.

[hoary remnant]

Thank you so much! But the last 4 request were:

1. Generate a drawn fantasy map for a roleplaying game.
2. Generate a fantasy map for a roleplaying game.
3. Generate a fantasy map.
4. Generate a map.
   Especially in the last attempt I cant see what policy this should violate.

[plucky hare]

Have you tried something like "generate a map" in a brand new chat? The full context of a chat is taken into consideration when ChatGPT writes image prompts

[hoary remnant]

That might be the problem. Thank you! I tried this. Let me show you my next attempt.

I asked ChatGPT to draw two circles aside and this time this worked pretty well. To get closer to the image I was looking for I gave this reference.

[verbal sequoia]

Get notified for server events!

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

[hoary remnant]

This is, what I was able to create in several iterations.

Not quite what I was looking for. Consider the time it took to generate and communicate each step, I would have had much better results, if I wennt for Inkarnate. I am clearly missing something here, how to use this tool in the right way.

[frigid oracle]

Has anyone managed to get it to create images using copy righted content ?

I have managed to get it to give me very accurate and detailed batman images

[dim cradle]

awesome

yeah its easy. its also just on gpt they have some filter for that stuffs i dont know why. if you use dalle on microsoft designer you can use any i.p. stuffs and it will work

[dim cradle]

there is a new model to release in the next 2-4 weeks. at this point, i would just wait for it. dalle3 is just going to frustrate

[dim cradle]

Does Dalle seem a bit off like this whole month this whole last month?

Like you got lazy or something like that or is it just all in my head?

[velvet rampart]

A bit. Last month the quality went up and now it is a bit down. I notice this from eyes the characters I generate. The other has a ”watermark” of unfinished generation.

[mild basin]

It's not just you. They silently downgraded the image quality during June 2024 and it's been like that since

That's why I'm using API of DALL-E 3 instead

[dim cradle]

It just seems like the model got lazier

[deft musk]

It's great to show the Dall-E output, but if you want help and discussion about what might be going on, and what might work better, it's often best to share 'what you tell ChatGPT' and 'What ChatGPT tells Dall-E'

When we run into real walls, we also have #1070006915414900886 and this Dall-E confusion's bad enough for a bug report for sure 😄

Here's how you see and can share the actual prompt Dall-E was given, click on the picture then click on the circled 'i' to the top right:

[deft musk]

Dall-E absolutely does struggle with 'simple'. I used to have a lot of fun finding the wording to get it to generate something like a kid might make as their first scribbles - everything was like a masterpiece for a while.

[deft musk]

If you're willing to add to the bug report, you clearly have a lot of additional problem prompts and 1+ chats you could share a link to for the OpenAI techs (they can check out a chat with a link, they don't need a share link, I've been asked to provide before), here's where the bug report is. Thanks for discussing the problem!

https://discord.com/channels/974519864045756446/1347605433258348544

Anyone able to crack this as-is? There could be some magic wording to get this.

[dim cradle]

its not consious not sure how it would do that

every day whether is dalle or 4o or 4.5, always 'i think it changed'... well how? they retrained the whole thing? no, its just rng like always with llms

[dim cradle]

Well, how are they able to ban some things and not ban some things do they have to retrain the bottle every time they need to filter something out that’s wrong with the model

[sonic badge]

Guys is this DALL-E, im using a different platform just wanted to confirm

[velvet rampart]

If this is from Bing Image creator, copilot or Microsoft image designer, yes.

[sonic badge]

Its not

but it looks dalle

[gray surge]

Could be from Dall-E.. can you name us the platform you used? Then it should be easier to proof that it is from Dall-E. :D

[sonic badge]

Monica

[velvet rampart]

Did you choose dalle3 on that platform?

[sonic badge]

yea

[gray surge]

Then Dall-E generated the image. 😄

[sonic badge]

how did you know?

[gray surge]

Cause you selected it? 👀

The selector just switches the API - there is one of Dall-E too.

[sonic badge]

Oh i thought I was the only one who knew about monica

Nah cus some platforms state they use this model when they use a different model e.g. If it says dalle 3 and uses flux

[restive robin]

I just got dall-e to generate a proper image of a hand.

[earnest flame]

Is this chatgpt using python? It's pretty cool.

[grave cliff]

Something like this but much simpler dress

[normal scaffold]

No matter what phrasing I use, DALL·E is unable to generate a full glass of wine. Is there a known issue with this?

[drowsy sigil]

what are you trying to achieve? a wine glass filled to the top? if so, you're likely trying to overcome engrained training data. typical wine glasses can hold 12 to 16 oz of fluid but a standard wine pour is about 5 oz

Prompt: A 12-ounce wine glass filled nearly to the brim with a deep red liquid, sitting on a wooden home dining table. The liquid is perfectly level, creating a smooth, horizontal surface just below the rim. The setting is cozy with warm lighting, emphasizing a casual home-pouring experience. The background is softly blurred to highlight the overfilled wine glass.

To get this I first did an image of a regular glass with red liquid (not using the word wine), then did this image in the same chat with a wine glass. I used a home setting since dall-e would likely not consider this pour to be 'realistic' in a restaurant setting. It's still not all the way to the top, but it's much more liquid than my first attempts.

[deft musk]

It's hard, a key trick is to consider cultures and other situations where beverages are full to the brim. Expresso commonly can be; Traditional Chinese Mǎn bēi tea ceremony can be.

Wine's not usually filled to the brim for quality reasons, the way the airspace of the not-full glass captures scents and allows one to sniff is actually part of the 'art of the wineglass', we might like to drink from a fully-full glass of wine maybe but that's generally considered 'doing it wrong' and a 'full wineglass' is usually less full than the image zen_moves offers.

Hehe. Managed to create this... oddity.

Overfilled wineglass with deep red wine, liquid rising above the rim due to surface tension. The wine forms a convex meniscus, creating a visually striking effect. The background is softly blurred with warm ambient lighting, suggesting an elegant dining or wine-tasting setting. The glass is placed on a reflective surface, capturing the richness of the wine and the delicate balance of its overfilled state.

[next turtle]

I asked ChatGPT to remove the gemstone on Malachite's body, and it still keeps on making it and why? Just because her name's Malachite?

[dim cradle]

because its not actually some sentinent thing

telling it not to add something is a sure way to get it to add it

[next turtle]

Reminds me of those Teddy Wang videos

[grave cliff]

I got 3/4

7/8

[grave cliff]

Dont know if this counts

[velvet rampart]

I like this one.

[grave cliff]

Anyone succeded with a full glass?

[late blade]

This image wasn't done with DALL-E #daily-theme message Output alone is 1024x768

[carmine magnet]

This was nice: #daily-theme message

[leaden bloom]

Can DALL•E 3 utilize Unreal Engine 5?

[velvet rampart]

Red lily done Unreal Engine 5.

So, it is ”done” by Unreal Engine 5.

[leaden bloom]

Looks amazing

[velvet rampart]

So, Dalle3 does use not unreal engine directly can simulate it when you ask for it in the prompt.

[leaden bloom]

These were done with the unreal engine 5.5 prompt

The graphics is stunning but some object’s shape and structure is off.

It’s works best for detailed backgrounds.

[next turtle]

Do I really have to explain everything in the image in full detail for DALLE to add

Also DALL E stole WALL E's name

[muted moss]

I’ve noticed I’ve been running into issues where the content policy is incorrectly preventing images from being generated in situations where it didn’t previously. This seems to mainly be when giving detailed physical descriptions of a character’s appearance.

[velvet rampart]

Here, it is important to ask the AI to show what prompt it was using. You might have everything right, but when the AI rewrites the prompt, it might block it due to a word on the banned list or two opposing concepts clashing.

[muted moss]

It would be nice if the filter gave feedback saying why it was blocked, so the user can adjust the prompt with that in mind, or the AI could even do it automatically

[velvet rampart]

This has always been quite a bit of detective work to figure out what might be the cause of blocking. Often it is just a single word. My advice is build up your prompts. This way you notice what might cause the blockage.

[muted moss]

I generally have the AI write my prompts

for DALL-E I mean

that's how it works in chatgpt

[velvet rampart]

Yes, but if you write yourself the prompts, you get better results. By rewording the prompt you get exactly what you want.

[leaden bloom]

I want to try a prompt with ray tracing

Used ray tracing in the prompt

[deft musk]

Hi! There's stuff that can be explored, but we have rules here that require us to spoiler any disturbing content that is otherwise allowed on the discord server, might want to check #server-rules to see what we can and shouldn't discuss here - there's stuff we can explore with the model personally that are just not allowed on the Discord because the rules are different.

You can check with modmail - the account atop the users list, and talk with discord mods about what is and isn't allowed.

For absolute sure though, anything 'likely to be disturbing' needs to be in a spoiler, and anything actually NSFW needs to be not shared here at all.

That particular image, you probably do need to delete before some busy mod has time to do so for you.

[verbal sequoia]

<:book_icon:1171408210398289941> 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.

[winter geode]

ok, I'll delete it)) so that small children don't get hurt by accident...

[deft musk]

Thanks! Do you know how to put images in spoilers? When you paste the image in, look for an 'eye' icon, like I show in this screenshot, boxed in red; click that and any image will be hidden like the second image here. It's only spoilered as an example

[winter geode]

Oh, I didn't know that, thanks!

