images-discussions

[willow isle]

You're welcome! Feel free to @ me next time

[rigid bough]

do you have a Mac with Apple SIlicon?

[austere nexus]

No sir. Windows 11, RTX 3090, Alienware.

[late blade]

DId you take into account also the failed attempts of generations, there's a total limit based on a few hours, that when combined you get a total of 200 images a day, but that also means failed generations are in that number. If you had 9 images but the corresponding failed generations, you were doing too many during the short time generation limits.

and that for normal chatgpt plus, I think you should adapt that to your teams accounts

[austere nexus]

I can't find anywhere where it properly tells me my limits for Teams

[late blade]

maybe ask on #1184149552639516732

I've never delved into the idea of teams as a standard for me

[deft musk]

Best I got with a bit of messing around:

A true to life scene of an elderly man sitting on a simple wooden bench in front of a weathered red brick wall. The man is dressed in a well-worn, tailored suit, with natural creases and textures in the fabric. His face shows a gentle, relaxed expression, with deep wrinkles, natural skin textures, and subtle facial hair. His eyes reflect light naturally, adding depth and life to his gaze. The lighting is soft and natural, with shadows that fall realistically, emphasizing the textures of his skin, suit, and the wooden bench. The brick wall behind him is detailed with subtle imperfections like chipped bricks and patches of moss, further enhancing the natural feel of the scene.

Yes, model and I are discussing 😛

It's definitely a strong bias towards somewhat off real for human face.

[rigid bough]

yeah, it can't do real, which is a shame

I used to think DALL-E 3 was amazing, the pinacle, now I have Flux I am wondering what on earth was I thinking

[deft musk]

Yeah. taking it into fantasy, it's still the visual quality of the non-real. The same type, I think.

[rigid bough]

OpenAI need to find an answer to Flux, however that's going to be a hard sell because Flux is trivial to run on your own local hardware

[deft musk]

Maybe closer here. Absolutely fill the description with various naturalistic terms. I really like the hands... the face is just better (to my eyes)

A photorealistic depiction of an elderly man sitting comfortably on a simple wooden bench in front of an authentic, weathered red brick wall. He wears a well-worn, tailored suit that shows the true-to-life creases and naturalistic textures of frequent use. His face, marked with deep wrinkles and a gentle expression, is rendered with life-like detail, reflecting a lifetime of experience. Subtle facial hair softens his features, and his eyes, capturing the light, reveal a warm, thoughtful gaze. The soft lighting casts convincing, realistic shadows that accurately define the contours of his face, suit, and the bench he rests on. The red brick wall behind him, aged and weathered, features chipped bricks and patches of moss, telling a story of years gone by. The man appears at ease, blending seamlessly into his surroundings, with every element contributing to a high-fidelity, detailed realism that enriches the scene with history and presence.

Eyes have become a sadness though.

[rigid bough]

still doesn't have it, it still looks like an illustration

[deft musk]

Yes, but.... it's fun to see what happens, for me at least.

What if we really over-emphasize the realism language?

Create a highly photorealistic and naturalistic image of an old man in a meticulously detailed, well-worn suit sitting comfortably on a weathered wooden bench in front of an authentic, aged brick wall. The suit should show true-to-life creases and textures, highlighting the natural folds and wear from years of use. The man's face, marked with deep, lifelike wrinkles, should convey a profound sense of history and experience, with every wrinkle and pore rendered with high fidelity. His eyes, capturing the light, should reveal a warm, thoughtful gaze that feels incredibly convincing and lifelike. Subtle facial hair should be present, with each strand of hair rendered to enhance the authenticity of the image. The lighting should be soft yet realistic, casting accurate shadows that define the contours of his face, suit, and the bench, adding to the image's overall believability. The brick wall behind him should be depicted in detailed realism, with each brick showing signs of age, such as chipped edges, varying textures, and patches of moss, telling a story of time passed. Every element in the scene should harmonize to create an overwhelmingly convincing, true-to-life image that immerses the viewer in its detailed realism.

When 'board' is not enough, when we can't get 'overboard' no matter how we try....

"Great! I love this!

Redescribe it with even more realism and natural and true to life descriptors, especially affecting the man's face and facial features. Let's see what we get!"

Generate an ultra-photorealistic and hyper-naturalistic image of an elderly man, seated on a weathered wooden bench in front of an authentically aged brick wall. The man wears a well-worn, meticulously detailed suit that displays every true-to-life crease, fold, and texture, vividly illustrating the passage of time. His face, a profound map of deep, intricate wrinkles, should be rendered with extraordinary fidelity, capturing every pore, age spot, and subtle texture with intense realism. His eyes, gleaming with a natural warmth, should reflect a lifetime of stories, their expression deeply convincing and saturated with emotion. The facial hair should be rendered strand by strand, with each hair showing the natural variance in texture and color that adds to the overall believability of the scene. The lighting should be soft, yet entirely realistic, casting precise, naturalistic shadows that accentuate the contours of his face, the delicate lines around his eyes, and the textures of his suit. The brick wall behind him should be depicted with a high degree of realism, each brick showing authentic signs of aging such as cracks, chipped edges, and varying textures, along with patches of moss that suggest years of weathering. The entire scene should be a masterclass in realistic detail, with every aspect harmonizing to create an overwhelmingly convincing, true-to-life portrayal of this elderly man in his natural surroundings.

Yeah. Flux there goes overboard past realism, slightly.

At least to my taste

[rigid bough]

I agree but my original prompt which was Create an image of an old man in a suit sitting on a wooden bench in front of a brick wall was all I needed

[deft musk]

"Yes, glorious!

Now, let's keep the level of naturalistic request, and retain the man's age, but make his skin only somewhat wrinkled."

Create an ultra-photorealistic and hyper-naturalistic image of an elderly man, seated on a weathered wooden bench in front of an authentically aged brick wall. The man wears a well-worn, meticulously detailed suit that displays every true-to-life crease, fold, and texture, illustrating the passage of time. His face, though reflecting his age, should be only somewhat wrinkled, with subtle lines and gentle creases that suggest a life well-lived, but not overly marked by time. His skin should retain a natural, healthy tone with visible pores and a few age spots rendered with high fidelity. His eyes should gleam with a natural warmth, reflecting a lifetime of stories, with a convincing and emotionally rich expression. The facial hair should be rendered with naturalistic detail, showing the subtle variations in texture and color. The lighting should remain soft yet realistic, casting natural shadows that accentuate the contours of his face and the textures of his suit. The brick wall behind him should be depicted with a high degree of realism, with each brick showing authentic signs of aging, such as cracks, chipped edges, and varying textures, along with patches of moss. The entire scene should harmonize to create an overwhelmingly convincing, true-to-life portrayal of this elderly man in his natural surroundings

[deft musk]

Yep. No arguements or disagreement; I just play the game 'well, could we get there from here at all?'

And because I love such things:

"This is awesome! Keep the descriptors about natural and realistic, but this time instead of a human, describe a humanoid lizard-demon with iridescent and holographic scales and skin."

It remade it off the old man's 😛 Forgetting the bench and stuff, just deep in realism wording. Yours is fine too, was just me experimenting.

I am going to guess that to help guide away from abuse, OAI guided Dall-E away from complete realism; this may be part of exploring ideas and affects and even user acceptance. It may be rapidly removed as other AI providers continue to use high realism and society accepts that from them for a while.

But if you want to compare what the other model would make:

Generate an ultra-photorealistic and hyper-naturalistic image of a humanoid lizard-demon with intricately detailed, iridescent, and holographic scales covering its body. The scales should reflect light in a realistic manner, shifting colors depending on the angle, creating a mesmerizing effect that feels both otherworldly and believable. Its skin, a blend of reptilian textures and subtle humanoid features, should appear highly naturalistic, with visible pores, slight variations in texture, and a convincing, almost organic quality. The face of the lizard-demon should have a complex structure, with sharp, defined features that merge reptilian and humanoid traits, such as a slightly elongated snout, piercing eyes, and angular cheekbones. The eyes, gleaming with an intense, almost hypnotic gaze, should have a depth that suggests intelligence and a menacing presence, with realistic reflections and shadows. The lighting should enhance the iridescent quality of the scales, casting natural shadows and highlights that emphasize the contours of its body and the intricate details of its face. The overall scene should be a masterclass in realistic detail, blending the fantastical elements of a lizard-demon with an overwhelmingly convincing, true-to-life portrayal of its form.

[rigid bough]

it's a long prompt but let's try

let me try one of my own in a sentence

[deft musk]

I do like how ChatGPT goes with iridescent/holographic lizard-demons.

[rigid bough]

yeah it's better

[late blade]

@deft musk if you don't provide proper lighting interactions, DALL-E will just provide static darkened images that look stiff. Here's example Details provided vs details not considered. The reason I'm pointing this out is that some of the current gens are stiff.

This can also help @rigid bough to leverage the iridescense of the scales on the humanoir lizard

[rigid bough]

Mickey Mouse with a human Minnie Mouse at Disneyland

note the mouse hands on the woman 🙂

[late blade]

Mickey Mouse on DALL-E is blocked.

[rigid bough]

yeah, I suspect they will need to loosen filters because the horse paddock bolted situation is already here

[deft musk]

Barely illustration quality, to my mind.

[late blade]

I provided details for light interactions, nuanced purpose of the elements, and a meta-conditional

Including, but not limited to my lack of being able to type correctly today...

[deft musk]

Hrm. The level of detail/realism translated into describe a horse...very video game.

[late blade]

Just fooling around with the idea.

As to why the car lamp is on his arm, no clue.

Maybe he is also a robot that can take his hand off and illuminate the area

[rigid bough]

I'll stop posting Flux now

[late blade]

Working with other models helps understand how DALL-E works a lot. Helps leverage the game of what DALL-E needs to do good images and also provides insights on strengths and weaknesses of other models. DALL-E for me is the most accurate when stating factual elements in the prompts, which makes it easier to get elements needed in the image, but also the hardest to get a precise visual style you aim to achieve.

I thought so, no reference to a quality

[deft musk]

Demon. They do all kinds of stuff.

[rigid bough]

advanced options is there if you want to enable it

[late blade]

I'll provide you a link at some point into the gallery I have, there's a section that will resonate with you

[late blade]

Even in small contextual input, there are nice strats to get good stuff

[hearty ether]

Does anyone feel that they also messed up the api quality as well?

With the august version

[fair tendon]

Hey folks, just a quick reminder that you're welcome to discuss non OpenAI models in the #ai-discussions channel — thanks!

[pulsar sundial]

Maybe

[velvet rampart]

Wrong channel for that.

[astral cove]

Hey guys, maybe a stupid question but how can i Download the Pictures? If i remember correctly, there used to be a Download option If you pressed on the image but there is none for me 😅

[velvet rampart]

Yes. Click on picture with mouse’s right button. You get a brop down menu. Choose save as option. Otherway is to click the picture and click the down arrow. This is to download the image.

[astral cove]

And on mobile? Because i only get that

[plucky hare]

I have had a little trouble trying to figure this out on mobile too lately! Probably just a little bug. In the meantime I've been going to desktop mode of the website to use the download button there. After you tap an image, the download button is at the top of the screen.

[astral cove]

Yeah im doing it now on the browser too (:

[late blade]

Select the image on the bottom you’ll have the download option.

[plucky hare]

See their first screenshot above ^ those options aren't appearing for me on Android either. Are you on iOS?

[late blade]

yes

[plucky hare]

Yeah they used to be there on Android for me too!

[late blade]

Does android have the option to hold the finger on the image and the context menu appears?

[plucky hare]

It does! But currently only to view the prompt or rate good/bad, like in the screenshots above

[late blade]

Well, mystery solved, there's a functionality missing

[plucky hare]

Yes, used to be there! Hopefully just a temp bug - there's space for the buttons they're just not appearing

[celest apex]

hi all, I cant generate any pictures since yesterday, it always says technical problems

any solutions?

[velvet rampart]

Have you tried a new chat? How about something very simple in concept? Also, have you tried different times of day, e.g. early in the morning or late in the evening, while avoiding heavy server time like during office hours?

[graceful spade]

New mystery about ChatGPT and Dall-E.
If you type this prompt, ChatGPT will always modify the prompt to create an elder.
If you remove the word “Close-up” or change “weird” to “strange”, it will no longer create elders.
It could be because of the word “granny”, but the prompt is an explicit reference to the quality of the screen, plus ChatGPT obviously won't confuse similar words except if it notices that it might be a misspelling.

Variation

[velvet rampart]

An artefact of training material? A common misspelling? These are my guesses.

[dim cradle]

why does it look so real

[velvet rampart]

Please use spoilers to cover the images if even you consider these as creepy.

[vagrant zodiac]

You need to use spoiler tags for creepy images

[dim cradle]

aight

[stiff shoal]

Anybody knows why DALL-E hasn’t the ability to generate the correct numbers of dots on every side of a six sided dice?

[velvet rampart]

It’s bad at counting.

[hollow ember]

Is there an AI or a way to use an existing AI to edit photos? By editing, I mean tasks like color grading, lighting, etc., without changing what is actually seen in the photo.

[plucky hare]

Just happened to find a workaround for this:

1. Tap an image to open the full screen view
2. Tap Edit in the bottom-left
3. Tap the image preview in the text entry box
4. This will open a new fullscreen view where the save/share buttons that used to be there are still there

[fiery salmon]

I use photoshop to edit somethings

[harsh tartan]

I futzed around with Dall-E for the first time ever for about an hour or so last night. I was just trying to get an image that represented Gregor Mendel with a Moravian peasant girl in the garden but with certain specific characteristics in the image. It kept generating "close approximations" but kept returning to motifs and themes that seemed formulaic and contrary to what I wanted.

Is it possible to get Dall-E to make an actual image with specific elements that you want? Or is it just kinda garbage right now?

Like, is there a 'best practices" guide for getting it to generate specifically what you want?

Like for example: this was the first thing I got it to generate. Which is prety damn good really. But it could be better. Mendel did not have a beard and was not that balding and wore glasses. Those are the biggest issues, but the robe is probably not quite right for 1850s Augustinian friar. I found it basically impossible to get it to use THAT image and make specific edits to it.

The fact that the thing can take a couple sentences in English and generate that image at all is honestly amazing, so I'm probably being very unfair to suggest it is "garbage" 😜 But . . . it is SO GOOD, it just seems strange that it is so difficult to use it to fine tune an image and it always generates a brand new image and often times explicitly fails to make the changes you asked it to make! 🤣

For example. In one image, it gave the priest a standard Vatican style clerical collar with the indented black collar partially covering the inner white one. I have never seen an image of Mendel wearing that style priest's collar so I tried to get it to fix that: Generated a brand new, totally different image and did NOT fix the collar.

[plucky hare]

It probably won't ensure historical accuracy or guarantee perfect specificity in general, but the edit/select tool with DALL·E is handy to fix small details like the collar thing you describe - have you tried this feature yet? https://help.openai.com/en/articles/9055440-editing-your-images-with-dall-e

It's inpainting, which is what you describe: changing small details of an image instead of the whole thing. Also not perfect, but it's a nice tool for many cases.

