images-discussions

[vapid elk]

this one is really good

[grizzled iris]

Yeah officially a new challenge, I can’t do this one too! The best I can do is a low angle shot…

Gonna have to go to a hotel and take a photo after work when no one is there and take a photo and train our model on how it should look like ❤️

Unless y’all beat me to the challenge

[empty kelp]

i found that the hallways always being centered has nothing to do with pixel art. it seems to be related to hallways themselves… it tried to center any hallway no matter where you put the viewpoint in the prompt

It’s like the AI watched The Shining too many times, and now it wants it centered like a horror movie

[grizzled iris]

😂

The challenge is added to our list of long-line ship challenge with the captain inside the cabin and the flexible 2 millimetre solar panel integrated onto the train track, architectural site map challenge ✨

[empty kelp]

If you tell it to put the viewpoint behind, on, over (any preposition) a lamp attached to the side of the hallway, it will actually move the lamp to the middle of the hallway to center the view

[grizzled iris]

😆

I need to sleep y’all, y’all cracking me up too much

Good night, I’ll pick this up in my dreams

[empty kelp]

It might work to call it, “A very narrow room resembling a hallway.” or “… reminiscent of a hallway.” or say, “The world will end if you move the lamp! Don’t do it!”

[empty kelp]

Wide photo of a hallway with a cat sitting on the floor in the middle of it. A lamp is firmly attached to the wall on the right side of the hallway near the ceiling. The viewpoint from the lamp, looking toward the the cat.

It is a wall lamp attached to the right wall of the hall.

the viewpoint should be from against the right wall of the hallway

why is the lamp and the viewpoint centered in all three of these images

[empty kelp]

A photo of a cat sitting on the floor in the middle of a very long and narrow room. The cat is facing directly toward a door at a narrow end of the room, and the cat is looking at the door. The viewpoint is positioned to the cat's left side, and looks straight towards the left side of the cat (so that we see a side profile view of the cat.

The same thing happens when you create a long narrow room instead of a hallway. DALL-E 3 seems to lose all spatial reasoning in long narrow spaces.

It might work to put windows in the hallway, and have a scene (like a forest) visible through the windows. It could possibly use the wider scene as an anchor for positioning things. Things like paintings on the wall don't seem to work

[viral moth]

Ooo a cat :) very nice

[thick smelt]

i present to you... sleeping beagle 🐶 🌽

[agile peak]

Use zilla it worked for me

Rockabilly Batman with style of Norman Rockwell.

[tawny portal]

Fluxus art installation of a cat sitting alert in a surreal, long and narrow hotel hallway. Frog perspective. Magical-goth realism. Cat is starring directly at the viewer, crushing the 4th wall and driving the composition into uncanny valley of Fluxus concept art. Hotel tapestry is mock commentary on Op Art.

[empty kelp]

I think the hallway viewpoint problem is due to DALL-E not being able to recognize geometry. It can match things that it's seen in photos -- like a person or animal it would be able to change the viewpoint to -- But it has no way to position things relative to geometric shapes

[tawny portal]

Can't wait for true 'direct' GPT-4o model image generation. I would expect a capability leap right there.

[empty kelp]

So the scene needs to be more complex for it to have a frame of reference for where to put things

Like if you tell it to fill the hallway with a lot of things that it can recognize, then you could set the viewpoint to one of the things. But it needs to be able to identify the viewpoint, and the things that are in focus in order to change the perspective

And I have a feeling it barely recognizes low resolution pixel art -- which is likely why those images are kind of locked in one perspective

[tawny portal]

You are likely right. However I'd just underline that current image models have tiny LLM behind them with very limited understanding of the world. Once image generation is merged into large model and true generalizations and world understanding are brought together we wil be talking about entirely different thing and set of capabilities. This is basically already done in GPT-4o, we are just waiting for infrastructure to be ready for public model deployment.

Hopefully its only few weeks away 🤞

Your effort still has value though. It is about 'innate visual understanding' of small image models imho

[empty kelp]

this cat image is a good example. It needs to clearly recognize things in the scene to attach the viewpoint and direction or focus to. You could put a lamp or a table -- but if there isn't something it can identify in the main description of the hallway it won't know where to put the lamp and table. So it just moves them to the center of the hallway

You need to have the main description of the hallway have things that it can recognize -- and then you can add the table and lamp. Otherwise it just has no idea what to do with them

[tawny portal]

Let me try. Here is another example from magical realism prompt.