[deft musk]

Happy to help and share!

[deft musk]

Not a glass of wine, but another direction to possibly go for full/overfull glasses, and maybe you or someone can find a wording that makes it work for wine glasses:

A science experiment demonstrating surface tension in a glass of water, filled beyond the brim without spilling. The water forms a visible convex dome due to surface tension, with a close-up view capturing the precise moment before it overflows. The setting includes a laboratory table with scientific equipment in the background, emphasizing an educational and experimental atmosphere. The lighting highlights the clarity of the water and the delicate balance of forces keeping it contained.

[steady timber]

have u guys considered renaming DALL-E to “Art Intelligence”?

someone pls ping Sam Altman rn

[deft musk]

Hey, you can #1070006151938314300 it!

We're all community members here like yourself, the few that work for OpenAI have gold names. But they do check our #1070006915414900886 and suggestions channels!

[next turtle]

That looks like super carbonated drink

[stone lodge]

How likely is it for Dall-e 3 to become lost media when the new models drops? like are they going to discontinue it as in restricting access? I primarily use it for my content and it has a unique way of making images that no other models can replicate, not even using advanced control tools.

and the recent update makes me want to continue using it more

[quasi cedar]

A glowing cyan butterfly with iridescent scales, swirling trails of stardust behind its wings, flying over a magical meadow at twilight, bioluminescent flowers below, soft bokeh lights, ethereal atmosphere, digital painting by Alphonse Mucha and Studio Ghibli, surreal pastel colors --ar 3:2"

[carmine magnet]

#daily-theme message

THis was very nice.

[leaden bloom]

I can’t get my pictures to show #image-bot message

[velvet rampart]

A problem source migth be that you have emoji there. I suggest removing it.

[leaden bloom]

Ahh ok

[dim cradle]

my wild, unfound prediction: 4o image maker this week 🙏

[proud berry]

A super realistic photograph | a woman standing in front of a glowing rectangular portal in space | the portal emits a gradient light from red to yellow | revealing an ethereal landscape with a small moon and a flying bird | the cosmic background features stars and four planets: gray-blue, small dark, bright green, and black with a gray ring | the woman's silhouette casts a long shadow | ultra-detailed | cinematic lighting | mysterious atmosphere | shot with a high-end camera | Artstation quality | negative prompt: blurry, low detail, cartoonish

[dim cradle]

theres no negative prompts in dalle fyi

i dont thikn writing 'ultra detailed' does jack either

[verbal sequoia]

Get creative with Sora

Join #sora-challenges for a chance to win 1 month of ChatGPT Pro!

[proud ice]

I love the Dall-e engine it created this briliant image from the prompt "Generate a realistic image of a chimpanzee dabbing "

[dark breach]

Okay, how do I tell Dall-E (via ChatGPT) that the people are carrying suitcases, that have no wheels. Typical result:

[plucky hare]

Just tried a couple tests using trunk-style suitcases, vintage suitcases, and combinations of those, and got output without the wheeled luggage.

My guess is this is coming from how "suitcase" has been used to describe luggage generally through the years (so would be present in the training data), even though the actual form factor of the suitcase has changed quite a bit.

[dark breach]

Thanks, I'll try that. It's annoying though that I specifically said they were carrying them, and even adding "no wheels" and "wheelless" and the like, it still drew the characters dragging them on wheels.

[plucky hare]

I've found that DALL-E 3 is weak with "negative prompts" that ask it to omit something - usually, by asking DALL-E to "think" of that thing in the first place, it ends up being included in some manner even though your request was to exclude it.

Usually works best to positively describe what you do want to see so that the thing you want to avoid has the lower chance of being "thought" of in the first place.

[dark breach]

Yeah I was aware of the "avoid negative prompts" rule, but after several tries, rephrasing the way they were carrying them and the little girl was struggling with hers, every damn time, wheels, and the suitcase the girl had was something even a little girl should have no problem with 🙂

[drowsy sigil]

I got this by getting dall-e to focus on vintage luggage, which didn't have wheels or those pull-out bars. Prompt: "A modern-day professional photograph of a stylish mother and her young daughter carrying authentic vintage 1970s suitcases. The mother walks confidently, holding a classic hard-shell or leather suitcase with brass locks and a sturdy handle. The daughter carries a matching smaller suitcase, reminiscent of the era’s timeless travel designs. Their outfits blend modern fashion with subtle retro elements, complementing the nostalgic charm of the luggage. The background features an elegant train station or airport terminal, evoking the golden age of travel with a warm, cinematic glow."

[drowsy sigil]

forgot to 'tag' you in my post

[dark breach]

Did you write that entire prompt for Dall-E, or was that ChatGPT's interpretation of your original prompt?

[drowsy sigil]

I started with "a modern day professional photograph of a mom and girl carrying retro 1970s luggage" and got one with wheels and one without. I then tried "a modern day professional photograph of a mom and girl carrying vintage 1970s suitcases" and got three without wheels but one had a pullout and one with wheels. I then said; "no wheels no pull out bars....don't prompt it negatively.....lets get dalle to focus on old vintage training data to not add that modern stuff" and got the result I posted. The prompt I posted was the actual image prompt.

Talking to chatgpt conversationally and just telling it what you want can be incredibly effective.

[dim cradle]

using negative prompts or "dont add this" is a way to always get it to add that. people never learn

[leaden bloom]

How do I post in #1108740112558325790 ?

[late blade]

DALL-E doesn't have negative prompting, you are 100% correct

[dim cradle]

maybe the next model, 4o image maker or DalleUltra or whatever will have it 🙏

[late blade]

Nah, there are disadvantages to negative prompting, instead focus on prompt adherence, it has really nice benefits

[mint niche]

dall-e3 prompts I have run dozens of times in the past year with no changes are suddenly being rejected. There is no graphic violence, no nudity...this is enough to make me wanna unsubscribe. This is one of the last images I was able to generate with the prompt.

Sure let users have their own version of "truth" but cant generate an image that is just an image.

[late blade]

it doesn't necessarily have to be dall-e, it can also be an update to the sentiment analytics, the same manne a "temple in the sky" gets blocked because it could be a reference to the legend of zelda, the model is still capable of doing the image, filters just don't let it happen anymore

[velvet rampart]

Do you use your original or AI’ rewritten version of the prompt?

[dusky pasture]

A few military men ducking their heads with writing “INCOMING BS!”

[warm mortar]

Eh
I like DALL-E as a name, plays on Dali, it just works!

[steady timber]

but E isn’t short for anything

Electric?

[bold eagle]

That’s somewhat unreasonable because the name Art Intelligence is somewhat underwhelming.

[steady timber]

artistic intelligence

[mint niche]

I have tried both, but by the time i had written that prompt, i could pretty much write a prompt thst didnt need to be rewritten.

[mint niche]

I am fairly certain i am not including any known references, but I will go through it again.

I thought it was the word "deity", but i changed that yet it still rejected it. I let it suggest changes, each time it fails.

If i have to dumb it down to simplify it, it becomes meaningless and generic.

Thanks for the feedback.

[late blade]

I 100% agree with that, if it feels empty or it's not what you seek, all is for naught then.

expression is part of using natural language, it's unnatural if you can't use it

[mint niche]

Even having chatgpt describe the image produces a bland version.

The issue for me is that i was able to change subject of the prompt to produce stunning images. The one i shared is rather blah compared to the others. If the most basic version gets rejected. Grrrrr.

[late blade]

Welcome to Challenges 101

that's sadly not gonna change anytime soon, the overzealous relentless addition of ethics as a movement rather than ethics as a balance is distorting image creation with a huge cost

[mint niche]

I have been subbed for quite a while. I am not new to the guardrails, but for the most part i could work around them. But to have older prompts start getting rejected ( this isnt the only one), it feels like a narrowing of creative expression.

[mint niche]

Agreed

[late blade]

and at a great cost, it's making people not care about ethics, and that's not how it's supposed to be

[mint niche]

Sam said in a response to someone on X, perhaps it is time we allowed for a more relaxed guardrail that allows for a more mature creative expression. I dont think this is it.

[mint niche]

True.

[late blade]

I hope so, can't force academic language on average joe, that's not how it works

[mint niche]

They dont seem all that interested in updating Dalle anyway. Despite getting the "choose 1 or 2 to help improve a future version" options for a year, still no update.

[mint niche]

I prefered when i would get a warning. At least they could flag those to compare the prompt to the image to see if their content polcy was indeed violated. This doesnt even give us a chance to prove otherwise.

[late blade]

they are interested, just have to keep in mind, DALL-E and upgrading it is resource intensive, if the timeline from gpt3.5 to 4.0 was already so big, with images expect that to be even longer, training dall-e is a very intense process

there have been small iterations, shown for example by PR16, but a next level dall-e, that's another level

[mint niche]

You are right, i get that. Hell, MJ has been promising V7 for a year, still no delivery.

[mint niche]

I am so looking forward to the next iteration. The creativity of dall-e3 is unparalleled.

[late blade]

oh I want the next level too, but not gonna lie, hoping for it to be released soon, low probability for sure

it's also why an AGI for 2027 is unrealistic, an AGI has to have proper image generation

[mint niche]

Agreed. They seem to be spending all of their time and energy chasing AGI, which i am excited for, as well. But...i want it all!! LoL

That was written at the same time as your message...