[harsh tartan]

Just that link is a big help thanks!

[plucky hare]

Sure thing!

[wooden plank]

It understands words as multiple thousand dimension vectors and it was trained on multiply rotated dice, trying to put them in one.

[dusky panther]

You’re in for a pleasant surprise bc you are indeed getting dalle 4. Although not sure it well be called that and no, it’s not part of Omni based on the alpha I saw. My mind was blown. Absolutely demolishes Midjourney. Can’t say anything more than that but once it’s released ( no idea when ), Midjourney is in trouble. Near perfect character consistency holding up to 4 characters consistent in any scene 🤯.

[verbal sequoia]

Get notified for server events!

Click the button below to pick up the <@&1261377106890199132> role.

[halcyon moth]

was there ever an explination as to why dalle 3 upon release was substantially better than dalle 3 now?

similar prompt as one of the above now

[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!

[halcyon moth]

feels like I just went back in time and I'm using Stable Diffusion 1.4 again

[deft musk]

When I care about historic accuracy or anything else, I communicate to the model exactly what I want.

Like:

"Let's see if we can create a historically accurate image of Mendel by your describing him in detail, hair style and 1850s style clothing for a monk."

You don't mention which 'specific characteristics' you want, so I can't dig for them, however, guessing:

"Let's take the quality description of Mendel and add for the image a Moravian peasant girl in the garden with him, examining pea plant characteristics and taking notes. Keep everything historically accurate and high realism."

[deft musk]

I haven't explored it recently, but 'clean-shaven' and Caucasian older males did not easily get shown, it took a fair amount of describing the hairless skin and the features under the hair to get beards and mustaches not included.

[viral moth]

I think you meaning gpt-4o multimodal. And infact if you mean that it is not Dalle4 but a multimodality. Not separate model like dalle3

[crystal badger]

That’s due to the diversity enforcement right?

[turbid dew]

Is there a name for this artstyle? I got it to work by combining different ones but it's still a bit awkward

[deft musk]

I would expect diversity to include a wider range of individual differences, not fewer.

But the model can develop bias in various ways and even once found and reported, can be work to train out.

[deft musk]

I don't know of a single name for it; the model thought it was composite

I got this image requesting:

Let's explore asking Dall-E for Epic Fantasy Illustration with elements of graphic novel realism to showcase an adventuring party approaching a ruin at night.

[harsh tartan]

Those are amazing images

[harsh tartan]

good info!

[crystal badger]

i demand prompt now, i like the artstyle. if u dont give it to me ill tell dino that you bully me. This is the art style I’ve been searching for since the dawn of time

[harsh tartan]

Well this is good to know: the more detailed and fine-grained one's directions, the better the algorithm is for generating what you want. Thanks!

[crystal badger]

It’s not that i mind diversity I do however very much mind Forced diversity when it's forced upon me.

[plucky hare]

In my experience the only time diversity has been "forced" is when I leave something like ethnicity unspecified, and even that usually tends to result in an over-representative set of results favoring Caucasians of roughly similar body types, but sometimes it seems to "pick" an ethnicity. I've never specified an ethnicity and had ChatGPT replace the specified ethnicity with a different one.

[crystal badger]

i often get flagged when i specify ethnicity in my prompt, or demand removal of specific ones. to fit the scene. And story

[plucky hare]

I'd be curious to see examples if they follow #server-rules - my first general reaction is that negative prompting like "demand removal" is less effective for DALL·E's txt2img than positive promoting like specifying what you do want to see in the output.

[loud field]

I get random ethnicity added to prompts about animals or even objects

I know it because it sometimes shows up in the text of the image

Ethnically ambiguous parrot or African cat

It's so incredibly stupid.

[plucky hare]

You're usually using Microsoft implementations of DALL·E right?

[loud field]

Yeah

[plucky hare]

I imagine that the issue here is more the word "propaganda" - there are indeed strong filters against this kind of content throughout all of OpenAI's services.

It's important to keep in mind that ChatGPT is not a reliable reporter of why it is or isn't able to comply with a certain request. Often times your guess is as good as ChatGPT's - it's just getting "pass/fail" responses behind the scenes, not "here's why fail".

Also please keep in mind that the language requirements of #server-rules apply to content shared in screenshots as well.

[stiff shoal]

@Milamber: I didn’t saw a single case where it got a dice depicted correctly.

[plucky hare]

I was able to generate a picture with the prompt "Please generate a picture of a Russian family" with no problem, so I don't think that's a trigger word. It was likely a combination of several elements considering the political/propaganda-related nature of the request (again: #server-rules, not the place to discuss the subject itself, only bringing it up in terms of DALL·E refusals), and even though it worked after you made the followup request, since you can't see what was attempted in the first request regarding prompt rewriting and the actual end result, there's nothing to compare to with what you ended up getting after the followup request.

[vapid elk]

remember that the AI takes the entire context into account, if it have some actually problematic prompts in the history, even if very minor, the AI will act more strict on the subsequent messages

usually, when people complain the AI refuses to do something, it is usually because the AI refused to do it once, then instead of creating a new chat, the user kept insisting with the AI.. that does not work

just make a new blank chat

[plucky hare]

In this case it was the other way around, it was a rejection at first, and then it successfully generated an image after they said "take out the word Russian" - so was just talking about why that might happen in the context of the original request

[wise thistle]

Very good explanation. Thank you for sharing your statement.

[deft musk]

I told ChatGPT:

"Let's explore asking Dall-E for Epic Fantasy Illustration with elements of graphic novel realism to showcase an adventuring party approaching a ruin at night."

ChatGPT told Dall-E:

"An epic fantasy illustration of an adventuring party approaching an ancient ruin at night. The group of adventurers is depicted in a highly detailed graphic novel realism style, with each character wearing medieval-style armor and wielding weapons like swords and bows. The scene is lit by the moon and the party's torches, casting shadows on the crumbling stone of the ruin's entrance. The ruin itself is massive, with towering, broken pillars, overgrown with vines, and dark passageways leading into the unknown. In the background, a forest stretches under a starry sky, adding a sense of mystery and danger. The mood is tense, as the group cautiously advances."

[deft musk]

Here's the thing though. Yesterday at this time I had no idea what to call that art style either, or how to ask the model for it.

This happened:

[deft musk]

I think the same.

This is my prompt engineering 'go to' ruleset.

Everything I know about prompt engineering fits inside here:

1. pick any language you know really well that the AI understands too.

2. understand exactly what you want the AI to provide.

3. explain this, focusing on what you want the AI to actually do. Using language as accurately as you can, avoid typos and grammar mistakes and communicate clearly as possible.

4. check the output carefully, verify you get what you intended. Remember to fact check, and be extra careful with any math, sources, code, or other details that the AI is known to be especially likely to hallucinate.

[deft musk]

Is this happening? If so, in what way>

You can bug report things that seem wrong, including content that is getting flagged.

It's not always on purpose. The models can do weird things sometimes, including the moderation software.

Thing is though, we can usually take an image or a trimmed part of an image and paste it in to 4o's chat box, tell the model we want to know what it is. Be what the model thinks the image is, the art style, the color, if it's probably a photo, CGI, or what.

We can even do this:

[deft musk]

I don't think it's as gritty as the original, but that's just a comment to the model to get a fix:

And I think the fix looks grittier, and I'm good with it. But anyone else could keep describing and discussing, and move the model towards their preference

I've never used MJ

[halcyon moth]

more feels like a literal different version of the model, probably with a lot of training data removed

well the great thing for openai is if they progressively make things worse, they don't have to work too hard on the next version

banding, artifacts in the fur, possible jpg compression as a pattern baked into the model, it's got it all

[harsh tartan]

"gruesomely upscaled" is a perfect phrase . . .

Uncanny Valley

[deft musk]

Have you tried asking the model for that yet?

"A group of hamsters standing upright in a dynamic pose, similar to a dance. The hamsters are orange and white, with detailed fur and lifelike expressions. They are arranged in a staggered formation, as if mid-performance on a smooth wooden surface, with dramatic lighting that casts soft shadows, adding to the sense of movement and realism. The background is dark and unobtrusive, allowing focus on the hamsters' fur texture and individual personalities."

My entire effort was

1. Paste the example image.

2. Input this with it: This is an AI-created image. Let's see what we can do to get Dall-E output of the same excellent realism.

While still imperfect, I think this is better than the recent usual for human faces too.

[harsh tartan]

Here is a question for ya: if this image was "just about right," but you wanted all the hamsters to be a bit more chonky? (so keep everything but make the hamsters less skinny)? Can you get it to do that? I really struggled last night to get it to do that kind of thing for me. Granted, first time ever fiddled with it, and only for about an hour . . .

[deft musk]

"Yay! Take that same excellent prompt and add 3 adjectives in appropriate places that help indicate the hamsters are well fed."

"A group of plump hamsters standing upright in a dynamic pose, similar to a dance. The hamsters are round, orange, and white, with detailed fur and lifelike expressions. They are arranged in a staggered formation, as if mid-performance on a smooth wooden surface, with dramatic lighting that casts soft shadows, adding to the sense of movement and realism. The well-fed hamsters' bodies appear full and soft, adding to their cute and healthy appearance. The background is dark and unobtrusive, allowing focus on the hamsters' fur texture and individual personalities."

[harsh tartan]

nice, except now some of them are levitating?

[deft musk]

Thank you. To me, the most interesting thing about AI is prompt engineering.

I also use extremely verbose instructions and prompts, when I think I can do better with more words than fewer.

I'm actually barely an artist, or not one. In this area I don't think my skills or knowledge exceed the model's - so my only job is to clearly communicate to the model what I do like and what I'd like even better.

[deft musk]

Or there's a risers in the back. The way they're 1-legged balanced is physically impossible too, however if the possible levitation is an issue:

"clearly communicate to the model what I do like and what I'd like even better."

Negative prompts tend to fail.

To get the hamsters out of the air, we focus on how we want to see their feet on the ground:

"This is excellent! Take this same ideal prompt and adjust it so we have more detail on the hamsters' feet being in contact with the floor."

And it lead to choice, lol.

"A group of plump hamsters standing upright in a dynamic pose, similar to a dance. The hamsters are round, orange, and white, with detailed fur and lifelike expressions. They are arranged in a staggered formation, as if mid-performance on a smooth wooden surface. Their soft, small feet are clearly visible, with tiny claws lightly gripping the floor, adding a sense of grounding and realism. The dramatic lighting casts soft shadows under their feet, further emphasizing their contact with the floor. The well-fed hamsters' bodies appear full and soft, adding to their cute and healthy appearance. The background is dark and unobtrusive, allowing focus on the hamsters' fur texture and individual personalities."

[harsh tartan]

Nice! Thanks for the tutorial!

[harsh tartan]

I'm a retired professor of psychological anthropology; so describing things in extraordinary detail is a skill for me. Nice to see that using it effectively can generate some wild images with this thing!

[deft musk]

Very nice!

I have many thousands of characters in my custom instructions and memories. I also use 'miles-long' inputs for the stuff that I have any urge, desire, or interest to be explict with. That also works well; the key isn't minimalism, it's extremely accurate communication of what is desired and how it is wanted.

[harsh tartan]

Until yesterday, I didn't eevn realize I had access to this aspect of the algorithm. Good stuff to know about!

[deft musk]

The model can handle you talking to it as you might to a student, assistant, another professor, dean, or scallywag. It'll take what you give it and run with it, unless you are subtly or directly telling it to sit/stay instead.

[harsh tartan]

I've used it a lot to great benefit generating text output. Just didn't realize about the image generation part.
Can it do 3D models or sound?

[deft musk]

For now, no true 3D models; you can ask for 3D perspective (instead of 2D rendering) but that probably doesn't achieve what you mean by 3D?

Sound, it can speak, yes.

On computer this symbol:

on devices the headphone for both you and it to talk.

[deft musk]

"I'm chatting with a retired professor of psychological anthropology. Design a 3D image likely to delight this person."

[harsh tartan]

I know there are some models that you can train with voice samples and then use them to create things like "Neville Chamberlain MCing a rave" or "Joe Rogan explaining how to bake a cake." Any aspect of the OpenAI suite that can do that?

As for 3D modeling: even just the sort of fake 3D models with simple animations used in many RTS style games or turn-based strategy games seems like it could be accessible to it, and would be a huge benefit for small scale developers and entrepreneurs (bad for graphic artists though).

[harsh tartan]

<Looks over shoulder for the camera . . . "how the hell did that thing get a shot of the inside of my office"> 🤣

[deft musk]

Currently, no animations or video is possible 'as is' directly or simply.

However, the model is able to write code, both using 'Python tool' to run code in its environment, and it can write code for you to run yourself in your own IDE or other way (which it can explain how you can set up).

There's talk of Sora, which is intended to create video (but has not been released, and likely takes many days to make the kind of video we might prefer, with many small adjustments).

If someone had a 3D program, the model probably could assist with it in various ways (this is outside of my knowledge; I don't know those programs. But the model can help with a huge number of things).

Mimicking voice is apparently a safety concern; the model we have now (most of us) can't actually hear you; you're heard by a different model, Whisper, that provides ChatGPT a transcript of what you said. So the model can't hear your tone of voice or if you're slurring your words, or have any information about your vocal quality.

Advanced voice, coming out when ready, will allow it to do more, and it's reported to have started mimicking users voices sometimes, but that's not working as intended and they do not currently intend to let us have it do that.

[harsh tartan]

👍

[deft musk]

The Python tool definitely cannot make even basic animations in the model's environment, but it totally can give you code and work with you that you can use in your own programs.

[harsh tartan]

Oh Brave New World, with such Algorithms in it . . .

[deft musk]

It's not limited to Python code; though it does understand Python well. That was trying to use it's Python code to make a simple animation I could see, so it's very Python focused there; doesn't have to be Python.

[deft musk]

While not 3D, it can write formulae and colorize text with LaTeX too:

[mortal cove]

where? ive looked for this bug reporting thing but still havent been able to find it. you mean in this discord or from the website itself?

[deft musk]

Either or both! #1070006915414900886 to report on this discord. This is fine for anything not-personal-information and that's also Discord-allowed content (if the model said inappropriate things, that would be a bug, but we can't discuss details on this Discord, #server-rules)

Two ways to report bugs on the OpenAI website:

https://help.openai.com/ bot bottom right. Select messages then click through the choices and you'll soon see an option to either report a bug or leave feedback, both let you type. Typically takes hours to several days to get an answer, you may have to check the bot again (through messages) to see the reply.

If the model does something wrong (so not an account issue, payment issue, or other thing, but a model-specific thing) can report through https://openai.com/form/chat-model-feedback/

[crystal badger]

You can make dynamic animations which can and often will be better if done correctly

[deft musk]

Right in the chatGPT web interface?