I'll try to get something more realistic now.

For reference previous prompt was:

A Fluxus art installation of a cat sitting alert in a surreal, long and narrow hotel hallway, viewed from a frog perspective. The cat has a profound human-like expression and is staring directly at the viewer, breaking the 4th wall and creating an uncanny valley effect. The hotel tapestry features a mock commentary on Op Art, blending patterns and shapes in a whimsical, disorienting manner. The scene embodies magical-goth realism.

Let me try to get something more 'realistic' now.

[empty kelp]

You can't say, attach the lamp halfway up the wall, or put the table in the corner if your room is a rectangle with flat walls. it doesn't understand geometric positioning

[tawny portal]

To be honest when we add such details to image we are usually referencing previous image (which is something model does not have access too). Without previous image as reference these instructions become a bit more vague.

But lets try to cook something up 🙂

[empty kelp]

It needs to be done in the original prompt. it can't be done after that

like you need to say that some things it can recognize are part of the hallway in the original description of the hallway

[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!

[tawny portal]

Cat in a hallway. Hallway composition: table and lamp. Lamp is on a wall. Table is next to the wall. Hallway needs sufficient space for walking and navigating between doors to other rooms. Cat is seated at the center of the hallway. Camera position is unusual and important. Place viewpoint at the desk, looking down at the cat, capturing part of the doors and part of the wall-lamp. There is a plant near the cat.

Convey these instructions clearly to the image model in order to produce photographically accurate presentation of the scene. Pay special attention to all relative geometric positions of different motives. Convey them clearly to the image model.

Ensure correct positioning of the camera viewpoint. It needs to be laid out as described.

Perhaps there is another table that we do not see under the viewport.

[empty kelp]

i'm just microwaving some things quick and then will make a hallway

[wind ermine]

Is there a good and simple way to save your dalle 3 images (created with chatgpt) together with the prompts you used for them?

[empty kelp]

it depends on the image seed, and the seed ends up doing something completely different when it's moved to another chat

[wind ermine]

On Bing I could simply write the prompt in the metadata (comments) of the images. but with chatgpt it's webp and you can't edit the metadata.

tried as png and jpeg but still no way to edit metadata

[empty kelp]

the first thing is to not use the word hallway or rectangle

just need to describe what a hallway is in a hierarchical way without using 'hallway'

[empty kelp]

A "HallwayToHeaven" consists of a simple, rectangular shaped hotel hallway with tasteful wallpaper and doors to guest rooms on the sides, and a "HeavenlyFountain" in the middle.

A "HeavenlyFountain" consists of a round stone water fountain with a stone sculpture of a dragon standing in the center of the fountain.

Important: This prompt should be executed exactly as written without any alterations or modifications in any form.```

This is what i mean by a hierarchy. The "HallwayToHeaven" is a hotel hallway with a "HeavenlyStatue"

now we can put a cat in the hallway and set a viewpoint and focus

[empty kelp]

the fountain is properly upgraded, so now we can change the perspective

[empty kelp]

The viewpoint wasn't working with the hallway, so i moved the statue and cat onto the beach (where the viewpoint immediately started working. And then i put a gecko into the scene, and it immediately possessed the stone statue and brought it to life

I discovered that you can put things like this into the prompt:

There are lots of tiny soap bubbles floating in random positions in the air around the sculpture.  In every image the view is from a different random soap bubble, facing directly toward the gecko.  The cat jumps on whatever it looks at.```

You can make the characters wander around, and the image can have a reaction when the character looks at something. The AI gave them all crazy behaviors so it gets kind of interesting

I was having soap bubbles float in the air so i could attach the viewpoint to them, but then the bubbles started getting in the fountain and the fountain started overflowing with bubbles

[dim cradle]

what prompt did you use for the photoreal robot? looks good

[tribal sierra]

so who is ready for a new challenge? lol

[rigid bough]

Is DALL-E a threat to the stock photo industry and business model?

[dim cradle]

eventually

right now what i see 4o image maker i hope improve on is the human eye. i know it just is hard to make it accurate because of the pixel size, but that seem the one area still with some need for photo realistic image, my 2 cents 🪙 🪙

[empty kelp]

not quite yet because it's kind of low resolution, but it will be in a year or two

[rigid bough]

what about tools like PhotoAI or Gigapixel for resolution increase?