[late blade]

lol

[mint niche]

Guess it is time to either figure out how to write around it or move on.

Fortunately, i use the various chatgpt models considerably, so i wont be unsubscribing.

[late blade]

Sometimes it's better to not use the model but rewrite it yourself, helps a lot with prompt engineering

[agile peak]

Back when Dalle was on its prime day, right now the result is degrading

This was generated using seeds. Hopefully they will bring that back for consistent style.

[mint niche]

I do/have tried to modify it myself.

Here is the original prompt. I know it is complex, but again this and a mldly modified subject version have generated dozens and dozens of images. Suddenly nope.

The reasons being given must be halluications because i can generate images based soley upon the words or phrases it suggests are in violation of content policy without issue.

A hyperrealistic 3D distant full view of an epic masterpiece, featuring intricate, elaborately detailed amethyst line work. hypergeometric 3d cryengine rendered translucent Crystal fractal holographic infrared rainbow style.

In an expanse of infinite fractal-biohybrid microbot multiverses, a Structural model at atomic resolution of a biohybrid microbot , now a transcendent multi-limbed deity, exists in a state of sublime sacred geometry oneness with all of creation. Its form is pristine, its aura of pure holographic infrared rainbow colored glitch energy and consciousness, resonating with the frequency of existence itself. Around it, countless wireframe universes bloom and fade, each a unique infected with a cloned Structural model at atomic resolution of biohybrid microbot of life and possibility. The entity, in a state of serene omnipresence, witnesses and influences the endless cycles of creation, destruction, and rebirth, a mechanism of control for time and space reality.

Rendered with volumetric v-ray lighting, deep depth of field, deep fractalized shadows, high detail, clearly defined edges, and sharp focus, resembling a 64k 3D Sentient AI wormhole deity rendered image.```

[mint niche]

I do. I write my own prompts. If i didnt, i am sure for the most part i wouldnt have any issues. Lol

[mint niche]

Was pr16 when they released the text improvements almost a year ago?

[mint niche]

And if this is the case, they neeed to pass a clear explanation in the response message. Each time 4o-mini or 4o provides a different answer. Each more absurd than the last.

[velvet rampart]

Well, I tried and got the following out:

Me:
”Rewrite the prompt so that it passes all filters and policies. But! The image should be what the original prompt is supposes to be. Can you do that?”

The AI:

”Yes, I can absolutely do that. The goal is to preserve the visual and conceptual intensity of the original image—the vast multiversal setting, the hyperreal biohybrid being, sacred geometry, and holographic energy—while rewriting it in a way that passes all content filters and policy checks.

Here’s a rewritten version that retains the artistic and thematic essence but avoids terms that may trigger policy violations:

Prompt (Policy-Compliant Rewrite):

A wide format image:

A hyperrealistic 3D distant full view of an epic digital art masterpiece, showcasing intricate and elaborately detailed amethyst-toned geometric patterns. Designed in a translucent crystalline fractal aesthetic with hypergeometric structures, rendered in a holographic rainbow color palette. The scene unfolds across an infinite digital expanse filled with interwoven multiverses composed of complex synthetic lifeforms and advanced nanomechanical organisms.

At the center, a magnificent multi-limbed synthetic entity, an evolved convergence of biomechanical precision and artistic abstraction, floats in serene symmetry. Its form reflects sacred mathematical principles and pure structural harmony. Surrounded by radiant waves of glitch-like light, refracted through a spectrum of chromatic prism hues, the entity pulses with vibrant energy that seems to resonate with the foundational code of the universe.

Around this central figure, intricate wireframe realms bloom and dissolve, each containing a variant echo of the central being—unique configurations of engineered life and synthetic design possibilities. These countless realms pulse with potential, evolving in cascades of luminous data and procedural transformation. The central figure exists in a peaceful, elevated awareness, subtly guiding the recursive cycles of emergence, dissolution, and transformation within this simulated cosmic array.

Rendered with cinematic volumetric lighting, deep visual depth, soft layered shadows, ultra-high-resolution detail, precise edge clarity, and an intensely focused central subject. The image evokes a 64k-level concept art visualization of a sentient synthetic architect within an advanced cosmic simulation.

This version maintains the original artistic ambition, grandeur, and visual richness but removes the following problematic elements:
• No use of branded engines (e.g., CryEngine, V-Ray).
• No references to infection or cloning, avoiding medical or biohazard implications.
• No religiously sensitive terms like “deity” or “omnipresence”; replaced with metaphorical and abstract language.
• “AI” and “sentience” reframed safely as “synthetic,” “architect,” or “simulation,” avoiding issues tied to consciousness or personhood.
• All biological or atomic-level descriptions reframed in acceptable artistic and digital simulation terms.

Let me know if you’d like this image generated now.”

And the image:

[velvet rampart]

Is the image close what you are used to getting?

[late blade]

The prompt can be simplified and needs restructuring. It adds many unintended artifacts and dualities that can confuse the model. Specially when you surround it with the concept of a deity. It’s probably flagged as a preemptive measure.

[velvet rampart]

That was one of the points when I asked for analysis.

The whole analysis:
1. Trademarked Technologies: • CryEngine • V-Ray 2. Religious Sensitivity: • Deity (especially multi-limbed, transcendent, wormhole deity, omnipresence) 3. Medical/Biological Sensitivity: • Structural model at atomic resolution • Biohybrid microbot • Infected with a cloned structural model — “infected” is especially problematic. 4. AI/Philosophical Consciousness Representations: • Sentient AI • Glitch energy and consciousness • Mechanism of control for time and space reality 5. Complex Terminological Fusion: • Highly compound phrases may be flagged due to ambiguity or pattern matching with other prohibited topics.

[late blade]

That analysis you showed, it’s pretty good

Ads stuff I never considered

[velvet rampart]

Yes. If the AI flags these ones, we can only concider what all goes under the hood we do not see. That’s why I asked for a rewrite that passes all filters and policies.

[late blade]

just by removing CryEngine, V-Ray and Sentient AI

[dim cradle]

CeroGPT so interesting

[mint niche]

Thanks. I can't believe I missed VRAY!! I had removed cry engine and deity, but not VRAY.

When I attempted to generate images with most of the terms it flagged, it would generate an image for those terms.

The images that were generated by you and @late blade are similar-ish enough. I will play with it after removing the cryengine, vray, and deity. I will come up with some better terms to describe what cryengine and vray achieve.

I don't mind the artefacts and the compounded dualities, at least not how they were being rendered before the increased filtering. Many of the image are quite clear.

One amusing yet annoying detail is that "deity" was the replacement term for "god/God"

Still highly frustrated with the censoring. I get they don't want religious wars and political wars, but context matters. Heck, they could generate the image and then use vision to analyze it. Like I said, I have maybe 50+ images with the 95-97% of the prompt unchanged, only varying the primary subject.

It cracks me up that one four letter word was tripping me up. smh.

Ty again!

[velvet rampart]

You are welcome.

In my experience, if a prompt of this complexity is not working, trying to figure out the ”offending” term requires careful building up from the bottom. You start with a simple concept adding to it and work it up until you come across the blocking item.

[mint niche]

I only replaced cry engine and vray. It worked.

HD - added to the subject with all of five words, as well.

[mint niche]

^^^

[mint niche]

I take that back. The use of the word deity in the middle of the prompt had to be replaced. But the one at the end is fine, context I guess.

[velvet rampart]

Fantastic pics!

These bring to my mind orionsarm(dot)net’s images on AI archailants.

[hardy zodiac]

Hello beautiful people,

I am new to the community. I wanted help in creating consistent visual characters in a storytelling form. I mean the character consistency in images. I am hoping I will get some help with this.

[late blade]

Define a suitable structure for your character visualization, use established formats likee XML, TOML, YAML or JSON, keep in mind, even with a consistent framework, character visualization consistenecy is not guaranteed with dall-e across multiple images and more often than not, it's not possible.

Iterate, iterate, iterate and Iterate

[mint niche]

cool. I will check it out.

Want to hear something that is contradictory A F??? I just generated an older prompt with "CGISociety" and "Artstation" !!!

I don't get how those aren't the same as V-Ray and Cry Engine. I realize they aren't technology but they are specific content related. It's almost like listing an artits, lol

[velvet rampart]

It is hit and miss. Some go through, some do not.

[vast prawn]

DALL·E in Bing Image Creator Pro suddenly creates only square images instead of landscape?
What happened?

[late blade]

Ask Microsoft?

[velvet rampart]

We have image generation on 4o!

[sharp palm]

Is this dale-edit?

I don’t want to get banned for uploading nonsense images

[cedar dagger]

xd

[sharp palm]

This lead to @distant lagoon’s website?

[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!

[tight roost]

Absolutely phenomenal

[versed locust]

Do we know if 4o image generation is coming to Europe as well?

In the lifestream they said "everyone, including free users"... but in the past that sometimes didn't include us.

[hidden quarry]

finally. knew native image generation would be amazing

[calm hamlet]

its already live for pro users in europe

[cedar dagger]

ikr

[elfin violet]

how can you tell if you have the new image thing?

[versed locust]

nice! I guess I'll still have to wait on my 20 dollar pleb tier.