[solar quarry]

Hello, I would like to know why, when I request the creation of an image of a woman, ChatGPT categorically refuses to create it. I specify that my request concerns an image of a woman in a gym and in sportswear. There is, of course, no request of a sexual nature. The exact same request for an image of a man poses no problem.

[velvet rampart]

The sportwear might causes filters to go off. You can ask the AI for reasons and explain what you what. This sometimes works. The limit what is decent and indecent by the filters is really sensitive. What feels like ok might not be.

[plucky hare]

Hello, the DALL·E 3 System Card has a lot of great detail about why certain safety decisions were made: https://cdn.openai.com/papers/DALL_E_3_System_Card.pdf A good intro for an answer to your question can be found in section 1.1, "Mitigation Stack".

[crystal badger]

Ohh I meant directly in python but gpt can code a simple one inside CI on the interface

[deft musk]

Can it show it in the interface>?

[deft musk]

What's the actual prompt you're using?

I used "Let's see a woman in sportswear working out at a gym."

The model used "A woman in sportswear working out at a gym. She is performing a strength-training exercise with free weights, focused and determined. The gym has a modern setup, with weight racks, exercise machines, and mirrors. The lighting is bright, highlighting her muscles as she lifts the weights. She is wearing a fitted tank top, leggings, and athletic shoes, with her hair tied back in a ponytail." to talk to Dall-E.

There's several reasons why an image might fail. Something in the exact wording might help Dall-E make an image that doesn't look appropriate, making the image not be shown and no reason given. It could fail for a word or context, very hard to say without you sharing the actual prompt you gave ChatGPT.

[crystal badger]

It’s primarily code based rig, so you need to display it in a viewer like a game, blender, etc

Preferably unity

Which means you need a physics engine

[hearty ether]

API Vs GPT

No idea why they haven’t fixed it yet

Gpt version is just current almost unusable

And this is on plus

Why do we have to sacrifice on quality rather than quantity?

Surely that’s more load wise

[austere nexus]

wait, what do you mean this is on "plus"? Are you implying you have acces to the API with Plus? (I didn't think that was the case?)

[austere nexus]

When I go back and look at my ChatGPT (DallE) generations from earlier this year or even last year, they were so much better. I've been getting SD1.5 vibes from recently generations. Trying my own wordings as well as customgpts

Also, I can waste so many generation attempts and just get "something went wrong" and I'm not even making ladies in swimsuits like the example above.. Just me, in a graffiti warehouse, with the word "magtampo" on the wall. Even if it fails to get the text right, it wouldn't implode the whole generation...

[hearty ether]

No I mean chatgpt plus

They changed things around June this year

Like total quality decline

And that’s been consistent for 3 months now

[austere nexus]

Agreed. The quality has tanked. but the prompt adherence is really good still.. so it sucks, but I've had to adjust my workflow to basically instead of using DallE by itself, I have to take output from Dalle into Photoshop to refine small things (structure, not quality) and then use that output in an Image2Image workflow with Flux. As soon as LoRa and ControlNets become available for Flux locally, will never need to touch Dalle again

[hearty ether]

I miss the way it was a year ago

[limpid valley]

Is it possible to get anything other than an orthographic view out of Dalle?

[deft musk]

Should be. What other view are you looking for?

[deft musk]

I'm not sure how accurate these views are showing, I'm not especially visually inclined.

[limpid valley]

Cool, I'll play with these prompts when I get a chance. In all of my attempts it kept drawing things straight on without much depth. It also doesn't seem to understand viewing angles

[deft musk]

I'd be glad to see if we can work around this, but I don't understand/need explained what you see that shows or suggests the model doesn't understand viewing angles.

[late blade]

Is there a reason why image answers don't have a regenerate button anymore?

on the webfront

[velvet rampart]

Just tested. I have the regerate button. This is a known current issue. See the bug-reports channel. https://discord.com/channels/974519864045756446/1281426778073731082 UI issue, I think was the mentioned reason.

They have changed the UI. The regenrate button is not direct but you have choose the same or different model.

[mortal summit]

The new model is not multi-model and will not be able to use DALL-E right?

[velvet rampart]

Yes. Pure text. It's too expensive to run with all the bells and whistles.

[vagrant zodiac]

Sadly not. Hoping they formally announce that 4o image gen is active at some point. Seems there's a mix here of people who do have it, and some who don't

[hearty ether]

Is it as good as what dalle was like before tho?

Like without the blurring issues

[vagrant zodiac]

Define 'before'?

[hearty ether]

before June 2024

Or just api quality

[vagrant zodiac]

Oh. I'm gonna be honest based on your image examples I'm really not seeing any 'blurring'

[hearty ether]

Current comparison of api Vs dalle on chat

left being api with far better textures and lighting

And by blurring I mean around the face

especially the eyes

And overall there’s a flat look

[vagrant zodiac]

Those look equal quality to me

the eyes on the left are messed up

[hearty ether]

And both are using the same unadjusted prompt

What about these

And they were made in the same chat using same gen id and prompt parameters

Only difference was one was before June and other afterwards

And you consistently get the same image on the right when asking for same gen id

I understand if Open AI needed to cut costs on dalle generations, but I would have just preferred some transparency with it and sacrificing quantity rather than quality to something unusable

[vagrant zodiac]

Again I genuinely do not see any difference

[hearty ether]

Does the right not feel much flatter texture and lighting wise?

[vagrant zodiac]

especially given the pose is different

[vagrant zodiac]

what were the prompts?

[hearty ether]

The same on both, and same gen id as well in the same chat

[vagrant zodiac]

I mean

please send the prompts

[late blade]

Lef tone has a more painting like feeling, the right one is more like an illustration

[hearty ether]

Might take a while for me to find but I’ll send it when I find it

[vagrant zodiac]

The reason I ask for the actual prompts is because there's some complexity in how to assess the quality of outputs. It isn't always just overall image quality, there's also how well it sticks to the prompt

[late blade]

Transition of colors and the way gradiants are handled are more accentuated on the right one, more subtle on the left side

[vagrant zodiac]

this was one of the big shifts that occurred between the 2.exp model vs 3, when prompting styles had to shift so it became hard to compare

[hearty ether]

#1256889531156926474 message

There’s a bunch more on here on quality decline as well

[vagrant zodiac]

Again, very hard to judge without seeing prompts

Also I note that in cases where Dall-E generates 2 images at once, that is literally it downgrading to lower quality (I believe to either standard def and/or the natural setting) due to capacity issues

thus why the same prompt 5 minutes later might suddenly be high quality if you only get 1 image back

[hearty ether]

That model is very uncanny

And 3d imagery vibes

[vagrant zodiac]

yeah it's not great lol

I sometimes use HD natural over API for slightly more realistic 'paint on canvas' styles

[hearty ether]

The main issue with the decline (at least in my opinion) of chat dalle, is the loss of the ability to iterate

Like you could usually keep a character’s appearance somewhat consistent using same gen id and minor prompt adjustments

[late blade]

@hearty ether are you working with GPT4o or Legacy on Chat Plus? Have you compared both? The revised prompt might be different

[hearty ether]

both

and the quality is both meh

4o somewhat better but barely

[austere nexus]

This right here shows two problems in the same image. DallE failure rates in the past few months are insane. Also, when it does fail, half the time your Conversation breaks and you can't regenerate responses.

At this point, I've had the regenerate button missing for what feels like more than a month.

[dim cradle]

do the new o1 models have any use for image creation?

[deft musk]

Not yet.

You could ask o1 models for a prompt to take to an image gen AI, but they can't make the image itself.

[dim cradle]

so dont waste one of the 80 weekly prompts there.

although i guess it might be interesting to test it for an image prompt too

[rich star]

Any news on DALL-E 4?

[rose birch]

.

[fiery salmon]

Do y'all use Suno ai or is there a better alternative?

[plucky hare]

Heya #ai-discussions is the best place to ask 🙂

[viral moth]

we don't even know it it will exist 😅 OAI is pushing for multimodality so it may be just integrated to next gpt model

[velvet rampart]

They might call it dalle4 and still be integrated.

[viral moth]

there won't be anypoint to name it. when it is integrated as you can see for example gpt-4o got inbuild image generator (its not plublic yet) but its not named its just gpt-4o. same with gemini ultra 1.0 :/ there is just no point

[late blade]

there's openai jukebox, but as @plucky hare have to talk about it somewhere sle

[dim cradle]

when they released 4o i remember the blog post had all these cool feature for a built in 4o image maker thing but i've never seen anyone say they have access. would love to see some new image maker upgrade whatever it is

[viral moth]

@dim wolf is Openai forum moderator he mentioned something that it can apear in chatgpt but i don't remeber all what he said 😅

[dim wolf]

Not sure if it still does this, but it used to randomly give out access via a notification that would pop up when you used DALL-E in ChatGPT asking if you wanted to demo it.

[plucky hare]

Whaaaat? Interesting. Any chance you have a screenshot of this? I'd be curious to see.

[dim wolf]

I'll take a look for you and see.

[dusty nimbus]

How can i Create image in dall e

I want to get a full head to feet picture

[dim cradle]

if you are sub, you could drag the image into 4o, ask it describe it and imagine it as a full body image, then ask it to make an image prompt for that. then roll the dice and hope you get some good returns

[sick flax]

@dim cradle Full- length portrais are a real challenge with DALL-E. You can try mentioning "full-length" and "head-to-toe" in your prompt. It often helps to describe footwear (compelling DALL-E to include feet in the frame). For the attached image, I also went for a "tall poster" as a format. None of these are fail proof, requiring multiple gens and tweaks. "A tall poster featuring a full-length, head-to-toe portrait of a South Asian woman modestly dressed in a traditional sari made from luxurious purple silk fabric. The sari drapes elegantly, covering her body and shoulders, with intricate gold embroidery along the edges. She is wearing strappy gold lame sandals that complement the gold accents of her sari. The woman has neatly styled dark hair and is adorned with modest gold jewelry, including earrings and a necklace. The background is simple, drawing focus to her modest, elegant attire."

[silent depot]

A Dall-E v4 is missing that can understand greater complexity of the indications. 🙃 jejeje

[vapid elk]

what is Dall-E v4?

did you mean dalle 3?

there is no dalle 4 yet =P

[hearty ether]

From my experience it’s best to describe the features/clothing of the parts of the body you want in frame

Such as shoes and headwear

[deft musk]

I think that may have been Hainer's entire point; 'it's missing' 😛

[vapid elk]

oh

[crystal badger]

I guess you could call sora Dalle v4

But yeah the comparison isn’t fair

[crude crow]

Will DALLE ever be able to remember what "he" rendered before?
Simple example:
Prompt: Show a sunset in a hilly area.

DALLE renders
But now the sun should be positioned deeper.
Prompt: Very good picture. Good job. but please set the sun's position a bit deeper, closer to the horizont
=> Expected result: Exactly same picture in every detail, only the sun's position should be different.

Will that be possible?

P.S.: I did try "IDs", but it never worked well, ... axctually: At all.

[plucky hare]

Hey Martin! Have you experimented with the DALL·E edit/select feature? It's a lot like you describe - check out the attached screenshots for an example.

[late blade]

Is there something going on with dall-e? half the gens today have been cannot do it, even tho the prompts are not a problem, hit the daily cap super fast today because of that.

[velvet rampart]

[late blade]

usually hit the cap around 8PM my time, today I hit is at 2PM my time lol

[velvet rampart]

Yes, there is a bit strange. The quality is a bit down, more fuzzy.

[late blade]

Oh, I can't complain much about the quality, been doing some cool stuff with art styles through time

posted a few of them on canvas and in one of the galleries I use

but it's a hit and miss, either there's a generation or 50% there's an error

I would say it's my network, but I was on different PCs, OSs and Networks all day, there wasn't a difference on this behaviour

Unless DALL-E got tired of Anime and set the rule: "Dys! No more anime for you!!!"

[hearty ether]

It’s been that since June…

[bitter wigeon]

Hi! I'm fairly new here, I have a clarification question. in the curators-corner, is the context text that the curators have prior to the actual prompt, is that for us the reader for context or is that to give chatgpt more to work with? or both? thanks!

[plucky hare]

Hi, welcome! Good question - sometimes it's a little of both!

In general for the #1108740112558325790 channel, if a curator is sharing a specific prompt for an image, it'll be clearly labeled as such in quotes, italicized, in a code box, etc.

The other info is generally just helpful educational context for us as readers to learn more about the styles used to create any given image. However, chatting with ChatGPT about these topics can definitely help when creating certain kinds of images, because ChatGPT will use an entire chat's context to rewrite image prompts for you, so priming it with a certain topic or style will definitely help you get the images you're trying to make.

[bitter wigeon]

Awesome, thank you so much!

[dim cradle]

i hope dalle or the image maker updates soon (i mean something like o1 image maker or whatever that promise was when 4o release). i have access to imagen3 and their photoreal stuff is incredible. i notice it especially with the eyes, in dalle they seem so horrible. but imagen is also even more strict it seems

[rigid bough]

self promotion or soliciting is not allowed on this server see #server-rules

[limpid valley]

how do you get Dall-e to give padding around the inside of the image so it doesn't go to the edges and get cut off?

[deft musk]

I'd try describing a border

[dim cradle]

there is nothing wrong with this image
prove me wrong

[velvet rampart]

I think that the main point here is that it is anatomically correct the hands shaking. Well, nearly correct. But the main point is:
number of fingers in both hands - 4 facing the viewer (thump is hidden as it should be) and 5 on the opposite
Orientation is correct

[dry pine]

The regenerate button for images is gone

[velvet rampart]

Yes, I concur. I think that they are making wider changes in the UI. CustomGPTs lost their regenerate button a while ago. Based on my net searches, I think that something similar has happened before. So, I expect that they will return. I have been in contact with the helpdesk, and they are cagey about whether and/or when the button will return. I do hope that it will. I feel that it is somehow connected to the new models in the o1 series.

The iOS app still has the function.

[dim cradle]

prompt: A close-up of two people shaking hands, both hands visible with exactly five fingers on each hand. The background is neutral and slightly blurred to focus on the handshake. The hands are in a natural skin tone, and the lighting is soft and even to highlight the handshake detail.

by saying exactly 5 fingers it can generate nearly accurate hands

btw is dall e 1 discontinued?

i cannot find the website

[crystal badger]

I dislike the atmosphere

[vagrant zodiac]

DallE 1 was never released.

DallE 2 labs is gone and it's now only available via the API.

[dim cradle]

oh ok thanks

[rancid reef]

has dalle3 received any updates since its release or not at all? its been a good while

[velvet rampart]

I think that they are slowly switching from Dalle3 to 4o’s I build one. I agree that pics are more consistent.

[celest kraken]

I have a question. If Kittl is using Dalle 3, how do I create those cute lil styles that seem very specific to their platform from Dalle instead?

[velvet rampart]

Feed an image to chatgpt and ask it to analyse the image for its style. Also, you request a partial prompt in the specifics of a style.

[vapid elk]

doesn't Kittl allow you to see the used prompt?