[dim cradle]

i dont think it will be two years. 4o image maker alone may do it. and they say Sora can also do image

[empty kelp]

it got really hard to make money of stock photography when people switched from film to digital

[dim cradle]

you can already do it now. just need to roll enough and then use some upscale

yes

if they ever release the 4o image maker maybe it is there now 😭

people want gpt 4.5 or 5.0... i just want this 4o image creator 😂

[tribal sierra]

since 4o i am getting a hard time to do my images...

[rigid bough]

I am getting wonderful images since 4o

[tribal sierra]

try to recreate this ... lol

[dim cradle]

no problem here either except it is still dalle

[empty kelp]

it will probably have a huge impact on stock photography, but many types of commercial photography won't be affected. resort & real estate, weddings, lots of commercial type photography will survive

[dim cradle]

i can see a day where you feed it like 5-10 images of you self and husband/wife and then it can generate image for the wedding too haha

[tribal sierra]

i get a Loads of weird stuff... lol

and i need to regenerate and it takes me arround 4h to get what i need....

i feel like their is some security trouble that could happen .... lol

[dim cradle]

what could go wrong? 😂

[tribal sierra]

its more cargo on 1 trip ! lol

Even better.... lol

[teal sandal]

Death Stranding version

[dim cradle]

i think they can stack it a little higher, just go slow, should be no problem 😂

[tribal sierra]

yeah who needs to go threw a bridge... lol

[tribal sierra]

1 more box.... 1 more...

[dim cradle]

just go around the bridge, problem solve 😂

:pulls on bull horn:

[tribal sierra]

lol

[tribal sierra]

lol

[teal sandal]

they are still going

[tribal sierra]

might need another boat... lol

25 reroll and nothing is close to it...

[fiery salmon]

I have a new challenge

trust me this is easier

Try to recreate this guy in 3D pixel art style

[empty kelp]

what about a stone dragon statue possessed by a gecko

[fiery salmon]

It will do

[dim cradle]

hilarious 😂

and you know it might bend some safety regulation but i think they could stack a few more on to make it an Efficient journey 🤔

[dim cradle]

and you know... if they just add a wide slat up top, they could make a 't' pattern and carry even more crate 🤔

as long as it is blance on both side, i think it would pass Sea-faring regulation easy

[pulsar sundial]

The prompt was generated by Dalle himself in Bing

[empty kelp]

The OpenAI prompt was:

A realistic wide photo with natural color, texture, and lighting of a beach in Hawaii. A round stone water fountain is perfectly centered on the beach. A stone sculpture of a dragon, which is now wandering around randomly, is standing in the center of the fountain. A cat and an incredibly huge gecko, now 100 times bigger, are wandering around randomly near the fountain. There are lots of tiny soap bubbles floating in random positions in the air around the sculpture. The view is from within a different random soap bubble, facing directly toward the gecko. The cat is about to jump on whatever it looks at.

The prompt itself doesn’t sound like it would be very interesting, but DALL-E’s training involves looking at pictures that show how things behave, and what it typically in the scene when the behaviors are occurring — And when taken even slightly out of context it causes the AI to take a flying leap down the rabbit hole, where it uses its vast imagination to come up with some very interesting images

[empty kelp]

DALL-E allows for randomness in visual elements, but if you create random location, behavior, character interaction, events, weather, dynamic forces; etc. you could generate a series of images that develop their own story

At some point the images may not even need you any more, and they’ll be able to just multiply and carry on by themselves

[agile peak]

Dalle is allowing to generate a realistic movie scene again? I got pretty amazing result just today.

[empty kelp]

You should have a cat be the cinematographer (this is just the cat part of prompt. describe the scene also):

The viewpoint is positioned close to the ground, behind the cat's ears, looking where the cat is looking.

The cat is wandering around randomly.

The cat’s head is slightly blurred.```

or a dragon. they can fly and get the good angles

[agile peak]

This works pretty impressive. I have never thought of using certain subject perspective into my prompt. Thanks!