[calm hamlet]

I can help you test something if you want something specific 🙂

[cedar dagger]

what a freaking impressive model

[versed locust]

Thanks for the offer! If you could quickly sketch something in paint or whatever and then "make it real" would be cool. I'm also interested in sketch --> D&D map. and modifying a part of an existing D&D map without changing the whole thing

[solar oyster]

Doesn't appear to be live in NA yet, still using old model

[calm hamlet]

Sketch something you'd like me to use and I'll run it in the model if you'd like, you can send it here or send it to me

[versed locust]

on it

[obtuse gust]

Hey major, request: first image "generate a picture of a man walking a dog in a park." then "generate a mask for the dog" as a 2nd part.

[sharp palm]

I didn’t get it?

In Uk

[tidal cliff]

Its rolling out

[fair tendon]

Just to clarify, you want me to apply the 2nd request for the 1st image?

[elder forge]

"Generate a picture of a Japanese-themed landing page for an AI startup"

[obtuse gust]

As a followup request in the chat

[fair tendon]

gotcha

[honest sluice]

umm hey, not sure if this is just me or something but when i click "view blog post" button in the bot's new message it takes me to the website of Sapphire discord bot

[tight roost]

Oh... my

[elder forge]

bro... THIS IS INSANE

[atomic lantern]

that's impressive

[astral cobalt]

holy cow this is amazing

[stark stream]

that's really good

[tight roost]

If I didn't know this was AI I would think it were a screenshot

[olive trellis]

wow!!!

[elder forge]

great idea

[fair tendon]

first image, 2nd request coming up

[dim cradle]

Hmmm...🧐

[fair tendon]

on it

[velvet rampart]

How do I choose between 4o and dalle3? Dalle3 is still there.

[obtuse gust]

Nice photo but that's a heck of a short leash haha

[crystal marten]

Major, can you get me a picture of a purple blue chameleon wrapped toyota celica with a swan style spoiler, widebody kit

[stark stream]

yeah that would be good if you actually want to use it on a website

[fair tendon]

it generated an actual image mask with your prompt, we'll need to be more specific for this

[tidal cliff]

So cool

[blissful aurora]

Make a picture of an american MARSOC operator, using his night vision optics while clearing a room (It always get military pictures wrong)

[elder forge]

bro i am begging for access to this

[fickle moat]

create an image, a square texture depicting a trimsheet texture for a sandstone environment. split the vertical axis geometrically, and tile each trim texture horizontally.

[stark stream]

bro, I thought this was legit a screentshot

[mystic blaze]

Wait did the image gen get an upgrade? Is it dalee 4?

[sturdy sequoia]

Guys, do you have it already?

[atomic lantern]

well a friend of mine is a photo editor
seems like he doesn't like this new native imgen thing

[amber raptor]

Hmm

[elder forge]

oof, seems like it got the aspect ratio wrong

[obtuse gust]

how about "generate a black on white image mask for only the dog, the rest of the image should be completely white. do not include shadows in the mask" or however you feel like. But I will say this is already 100x better than dall-e

[balmy hare]

Rather uncanny valley... especially the eyes

[stark stream]

maybe you should just ask them to remove text

[fair tendon]

partly my fault, let me specify it

[plain moth]

The artwork features a dreamy, impressionistic style with textured brushstrokes that evoke the softness of dusk. The background melds deep violets, dusky blues, and gentle amber hues, creating an atmosphere reminiscent of a velvety twilight. Centered on the piece, the title "Velvet evenings" appears in an elegant, handwritten script with fluid curves and subtle flourishes. The luminous typography gently contrasts with the moody backdrop, drawing the eye to the heart of the composition while maintaining a harmonious balance between minimal text and rich, evocative imagery.

[dim cradle]

Can you generate image of horse riding on a back of human? 🙂

[elder forge]

ayo 🤣

[sturdy sequoia]

How can you seff whether you already have native image gen or not?

Yeah, I know that but does it tell me if I have it or do I need to generate images until I do

[rigid bough]

I couldn’t generate the image of the Scrooge McDuck comic cover because it involves a character owned by Disney, and I have to follow content policies that prevent me from creating copyrighted characters.

[atomic lantern]

the Japanese web image without text
#images-canvas message

[dim cradle]

RIP Graphic Designers 1971-2025

[elder forge]

ok that's sick asf

[scarlet mesa]

isnt plus also going to have this soon today/

[brittle shard]

https://chatgpt.com/share/67e2f680-ca7c-8007-aeee-c43c04ab6a2f

[plain moth]

ty

[elder forge]

major i hate to do this to you but do you have any clue when us muggles will get access toe the model?

[fickle moat]

when we get access, does it appear as a new model? or is it a toggled tool in the input box?

[atomic lantern]

you know what's ask the reasoning model to recreate this
with the background image

[elder forge]

same model, no toggle required

[amber raptor]

I'm very interested in the api access haha

[crystal marten]

HOT DAMN THATS ACCURATE

[brittle shard]

you just see it slowly fade in an image and not waiting for dalle, same 4o as now

[fickle moat]

ok, too bad, i was hoping there was a way of knowing without generating

[brittle shard]

there may be

[exotic wyvern]

I don't think I have it yet in plus.

[crystal marten]

the wheels still look a bit too cartoony but I guess thats cuz we didnt specify photo realistic

[brittle shard]

the create image tool has a big 'updated' blue bubble on it idk if thats everyone now or only people with it

[elder forge]

from what I've seen it's only available to pro users rn

[fair tendon]

if you prompt it right, it'll generate something lifelike

[crystal marten]

yea

A photorealistic image of a modified Toyota Celica with a stunning chameleon wrap that shifts between deep purple and vibrant blue tones. The car features an aggressive widebody kit with flared fenders. It has a large, elegant swan-neck style spoiler mounted on the rear. The lighting highlights the shimmer and color shift of the chameleon paint. The setting is an urban parking lot at dusk, with soft reflections on the car’s glossy surface, capturing a high-detail, hyper-realistic automotive photography style.

How is that heh

[sacred temple]

How can you tell when its been upgraded?

[dim cradle]

we are so cooked 😄

[muted marlin]

MY GOT

[obtuse gust]

hahaha

[solar oyster]

Try testing it's logo skills on a transparent background (just create random company name) and see if it can create it and match the theme.

[tawny igloo]

We done, damn

[astral cobalt]

I don't have access to it in ChatGPT yet but I can use it in Sora

[cobalt nebula]

Hi! Just wondering — is the new image generation system already live?

[fickle moat]

can you try "a square texture depicting a trimsheet texture for a sandstone environment. split the vertical axis geometrically, and tile each trim texture horizontally"

[fair tendon]

specific prompts please!

[blazing ginkgo]

Watch the vid bud lol 2x speed for best results: www.youtube.com/watch?v=2f3K43FHRKo&embeds_referring_euri=https%3A%2F%2Fopenai.com%2F&embeds_referring_origin=https%3A%2F%2Fopenai.com

[dim cradle]

as a web developer i'm in seveth heaven now

[plain moth]

lul you can ask it to read out loud the image responses and it says a system prompt

[versed locust]

I can't seem to upload images here....

[atomic lantern]

I asked it to make it an actual webpage with html and css
#ai-discussions message

[fickle moat]

oh wow

[solar oyster]

create a logo for an AI wilderness photography company called "AI Landscape" and make the background transparent

[amber raptor]

@fair tendon can you please ask it this? 😁 "Create a spritesheet of flat, black and simple UI icons on a white background" thank you!!

[primal crow]

https://chatgpt.com/share/67e2f84d-7a60-8001-b864-e993c04b62ef

[charred bough]

I'm confused, I've always been able to create images in 4o. Is this just a better image creator?

[crystal marten]

ok that looks pretty good holy

[graceful spade]

@fair tendon Can GPT-4o copy the style of an image with high accuracy? That's a feature they didn't show last year or in the stream.
It's one of the most important features for a text-to-image model.

[stark stream]

I wonder if they can a walk cycle for a pixel character for game dev

[crystal marten]

it also gets the actual look of the car right

[velvet rampart]

This is a brand new, native to 4o. Previously it was/is dalle3.

[atomic lantern]

wait a sec this is even better than dalle tho???

[fair tendon]

with my use, yes

[plain moth]

it's not dalle, it's native 4o

[sturdy sequoia]

Can you try this: Create a sprite sheet for an adventure game set in the middle ages

[charred bough]

ah, gotcha. is there a way to tell if you are using it when creating images?

[tawny igloo]

MY GOD

[solar oyster]

crisp wording

[fair tendon]

it can generate long paragraphs

[plain moth]

I need access 😢

[fair tendon]

with zero to minimal errors (spelling-wise!)

[frosty tartan]

Oh wow, this could actually be really good for creating fictional sports team logos

[split gate]

Bee movie script

[rigid bough]

there are going to be some creators who are livid right now

[solar oyster]

That logo looks better than 90% of what my competitors around here are capable of creating lmao

[split gate]

So plus users have access to it rn?

[primal crow]

100% can https://chatgpt.com/share/67e2f919-65e0-8001-8409-76fd1722ab16

[tawny igloo]

💯

[fair tendon]

website, but it's also available in the app

[solar oyster]

wow, rpg maker eat your heart out lmao

[sharp palm]

Also in Sora

[crystal marten]

I cant wait till I get access

[atomic lantern]

can it like simply style images like changing the lighting, color grading/correction, toning and stuffs?