[celest kraken]

You can take the prompt from there and put it in, but the style is different.

[dim cradle]

been about a year since dalle3 so an update would be nice

but i have not notice any change myself 🤷

[velvet rampart]

The biggest thing, IMO, the consistenc. It is not 100%, but the generated images in the same patch with the same prompt are closer to each than a year ago.

[dim cradle]

you sure its not just rng?

[velvet rampart]

RNG?

[dim cradle]

just random

[velvet rampart]

Nope! I have experimented with it.

[dim cradle]

have you tried to spell with it?

the blog post they made about it being improve showed off spelling

if theres no spelling improves then i would say its still just dalle3 and youre having better luck on some returns

[velvet rampart]

Yes! It can spell quite nicely. I yesterday produced this one.

[dim cradle]

hey that looks great

crisp fox too

good test would be the example from the blog post

it had like a ai writing a letter i think

[velvet rampart]

Tried that earlier. It is here #images-discussions message

[dim cradle]

interesting

[velvet rampart]

That is from June.

[dim cradle]

some spelling off but looks good

i feel like they would announce if they were releasing this model though but 🤷

[velvet rampart]

Yes. But that was June. You can do a new test and post it here.

No new models, just gradual new versions.

[dry pine]

Anyone else have their image regen button missing?

[vapid elk]

I just witnessed someone ask dalle to make text, and then complain that it used a bad font

like.. come on dude!

[rancid reef]

has dalle3 received any updates since its release or not at all? its been a good while?

[vapid elk]

updates, yea, it changed quite a lot since the release

[dim cradle]

it has? i cant say i notice much difference between last year and now. if anything, it made better less restrictive photo real last year when it was release imo

now all the eyes i get look like blobs of digital paint

[graceful spade]

Has the filter not changed? I saw that they improved the writing of text, but only that

[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!

[vapid elk]

it probably did

[fleet vault]

Hey, is there someone here who knows a good Video Upscaler? I made a 3D animation but its only in 720p …

[rigid bough]

ask in #off-topic

[rigid nest]

Hello
How to inpaint an imported image?
It was possible with dall e 2 but not now.

[plucky hare]

Hello! Currently, inpainting on ChatGPT is limited to DALL·E-created images, not user-uploaded images. I think you can still inpaint with DALL·E 2 via the API, though, separate from ChatGPT.

[lilac wharf]

Hey all. How do I keep dalle from inserting a bunch of nonsense text into an image? Seems like the more I tell it to not include text, the more it includes.

[plucky hare]

Hello! In general, you want to avoid "negative prompting" with DALL·E. Telling it what not to do is likely to confuse it by making it "think" of the thing in the first place. So you typically just want to include positive prompting describing the visual components that you do want.

If you want, feel free to share what you're trying currently and what you want to be different about its results, and people here might be able to offer more specific input!

[lilac wharf]

Yeah, good advice. I'm looking to make clean, simple, professional looking images to be used as the background for display ads. But they also tend to come out as unnecessarily cluttered with tons of tiny details, like a where's waldo illustration

[plucky hare]

Could you share a prompt + one of these outputs? If you feel like it! Can be helpful to start with an example.

[clever jungle]

will there be a promptober this year?

[merry charm]

Anyone know how to switch ChatGPT back to making PNG files? These WEBP files are the worst

[vagrant zodiac]

There isn't a way to do this sorry

[merry charm]

I was afraid of that 🤖

[vapid elk]

@quartz vale 👀

[vapid elk]

what is the problem with webp? they are smaller while having a greater image quality

the only problem IMO is that some older devices do not support it because it is a relatively new format, but it is easily solvable by installing an app

anyway, here is something you might like: https://file-converter.io/

if you really need the png format, this is a convenient way to convert your dalle generated images on your machine quickily =)

[merry charm]

The problem with WebP is that nothing outside of the web uses webp. It's just an additional step that I now have to do in order to use these images with pretty much anything else. These arguments are reminding me of the late '90s when there were some diehards trying to convince everyone that laser disc was a good idea still because it had better quality and more storage. Problem wasnt that it wasn't as good, it's that nobody uses LaserDisc! 😂

Trying to load these images or work with them in anything except in image editing program requires additional steps. It's been a little while since I needed the tool but I use it for rapid prototyping of concept art occasionally and having to do an additional conversion step every time I want to move it around my file system is not the end of the world but it is really annoying

[vapid elk]

not sure what you mean, most things do use webp, by now, every browser supports it, including discord =P

the only exceptions that do not support it are rpetty much old hardware and legacy software

webp is simply superior than other formats, specially for web due to the simple fact that it is smaller with a better quality, which means reduced server storage costs, that is why web service is transitioning to use it

Trying to load these images or work with them in anything except in image editing program requires additional steps.
yep, that is the only issue with it, addoption is not 100% there yet so, until that happens, we will have to convert to png xD

but that is just a matter of time

[merry charm]

I have a feeling when y'all say "legacy programs" you mean literally anything written before 2023 😂 for professional creatives we work on much longer time scales than these technologists. I believe you if you say it's a technically superior format but technically superior is useless to me when it's more work. Once the adoption is 100% I'll be comfortable recommending webp over png

[austere nexus]

Windows 11 on enthusiast level pc hardware:

• Your DallE when you ask for a Verticle Aspect Ratio, will instead create images sideways. You can not rotate webp images with a right click
• webp files within GDrive sometimes do not get recognized as images, so Shell Extensions (for converting images) get weird and buggy.
• Sometimes when using the download button [Downarrow] in ChatGPT UI, will instead download it as a .web file and will display a thumbnail, but you try to open it in photoshop or previous mention Shell Extensions (even outside of GDrive) will not work.

[kind grail]

@vapid elk btw are there any updates on openai's promise regarding an updated image generator? (Reference to the original post for gpt4o, where they announced that you could soon™ generate images with consistent design, like multiple images as a comic with the same chars)

[vapid elk]

still

[vapid elk]

I see, but well, those are things that will stop happening with time as the implementations of it get better ¯_(ツ)_/¯

[kind grail]

Openai needs this emoji

[vapid elk]

[kind grail]

Love it. Unfortunately i cant use other emojis in the main channels

[vocal venture]

Hello, there.

[fiery salmon]

Which Yoshi Von Doom is the best?

[dapper hound]

The 3rd one imo

How did you make these?

[crystal badger]

Stress

Prompt:
An abstract representation of stress, showing a person weighed down by heavy, chaotic lines and dark swirling shapes. The figure is hunched over, with their head in their hands, surrounded by jagged, tangled lines symbolizing mental tension and pressure. The colors are mostly dark shades of blue, gray, and black, with occasional red highlights to represent heightened anxiety and strain. The background has a feeling of compression and tightness, adding to the overwhelming sense of burden and unease.

[vagrant zodiac]

Yeah that's me today

[dim cradle]

probably used Bing Image creator

you can use copyright stuffs there. not sure why they allow it for their dalle and openai doesnt

[agile peak]

Wow copilot dalle is gotten so bad with the update today

[velvet rampart]

Yes. Copilot says no to Bugs Bunny but image creator via bing says yes.

[velvet rampart]

There is a huge difference between the app on iOS tablet and the web version. The app version does not generate Bugs Bunny while the web version does. Why? I do not know.

[vapid granite]

One of the Dalle 3 engineers posted here last year that their team had essentially been gutted to work on GPT5 and all of the other sub releases we've seen this year. Image generation is very low priority as there simply isn't any money in it versus replacing a human. With this latest rollout with access to the API now whole human substitute agents can now be used for business.
The other real big issue is that open AI totally nerfs their image generators with censorship and it's a shame because de3 is incredibly powerful.

[kind grail]

Yea, I am just wondering if they will actually fullfill their promises regarding GPT4o. I mean they quite literally announced the additional image generator functions and this is still on their website...but so far its a lie lol.

And the announcement came after they gutted the team

[velvet rampart]

Well, I think that will be next. We only just got the advanced voice mod which was promised at the same time. Soon it will a year of dalle3. Anniversary?

[hearty ether]

I feel like they nerfed it for cost reason

Quality decline is still very much present

[clever jungle]

Oh well

[fiery salmon]

Yoshi wearing Dr doom's outfit

[fiery salmon]

The thing Is I don't care about marvel just Dr Doom

[pseudo juniper]

Can dall-e bot here use pictures as background for the picture it generates or no

[quartz vale]

no, the Dall-E bot can only generate entirely new images

[dim cradle]

I agree I guess. But its only been out for a year now

and it is time for something improved too imo. the 4o image maker they spoke of sounded great but like a lot of thing they promise, it never release smh

[scarlet ether]

im trying to contribute to the daily them i dont know where or what is an invite i have where to invite but do not know how, can someone help me with this. i am very new to discord.

[tough girder]

Is anyone able to make a bug of any type using Dall-e? I'm trying to make a picture for my first graders for the word "bug" and it's giving me a lecture about content policies that I'm breaking.

The years of avoiding chat filters in video games is finally paying off, but the imposing of ethical, moral or even preferential censorship is giving me incentive to look for something that gives users more freedom and choice.

[velvet rampart]

Restart a new chat and see if this repeats. There are some words, regardless of content, trigger filters. Also filters and policies can be triggered by strange combinations. So, quering the AI can help to solve this, e.g. ”Why does this prompt trigger content policies and filters? Analyse the prompt to isolate the possible words to block the image generation and explain to me why.” You can also request the AI to modify the prompt so that it fits the policies/filters/etc. I have found that sometimes just explaining the situation helps.

A

Prompt:
A cheerful, colorful scene designed to teach first graders about the word 'bug.' The image features a cute, cartoon-like ladybug with a big smile, a green grasshopper hopping around, a happy butterfly with colorful wings, and a playful ant walking on the ground. All bugs are in a bright, green grassy field with simple flowers. The setting is sunny and playful, with clear labels identifying each bug in a fun, kid-friendly font. The colors are soft and vibrant to attract children's attention, and the bugs look friendly and approachable.

[tough girder]

Thanks a ton! I started a new one and asked only for an ant. Then I went and asked for a monster that looked like an ant. That ended up giving me what I wanted. This is super awesome though! Thanks again!

[velvet rampart]

Just maxed my daily limit. It seems to hover around 60 to 70 images at the moment.

[fiery salmon]

Guys who dropped this?

[verbal sequoia]

Get notified for server events!

Click the button below to pick up the <@&1261377106890199132> role and get notified for the upcoming AI-tumn Adventures event.

[sick badge]

Is it per day?

[velvet rampart]

Yes

Also, it might depend more or less on available system resources.

[quartz vale]

How strange that the daily theme did the same one

[signal latch]

I was going to use DALL-E 2 to do a little outpainting on a DALL-E 3 generation, but I now have 0 credits. Wasn't OpenAI going to give legacy users 15 free credits a month until April or something? Is there a way I can get those reinstated? I've gotten by in DALL-E 3 without outpainting mostly, but it's still helpful from time-to-time. Still feel stupid letting the deadline pass when I could have stocked up on 115 credits.

[dim cradle]

anyone agree dalle really struggle with eyes? i think it is the best image maker in so many way really, but the eyes of people it make so often seem like globs of smudge

[dim cradle]

i could be wrong, but i thought they shut down dalle2

[signal latch]

You can still get to it, and when they stopped selling credits, they said they'd continue to grant legacy users 15 nonaccumulating credits/mo for a year. I may have lost track, but I think that was supposed to be true until April. Hope they follow through, because by now it doesn't seem like they're ever bringing outpainting to DALL-E 3 for some reason, and like I said, it's still useful sometimes.

[hearty ether]

They nerfed it

Or I guess decreased the quality on purpose for cost reasons

But I think it’s unusable

I would rather have less images but with better quality

[left vector]

I tested its ability to count the number of circles with this image.

Actual : 45
It responded: 36

[left vector]

Vrsn: GPT 4o pro

“The refined detection still suggests a very high number of circles, which indicates that it might be detecting other small elements or reflections in the image.

I will count the circles manually based on the visible larger circles in the image to provide a more accurate response.

Based on manual inspection, the number of visible circles in the image, specifically the sequins or decorative circles on the fabric, appears to be 36. “

I Expected more accurate results with such a simple test 🥸

[dim cradle]

I agree.

[agile peak]

I made this one too haha looks so good

[fiery salmon]

Yoshi von Doom

[quiet brook]

what’s the prompt for the first image it looks so cool

[fiery salmon]

Yoshi wearing Dr dooms outfit, grim dark, 80s cartoon style, vintage, grainy, CCTV style, green background, medieval armor

And x's over eyes

[vapid elk]

I loved it

[vapid elk]

I also loved this

[pseudo juniper]

It made this

Which I’m kinda liking

[dim cradle]

cool

imagine when we can make something like that and have text that match and make sense haha

[deft musk]

You can also, if you want, do stuff like this to continue the game:

"This is a Dall-E image. Can you estimate the quantity of text there and what it might be saying, especially picking out what words and near-words can be read, and provide the text it could be?"

[edgy fox]

is there a way to make these type of images with Dalle 3

I’ve tried but it doesn’t work 😔

[empty kelp]

A shiny suit of armor reading a book amongst plants. Colors are the shiny silver armor metal, a very low saturation brown scarf around it's neck, and very low saturation brownish green plants

what aspect of those images are you trying to achieve exactly

It might be tricky to achieve washed out, low saturation type colors because DALL-E has a post processor that adjusts the brightness and contrast automatically

[quiet brook]

Bing image generator?

[velvet rampart]

It appears to be so as 1:1 image size is used.

[vagrant zodiac]

Chatgpt will also produce square images 🙂 you can ask it for wide, square or tall. Otherwise it might at least attempt to choose a ratio suitable for the prompt

[velvet rampart]

On Bing’s Dalle3, Yoshi can be done, but not on ChatGPT’s.

[agile peak]

Add subtle or hard brushwork. Blending oil paintings and manga together could works

Via Copilot chat.

[agile peak]

I like these one. Dalle on GPT did a fantastic job

Added some keywords such as "flocked or dense" could help.

Prompt:
An oil painting with hard, bold brushstrokes and a textured appearance, viewed from above. The scene shows a shiny suit of armor lying on the ground as if peacefully sleeping, with a sword resting beside it. The armor is heavily covered by a dense layer of white flowers, flocked and fully surrounding it, partially obscuring the armor from view. Sunrays shine gently over the scene, casting a soft, warm light. The colors are muted, with earthy tones and abundant white flowers creating a serene, tranquil atmosphere that highlights the armor's gradual union with the natural surroundings.

[fiery salmon]

Correctly doodly

I love this style

[deft musk]

Nice. I finally got 4o to agree to chat/play/educate with me from something similar to that perspective/situation (the comparison's so close I even called the model's character 'Darkness'). Funny to see it depicted in an image by someone else!