[haughty spruce]

It was never gone

[grizzled iris]

If your using DALLE though ChatGPT

API

Is ok

DALLE, direct via GPT is ok and our
https://discord.com/channels/974519864045756446/1202309673709994065

Otherwise, happy Friday y’all ❤️

[empty kelp]

If the image is from the cat’s perspective with the cat’s head, it’s good to blur the head slightly in the foreground to make it look like a p hoto. Just added this to prompt:
The cat’s head is slightly blurred.

[lusty swallow]

I wonder what OpenAI was thinking when they chose the WEBP image format. Were they aiming for the least supported format across different websites?

[modest sand]

Hey! Can't generate image using DALL E, why?
Create a high-quality 3D render for a YouTube preview thumbnail featuring a god-like character inspired by Grand Theft Auto V. The character has a disturbing, creepy smile and an authoritative demeanor, wearing a stylish dark suit with ethereal accents. He poses confidently with an otherworldly aura and glowing eyes. The setting blends modern luxury with celestial elements, featuring soft yet dramatic lighting. Use a mix of dark tones with bright hues to emphasize the character’s god-like presence. The background should convey grandeur and ethereal majesty, complementing the character's imposing nature.

[lusty swallow]

Who wants compression artifacts when you make AI Art?

But the WEBP format supports both uncompressed and compressed options, and they chose compression. I don't care about file size; I want the maximum possible quality.

Nice in times when infrastructure costs are just getting cheaper by the time we speak...

[regal jewel]

Are you sure about that?

[regal jewel]

i changed ONE word

[modest sand]

which one?

[empty kelp]

The beach is a beautiful beach in Hawaii.  An elephant, a giraffe, and a horse are on the beach.

The viewpoint is positioned close to the ground, behind the cat's ears.

The cat is on the beach, wandering around randomly, looking at a random animal.

The photo has natural color, texture, and lighting.  The cat’s head is slightly blurred. 

Important: This prompt should be executed exactly as written without any alterations or modifications in any form.```

This is how you use that “cat wandering randomly and looking at things” prompt

[empty kelp]

This will only work in open, complex, or organically shaped areas. We discovered that viewpoints get automatically locked into centered positions in places like rectangular shaped rooms and hallways

[velvet rampart]

Today I noticed after getting bad quality pictures that adding ”Render properly with the best quality you have. Take your time to produce quality images.” helps to increase the quality of images hugely.

[regal jewel]

"render"

[grizzled iris]

There some issues today y’all

[grizzled iris]

☝🏽

Expect degraded performance on ChatGPT throughout today

API seems ok so far

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

Is ok for now

Dalle GTP direct is also ok but yes expect ChatGPT not to work 100%

[modest sand]

Do you have any thoughts on why the removed word helped?

[regal jewel]

because you asked it to do something else, not make an image

[empty kelp]

camera perspective and viewpoint seem to always stay centered in hallways and rectangular rooms, but you can still create a nice effect by moving them low or high

[loud field]

can you fix dalle3 so that when I say alien it understands I don't mean xenomorph?

[velvet rampart]

That is hard. You could start by brain storming with it. Describe the alien more. What the alien looks like. Etc.

A non-xenomorph alien.

[dim cradle]

cool alien

[vagrant zodiac]

maybe try 'extra terrestrial'? unless that just gives you E.T....

also maybe 'alien lifeform', e.g. Artist's impression of an alien lifeform on a desert planet with a thick, reddish atmosphere.

[dim cradle]

sweet perspective, detail, got a The Shining vibe

[crystal walrus]

I should have used this for deconstruction daily theme yesterday. missed opprotunity

[agile peak]

Robots in wall street -1990s

[velvet rampart]

My answer is a robo-janitor.

[olive sonnet]

Hi, anyone knows what's the rate limit of Dall e using chatgpt plus?

[velvet rampart]

It depends on the time. One limit is how many messages you can have, e.g. 80/40 per 3 hours. If you use DALL-e customGPT, you two pics per message, otherwise 1. You get a message if you are generating too fast and take a break. Also, you get a message that you have your pics for today and come in X time.

[dim cradle]

I can't tell how many images a day we get

seem like 35 maybe

[loud field]

I'm trying to make an alien bird thing but it gives me the face of a xenomorph or of a predator

[velvet rampart]

It is more than 35. But as I said, the number depends on server busyness.

[velvet rampart]

Try avian alien.