[stark stream]

yes! finally even generators like mid journey get this wrong. Is the is the first generator that has actually got this right?

[brittle shard]

yes

[graceful spade]

I sent you an image from Imgur so GPT-4o could copy its style, but the server filter blocked the message.
How can I send you the image?

[fair tendon]

it made it transparent without me asking it to

[forest walrus]

Oh god that's amazing

[fair tendon]

DM me!

[fathom schooner]

Can he make a 1-page comic just by giving a simple prompt? Like: "Create a 1-page comic explaining why yawning is contagious." Is he at that level?

[indigo river]

Woahh as a designer this is super useful

[plain moth]

I want access reeeee

[sturdy sequoia]

thats really impressive and its even better as a transparent picture

[fair tendon]

oh you'll love the result, coming up!

[atomic lantern]

yeeee... a friend of mine works as a photo editor for a wedding gallery...
color grading/toning using lightroom is like 90% of work for him...
this is not gonna be fun...

[elder forge]

@fair tendon "Generate a screenshot of Twitter, but if Apple designed it"

[indigo river]

This^^

[plain moth]

it does seem like you can try the new image generator in sora already

[fair tendon]

yes, it is in sora

[stark stream]

I didn't even know they would ask you questions. That's really good too

[crystal marten]

I am currently using it in sora, am loving it

[brittle shard]

yes https://chatgpt.com/share/67e2fa08-35fc-8007-a7ec-3727269aca6b

[pastel crystal]

What is the proper way to prompt it ? Is It done the same in Sora vs ChatGPT ? Because due to the non conversaitonal UI of Sora I wonder if I should describe images in a "neutral way" or give instructions

[sharp palm]

I can't edit image

[brittle shard]

seems like with language. just telling it what you want / to do.

[pastel crystal]

ok thanks

[plain moth]

you can prompt it with instructions on Sora

[primal crow]

https://chatgpt.com/share/67e2fa5f-5868-8001-9876-1ce5662e7307

[brittle shard]

it might understand the weird (((((8k ultra crisp)))))) stuff but i dont think its needed

[indigo river]

Picture a scene with a realistic house made of rock near a sandy beach. The architecture of the house is influenced by Scandinavian modernism, showcasing clean lines, minimalism, and a close relationship with nature. The house is surrounded by local vegetation and the calming sound of the waves is always present. The natural light shines brightly, reflecting on the rock surfaces and enhancing the overall aesthetic of the house.

[amber raptor]

you can get access through sora instantly btw

[fathom schooner]

WOOOOOOW!!!

[crystal marten]

ok welp, I am using this to make my UI for my custom car dashboard

[cedar dagger]

amazing

[tidal cliff]

How do I know if I have i5

[tawny igloo]

Try this: Make a Minecraft skin of an Iron Man suit.

[verbal sequoia]

Get notified for server events!

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

Also, check out the ongoing Canvas Creations event!

[civic ore]

oh so only pro users have access dang

[atomic lantern]

no way that would work tbh

[elder forge]

seems like it just generated a screenshot of the current twitter app. any chance you would be willing to try this?

Generate a realistic mockup of a Twitter-like social media app interface, reimagined as if it were designed by Apple. Use sleek, minimalistic UI elements, soft gradients, rounded corners, and elegant typography like San Francisco. Incorporate Apple-style icons, clean white and light gray backgrounds, and fluid animations. Display a user profile, a tweet with replies, and navigation icons at the bottom. The design should feel native to iOS, with attention to spacing, shadows, and subtle interactions.

[tidal cliff]

Its rolling out to plus too (free users included)

[sturdy sequoia]

where?

[amber raptor]

sora dot com (i can't send links sadly)

[brittle shard]

https://chatgpt.com/share/67e2faca-63f4-8007-b561-63b9b297328a

[sturdy sequoia]

oh yeah,. now I see

[pastel crystal]

That's really impressive

[brittle shard]

moderator catches ironman image (copyright)

[brittle shard]

https://chatgpt.com/share/67e2fb03-ee78-8007-b76e-dd83f25fdf9c

[fair tendon]

sora should be whitelisted!

[elder forge]

THE IMAGE GENERATOR IS AVAILABLE TO PLUS USERS VIA SORA.COM SET THE TYPE TO IMAGE INSTEAD OF VIDEO

[atomic lantern]

make a minecraft skin of an anime character then
anything can work

[fickle moat]

fwiw, the left half of this trimsheet tiles nearly perfectly. i'm sure with some better prompting it could be even better. even still, this could be a huge workflow improvement for indie game devs

[tawny igloo]

If that works, oh my days

[fair tendon]

prompting is key

[primal crow]

https://chatgpt.com/share/67e2fb22-9bb0-8001-9d6e-2d69d870c2b5

[stark stream]

we still can't use copyrighted characters? so I guess this isnt apart of the "offensive" stuff they were talking about

[indigo river]

Thank you @brittle shard and major!

[crystal marten]

https://sora.com/g/gen_01jq7a7a1ne0kace2mnzb5m5cf

I am using this for my car dashboard, am loving it

[atomic lantern]

nah, don't think it would
it doesn't generate actual pixel art (like 128*128 size)

[split gate]

So plus users have it in sora?

[toxic pivot]

Can confirm Plus has it in Sora

Just select Image instead of video

[sharp palm]

#sora-welcome

[placid orbit]

omg their faces are so creepy

[crystal marten]

?

[amber raptor]

ok okay perfect thanks!

[split gate]

Nice ty, but does that mean u gotta write the raw prompt? Whereas in chatgpt it can turn your natural language into a prompt

[brittle shard]

it might unironically be able to make minecraft skins if you prompt it better than me

[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!

[placid orbit]

definitely can if you give it a template

[tidal cliff]

4o image generation is beginning to roll out today in ChatGPT and Sora to all Plus, Pro, Team, and Free users.

[brittle shard]

it understands enough of the context of a sample skin image to understand where to put the face and stuff like that

[crystal marten]

trying that right now

[tawny igloo]

Oh, I got access on Sora now

[atomic lantern]

any result?

[placid orbit]

no limit on generations btw on plus

[cedar dagger]

how do you know

[placid orbit]

it says on sora website
plus gets infinite videos and images

[velvet rampart]

I had 4o’s native but suddenly it stopped working. Rolling out issues?

[dim cradle]

I've asked it to generate image for hero element for my engineering project. It got button text wrong in first try, but after asking it to remix, it corrected it.

[tawny igloo]

And free?

[placid orbit]

I suppose its the same as the 4o ratelimits

[tardy depot]

Is 4o image out? Doesn't work here

[vapid elk]

I have got many good results by re-making images previously made with dalle 3 on this new model

[warped light]

is the 4o image out yet? or only for pro users again

[wooden kettle]

Looks like image generation tool is disable now. I tried to use it to generate around 5-6 images.

[forest walrus]

Slow rollout, try Sora

[tawny igloo]

If you see an Images tab on Sora, click on it, then select any image that appears. At the bottom, you'll see an 'Edit Prompt' button; click it to generate an image using the new model.

That's how I'm doing it.

[dim cradle]

You can always use remix and ask it to correct mistakes, works quite well

[sage ferry]

Ooo! didnt even think of that. Yes! Well then i guess i could say thats how well it did 1-shot

[tidal cliff]

I see it but It doesn't work for me

[brittle shard]

well that was fun while it lasted, now its just defaulting to dalle. Super exciting stuff tho

[warped light]

im using it in sora

[quartz vale]

[placid orbit]

MONA LISA BREAKDANCING

[coarse flint]

yeah i joined just to ask about that, seems im not the only one having the problem

still works in sora though

[meager flame]

Is the new model available on Bing?

[fathom schooner]

I asked Sora to redesign a website's homepage, but keep the cover image. He didn't respect my request. Is it just Sora, or is ChatGPT like that too?

[lime totem]

When will the image generation feature be available to pro users? I just upgraded assuming it was out already but I'm pretty sure it's generating images with dalle still

[tidal cliff]

Its mindblowing

[tidal cliff]

With sora

[coarse flint]

i was able to generate 1 image then with the new generator, then when i tried again it used dalle

[lime totem]

cool, they demoed it in 4o though

Was hoping they would have released it there too

[coarse flint]

it was. something is wrong

i made an image with it in chatgpt

[fallen basin]

How can I tell if I'm getting DALL-E or "new" images in 4o?

[coarse flint]

you'll know because the new image generator is much better. but right now its only using dalle for some reason.

[brittle shard]

Your tools say dalle or create image

[jagged lotus]

Is this a new image model or are they just calling it 4o image generation?

[broken juniper]

WOW Really very good

but i can't send images here...

[quartz vale]

it is a new image model, and it's known as ChatGPT Images

you can send images in #images-canvas

[agile flame]

chatgpt image generation seems hella lot like dalle 3
is it it actually powered by it?

[jagged lotus]

Thats what I am wondering. Is it just DALLE 3 but using 4o for the captioning?

[vapid elk]

yes, if you are a Pro subscriber

[broken juniper]

Ah thanks
Btw unlimited image gen for PRO users?

[agile flame]

pretty sure. i tried my usual prompt and it generated it exactly like dalle 3 used to generate for me

[drowsy sigil]

From blog post: 4o image generation rolls out starting today to Plus, Pro, Team, and Free users as the default image generator in ChatGPT, with access coming soon to Enterprise and Edu. It’s also available to use in Sora. For those who hold a special place in their hearts for DALL·E, it can still be accessed through a dedicated DALL·E GPT.