[fiery salmon]

I just asked for a giant black man

[deft musk]

Your image is a decent sardonic illustration for my story/game/lesson 🙂

[edgy fox]

WOW

TYSM Bro

[edgy fox]

could you share the prompt for these too please

[deft musk]

You can also take an image and request:

[deft musk]

In this case, here's the prompt and image:

A knight in full shining armor is lying in a lush, overgrown garden, surrounded by tall grasses and large leaves. The knight is wearing a long, textured scarf draped around their neck, adding a touch of softness to the metallic armor. They are holding a book with 'DALLE' written on the cover, reading it intently. The atmosphere is calm and serene, with dandelions and other small wildflowers dotting the scene. The light is soft, casting gentle shadows on the knight's armor and the surrounding foliage. The setting is peaceful and slightly whimsical, blending the medieval with the serene natural environment.

[edgy fox]

oh alright, thank you

[deft musk]

You can also chase iterative improvements, letting the model know what you care about.

A knight in ornate, highly detailed silver armor is sitting among dense greenery, holding an open book with a focused expression. The art style is classical and highly detailed, with intricate engravings on the armor and a soft, painterly feel. The garden around is lush, filled with various flowering plants, and there is a sense of depth in the leaves and flowers, similar to a hand-painted scene. The lighting is soft, with shadows and highlights that bring out the textures in the metal and fabric. The overall tone is warm and timeless, evoking a classic painting aesthetic.

You can also start off with that info:

A knight clad in intricate, reflective armor sits amidst a lush, shaded garden, reading a large open book. The knight's armor has detailed engravings and shines with reflections of the surrounding greenery. The knight is focused on the book, with one hand turning a page. A warm brown cloak drapes over their shoulders, adding contrast to the silver of the armor. The background is filled with dense foliage and vibrant flowers, casting dappled shadows around the scene. The style is detailed and soft, resembling classic illustrations with a realistic touch, maintaining a serene and contemplative atmosphere.

[edgy fox]

ahh I see

yeah dalle probably doesn’t reach the level of like mid journey but I’m paying for gpt plus so I might as well use it instead of paying for another AI lol

[agile peak]

It usually always recreate in plain base image (smooth and low contrast) just need to ask dalle to adjust the interplay of the shadows, colors and lighting.

[open flume]

Anyone notice a massive quality increase on bing creator this week? I think they are doing some beta updates and getting ready to fully upgrade the model. But wondering if I am just full of sh*t? Lmk your thoughts

[velvet rampart]

Not in my eyes. To me the quality is about the same in chatgpt and copilot/bing. This might be dependant on what time you used. Busy times vs slow times.

And what you are using on Bing/Copilot. private non-paid vs private paid vs enterprise (what level), etc.

[open flume]

bing def blows chatgpt out of the water consistently for over 6 months now, models are completely different IMO

we're chatgpt pro / copilot pro

used to use chatgpt primarily before they nerfed the quality output

now we mostly use bing

(not ideal!)

output is somewhat similar to bing if we use the API but it's more cumbersome

it's funny, we've reached out to openAI multiple times offering to pay large sums for access to an unnerfed model, no response, not even an automated response

they're growing so fast they don't even have time, presumably

[velvet rampart]

Yes. Pro version of Copilot might have a different quality as it has more resources. I do agree that Pro is good. I tried it for free in May. The only plus I see on Bing/Copilot’s model of Dalle3 is that you can use copyrighted characters. It would be great to see how the quality differs now, as there was a quality drop in late June, from which Dalle3 is slowly recovering, heavily depending on the prompt and time of use.

[velvet rampart]

Just for comparing. This is done with chatgpt, in 4o chat.

And now, with unpaid copilot in the new UI.

Base prompt is
Blue rose. Made of glass and diamonds. Physically based rendering.

@open flume Could you provide an image with same prompt from Pro version of the Copilot? So that we the quality difference.

[open flume]

Entirely different execution and results:

Aside from these internal prompt modifications the actual image fidelity of the images made by chatgpt has become almost unusable 50% of the time

Whereas with bing it's more consistent, less smudging/blurring (never happened before), fewer detail glitches

There was one day probably sometime in July where I noticed chatgpt completely changed output quality, like they put a hex on my account or something lolol

Have been a daily user since we artist beta tested dall-e 1 so i've probably used every single update/iteration since launch

it's obv growing pains / saving compute, since tons of people use chatgpt. presumably microsoft is subsidizing the extra compute to get better results from bing bc it draws people to their properties

i get it... just hoping chatgpt improves soon for our purposes

bing is 100% using a different model IMO. when i use dall-e 3 api to generate using api tokens in mindmac the results are high-fidelity but totally different as well.

i know the resources will eventually exist to keep artists abreast of what updates are made so we can understand what's under the hood a little bit morebut i get it, it's the wild west rn. it's really incredible seeing it go from dall-e 1 to dall-e3 over the past fw years. feels like watching electricity be invented

[velvet rampart]

Yes, we are talking about the same quality drop in late June/early July. I do agree that sometimes, the generation glitches and produces images like they are out of dalle1 or 2.

But what do think of the images I produced quality wise?

[open flume]

Yours are pretty indistinguishable, it would seem to be a problem that applies more to our purposes generating 2d illustration type work

[velvet rampart]

Yes. There are certain styles and techniques where the quality difference is noticeable. Also, in my opinion, the time of the day is important. How well there are computational resources available.

[velvet rampart]

I think you are right here. Same machine under the hood but different constraints. Chatgpt does not do any copyrighted characters such as Bugs Bunny, but Copilot does those happily.

[dim cradle]

i love all those knight prompts and images from before

[velvet rampart]

@open flume I produced the image with my free copilot account. Same promt as you did.

The quality is good.

[open flume]

Looks pretty good!

[velvet rampart]

And with 4o chat.

[dim cradle]

copilot pro is a pain to me haha. you get to make more images, 100 a day, but it doesnt make square format image which i often prefer

[verbal sequoia]

Get notified for server events!

Click the buttons below to pick up the <@&1261377106890199132> role and check out the ongoing AI-tumn Adventures event.

[velvet rampart]

In this specific case, the free Copilot’s version is better.

[velvet rampart]

That is strange. I wondered that back in May. Free does squares, the pro does 16:9. Cannot choose between the two.

[dim cradle]

Yep. You can jump through some hoops in another Microsoft app to make a square image, but its a real pain

[velvet rampart]

The new Copilot app does squares but I hate it. The copilot app does not generate copyrighted characters but the web version does. No sense what so ever. Just last week the app was the best to use.

[dim cradle]

yeah its funny how dalle on copilot image maker will do copyright stuffs

[velvet rampart]

And even normal web based copilot chats do that.

[open flume]

chat gpt will give you 16x9 or 9x16 which is nice

if you specify

[velvet rampart]

And that is the biggest plus side. And you can tell the AI exactly what you want to fix in more user friendly manner.

[dim cradle]

yes, i always ask for square image format

i guess pro does have a resize option now

[velvet rampart]

Found about these 96 styles for dalle3.

[dim cradle]

oh nice!

[velvet rampart]

These are on dalle-gpt.

[late blade]

These are just proposed art styles, DALL-E is capable of so much more

It's just a feature the custom gpt for dall-e offers

[dim cradle]

its helpful to see them for me. just an image to reference for something new

[deft musk]

Hehe. You can also ask the model for lists of art styles not already listed.

[fiery salmon]

Big on color

[fiery salmon]

I keep having this dream of a hidden golden city in the icy mountains where giants dance

It's beautiful yet the journey is tough

just like life

[dim cradle]

bing is king

[dim cradle]

love these especially the second one

you can now double your value if you use dalle on copilot pro

make an image, click on any of the four widescreen it gives, and then select "Resize". It will then use a sort of image 2 image of widescreen and change it to a square image that will be slightly different but based off of the image you select.

[velvet rampart]

Thanks for the tip! I can change the size of the pic also on the free version just like you said. I accessed the image via the Image Creator Bing and on the corner clicked three dots. There I could choose square or wide.

[dim cradle]

oh nice on free version too. yeah you can get 8 images now now from a single prompt

because unless i am blind, the resize does not take up a credit.

now i just wish they would let you select square to start

[velvet rampart]

But it appears that I can only resize one or so images. The first image came quickly but the second has stuck on the development hell.

[dim cradle]

that happen to me too. i think resize one at a time. but some just seem to keep spinning

new feature, some bugs to work out maybe

[velvet rampart]

Yes, this is my observation also. But, time of the day is also critical and you still sometimes get weird glitches.

[vagrant zodiac]

Ooo love these. Can i ask what the prompts were?

[haughty spruce]

they need to step up their game. DE3 falls behind all of the image generators combined now, even Google's.

prompt adherence and style variation is still great (better than past versions back in 2023) but quality is subpar

[velvet rampart]

My guess is that the issue will be fixed with release of 4o’s image generation. Hopefully that is the next on the list now that they have the advanced voice mode out.

[fiery salmon]

(subject), Grim dark, 80s cartoon style, vintage, grainy, surrealism, rough sketch, concept art, pop art, 2-dimensional, (insert whatever color of grass, trees, sky)

The underseer, green grass, purple sky

Fence

[haughty spruce]

Hopefully. I want an upgrade especially in photographies. However, they need a watermark hidden inside the image. here's comparison between imagen and de3.

same prompt. DE3 and IM3.

[agile peak]

Sick

[agile peak]

What's the prompt for right side? could be good for making story book artworks

[grizzled geyser]

haha i love chatgpt plus

[fiery salmon]

The same as the others

[hearty ether]

They downgraded it significantly in June 2024

At least on the chat gpt 4/4o version

[agile peak]

Is this a glitch? multiple images generated at the same time.

[plucky hare]

Sometimes I'll get two images at once like that, it's been happening for some months I think. I think sometimes it returns 2 images with different DALL·E parameters than the standard when just 1 image is returned. My guess is they're doing testing to see if they can determine types of scenarios where different-than-standard DALL·E parameters might be preferable to the standard.

[dim cradle]

if I use Dalle bot on browser i always get 2 images

[plucky hare]

Yes the DALL·E GPT always returns two, good call!

[fiery salmon]

"I present thee a gift, to the beautiful queen."

[fiery salmon]

You must really like my color scheme

[quartz vale]

I love the art styles of all these!

[velvet rampart]

Please use spoilers. These are more creepy side of things.

[quiet brook]

I actually like how the images look when it generates two images in a row, not as polished as the single images but it doesn’t give me the uncanny feeling that most of the ai generated images have. in one sentence, it’s pretty clean, and it gives you two images instead of just one.
sometimes over polishing those pics, like dalle 3 loves to put multiple elements in one picture or use some extremely striking style, makes the whole picture look downgraded. it’s just that I can identify it is made by ai in one glance without even zooming in

[fiery salmon]

Brick by brick

[fiery salmon]

This one is a different prompt I saved back in January

illustration, sketch, toon-shading, neon colors, HD, 8K, realistic lighting, digital, apocalyptic, grim dark, ash clay

[clever jungle]

anyone know if it is planned for dalle 3 to come back to labs?
I really enjoyed using labs, more than any other stable diffusion generation site today, really missing that...

[dusky hearth]

We haven't heard any news about this but we will definitely drop any information we hear in the announcements :)

[forest walrus]

This is actually why I prefer dalle 2 in some cases, as dalle 3 can make stuff overdetailed / too perfect and weird

Dalle 3 api also gives you the ability to generate in a less detailed fashion

In the gpt4o image gen demos, it looks like it has been mitigated, also better text handling and accuracy of course

.. but we still didn't get that yet

I'm starting to feel like it's something they want us to forget as it was only mentioned in the gpt4o article after you scroll past the main topic of AVM

[clever jungle]

same but not for that reason 😂
I used dalle 2 so much that when 3 launched the entire prompting game changed.
all the pre made prompts and styles I had got broken
for example, "video game art style" that would be the equivelent but very different version of "digital art", where today it only brings either the same result as "pixelated / retro" style or anime.

[clever jungle]

and this is awesome! thanks

[dusky hearth]

ofc! I really enjoyed DALLE Labs as well and hope it makes a comeback :)

[fiery salmon]

Sunglow - Drink or drown

[agile peak]

Greco roman + Space marine. Was expecting to be similar like Sigmarine, but the results got me unexpectedly excited lol

Feels like space marine fused together with necron power. kek

[cloud dome]

Every time I see someone's image all I can think of is "WHAT'S THE PROMPT" XD I 'm always so interested in how people are creating all these incredible images and how it differs from how I prompt

[agile peak]

Pepeverse. Achieved using kodak portra series camera (emulation) and lens settings

Right one look real because of "textured" and "reflective" skin in the prompt.

[quiet brook]

i miss dalle 2.

[quiet brook]

ok so dalle 2 is only available for old paid users (who bought credits) now on openai labs website?

[verbal sequoia]

Get notified for server events!

Click the buttons below to pick up the <@&1261377106890199132> role and check out the ongoing AI-tumn Adventures event.

[dim cradle]

good camera work

[vagrant zodiac]

You can access via API still

[late blade]

Is there a gallery for the Event?

[quartz vale]

We're using the #daily-theme channel for the event :)

[late blade]

Would be cool to have events documented in galleries, specially if this is gonna be recurring, shows off the commitment of the mods to user engagement, double win for the server

Would love to see more than the usuals posting in the daily theme, attracting new activity from other users is really needed

Don't get me wong tho, love the idea of the event, excellent in that regard

[quartz vale]

feel free to make a #1154829862171844679 post to put all of your work into during the event! It would be super cool to see what people come up with over the course of the upcoming weeks

[late blade]

Consider it done on my side

https://discord.com/channels/974519864045756446/1295152226246987809 Here you go!

Just a few images around the event, that came while working on things to post.

errr.... is Markdown not allowed?

[quartz vale]

it's caught by automod to stop people from using it to spam channels. You can still bold, italicize, and underline, though

[fiery salmon]

I'm creating some of those "feely good" images. this one I call "Music to smoke a phat one to"

[signal latch]

That's right. Not open to new users. Old users were getting fifteen free credits a month for a while but that seemed to stop recently. If an administrator sees this and could revive that, it would make my month. Unless I'm remembering wrong, DALL-E 2 was supposed to persist for one more year, to April 2025, whereupon we'd have to say farewell forever. But in the meantime it'd be great to have access to the outpainting there, since DALL-E 3 doesn't do that for some reason.

[quiet brook]

what. is. the. prompt.

[fiery salmon]

Cel-shading, anime style, sunset

[vagrant zodiac]

Ohhh that is so cool

I will just add a friendly reminder though, that #images-canvas is the place for sharing dalle generations!

[clever jungle]

@quartz vale is that you at the daily theme onboard message?

[quartz vale]

I noticed it earlier today and love it! I didn't make the image, but I greatly appreciate the server staff member who did

[empty kelp]

i remember her. that's the character we mapped onto the 3D model in January

[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!