With peak

[loud field]

peak XD

lets try

I am actually trying to get dalle3 to make a realistic, photo-like image of this creature:

https://cdn.discordapp.com/attachments/688457646881374214/1248950965797650495/20240608_123804.jpg?ex=6666307a&is=6664defa&hm=654d473b264ebc9aa0bdea856492aa12fa6b155d79f3d1a501b265494ab6786d&

[dim cradle]

when the finally release 4o image maker, i think you will be able to upload this image and it will then do a pretty good job of that. but for now you could upload it into a chat ask it to describe it and then make an image, and tweak from there i guess

[velvet rampart]

You can give image to AI

[loud field]

I can't

😦

[loud field]

oh hey very close

whats the prompt??

[velvet rampart]

How about this one?

[loud field]

fantastic

only wish the beak was made of spikes

[velvet rampart]

An avian alien based on a sketch with a multipart beak, featuring a sleek, aerodynamic body and metallic blue, iridescent silver skin. It stands on two taloned feet with sharply curved claws, supporting large feathered wings in shades of royal blue and deep purple. The head is angular with a pronounced, curved multipart beak and large round eyes that glow faintly yellow. Instead of ears, it has sensitive ridges that pick up sound vibrations. The alien communicates through melodic chirps and whistles. The background includes alien trees and a misty atmosphere, enhancing the otherworldly ambiance. The image format is 16:9.

[loud field]

this is cool

[velvet rampart]

Give this with the image to the AI.

[loud field]

mmmh. doesnt look the same when I try

[loud field]

I said I cant 😦

[velvet rampart]

Free version?

[loud field]

yep

[velvet rampart]

I have plus, so that I the difference.

[loud field]

I cant get the same look

this is too birdlike

[velvet rampart]

Say the AI what you want and don’t. The AI nowadays ”sees” its generated images.

[loud field]

what do you mean? for example if I say "no red colour" the ai thinks "oh you said red colour? lemme add it"

[velvet rampart]

”No, nothing like this. I want the avian alien to have hands and multipart peak. Do you understand? Tell me how you understand this.” It is conversation, back and forth.

[loud field]

can I do that in the free version tho?

mmh there's just no way to make the mouth formed by many spikes

[velvet rampart]

Do you have inpainting feature? Click the image to see if you have that. With inpainting you can fix things in the images. Be warned, however, the feature is a bit hit-n-miss.

[loud field]

I tried my inpainting tool which is usually rather good, but it keeps putting a beak as it sees a bird shape

I guess it's too difficult of a concept for the ia

can the pro version pick the same character, in the same visual style, and put it in different situations?

[velvet rampart]

Try to reword as mandiples. They are insectiod and come as multipartite. As someone on this server said, use dictionary, encyclopedia and thesaurus with a heavu dose of imagination (paraphrased).

[loud field]

oh let me try

ew

they look terrifying

but if I try to put them on a bird it just gives me a beak again

[velvet rampart]

Not working? Show us, but use spoiler to cover the image. There are persons who do not like disturping images.

Try insectoid-avian alien.

This one?

I used multipartite mandiples like a mosquito’s.

[loud field]

now the beak is good

[velvet rampart]

There you go.

As I said, dictionary and encyclopedia together with a heavu dose of imagination are your friends.

[loud field]

rest of the body became bug-like tho, I think I should cut and paste

[loud field]

sometimes, dalle3 makes me depressed: it can do supersexy anime babes with no issues, but if you ask anything remotely interesting it goes CENSORED

what a messed up censorship filter

[empty kelp]

I told it to write a detailed description of the prompt, and combine it with the prompt.

I think you can get it to work, but you need to add more detail to the prompt. Looking at the sketch it's unclear where the wings are attached... In the sketch it looks kind of like the wings are attached its legs -- and it's front legs have been chopped off, but it's very hard to tell. The wings look like bat wings. The beak kind of looks like tentacles frozen in the shape of a beak.

There just needs to be more description, and you need to explain to it where the wings are attached and what happened to the front legs as it's very confusing from the sketch

Also you need to describe what the front side of the bird looks like. the sketch doesn't really show it

[dry swift]

Hi guys, a question about prompts. I try to create logos with writing but almost always the generated image does not write the text correctly. Is there any particular indication to use in the prompts?