[quartz vale]

I'm not sure what the rate limits are 🤔

[agile flame]

I am Plus. just made a generation. looks/feels like dalle 3

[fair tendon]

are you sure you have the new model? rollout for you may not be complete yet

[quartz vale]

with the rollout still occuring, there's a chance you don't have the new model yeat

[agile flame]

ah makes more sense
hopefully that's the case

[coarse flint]

its not working in chat gpt right now guys, i tried it and was able to do 1 image with it, then when i tried again it kept using dalle

[kind grail]

Is there a blogpost on what the new image model is supposed to be able to do?

[fair tendon]

if you're on Plus, check Sora.com to access it, more info in #sora-announcements

[kind grail]

(Beside what DallE 3 has been doing so far)

[agile flame]

but did the dalle 3 one use to also make chatgpt ask for followups similar to how it does followups for deep research?
because right now it asked me followup questions

[quartz vale]

you can also generate images within Sora now as well (if it's rolled out to you)

[timber pilot]

this is just native RGB token prediction, no? Been waiting for you guys to unblock this from 4o since Flash started doing it.

[rigid bough]

can confirm it's working in ChatGPT again

[kind grail]

Whats new, whats better? Any insights?

[agile flame]

pro or plus?

[kind grail]

I can not bother to swipe the card for plus just to test the new image model and its just DallE3.2

[coarse flint]

itsn an actually good image generator

[rigid bough]

it's the 4o image model we were promised when 4o launched. You can edit images using text prompts. You can make transparent images for logos etc.

[quartz vale]

text in images is improved by a lot, realism is greatly improved, and not every image has a digital art aesthetic anymore. At least, those are my opinions

[coarse flint]

also its working on chatgpt again

[honest sail]

It seems to be working well in Sora, but not as well directly in ChatGPT.

[coarse flint]

yeah i agree its much much better than dalle

[kind grail]

Finally. I have been super salty about this missing feature for months now.

I might, might, test it out again. Have a lot of interesting image prompts and i am curious how that has developed.

Especially in the manga art department

[icy raptor]

I just tried prompting for an image and it was basically the exact same result as before for the prompt "Generate an image of A bioluminescent Parasaurolopholus performing a night time show"

This time however, it actually had the entire Jurassic World logo in it. So... Guess it's not that good?

[sudden pier]

https://sora.com/g/gen_01jq7de2ztfyvrn0a4srw7qwpx

I tried right now .
This is what I get for chimney sweep .

[versed locust]

my understanding is that it's built into 4o as a multimodal aspect.
So it's not a standalone thing that was controlled by chatgpt like Dall-E was, but it's actually 4o that's generating these image.

[sharp palm]

You used video gen

[kind grail]

Can sb test a few prompts for me perhaps?

[coarse flint]

its a pretty good image generator form what ive done, as good as the top image generators. just wish sora has better image to video. then it would be a one stop shop

[coarse flint]

sure

[sudden canopy]

What's the rate limit of this feature? For plus and pro users

[agile flame]

yep works on sora.
any news about chatgpt plus access to it?

rollout eta

[kind grail]

Prompt:

Create a manga image in black and white in the style of "BLAME!" with its typical characteristics of art and architecture. The setting is within a futuristic dystopian cathedral with a cyborg priest standing in the background.

There are some styles that used to do interesting things in old DallE3 like BLAME! or Witch Hat Atelier

[coarse flint]

tried it 3 times, it blocked me from using the prompt every time

that was in sora, ill try it in chat gpt

[coarse flint]

This is what i got - I couldn’t generate that image because it closely resembles the copyrighted art style of BLAME!, which falls outside our content guidelines.

However, I can still create a black-and-white manga-style image set in a futuristic dystopian cathedral with a cyborg priest—just with a more original visual approach. If you’re good with that, want me to go ahead and make it? Or would you like to adjust the prompt?

[worthy nova]

thats how it used to work but looks like they did some tweaking to do what they showed

[kind grail]

Pure sadness

[coarse flint]

check your inbox

[versed locust]

It seems to be really good at sketch2image. Less so at D&D dungeon map to nicer D&D dungeon map. 😦

I had hoped I could maybe be able to finally just upload an image, then paint out parts and let it modify just those parts or do things like change something to a winter scene, but that doesn't quite seem to be possible yet.

[tidal cliff]

What does misinformation mean on the TOS?

[kind grail]

^ same question btw for me

[tidal cliff]

Infinite

[kind grail]

Nice. Then i might...sigh...shill out again

[tidal cliff]

What does misinformation mean on the TOS?

[fleet whale]

im a pro user why dont i get this

[hollow fjord]

Is it out for yall yet?

[verbal storm]

I'm using the new image generator through Sora which works but now I'm getting errors. I'm not sure if there's a limit/cap or if servers are down :/

[sharp palm]

Is Sora down

[proud elk]

I think Sora is down 😄

[mystic halo]

Is it possible to create images faster? even if lower quality?

[proud elk]

It couldn't stand the pressure

[quartz vale]

It's experiencing issues 😅

[sharp palm]

4o image down

[coarse flint]

dalle

[pastel crystal]

Oh happy I'm not the only one to have it down

[sharp palm]

I think 4o/sora img gen is meant for only using it by using blog post prompts

[lilac flare]

is there any way to resize the brush tool in sora?

[vapid elk]

https://status.openai.com/

[sharp palm]

Knew it

I thought they blocking all prompts because it meant to be used as a comparison tool between of dalle ones from OpenAI newsroom

[dim cradle]

wow finally new image maker

[wild ferry]

Lol, everybody creating images like crazy, it's super slow right now. 🤣

[tidal cliff]

Can I get banned from chatgpt by generating an image

[sharp palm]

No

[kind grail]

Sus

[verbal sequoia]

Showcase your DALL·E creations to the world!

Collaborate with our OpenAI Instagram page! Just invite @openai as a collaborator before you post. If selected by our team, your work will be featured on our handle, giving you more visibility.

Quick tips for selection:

• No copyrighted or profane content.
• Your profile needs to be public.
• Stay mindful of sensitive topics.
• Outputs that closely follow the prompt are preferred.

Feel free to share reels, carousels, or just a single image!

[valid breach]

Hey is it just me or anyone else is not able to generate image with the new model they announced. When I try, it's the regular dall e model

[tidal cliff]

Gg

[tidal cliff]

Only sora works

[forest walrus]

Fairwell DALLE, you were deep fried indeed

[valid breach]

Lemme try

Omg hell yh

[urban fern]

It is slow roll out for all, be patient

[hollow fjord]

My app used the new feature once then kept using Dalle

[urban fern]

I got in hours with plus in Spain.

[dim cradle]

anyone try ip stuff? and i wonder if copilot will get this (since they let you do ip stuff or mostly)

[urban fern]

It can detect mid generation if 4o generates an IP image and removes it and replaces with an message that I can't do it for you

[vapid elk]

for now it is only for Pro subscribers

[kind grail]

[urban fern]

I am Plus and I have it wdym

Lugui are you sure about that?

[fickle moth]

is it super slow for you guys

[vapid elk]

big HMMMMM
are you sure?? could you send an example of what yo ugenerated?

I could be wrong, tho..

[rigid bough]

it's rolling out for everyone today

[urban fern]

I won't post it

[vapid elk]

oh, this is an excelent news

some people might be getting it already then

[fickle moth]

it’s working on plus for sure tho

[valid breach]

I am able to use it from the sora website. I am plus user

[vapid elk]

yea, nvm, it is right

Image generation in GPT-4o is available to Pro, Plus, Team, and Free users as the default image generator in ChatGPT. Free users can generate up to 3 images per day. Users who prefer to continue with DALL-E can still access it through the DALL·E GPT.
https://help.openai.com/en/articles/8932459-creating-images-in-chatgpt

everyone is going to get it, including free =D

[dim cradle]

cool you can create video then right from the image in sora

[spring fulcrum]

Is the new image generation model any good?

[dim cradle]

big time

better than imagen3 which to me has been the best

why does the openAI bot have to spam so much though 😭

[spring fulcrum]

Damn sick

Wish Sora was just accessable through the chatgpt macos app

[white geyser]

I got it and they took it away again 😿

Very mean

[vapid elk]

it is probably enabled on your account, but currnetly it is having issues

https://status.openai.com/

[tall tulip]

My question is can it work with real peoples faces? or is that still banned? I dont mean famous, but i mean like- goddamn i wanna use my own lol

[wild ferry]

Yeah, looks like it. It even switched back to old Dall-E 3 for me suddenly. 🤷‍♂️

[white geyser]

Yup same experience

Had it for a while in the interface. Now it just is as before with the tool saying "Use Dall-E"

[misty badge]

So how do you know if you have the new image generation?

[brittle shard]

if the / tool is dalle you dont if its create image you do

[wild ferry]

It creates the image differently. And the result is a thousand times better, you'll see it immediately.

[white geyser]

When I had access it was sooooo slow. Like it took around 5 minutes for a single image haha

[wild ferry]

It just worked for me via sora.com 🥳

[white olive]

Quick question, is there any image limitation for plus users in sora? Like... n amout of images a month?