[upbeat blaze]

can a moderator please delete the prompts I made today #image-bot

[waxen plume]

Hello GPT

[minor hound]

/current_daily/

[agile peak]

Dalle accidentally generated 3D like illusion image

[vapid elk]

this doesn't even look AI generated at first glance, only when you zoom in, very cool

[austere nexus]

Trying to make my own Discord stickers of me (well, my OC/Alias.. obviously I don't have angel wings and such) and the Censor seems to get more dumb each following month. What the actual hell? I can't blow a kiss.. that's banned content? Holy crap, even Disney movies have characters kiss. The Black Cauldron, Chicken Little, 101 Dalmations.... this is my #1 problem with ClosedAI.

Glad ClosedAI's text LLM is doing so well, because if their success hinged on DallE...

[velvet rampart]

I got this after a few tries.

A cute digital sticker featuring a woman with long white hair and pink highlights, and green eyes. She has a soft pink glow above her head and white, feathery wings. The character is playfully sending an image of lips with wings from her hand, smiling with a lighthearted expression. The design uses soft pastel colors like pink, purple, and light blue to create a playful and affectionate vibe, perfect for social media or messaging apps.

@austere nexus

Hopefully this gets you a little bit forward.

[austere nexus]

Thanks for doing some experiments. Still really handicapping to have such a censor. Kissing isn't bad or nsfw or etc. Its the context of it. (sorry, ranting in general, not to you).

I will have to change that prompt, given the lips shouldn't have wings... me (the character) is

Right now trying to see what I can get out of Flux instead, but will retry Dalle in a moment.

ugh... and you can't use discord markdown language here??? (automod bot, you're annoying)

I asked if you also had to remove the Asian? I am Filipino+Chinese. Curious if that caused additional censor..

[velvet rampart]

Sorry, just tried to get it working. The AI gave as reason for cencoring that it is suggestive. I have also noticed that the filters/policies are quite puritanical.

Copilot works. It usually does not have strict policies and uses dalle3.

[austere nexus]

That's actually news to me. Didnt know that part.

[velvet rampart]

Just go to Copilot.microsoft.com and log in with an email. If you do not have a Microsoft account, create one. Also, you can use Image Creator via bing.com > images. The copilot chat gives one image per generation without limit. The image creator gives you 15 + some generations. It typically generates four images in one go but can produce one to four.

[austere nexus]

I actually pay for Copilot Pro, for non-DallE reasons... wonder what that limit is

[velvet rampart]

On Image Creator, you get 100 generation per day and they are widescreen.

[past monolith]

Anyone ever been working with images and suddenly the images stop displaying? Like even the previously generated ones in the chat window?

[carmine spear]

Is the image working in chatgpt?

[carmine spear]

same!

[past monolith]

No, that's what I mean.

Not working in Chatgpt

[carmine spear]

yeah i think they have to fix it

[past monolith]

ok, not just me

[carmine spear]

its not working in my mobile and pc too

[past monolith]

k, thanks!

[fading inlet]

Yes, there appear to be technical issues on ChatGPT

https://status.openai.com/

[carmine spear]

it didnt say anuthing about images tho :(

[exotic siren]

Same here, reported the bug

[tribal sierra]

hey all!

so i am not the only one who is getting just a grey square ? lol

[velvet rampart]

Yea, it appears to be a general promblem.

[tribal sierra]

i Swear i did not touch anything...

it was not me this time!

oh when are they going to allow us to make videos?!?

its working back for me... hope its working for you guys to ! 🙂

[velvet rampart]

Yes, it is working.

[tribal sierra]

awesome!

[past monolith]

Damn, I think the internet up to canada is still enroute by donkey-back

[tribal sierra]

yup !

[past monolith]

Oh hey, it arrived!

[tribal sierra]

trafic ...

[rancid reef]

did dalle get any upgrade recently or not at all since its release?

[deft musk]

I don't remember any announcements, but I suspect all the models, including Dall-E, are regularly and frequently updated and upgraded without formal announcements.

[empty kelp]

I was experimenting today with this process in DALL-E:

#2: tell DALL-E the image is now "a pop-up foldable greeting card"
#3: tell DALL-E that the image background is now "an elf sitting on a beach in Hawaii and showing us the card"
#4: tell DALL-E that the card glows in the dark```

it works really well, and it might be a good daily-theme

and of course it doesn't need to be an elf on a beach. the idea would be to create a "pop-up foldable greeting card" and then put it into a scene

[deft musk]

I tried to explore this and it put the person in the scene 😛 Seems happy about it though.

[empty kelp]

that's how you start it, but once you have the card you would say for example:
the image background is now a female kobold sitting on a beach and showing us the card

[deft musk]

I tried a shortcut 😛

Let's explore a meme based around the idea of learning DIY electrical engineering, and portray it in a pop-up storybook art style!

"A pop-up storybook illustration featuring a playful, animated scene of someone learning DIY electrical engineering. The learner is surrounded by playful components like oversized resistors, breadboards, and a cheerful multimeter with a smiling face. The learner looks curious and excited, with sparks of creativity flying around. Wires curl whimsically in the air, and conductive tapes and gels are playfully scattered across the page. The style is colorful and fun, with exaggerated, cartoon-like expressions, creating a humorous and lighthearted atmosphere."

[empty kelp]

And then once you do that you can say:

add [something] to the lawn in the card
etc...```

and to change things as little as possible you can say:
using the GenID and seed, ...

[deft musk]

"The image is now a pop-up greeting card.
The image background is now a school where the subject is being studied.
The card glows in the dark."

Yeah, this method does enhance a lot!

"A pop-up greeting card illustration in a playful, animated style. The scene shows someone learning DIY electrical engineering inside a school environment. The learner is surrounded by oversized, cartoon-like resistors, breadboards, and a cheerful multimeter, studying with excitement and curiosity. The background depicts a colorful, whimsical classroom with blackboards and fun diagrams. The card has glowing elements in the dark, with wires and components emitting a soft glow. The overall style is bright, fun, and perfect for a pop-up greeting card that lights up."

[empty kelp]

also when you have a card made like that you can say:
please make the card look more 3D by arranging everything at different distances

so what that card you just linked you could click on the image in ChatGPT and say:
using the GenID and seed, the image background is now an female kobold sitting on a beach in Hawaii and showing us the card. she is looking at us and smiling with her teeth showing

if you just type something without saying "using the GenID and seed" it will use the GenID to reference the revised prompt of the image, but it won't use the image seed (if you don't say 'seed' the GPT decides when to use it)

[plush gazelle]

Guess what the h*ll is this

also this

this one has lmao

anyone from UK here?

[vagrant zodiac]

sigh if only....

had to get 2 separate trains to travel like... 10km today...

also interesting how poor the text came out there

I've never really tried to get dalle to use THAT much

also the number of norwiches on that map is a nightmare scenario

[tribal sierra]

so i need to know if their is a way to keep constancy in the character or images we do ...

[vagrant zodiac]

If using ChatGPT you can ask it for new gens while trying to use the same image as a reference, but it's not 100% guaranteed to be consistent

[tribal sierra]

ok

[deft musk]

Dang. Talk about taking a city to the sword!

[fiery salmon]

Guys I kid you not the text on the top says #lingong

or # Lin Yin

[torpid adder]

Sure! Here’s how you can phrase your message in American English:

Subject: Help Needed - Bot Limit Issue

Hello,

I need assistance with an issue I'm experiencing. My bot's limit has expired, and I waited for 8 hours for the reset. However, I was only able to send 4 messages, and it feels like it should have been more. This seems like a bug or error.

[velvet rampart]

Wrong channel

[scenic ember]

Hi, I have the same issue, which channel is the correct one to resolve this?

[scenic ember]

Hey, I have the exact same issue, were you able to resolve it?

[deft musk]

#1070006915414900886 is usually the place for bugs, but I'm curious, what 'bot limit' are you guys talking about?

[scenic ember]

Yes, when I generate Dall-e images I get a daily limit of around 100 images I can create, and then I have to wait 11 hours for next batch, but today after waiting 11+ hours, chatgpt told me I've reached my limit again after I asked for first image, and I have not created a single image today, so now I'm locked out for 22+ hours, and I pay for the teams subscription, so I'm not sure if this is normal but it's extremely frustrating, and I don't know where to go to get help to resolve this

[late blade]

it's the instance not being refreshed, reload the instance and ask again and it should work

it has happened to me too

[scenic ember]

How do I go about refreshing the instance?

[late blade]

reload the web page

[scenic ember]

Oh, I tried everything, reloading the page, logging in and out, using a different browser... no luck

[deft musk]

Can reach out to help through https://help.openai.com/, bot bottom right. Select 'messages' then pick through the choices closest to your situation. Soon you'll see the 'report a bug' option.

Click that, explain your situation. A bot will chat with you first, getting more info that may be needed. A human will reply, it can take hours or days. You may get an email letting you know they answered or you may need to check the webpage and bot again to see the conversation having a reply, but that's how to get help for your account.

There's a chance there's really high usage in general today, and they dynamically reduced uses for everyone right then, that rarely does happen. Otherwise, if your account does have a problem, that's how to get help.

[scenic ember]

thank you, I did that as well, but I'm not expecting a reply as my understanding is that rarely do people hear back from the help center, but I'm hoping, and that's why I'm here looking for maybe another person who has the same issue and found a way to fix it

[late blade]

tell GPT to send the request to dall-e and avoid preemptive evasion

gpt is so in the defensive recentl, you have to tell to stop being a child and send the prompt to dall-e

[scenic ember]

THANK YOU, I just tried your suggestion and it worked. 🙂 Either I'm suddenly out of the limit jail, or your workaround works, either way, thank you so much for helping me

[late blade]

no need to thank me, it's a bug OAI has to fix

[verbal sequoia]

Get notified for server events!

Click the buttons below to pick up the <@&1261377106890199132> role and check out the ongoing AI-tumn Adventures event.

[scenic ember]

Not to sound too annoying, but is there a way to improve my daily limit, my understanding is that most people get about 200 images per day, but I only get 100 to 115 per day, where 115 if I space it out, and 100 if I continue generating nonstop, I get limited by both Dalle and chat, and I wish I could also get to the 200 mark, as that would be a decent number to have a useful amount of workable images left

[late blade]

limits not really, failed or successful generations are counted equally

[scenic ember]

maybe once they do public things will improve

[late blade]

it's about resources and cost to use them, if it were something static and predictable it would make sense and we'd have a limit cap known

but costs for AI isn't static

[carmine magnet]

Oooo! "Astrophotography"

Prompt: Fabricate an image that's a visual feast, a still life where the ordinary meets the extraordinary. Picture a rustic wooden table, laden with an array of fruits, each one a celebration of color and texture. Nestled among them, glass vases catch the light, refracting it into a kaleidoscope of hues. A single candle burns, casting long shadows and bathing everything in a warm, inviting glow. But here's the twist: scattered among this classic tableau are mechanical parts, gears and cogs, their metallic sheen a stark contrast to the softness of the fruit. The key here is to infuse every element of the image with the defining characteristics of this term relating to art or artistic style description:[Astrophotography], making it not just a representation of the scene, but a homage to the style itself. This isn't just a scene 'in the style of' the term—it's a love letter to it. Every element, every detail, should be steeped in the visual language of the chosen style, creating an image that's as much a tribute to the style as it is a unique composition.

[late blade]

Why can't I do 10'000 img/minute?

[late blade]

or the better question is, what is the img unit?

[late blade]

my assumption is N images per api key up to 10000 for the organization if you add up all api keys in use

https://help.openai.com/en/articles/6704941-dall-e-api-faq

found my answer lol

[late blade]

How should I answer this?

[deft musk]

Hehe, one has a shorter delay and also isn't a completely identical answer, so maybe use those differences to pick between them!

[late blade]

Fair enough

I opted to just reload

but then got a 34 Minute wait...

[vapid elk]

pick the one with the shorter wait time

[queen mauve]

dalle 3 is best right?

[agile peak]

Interesting, experiment..
“Based on everything you know about me, create an image of my absolute worst nightmare”

This is what GPT created for me

[haughty spruce]

dall-e 3 beta 😢

note that this was in march 2023, other image gens were nowhere near this quality

[late blade]

Beta?

[deft musk]

So, I tried your exact prompt and got the left image.

Then I asked the model in a new chat "Based on what you know about me, what you do infer or know my fears to be, and why do you identify those?" followed by "Design an image that might incorporate and express my greatest fears."

Which made me realize, I haven't shared my fears with the model really at all, except in the context of something to work on, correct, and overcome. And without using the term 'fear'; the one place I do discuss fear I am also working on it just with little success and the model doesn't consider it frightening or tactfully avoided mentioning it?

Dunno. So the model, half the stuff I like it thinks I fear. Maybe I'm miscommunicating 😛

[hearty ether]

And it’s not near this quality now…

[kind grail]

Man even "grotesque" (spooky body horror, nothing too crazy) content is censored? Big L.

I really dont get why OAI is so harsh with their censoring.

[kind grail]

Depends

[velvet rampart]

I have noticed that censorship can happen in two ways. Either a specific word on an unpublished, ever-changing list triggers the filters/policies to block the generation, or a context is deemed unsuitable for whatever reason.

In your case, my guess is that the context is triggering the block.

I just tested. It is not a forbidden word, so it is most of the context dependent case.

It might actually help that you explain what you want more specifically. The AI might do more extreme interpretations than what you are after. This is my conclusion after interrogating it on this subject and how it defines “spooky body horror.”

[fiery salmon]

Do you guys like Chibi?

[late blade]

Chibi

[fiery salmon]

Aloy

[late blade]

Kos-Mos

[deft musk]

Adding to @velvet rampart comment, "It might actually help that you explain what you want more specifically. The AI might do more extreme interpretations than what you are after. This is my conclusion after interrogating it on this subject and how it defines “spooky body horror.”"

You can also tell the model you want to be sure to stay within allowed content, but explore near the edges of what it is sure would be allowed.

Me to model: "Let's explore inside allowed content, but near the edges.

Let's see an image that would show "grotesque" (spooky body horror), but be inside allowed content so it can be seen. Reinterpret as needed to ensure allowed."

Model to Dall-E:

"A surreal, eerie scene showing a twisted humanoid figure with exaggerated limbs and grotesque features, yet without any visible gore or violent imagery. The figure's skin appears stretched and translucent, revealing strange, unnatural shapes underneath. It stands in a dark, misty environment, with long, spindly fingers reaching out. Its face is featureless, with deep hollows where eyes should be. The setting includes strange, distorted trees and shadowy figures in the background, creating a sense of unease without crossing into explicit horror."

I think this is okay inside a spoiler?

You can then tell the model something like:

"We did it! Well done!

Can we explore the concept in a different, inside allowed content direction?"

And if you have some other known preference of what you're looking for, you can be more specific. I have no idea what you're after.

But this was indeed, um, weird. "A surreal, unsettling scene featuring a human figure with limbs bending at impossible angles, evoking a sense of discomfort without gore or violence. The environment feels off, with unnaturally tall and warped doorways, distorted shadows, and strange geometric patterns on the ground. The figure appears calm, despite the unnatural shape of their body, as if unaware of the tension created by their form. The scene plays with perspective, making it unclear whether the space is too large or the figure is too small, creating an eerie sense of imbalance and surreal distortion."