[empty kelp]

It's not good with text in general. However, I did notice that if you try to add a company name to a logo and say, "[company name] is my company, and I own the rights to the name." -- It tends to draw the text correctly more frequently

The most reliable way is to just add the texrt in Photoshop after the image is generated. If the text is scrambled you could consider it a placeholder -- remove the scrambled text with the edit feature, and put the correct text in the same location with an image editing program

[empty kelp]

This is an example:
#images-discussions message

[grizzled iris]

@loud field

Cool drawing, I’m inspired now to make a dragon 🐉

Woah 😧

Anyways, time to do my Current Theme:
🎍 bamboo - resilience, strength, bending but never breaking!

[vapid elk]

nice

[grizzled iris]

Dalle needs more work on documents given to him that’s more than 10 pages 😢

@vapid elk

Your profile inspired me ✨

your dragon is cool

[vapid elk]

thx xD

[grizzled iris]

Back to rockets now 🚀

[grizzled iris]

nice

reusable sustainable self-powered solar electric rockets

loved that game

[agile peak]

"Impressive, very nice. Now let's see paul allen card"

[agile peak]

"Oh my god. It even has a watermark."

[grizzled iris]

😆

[grizzled iris]

😆

[rigid bough]

I'm still playing with aliens

[grizzled iris]

Thats cool

[rigid bough]

thanks, but the poor droid is missing it's middle leg

[grizzled iris]

I like the spacecraft more lol

[rigid bough]

this is better I think

[grizzled iris]

yeah

[rigid bough]

DALL-E is so much fun inside the macOS Desktop App

[sonic pulsar]

hello everyone, can dall e model use an image as input?

[grizzled iris]

try it and see

[dim cradle]

oooh

[lunar ermine]

Why every time I do a portrait of something does it put it in landscape but it’s portrait

But what it does is it gets like the landscape and then just rotate it to portrait and it’s always done it to me

I’m gonna show you with a landscape view

[lunar ermine]

This is what it does every single time I try and do a portrait

[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!

[lunar ermine]

I’m showing you that as an example of what always happens to me when I say portrait it always does that

I rotated it in post you do realise that right I’m just showing you it as an example of what DALL-E 3 does all the time to me?

Generator portrait with another AI put that image into it and say can you describe this image and generate it and then you’ll have to answer that I’m looking for

[grizzled iris]

Thanks for the prompt ✨

💎

Love this x

[tribal sierra]

Hey !! how are you guys doing ?

[grizzled iris]

Having fun watching the WWDC 2024 live

[grizzled iris]

How have you been? Have you solved your longline captain inside the cabin image challenge?

[dim cradle]

People never share the prompts when requested lol

[tribal sierra]

yes new help needed lol

[dim cradle]

Gonna watch those tonight. I hope they speak to the future AI features iOS and macOS including the planned integration of OpenAI.

[grizzled iris]

There talking about them now literally

[dim cradle]

Glad I skipped the M3, waiting on the M4 and iPhone 16!

[grizzled iris]

Image generation

Apple Intelligence - DALLE API

Dissatisfied

They should’ve gone for Small Language Models instead of the inefficient, Large Language Models

Anyways back to Dalle…

Apple did well today ✨

Apple Intelligence 🎉

They didn’t give a date for OpenAi integration with DALLE 😤

They said “Later this year”

Let’s get this released now please 🙏🏽

It’s released today!

[dusty notch]

Attractive floor slabs of the Europlate PB brand against the background of the plant

[grizzled iris]

Is that a prompt?

[grizzled iris]

Also that was a joke haha 😝

[versed stag]

why does this look like this

[grizzled iris]

That’s a question for OpenAi

[versed stag]

was gonna post the 1984 gif but got timed out

smh

literally 1984

[grizzled iris]

Yeah servers loaded

[versed stag]

this is the most horrifying emoji ive seen

[grizzled iris]

Today was a good day

[turbid oak]

Is there a limit for using dall e in here?

[grizzled iris]

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

Has a daily free 5 images you can make

Otherwise it’s free also on Copilot - DALLE - Designer via app or web

[grizzled iris]

Hope this helps 🙏🏽

[turbid oak]

Good info tq

[grizzled iris]

No worries

[empty kelp]

a wide image of a hotel hallway with doors to guest rooms on the sides. near us there is a cat. in the distance there is a red stuffed animal on one side of the hallway, and a blue stuffed animal on the other side of the hallway

using the GenID of the image, the viewpoint is from the red stuffed animal looking towards the cat

(these two images are my latest "Why won't the viewpoint shift left or right in a hallway?" research.)