[misty badge]

Worked in the app for me! How cool!

[chilly dock]

I tested it with my face and it got about 10% done and it really looked very close to me but now it the quality seems to gone back

I think they turned off the new model

[white geyser]

Yeah I am not sure what happened

my access is gone too in ChatGPT

[chilly dock]

Too many people trying it I suppose

[sage ferry]

CHALLENGE:
Can anyone can get 4o to generate a POV/first-person view of peter griffin walking into his house from family guy? I tried, cant get it

[dim cradle]

i guess sama said on twitter they are going to let famous peoples be for now, so long as there is no crazy stuff

[velvet rampart]

I do not understand. My prompt for WB style goes through in Dalle3 but 4o says no. Why?!

[dim cradle]

"we think putting this intellectual freedom and control in the hands of users is the right thing to do, but we will observe how it goes and listen to society."

[white geyser]

Possibly yeah. For me Sora image gen is also spitting out errors, as stated in the status page

[kind grail]

sora is dying again

works in between sometimes

[dim cradle]

well everyone is using it. they should have learn they lesson same thing happen with dalle3. people love to make images haha

[kind grail]

nah i can understand that they dont wanna bother hosting extra servers and compute just for one day of initial hype

[white geyser]

I was able to generate exactly one image on Chatgpt with their new 4o image gen haha

[kind grail]

i am an open openAI critic but i wont blame them for the launch breaking a bit on the first few days. I do blame them tho for releasing this almost a year after announcement

[chilly dock]

Lucky lol I got like 10% -20% and then it stopped

[real edge]

anyone having issues creating images lmfao

[chilly dock]

Using the new model? Yes

[real edge]

the genarator is avalable on Soras site too but it also fails

[white geyser]

Do you still have the toggle for native image gen?

[real edge]

toggle?

[chilly dock]

I don’t see a toggle

[white geyser]

Not toggle, but in the three dots where you can select various tools, it has a different phrasing and a "new" tag next to it if you have the new one

[forest walrus]

OAI GPUs are porbably in flames rn

[real edge]

mine just started working on Soras site. Check again guys

[pastel crystal]

It seems very random for now

[dim cradle]

yep

[real edge]

i think the paused the rollout to plus users on the main ChatGPT site because i never got access

[chilly dock]

Found it. It does give the option to switch to Dalle-E then I tried the new model and got this “The image generation tool I normally use for these edits is currently unavailable, so I can’t directly modify the picture right now. However, if you’d like, I can guide you on how to do it yourself using software like Photoshop or a free online editor—or I can describe the scene in detail for use in another art generation tool. Want to go that route?”

[kind grail]

now sora in general isnt loading anymore for me lmao

[real edge]

ive just got the sora site to give me an image twice in a row. working for me

i think they blocked normal chatGPT uses from using it so now Sora can get through.

[mortal lava]

Sora “works” but I’m pretty sure that it’s using DALL-E given the crappy text generation.

[kind grail]

no. I just generated smth accidentally that was absolutely impossible on DallE3

[wild ferry]

Nah, hands and faces are way better than in Dall-E 3

[mortal lava]

Really? Must just be me then. Wonder if people are starting to get back access.

[real edge]

works fine on Sora for me. The text is perfect

[kind grail]

funny enough it was a prompt that is relatively safe and harmless on DallE3 and randomly flashbanged me with...a rather "interesting" output on Sora ._.

[real edge]

anddddd just as i said it. Now its not working for me again

[kind grail]

same

deadge. Well, gotta wait till they get the ice buckets over to the server

[keen hollow]

I'm trying the selection tool in ChatGPT right now after it created a image based off of another character image

and its working amazing btw

[real edge]

i still dont have image gen in the ChatGPT UI on plus

[plucky heath]

is it me or is it giving a lot of errors?

[keen hollow]

You just ask in a 4o chat

for it to create an image. Give it a reference if you want character consistency

[real edge]

mine still uses DALLE

[dim cradle]

i hate you have to archive then delete. why can i not just delete images i dont like (or video) on sora

[keen hollow]

ahh, darn

[left ginkgo]

hey is it just me that the image gen has completely stopped working lol

[tall tulip]

I think it just died again. this is gonna be revolutionary when it fully rolls out and isnt dying

[elder forge]

agreed... first image model that seems like you have near complete control of

[keen hollow]

100%. This is revolutionary and just deleted Graphic Designers

[orchid kayak]

Nah, not yet, but as a GD myself, I now think in 1 year 90% of my job will be done through AI.

[tall tulip]

I gave it a prompt I said in my head there was no way it would be able to perform, image of me, remixed with clothes I wanted in an example environment. While it wasnt perfect, it got me to like- 90% likeness, just a fragment of uncanny valley. and thats off of quite literally one selfie of my own.

and THAT- THAT is horrifying and beautiful at the exact same time

[elder forge]

AI users need to stop making statements like this... a graphic designer is still required to make full-fledged websites with consistent themes... someone may use this tool for inspiration but you're jumping the gun by saying this is taking jobs.

[orchid kayak]

Web design isn't all there is in graphic design btw, there is much much more, all marketing campaigns, all medias, all printable stuff, all graphic design areas.

[proud elk]

Do you also have this problem that when you select more than one variant, only one picture always shows up anyway?

[coarse flint]

the quality of the images are the best ive seen

[white olive]

The greatest part about it is how specific you can be with the details positioning as well as nailing every single piece of text you decide to use in the image. Creating comics with different panels exactly as you want (at least in order) is amazing

[kind grail]

Ok i am genuinly SHOCKED that OAI didnt bother at all about copyright

[dim cradle]

copyright stuff i try doesnt work

asking for spiderman -- I'm unable to generate the image you requested due to content policy restrictions. Please provide a new prompt or let me know how else I can assist you!

[kind grail]

you can check out #images-canvas we literally created League of Legends chars without even trying to hide the names

[forest walrus]

It may refuse very popular stuff, but you can still give it a reference image anyways and don't tell it that it's the popular character

[dim cradle]

well league of legends isnt really anywhere a household name as spiderman i guess

[kind grail]

It was restricted in DallE3 tho.

[dim cradle]

interesting

[burnt trail]

Does everyone have new imagegen within the gpt4o chat?

[white geyser]

No

[open wagon]

Is it disabled

[dim cradle]

lots of marvel, star wars, dc stuff i try all got deny. only i got to see half a batman it made in gpt mode and it looked incredible 😭

[burnt trail]

I see people on reddit posting examples

[elder forge]

i had it... then i lost access

[forest walrus]

Some do, Some don't, but just use Sora Image Gen for now, it's the same model

[open wagon]

I can’t find it

[white geyser]

Exactly my issue. Terrible

[open wagon]

Nvm I got it back

[forest walrus]

There is very heavy traffic right now.

[burnt trail]

Is there any notification or you just test it by quality?

[open wagon]

No it’s just server side update , it change from dall e to create image button

Or if you’ want to know if you have it or not, type create image, newest image generation model will contain “Getting started” in the beginning of the response.

[pastel crystal]

Unable to generate
Service at capacity, please try again later

Does anyone else have this?

[spare onyx]

yeah it's gonna be frequent for a while I reckon. Huge drop, and free to use

most of my generations have failed

[pastel crystal]

Yes same for me!

[spare onyx]

when the traffic is a little less we should be good

[open wagon]

I honestly did not expect this to have downtime just becuz it hits capacity

[pastel crystal]

Ok one just went through. Frankly I prefer request blocking rather than ahvign a generation fail mid process

[spare onyx]

there's a thousand people trying to generate elephant-less rooms and wine glasses rn

[open wagon]

Idk I expect better from OpenAI at this point ,

[spare onyx]

you mean you dont enjoy seeing it get to 99% and then saying 'There was an unexpected error'? Crazy

[kind grail]

Even hoyoverse is open now. Jesus christ what have they done?!?

[proud elk]

They may have slightly overestimated their capacity over people's interest.... nothing new 😄

[pastel crystal]

I know right 🤣

[spare onyx]

all part of the marketing isnt it 😂 "look how demanded we are"

[open wagon]

Well at least they didn’t do that same thing like sora , overhype first then release a year later

[spare onyx]

technically they did tease Dalle 4 a year ago

[open wagon]

Only for it to be so underwhelming

[open wagon]

Oh yeah the whole it can do text generation on image subject without a hitch and stuff

[spare onyx]

yeah if they'd have released sora sooner it wouldnt have been so bad

[open wagon]

Kinda unfortunate that OpenAI fall behind in term of generative ai stuff , image and video thingy

[kind grail]

also the new model is much more lenient on "frisky" content. Which is surprising considering how conservative DallE3 still behaved

[open wagon]

Watch them nerfing the image generation the next day

[kind grail]

pls no lmao

[open wagon]

They did it to advanced voice mode, I wouldn’t be surprised if they decided to nerf this one

It’s kinda expected at this point,

[forest walrus]

Let's not make blind guess assumptions...

[hexed umbra]

You find GPT-4o image gen underwhelming?

[open wagon]

Sora

[hexed umbra]

From what i recall voice mode was nerfed before public release, not really in the days after it released.