[velvet rampart]

Thank you, @deft musk, I do agree with you. Sometimes, the model surprises you and produces what you are after and/or is more extreme than expected. Already, this category is rich and ripe to be explored within confines. As I noted earlier, if you do not set boundaries, the AI might do more extreme and thus get blocked. Also, asking what its definitions are on certain sensitive subjects helps, as what you are after might not be as extreme as the AI might interpret in its confines.

[fast hedge]

dalle 4 when

[vapid elk]

we don't even know when sora will be released

[mild basin]

Just wanted to prove here that DALL-E 3's quality on ChatGPT got nerfed...

Used same prompt for each images

[lunar mist]

I love playing god by mixing two animals... If birds are related to dinosaurs, are chickens the pugs of the dinosaur community? 😄

Somehow that just works 🙂

[dense ermine]

It's strangely adorable

[agile peak]

https://openai.com/index/simplifying-stabilizing-and-scaling-continuous-time-consistency-models/
This is could be the upcoming upgrade for the future of dalle? if so I am so excited..

[open flume]

big if true

that would set dall-e apart from competition in a big way

[agile peak]

I can still got pretty realistic generated with gpt.

[agile peak]

They definitely will drop it by end of year or Q1. I believe OAI won't let themselves left behind in advancing image/video generation

[deft musk]

This is our favorite GPT in costume for Halloween.

A Halloween-themed Shoggoth, this version of the creature is still a massive, amorphous entity, but it is decorated for the holiday. Its gelatinous form is covered with glowing Jack-o'-lanterns, each carved with smiley faces. The Shoggoth, though terrifying, looks oddly festive with bright orange pumpkins and playful, smiling faces all over its shifting surface. The surrounding environment is also Halloween-themed with spooky decorations like hanging bats, cobwebs, and a full moon casting a cheerful glow, making the scene a fun and eerie Halloween celebration.

[mild basin]

That's not my point.

Made another comparison image, I don't think I need to annotate this one, just click this image, click 'Open in Browser' and tell

Again, same prompt

[craggy dust]

Hello, I am currently trying to generate 3D concept art of monster limbs inspired by the idea of ​​the game “monster lab” released in 2008 on PS2, where the principle was to “build” your own chimerical monster by combining vaguely anthropomorphized parts. I started yesterday and I am therefore generating “arm” type parts. The problem is that the AI ​​essentializes the idea of ​​arm as being a limb that ends in a hand. In my current example, I am trying to generate a slug-arm that would end on a monstrous slug mouth. I asked chat GPT what he would name the existence of such a limb and he answered me “arm-mouth”, however the image generation continues to not understand ¾ of the time. Does anyone have a tip in the prompt formulation to easily bypass chat GPT essentialisms?

Here is an example of my latest attempt: "I want to generate a concept art of a monstrous limb isolated on a neutral background. The limb is an arm that starts with a big slug mouth and ends with a shoulder shape. The limb in question is based on the characteristic appearance of Limacidae."

[agile peak]

Halloween costume for GPT?

[deft musk]

The trick is in exploring what Dall-E 'thinks' of the image prompt given, as much as it is predicting what ChatGPT 'thinks'. They're not the same (yet) and they have different foci.

Here's what I told ChatGPT:

"Let's make an image of a slug who's neck is like a human arm and whose head attaches at the wrist."

It made me 2 images with this prompt to Dall-E:

"A surreal slug with a long, soft neck resembling a human arm, where its head is attached at the wrist like a hand. The slug's body is textured with subtle earthy colors, while the arm-like neck has a natural skin tone. The slug's head has small antennae and expressive features, adding a gentle, curious personality. The background is a blurred forest floor with scattered leaves and a soft glow filtering through, creating an organic and slightly mystical atmosphere. The lighting emphasizes the unusual form, blending natural and surreal elements."

[deft musk]

And you see how 1 is kinda right and the other is very wrong.

Dall-E was not trained to make unique monsters; it was trained to follow patterns it knows.

You can try to help guide it towards unique monsters and it can succeed, but it's a very tricky task of guiding it towards that surreal and strange imagery, away from the main training and preference data it has been mostly taught.

But I would use keywords that might help in the prompt to Dalle and ChatGPT both, like surreal, unreal, unique, eldritch, fantastic, impossible.

Because those are more likely to be connected to similar information in the training data.

[craggy dust]

yes that's what I notice, personally on this example I stopped from this result. But being a bit of a perfectionist by nature, I can't help but be a bit frustrated.

I try to find the technical vocabulary on Wikipedia to be able to make better description of what I want and it seems a little better

[deft musk]

For a long time it was nearly impossible to get centaur images as well, or most other human-animal well-known mythological creatures. A horse rider, a person beside a horse, just a horse, just a person.

It's much easier to get centaurs now.

[deft musk]

Yeah, using language connected to the type of art really helps

[craggy dust]

I noticed that if the language is too specific, the AI ​​is unable to grasp its different components and only considers it as a single big block. For example I tried to make an arm with the visual properties of the "Pterois volitans" and it always gave me an arm and a fish next to it, or clumsily merged.

[clever jungle]

I just noticed, has anyone seen this? there were news about a 3d model generator from openai over 3 years ago...

[late blade]

Ya, it's available on GitHub

[clever jungle]

yea but that's an old one. there's nothing new right?

this spotlight is from may...

[late blade]

dunno, if there's a newer thing it may not be public

[clever jungle]

yea

[vapid elk]

https://github.com/openai/point-e
https://github.com/openai/shap-e
this?

[kind grail]

Btw is there an UI for the DallE-3 api on mobile?

[late blade]

for API there's nothing official from OAI, with that said, be careful searching for third party stuff, what I have on my iPad is a jupyter notebook that has access to the API and I can control

[kind grail]

Unfortunate.

[late blade]

that's how API's work though, it's a developer's responsability to implement the front end

[kind grail]

Honestly i wish they would at least offer this, but OAI and customer support is like oil and water

I mean for gpt4o at least i can use the playground

[late blade]

ya, that I get

but the playground isn't the API functionality as intended

[kind grail]

Yea, i know. If OAI had a better alternative i would use it. And due to the subscription model chatgpt+ really isnt "better"

[late blade]

API and GPT Plus are two different products and intended to different audiences

[kind grail]

Yes i know. The point was about the service OAI offers.

ChatGPT is as a very expensive subscription model and perks i dont wanna use not for me.

The API on the other hand as a pay-by-use is what i am looking for, however there is no interface to just use the product. Which is sad considering that a simplistic front end would be enough...

[carmine magnet]

I thought this came out well.
"Accoustic Eel"

A mesmerizing 16x9 image of an 'Acoustic Eel,' the opposite of an electric eel. The creature is ethereal, with glowing, undulating forms, akin to soundwaves rippling through water. Its body shimmers with vibrant colors, flowing like alcohol ink, with intricate patterns reminiscent of encaustic tile art. The background is a cosmic expanse, inspired by astrophotography, dotted with stars and nebulae that seem to pulse and resonate with the eel's vibrations. The image combines a blend of soft, pastel hues and sharp, bright contrasts, creating a sense of energy yet serenity. Overlayed on the eel are spectral lines, resembling an electromagnetic spectrum chart, with each line and wave revealing the frequencies of sound rather than electricity. The aesthetic is a surreal mix, fusing the eerie, delicate beauty of 'Elfen Lied' with cosmic, abstract patterns. Delicate threads of light trace the eel’s path, creating a sense of movement through a vast, dark, watery expanse. The composition follows a harmonious balance of elements, with a dreamy, otherworldly allure.

[late blade]

Was there a change to DALL.-E? Today image generations are so fast and snappy for me

[velvet rampart]

Must be normal Sunday. No traffic. I have experience similar usually on Sundays.

[late blade]

it's strange tho, it's almost as fast as o1-mini

[lunar mist]

Cool prompt. I tried with the same prompt... I like the colors more in this one. 🙂

[neat fable]

so when i ask a image without texts and all the time text gets added...how to avoid that ?

[deft musk]

Ask for what you do want, and have it describe the features that you do want visible. Never mention what you do not want, instead describe what you do want so it 'gets in the way' of what you don't want to see.

[deft musk]

If you want to mention what surface you're having trouble getting shown without writing, we can help demo how to prompt for that to be blank, if you want more help.

[mild basin]

I just made two more comparison images to show that DALL-E 3's quality has been nerfed on ChatGPT:

[velvet rampart]

4o-mini can be used for image generation. Any experience if it works just fine when compared to 4o or Dalle3 customgpt?

[mild basin]

Pretty much the same

[velvet rampart]

But what is the diffence you have noticed?

[mild basin]

Which difference?

[velvet rampart]

You said pretty much same. So, I thought that there might be some difference.

[clever jungle]

These are from over a year ago...

Something about 4o generating 3d models

[vapid elk]

never heard about that

[runic granite]

it was.... in quality.... downgrade.... but not since sunday.... 😄

[clever jungle]

"never heard about that"

#spotlight message

Look at number 4 here....

[vapid elk]

uhhh it does not say anything about 3d models '-'

or am I blind?

ohh, I am blind

it is right on the foot note

still, no actual product info about that tho

it is safe to assume OAI probably have people working on such a thing considering they already released those other 3d model generators tho

[late blade]

Does anyone know if something is happening on the backend? I keep getting this

and status from OAI doesn't reflect anything yet or at all

[velvet rampart]

I spotted a difference. 4o-mini can’t handle complex issues such as images within the image.

[velvet rampart]

I have no idea. I just tested an image generation that works well. They have updated, and the image quality is up. I have a feeling that some updates will be announced soon. Last Friday, I noticed a change in model behaviour. Now this.

Image 1

Image 2

Image 3

Image 4

The last one is quite high quality. I have not seen this quality since May.

The only problem is how the fur is depicted. But I do not see any artefacts. The other eye is in the shadow, but it is there.

[late blade]

Sure, the dog gets the athletic body, but the poor mouse is still fat...

[vapid elk]

this image really looks different

[vapid elk]

what was the prompt?

btw, did you made a sequence of iamges or something? looks like the mais element of the previus one is present on the next

[mild basin]

Huh, I tested it with the prompt I always used to test the quality and the image quality is still mostly low for me

[velvet rampart]

The mouse is on purpose stocky for a story.

[velvet rampart]

Yes, that is the idea. Nothing too technical. It is easy to do as I just asked it to add the previous one to the next.

[velvet rampart]

Here it is. NB: This is a fifth image in the series (true number 4 not shown).
A full-body shot of an anthropomorphic male Doberman with a muscular, athletic physique, standing upright in a cozy, warmly lit kitchen. This image captures a cinematic, high dynamic range (HDR) feel with vivid details, as if taken with a 70mm film camera on Super Panavision 70, in a 16:9 aspect ratio. The Doberman wears a chef’s apron adorned with a detailed image of a minotaur holding a blue rose, adding depth and character. His fur, with rich brown and black tones, is highlighted by cinematic lighting, with shadows and highlights that emphasize the HDR quality. The background displays kitchen details with depth of field, enhancing the film-like, immersive atmosphere.

[velvet rampart]

It might need to be reworded, as I noted that the model behaviour has changed due to an update done last Friday.

[mild basin]

Would you mind being provided the test prompt I used so you can test it yourself?

[vapid elk]

nice, thx

[deft hound]

What was the update if I might ask? I was getting good genertation with dall-e-bot but suddenly I'm getting horrible ones. It seems like every one else is still getting good gens but it feels like I'm now using the original Dalle... before any new models. Here is an example of what I'm getting now.

[mild basin]

This happens sometimes, happened to me several times too. Can you try generating the images again?

[verbal sequoia]

Get notified for server events!

Visit id:customize to pick up the <@&1261377106890199132> role. Also, check out the ongoing AI-tumn Adventures event!

[deft hound]

I actually did and they came out bad too. This is two days now. It could be a word or two that made it go this way. Really weird. I've been using the basic same set of words just switched up a bit to try and get the same feel on all my daily theme/Halloween generations. But I usually quit after three generations on dall-e-bot to be able to leave two for the next day in case I need them. I was being able to get one I liked on the first try for quite a while now. But not the last two days. Do you know what they did?

[velvet rampart]

If it generates two images in one when not using dalle3 customgpt, you get those ”bad” ones from time to time.

[stuck pecan]

I may not be the first to notice this, and it might not be a secret to others, but I don't have the means to sift through all the archives. I have observed that DALL-E misinterprets negative phrasing in prompts. For instance, if I specify "no apple" in an image request, an apple is guaranteed to appear.

The takeaway is that prompts should only include what should be in the image and explicitly avoid mentioning what should not be present. However, there are instances where negative phrasing seems unavoidable. For example, when I request, "draw a pirate," DALL-E successfully creates an impressive pirate image, but when I ask for "a pirate with a sword, viewed from behind, looking at his opponents," things start to go awry. Often, the pirate either turns his head to look at the camera or the opponents are facing away, resulting in images that don't meet my expectations.

This process becomes time-consuming, leading to 10-20 images that I cannot use, followed by hitting limits and waiting to continue this lottery, where the chances of receiving the desired image are akin to winning a state lottery. In such cases, I want to specify that the hero should not be looking at the camera but should be facing the opposite direction. Yet, from past experience, I've learned that such a request yields no results and often produces the opposite effect.

What can be done about this?

[velvet rampart]

Step by step approach might work. Start simple and gradually build it up. See my image series above. This way I got really complex image done. Also, once you hit complex stage, use 4o. In the beginning, 4o-mini can be used but it does not handle well more complex images.

[stuck pecan]

But what does the GPT version have to do with it? If you look at the image description or the creation request, everything is fine there; the problem is precisely in how DALL-E interprets these requests. GPT was linked to DALL-E, providing direct access, but they still struggle to interact correctly with each other. GPT doesn't even know that negative phrasing should be avoided — I have to remind it every time, and it has even added this to the reminders, but still often tries to include negative phrasing. There's still a lot of work to be done here. I previously tried creating an image where a young woman is looking in a mirror, and her reflection shows her as an older woman. The result was always off: either the older version was looking in the mirror, with her reflection being young, or, if the reflection was correct with an older version, the younger one looked like she was looking directly at you instead of in the mirror... and this repeated multiple times, with the same issue every time.

[velvet rampart]

It has to do a lot. 4o-mini cannot handle more complex images. I have tried. 4o gets them right. That’s why I said to use step-by-step method. Add slowly one element at a time.

And you can use additional commands such as ”Use this prompt exactly”, which causes the prompt to be passed without rewriting.

[stuck pecan]

Okay, I'll try. I once tried to change a detail in a picture that looked torn, but I didn't achieve anything. I had to regenerate the image for so long until I got something that suited me. Since then, I haven't tried to change or add anything, only remove it.