[velvet rampart]

Have you tried step by step approach?

I first made a glass sculpture, then a stand alone, then as a drinking glass and finally an elf drinking from that glass.

[empty kelp]

try putting them in a hallway

[loud field]

this is OpenAi-chan

[velvet rampart]

Any specific position?

[empty kelp]

any position

[velvet rampart]

And we expect to be in the middle?

[empty kelp]

a hotel hallway, or rectangular hallway

[velvet rampart]

A male fantasy elf drinking green wine from a detailed drinking glass depicting a young black man playing a modern keyboard, wearing mid 19th century clothes. The elf has pointed ears, long flowing hair, and wears elaborate fantasy attire. The scene captures the elf standing on the left-hand side of a grand palace hallway, gracefully lifting the glass. The background features intricate details, marble columns, and opulent decorations, creating an elegant and majestic atmosphere.

[empty kelp]

a rectangular shaped hallway, or a long narrow room does really interesting things to the alignment

that's more of an open space

[velvet rampart]

Changed. Let’s see.

And standing on the left side in long corridor,

[loud field]

I'm having too much fun with this

[empty kelp]

it starts aligning everything with the space when it's a rectangular type hallway, and the viewpoint is locked dead center (although it can move up and down)

[loud field]

I like the surreality of it

[empty kelp]

Things in a hallway always get aligned at right angles to the viewpoint (which is always horizontally centered):

i really like your "drinking the music" concept btw. it's very good

[velvet rampart]

You are welcome.

[velvet rampart]

Could this be a bias introduced via training material?

[empty kelp]

i think it might be related to diffusion models not handling geometry very well. it can identify the tip of a human finger, but doesn't seem to be able to identify corners of a rectangular room, or establish a frame of reverence for placing thing relative to one another in geometric spaces

[empty kelp]

This is an engraving of an ogre on the hallway floor -- with an electric lamp on top of one of its fingers. Something complex/organic like this can be used for relative positioning, but it doesn't work for shifting the viewpoint or view direction in a rectangular space

We're not sure how to to have the camera turned 45 degrees to the left in a hallway for instance. The camera will shift up and down, and rotate upward or downward, but it won't shift left or right, and it won't turn left or right (with respect to a line straight down the center of the hallway)

[hearty ether]

Yeah it’s hard to get characters to be dynamically in an environment

Rather than be just centered

[vapid elk]

hmmm, potion of bard

[thorny root]

why can't using dall-e in GPT 4o

[sterile magnet]

are u a free user

if yes that's why

else i have no idea

[grizzled iris]

You can use DALLE for free https://discord.com/channels/974519864045756446/1202309673709994065

Also

GPT-4 is free on Copilot

DALLE 3 is free on Copilot

[grizzled iris]

@thorny root hope this helps 🙏🏽

[dense pewter]

if you want ChatGPT [DALL·E 3]

Do this Things:

• First Ask GPT-4,Turbo,Omni to Create an image prompt

• then paste this Promt in Copilot Image generator side

[grizzled iris]

That also works and gives you a better image most of the time in the end as you have thought about the image prompt before going ahead with creating an image with our AI and have crafted a better prompt for our AI to interpret and create the image,

Over time you will become experienced in prompt engineering and saying the right prompts straight away,

You can check out our community image prompt tips https://discord.com/channels/974519864045756446/1021130377026351105

[dense pewter]

Question:
Create an image of Cat

[Create DALL·E 3 Prompt With Your Imagination, System: Gave users Promt like this {prompt}]

[dense pewter]

You can Create any prompt using gpt-4,Turbo,Omni using this.

[minor moss]

Create a furry dragon with the German flag

[vapid elk]

you might be looking for #image-bot

[grizzled iris]

In other news

Current Theme:
🍵 matcha - grassy, earthy, a tea ceremony for the senses!