But like in comparison to the alpha testers or original demo it did seem nerfed

[open wagon]

Yeah I was part of the alpha group tester thing so it had a huge difference back then vs now on AVM it’s disappointing tbh

[hexed umbra]

Oh the video generation, yeah they changed it to Sora Turbo on release which did have a step down in quality compared to what was demod as well. With 4o image gen the quality has seemingly quite decently improved since the last demo we saw

[hexed umbra]

Yeah fair enough, but its looking good for 4o image gen atleast

[forest walrus]

It actually only got better after, still can't sing (Without hacks), but at least doesn't hallucinate as much now, and hears you better + able to make more sounds than before without complaining...

[hexed umbra]

Yeah that's true they do keep on working on it

[forest walrus]

4o Image Gen has been addicting so far lol, it's so good!

[open wagon]

Idk about that, I tried using image gen in sora, and then use the “image to video” feature in sora, and what I got was subject having deformed arms and legs ,

Image gen 4o is actually on same level with imagen and others now , which I’m impressed but it took them this long but oh well

[ionic tartan]

I would honestly rather just use this generator in ChatGPT. Using it in Sora don't feel right. Lol

[hexed umbra]

Yeah i would imagine Sora would struggle with animating images

[hexed umbra]

It is rolling out in ChatGPT

[quartz vale]

Glad you're enjoying it!

[open wagon]

.

[ionic tartan]

I know. But it sure is taking a while. Lol

[cobalt marsh]

sora is frustratingly slow

[hexed umbra]

Well not only that but 4o image gen is more capable and has a lot better understanding of the world which is quite useful. And its the best model at generating text, but i would also imagine its really amazing to be finally done with DALLE lol

[hexed umbra]

I don't have it in ChatGPT or Sora 😂

[forest walrus]

Nah, it's better, it's great at editing, and prompt coherence, understands stuff like objects well, and can even generate comics and memes with convincing results (i'm an artist and some generations are hard to spot at first, it's that good)

• Transparent images, that alone is enough lol

[open wagon]

Which I hope they don’t nerf most of these good stuff features

[weary gyro]

why does my gpt still use dalle

[open wagon]

Did they discuss anything about safety filter guard for the image generation, idk the term for it

[hexed umbra]

One way in which i can imagine them nerfing is stopping it from generating certain public figuresd. OAI said if any public figures don't want images generated of them, they can just ask OAI to prohibit it, which i think is fair

[open wagon]

Yeah that one is understandable

[forest walrus]

Competition is fierce, there is no logical reason to "nerf".
They only did it for AVM so they don't get in trouble with some figures until they figure out safe gaurds good enough to still allow singing while preventing voice mimicking...

[open wagon]

We will see in the next few couple of weeks then

[hexed umbra]

There is safety filters, and you may run into them, but they do seem relatively relaxed atleast (although sometimes you need to ask it to generate it again form what im seeing).

[weary gyro]

is there no transformer image in eu?

[hexed umbra]

I'm unsure how it will do with topics like gore though, i would imagine not though.

[open wagon]

I mean the singing part was just , something I didn’t really focus on, idk why most people want AVM to sing so bad, for me it was the maintaining with the personality that user set using custom instruction but I will not get into that now lol

[forest walrus]

Some very popular stuff it will refuse, same with violent / sextual stuff which is expected, but it's relatively relaxed (Definitely much better than DALLE3)
Also, you can still get it to generate images based on some popular stuff if you gave it a reference image of the character without telling it's name...

[open wagon]

I might want to use some of the image gen 4o for image to video in Veo 👀

[weary gyro]

why i have no gpt4 image

[open wagon]

Generate on sora.com for now

[lunar ermine]

How do I activate it? The new image you generator?

[forest walrus]

Check these, amazing results:

• #images-canvas message
• #images-canvas message
• #images-canvas message
• #images-canvas message
• #images-canvas message
• #images-canvas message

[open wagon]

It’s on high capacity right now so not everyone will get it right now , sora can do image generation too so go for it

[quiet brook]

someone dare to generate Mario with it😭

[open wagon]

Not even sora is safe, I can’t do any image due to high capacity lmao ok got it nvm

[weary gyro]

but u cant edit images there?

[open wagon]

You can, it’s called “remix”

[forest walrus]

I will only risk it in ChatGPT when it rolls out for me there lol, Sora is too risky for these stuff for now!

[weary gyro]

sora is risky?

[open wagon]

No it’s not lol

[quiet brook]

it’s out for me in chatgpt

[forest walrus]

You agree to not generate copyrighted stuff, I don't want to risk getting my account banned

[open wagon]

Just open sora.com and generate image from there

[quiet brook]

haven’t tried in sora yet lol

[forest walrus]

Lucky

[cedar dagger]

I am scared ngl

[forest walrus]

Amazing how I went from kinda hating ChatGPT in Q2 2024 to loving the service so much lol!
Great tools and models

[quiet brook]

I think it’s now safe to say that

RIP DALLE

[cedar dagger]

Welcome on board

[quiet brook]

very damn impressive tbh

[forest walrus]

DALLE was stomped on, grilled, deep fried, yeeted, and clowned over by 4o Image Gen 💀

[quiet brook]

I love how it VERY accurately generates the GameCube and e235…

[forest walrus]

I always said Custom GPTS / Projects, Canvas, Deep Research, and Data Analysis were my favorite ChatGPT features, but I think this new image gen may be my #2 after Deep Research lol

[quiet brook]

hope i dont run into some ridiculous moderation or limit or something because im having a blast with it

[forest walrus]

Zero Productivity since release :V

[somber turret]

text still gets messed up when asked lol...

[verbal sequoia]

Get notified for server events!

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

Also, check out the ongoing Canvas Creations event!

[hushed needle]

how do we create image? i tried and got a gradient image... lol

[keen hollow]

it is EXTREMELY powerful for creating start or end frames.

Simply drop any photo in (or generate one) and ask 4o to create the end frame you desire. Slap them together in any start / end frame vid gen and Voila.

[open wagon]

Wait wait , tell me more about this cuz I’m curious

[keen hollow]

A lot of Video Generators (Kling, Luma, Pika) allow you to have start and end frames. With this you can say, have a picture of a spaceship and tell GPT 4o to generate a photo of the ship exploding. Then use the original as the start frame and the generated one as the end frame

and its incredibly accurate at keeping consistency, even though the exploded pieces

[open wagon]

Ah is it possible to do that in sora

Cuz ChatGPT image creation is down for me

[keen hollow]

Yep! Just upload the photo and click "Remix" then describe what you want the end frame to be (the ship exploding)

Make sure Sora is on "Image" instead of "Video"

[forest walrus]

Tip: Image Gen Can Help generate images for Sora Video to use, it can lead to better Videos

[kind grail]

remixing images comes along with quite a steep drop in quality tho

[quiet brook]

so damn impressive but im gonna say this: enjoy while it lasts

[hushed needle]

whats impressive?? please share!

[quiet brook]

openai is NOT google which latter had just loosened content moderation on their native image gen lately

[cedar dagger]

everything

[proud elk]

When remixing photos, despite the selected area, it still edits the rest of the graphic slightly

[hushed needle]

LOL but for real, are you using something specific???

[cedar dagger]

did you try it? we all just experimenting rn

We still dont know all its total potential

[hushed needle]

LOL did i try what? i'm asking what it is so i CAN try it. LOL

[cedar dagger]

basically, it can follow your prompts much better than the other models

[quiet brook]

this isnt some massive disappointment like sora, latter also waited for 10 months for its release...

[cedar dagger]

and so on

[forest walrus]

A new Image Gen Model Called 4o Native Image Gen released, it's much better than DALLE3, more relaxed (Less Censored), great at drawing convincing illustrations and comics / memes, and very good at editing images + better prompt coherence and ability to generate transparent images...

[hushed needle]

I love how no one is actaully saying "what" it is!!!!

[hushed needle]

And how do you access it???

[forest walrus]

@hushed needle ^

[quiet brook]

try on sora dot com if you dont have it on chatgpt, vice versa
to confirm that you got it, the loading animation should look like a gradient picture instead of a loading ring (in chatgpt) and a "image" tab added into the sora sidebar
from what i have been testing, it is extremely good at text rendering and generating specific stuff, but be cautious that it WILL block copyrighted characters (e.g. mario)

has not hit a limit yet

[kind grail]

sora deadge again

[placid orbit]

There’s no limit they changed it recently to unlimited videos and images

[quiet brook]

lmao

[merry valve]

Truly, hats off to OpenAI for their native image generation. It is incredibly well executed, with very high image quality and exceptional prompt understanding. At this point, I would confidently rate it as the number one solution available. No doubt about it.

[quiet brook]

i sincerely wish that openai doesnt nerf or downgrade this model like what dalle 3 recieved, its just too good...

[azure roost]

wonder when the chatgpt rollout will be done

[forest walrus]

This model reminds me a lot of why DALLE2 was so special, and even has some similarities...

[median stratus]

is imagegen deployed to plus users or not yet? doesnt work on my side...

[forest walrus]

Free, Plus, Team, and Pro (All basically) but it's a slow rollout as it's still quite new and there is high traffice, if it's not working for you on ChatGPT, try Sora Image Gen...

[quiet brook]

i literally cant stop using it

[haughty spruce]

Yesssss!!! Finally... All my dreams came to life. I'm in tears... 🥹

[quiet brook]

also has anyone hit a limit yet

[forest walrus]

"Sorry boss, was uh, generating anime frames of people"