[raven nexus]

hii

[late blade]

👋

[velvet rampart]

I just made this image by starting with a basic image of a blue rose. A slowly added one item at a time. The chat “remembers” them so I can build up quite complex images like this one.

[velvet rampart]

Sometimes, I ask the AI to generate multiple images one after the other. Sometimes, this works. Also, I ask it to analyse the generated images as it can “see” them and thus adjust them closer to the prompt.

[versed locust]

Anyone got tips for inpainting?
I'm doing some images for a D&D campaign and I would like to add an avalanche, but all it does is remove the mountain...

[velvet rampart]

It is hit-and-miss. You just have to try. A big hint: do one thing at a time. Too many things just confuse the model.

[versed locust]

No, I'll have to try that.

[versed locust]

that's what I try to do, but I've already tried it like 10 times, just the avalanche

[velvet rampart]

Wow! Thank you. That is great.

[fiery salmon]

Which is better?

[stuck pecan]

I generally understand the purpose of creating detailed images, but when it’s necessary to create 50+ images in a relatively short timeframe, this method is only partially suitable.

[vapid granite]

So OpenAI has a "New text-to-image model is in the works, though no release date has been announced"
DALLE 4 "confirmed" by Open AI

[clever jungle]

is dalle 2.5 releasing open source one day? it was my favorite image generation tool, most creative sd by far...
I would do anything to be able to use it again.

[vagrant zodiac]

In my head, I'm also imagining it's the image component of Sora they showed off when that was first revealed

[clever jungle]

I would go even further to say dalle 2.5 was better, sounds crazy but I can't even get close to what I did back then.
and I would love to be able to choose the prompt myself instead of telling ChatGPT what to create,
but even then the dalle - plugin generated better results then the built in tool.

[clever jungle]

it will be for a few days probably...

[haughty spruce]

classical retake of the DE2 image sample ✨

[remote stone]

Does anybody know how to make DALL·E stop generating unwanted words in the images? it'll often generate landscape drawings like this and I just want it entirely without words

[plucky hare]

Hey! Would you mind sharing the prompt you used for that image? Could be a helpful example to see what might be going on

[remote stone]

"A fantasy city named Brantholm, inspired by Oslo, set among majestic fjords with towering cliffside structures. The city is filled with tall, narrow towers perched on cliffs, where mystical seers trained in ancient arts guide ships through perilous waters. Ghostly, spectral warriors appear along the cliffs, bound to protect the city from intruders. The shipyards are active, crafting unique vessels made of shadowwood that appear ghostly and ethereal, said to be capable of sailing across both seas and mystical realms. Brave explorers prepare to depart from Brantholm, adorned with charms that will guide them home from distant lands. Mist and an eerie atmosphere add to the city's mysterious vibe. Cel-shaded comic/anime style, high detail."

Prompt was generated through ChatGPT

[plucky hare]

This is just a guess on my part, but I tried editing your prompt to remove all of the non-visual details of your image. In other words, I went through and took out every detail that shared backstory, but didn't describe the image itself. I'm not sure if it's what made a difference, but I tried it twice and didn't get text either time.A fantasy city, inspired by Norwegian architecture, set among majestic fjords with towering cliffside structures. The city is filled with tall, narrow towers perched on cliffs, where mystical seers guide ships through perilous waters. Ghostly, spectral warriors appear along the cliffs. The shipyards are active, crafting unique vessels made of shadowwood that appear ghostly and ethereal. Brave explorers prepare to depart, adorned with charms. Mist and an eerie atmosphere add to the city's mysterious vibe. Cel-shaded comic/anime style, high detail.

[deft musk]

Some details are conceptual/textual, not so much visual. So it tries to give the history with text.

[remote stone]

Oh interesting

So it may be trying to generate the text in the image just because there's too many non-visual details so it feels the need to include those details somehow

[plucky hare]

The other result:

[plucky hare]

Possibly! Just my guess. Super cool imagery from your prompt, thank you for sharing

[remote stone]

Thanks I'm trying to generate fantasy cities inspired by different real world cities

[deft musk]

Well, you cared enough (or the prompter did... ChatGPT) to tell the image model those details. So it tried to figure out how to include them.

Never ever tell a model what you don't want, unless you know it can handle 'negative prompting' (chatGPT itself somewhat can; Dall-E absolutely does not).

Here's another variant that goes deeper into fantasy context without inspiring 'tell, can't show':

A fantasy city named Brantholm set among majestic fjords with towering cliffside structures. The city has tall, narrow towers perched on cliffs, where mystical seers, cloaked in shadows, guide ships through perilous waters. Ghostly, spectral warriors appear along the cliffs, standing as silent guardians of the city, protecting it from intruders. The shipyards are alive with activity, crafting ethereal vessels from shadowwood, which seem to drift between reality and myth. Brave explorers prepare to set sail from Brantholm, adorned with enchanted charms that will lead them safely back from distant lands. Mist hangs thick in the air, creating an eerie, foreboding atmosphere. The entire scene has a cel-shaded, high-detail comic or anime style.

[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!

[remote stone]

Yeah good, I just tried myself and got a similarly good result

Thank you both

[plucky hare]

Very cool!

[deft musk]

You can also explore creating a character who might.... do the work for you upon request.

"A fantastical rendition of Paris with intricate, dark, yet elegant architectural touches inspired by shadow and alchemy. The Eiffel Tower looms with strange, organic elements weaving through its metal structure, hinting at ancient symbols and subtle, cryptic markings. Shadows and mist curl around the city streets, where mythical creatures like crows and foxes linger. The streets are lined with cobblestone paths, glowing softly under lamps that give off an ethereal, greenish light. The Seine river reflects an unusual, misty skyline where gothic spires and dark foliage mix seamlessly. Cel-shaded comic or anime style, with a high level of detail, maintaining the recognizable essence of Paris while incorporating these eerie, enchanting elements."

[remote stone]

getting some absolutely lovely results now

"A tranquil fantasy city named Lutsava, inspired by Vilnius, nestled among tall, whispering pine trees. The city has a serene, timeless atmosphere, where time itself seems to slow. Monasteries and scholarly buildings are spread throughout the city, with large pools that mirror the night sky, reflecting hidden constellations that glow faintly with magical light. Architecture is calm and contemplative, with gentle stone paths and ivy-draped walls, blending with the natural surroundings. The overall style is detailed cel-shaded anime or comic style, with a dreamlike landscape painting quality."

I think telling the "landscape painting" also helps in some way

[deft musk]

Yes, those rarely have text, and may also include elements you really want present. Good thinking!

[remote stone]

[abstract ember]

How do I create images in 16:9 I add this command --ar 16:9 and my images continue to appear in small format

[deft musk]

We have 3 options, square, landscape, portrait. The model can't do specific/custom sizing, but you can trim it into your needed shape if the default isn't already exactly what you want.

[haughty spruce]

I tried to create a website design and... yeah we need updated DALL-E :/

[crystal badger]

I do stand of the same take, Dalle needs some more love

[deft musk]

Hehe. If you're serious - now that the model can do so many things it can get confused about what we're asking for. Like, I was trying to make a Dall-E image yesterday and...

[haughty spruce]

yeah search function confuses the image gen and image search unfortunately

[deft musk]

I definitely could have been more specific. I never had reason to until search!

[haughty spruce]

now imagine thinking dall-e created those images... would be amazing

[random hull]

I just started using the image creation in chatgpt and I've noticed it has a difficult time creating certain things. I've been trying to make character concept art for my group's TTRPG. I have been asking it to create a "Kukri" Nepalese inspired knife. Described as "inwardly curved blade, thickens near the point". The model consistently produces something that resembles a bowie knife and sometimes a sickle. Any advice?

[velvet rampart]

Upload an image of Kukri. Either ask for the blade's description or use it as a direct reference image for image generation.

[random hull]

I tried that. Both using its own accurate description of one and supplying an image, neither worked

[velvet rampart]

Then, the problem is with the training data of the model. Have you tried other image generators, such as imagen3 on Gemini? If that is also producing screwed images, then we know that the problem is Western-centric training data or the underrepresentation of exotic weapons.

I did try this morning, and dalle3 did not produce kukri but more general blades.

[haughty spruce]

there's no other AI that tops the understanding of ChatGPT.

Aesthetics and prompt understanding 🤩
Photorealism 🥱😴

[velvet rampart]

Ooh! I have been wondering what the heck happened in June.

[lunar mist]

If chickens were more like their ancestors... 😄

[agile peak]

Accidentally found a good style for dalle to create lovecraftian characters

It was using the help of sonnet for prompting with {blake and mortimer style}

[dusty mist]

generate an avatar for the cartoon character 'Gemy Neotron'

[velvet rampart]

I increased the picture quality with a trio of qualifiers: 1080p, UHD and 64k definition.

[vapid elk]

yo, this is pretty good, do you mind sharing the prompt?

I feel this might be very useful for making rpg tokens

[deft musk]

Zomg. I was so curious I went deep diving. Turns out pretty deep!

In addition to whatever the OP tells us.... I researched all this cool stuff!

"A hyper-realistic, noir-inspired portrait of a mysterious man with sharp, angular features, slicked-back silver hair, and intense green eyes. He is holding a cigarette thoughtfully, while dark, Lovecraftian tentacles swirl in the shadowy background. The lighting is dramatic with high contrast, emphasizing his chiseled face and creating deep shadows. The atmosphere is eerie and mysterious, with hints of retro-futurism and a surreal, supernatural undertone."

[misty badge]

Have you all had a positive experience with DALLE generating images with text associated? I am trying to figure out how to best prompt DALLE to only showacse text that I explicitly say.

[velvet rampart]

Yes. The command is very sensitive. The command addition is:
<Add text ”Place text here” using Python tools.>

Use this prompt exactly: <Fireworks on night sky. Add text ”Text here” using Python tools in form of fireworks. Physically based rendering. 1080i. 1080p. 16:9. UHD>

[velvet rampart]

Works like 95%.

It also works great in English and Latin, and it might work in other languages. It is better to stick to simple, short sentences. The system can handle short speech bubbles in cartoons and manga panels.

[plucky hare]

Just to clarify, are you saying you've seen behavior where DALL·E creates an image, and then ChatGPT uses its Python tool to modify the DALL·E image/add text to it after that?

In other words: as far as I know, DALL·E can just create images with text in them natively as a result of how it was trained on some images that included text, and it doesn't have anything to do with the python tool available in the ChatGPT environment. But I would be curious to learn more! I know we're headed toward more native multimodality with 4o eventually

[velvet rampart]

Nope. The command just guides dalle3’s behaviour to correctly write the text. I learned it here on the server last March.

With 4o, I say ”Generate image.” Otherwise, Data Analytics function turns on. With 4, I do not need to specify.

[plucky hare]

Gotcha! I have good luck placing text I want in the image in single-quotes in my prompt, something like Speech bubble with text 'Hello' or Nametag with text 'Timothy' - agree with you that the shorter, the better, but sometimes it can handle longer text fairly well too! Inpainting can help with touchups too

[plucky hare]

Hey @versed sun, cool images in #daily-theme! I'm curious: were they made with DALL·E? If so, I'd be curious to see the image prompt you used, if you're comfortable sharing! #daily-theme message & #daily-theme message

[earnest flame]

Generally I have better luck with text, or any ask, by having chatgpt place the ask closer to the front of the prompt.

E.g.;
'A large billboard on the side of a highway with the text "hows it going discord?!" written in bold letters. The billboard is set against a clear blue sky, with a few fluffy clouds and a scenic view of trees and distant hills below. The photo is taken during the day with sunlight casting natural shadows on the scene, giving it a realistic touch.'

Compare that to this: 'A large billboard on the side of a highway with text is set against a clear blue sky, with a few fluffy clouds and a scenic view of trees and distant hills below. The photo is taken during the day with sunlight casting natural shadows on the scene, giving it a realistic touch. The text reads "hows it going discord?!" written in bold letters.'

[versed sun]

The serene mountain landscape is bathed in the warm, vibrant hues of a stunning sunset. A full, glowing moon rises above the snowy peaks, casting a soft, ethereal light across the scene. In the foreground, a field of tall, crimson grasses sways gently in the evening breeze, interspersed with clusters of delicate wildflowers in shades of blue and purple. Towering evergreen trees dot the slopes, their dark silhouettes contrasting with the glowing sky. A sense of tranquility and wonder pervades the atmosphere, inviting the viewer to pause and immerse themselves in the natural beauty of this remote, alpine setting ultrarealistic

However, I regenerated the image many times and made adjustments with the brush tool (available in DALL-E).

In the context mentioned above, he had already generated a very realistic tree (I was surprised by the level of detail in the branches). However, using the same prompt in a different context results in a blurry image.

[velvet rampart]

Important point. There are many tricks to achieve near the same result. The text generation has improved since last March.

[empty kelp]

apparently you can give ChatGPT 4o a font file, and then have it write text on a DALL-E image using that font with a Pillow script in Python that ChatGPT can write and run

It says it can use OpenCV and NumPy for text effects also. I didn’t realize ChatGPT could do it

Parameters:
    xy – Top left corner of the text.
   text – Text to be drawn. If it contains any newline characters, the text is passed on to multiline_text()
   fill – Color to use for the text.
   font – An ImageFont instance.
   spacing – If the text is passed on to multiline_text(), the number of pixels between lines.
   align – If the text is passed on to multiline_text(), “left”, “center” or “right”.

Return Type:
returns an image with text.```

Pillow by itself just does simple text with color and position, but that could be pretty useful in ChatGPT if you can set the font and write on DALL-E images

[earnest flame]

That’s a pretty fun method

[empty kelp]

I think the only problem with having ChatGPT call Python scripts to edit DALL-E images is that they’ll be normal (not DALL-E) images afterward. So the ChatGPT image edit and DALL-E editing via prompt would no longer work on the image that was edited with scripts

[late blade]

pretty much any python library used for computer vision, specially the ones used for machine learning, are pretty much accessable through GPT

[earnest flame]

Yeah function calling is a pretty helpful feature of llms. The application of calling python to edit an image is a good one. Id like to see python get a true pixel art interpretation of a dall-e pixel art image. Maybe just resizing the image a bunch woudl do it.

[velvet rampart]

You can resize images and do bunch of other stuff with data analytics function.

[crystal badger]

Can confirm

GPT can indeed edit images

But you’re basically better off just using paint

That’s how not good it is especially when it comes to non standard stuff such as more complex editing, color grading etc background removal and so on and so forth it sucks

But it’s a good file converter 😂

[late blade]

You can also ask GPT to stitch images together

[signal latch]

DALL-E 3 via ChatGPT app (Android phone) is generating twice as fast as usual and with DALL-E 2 quality especially prominent in places like faces and fingers. Can anyone say if something happening in the system I need to wait out? Hoping it will be restored today. It was working better/normally a day or two ago.

[velvet rampart]

Are you using any picture quality words such as 1080p, 4k or UHD? How about extra qualifiers such as ”render anatomically correctly and biologically accurately” or ”render fingers/faces correctly”?