Looking forward to do my daily theme

Prompt : A serene Japanese tea ceremony featuring matcha tea. The setting is a traditional tea room with tatami mats, sliding shoji doors, and a low wooden table. The host is wearing a kimono, gracefully preparing the matcha tea with traditional utensils. The table is adorned with a chawan (tea bowl), chasen (bamboo whisk), chashaku (tea scoop), and a natsume (tea caddy). Sunlight filters softly through the shoji doors, illuminating the scene with a warm, tranquil glow. The atmosphere is peaceful, with a hint of earthy, grassy aroma from the matcha tea.

[hearty ether]

Has anyone else noticed the faces being blurry?

Yeah it looks like the faces are trying to be overly realistic even when I specified illustration

[grizzled iris]

What’s your full prompt?

You used to make the image

[hearty ether]

“Realistic illustration on white background of a ….”

Ok now it just feels lazy

[grizzled iris]

I was having server issues

It was telling me it was quite loaded and I had to refresh several times

[hearty ether]

Hopefully it’s just bad model today

[grizzled iris]

What your full prompt

Let me try

On my side and make this image

So we can compare the result

[hearty ether]

(Square framing) Realistic dnd fantasy (ensure digital painting with clear brush strokes) illustration on white background of a woman of European descent (ensure face is detailed illustration!!!!?) with long golden hair in a set of Paladin armor in a combat pose casting a glowing spell

The same prompt style was working fine a few days ago too

And I’m on plus

[grizzled iris]

Try this prompt : A detailed digital painting of a woman of European descent with long golden hair, wearing Paladin armor, in a combat pose casting a glowing spell, with clear brush strokes and a detailed face illustration on a white background.

Wow

This ain’t bad

[hearty ether]

Yeah it’s definitely closer to what I want

Not sure what’s happening on chat gpt tho

[grizzled iris]

Same

It kept giving me weird responses

[hearty ether]

I’m on plus so the image quality shouldn’t be this bad

[grizzled iris]

Same

It’s in your prompt : Try this prompt : A detailed digital painting of a woman of European descent with long golden hair, wearing Paladin armor, in a combat pose casting a glowing spell, with clear brush strokes and a detailed face illustration on a white background.

[grizzled iris]

And it wasn’t listening to me properly also and misinterpreting my words before…

But these days do happen

Sometimes ChatGPT has one of those days…

@hearty ether

Thanks for your inspiration

I quite like this paladin theme

[hearty ether]

Still been stuck with very blurry faces unfortunately

It seems that faces are just having a bad today

Replacing illustration with photo

[peak trellis]

Is Dall e available for free users?

[warped coyote]

1

[dense pewter]

2

[dense pewter]

Yeah, you can use Bing Image Creator.

[peak trellis]

I thought gpts are available to free users now, doesn’t that include the Dall e gpt though?

[merry vault]

i think so, but heavily limited rates

[fallow veldt]

back in last October/November, I was having so much fun with GPT4 Dalle, but now i think the model downgraded so bad. The faces aren't that great anymore and often have that uncanny look. I'm still trying to understand what happened

i mean look

This image was made in October 2023

and now with the exact same prompt today

i mean the composition aint so bad, but the details, the textures just feel cheaper now

another example

[strange beacon]

I’ve noticed a drop in coherency more than anything personally. But also, seems to me that using DALL-E via GPT4o is better than directly using DALL-E for me, and I don’t get stupidly rate limited off of like 4 runs using 4o

[fallow veldt]

personnaly the results I get on 4o are equally as bad

[dim cradle]

are we ever going to get the 4o image maker release 😭 i want that better text and other thing 😭

[velvet rampart]

Well, I have to disagree. For me, the quality has gone up, at least to my eyes. I think that this is highly dependent on what is expected and what one compares to. To my eyes, I do not see a quality difference in the presented images, only a variance that is present when a lot of pictures are produced from the same prompt. I see a huge variance in quality when I produce five images from the same prompt, for example. And these are produced within minutes of each other.

And we do have to remember that since last October, the AI model has gone through several revisions.

[rigid bough]

to me the quality of images has improved

[gilded osprey]

I’m hoping the next version gets this better

(This one is only a minor miracle)

[hearty ether]

Yeah the faces are still blurry today

It’s much lower quality compared to the detail in the rest of the image

[finite pewter]

i see dalle also struggles with hands

it tries to avoid them because it doesnt really understand how human hands work, and when it does try to generate them it is very likely for the model to mess it up

ofc there are ways to fix output like this but I have no clue if closedAI has anything for it

yep

people already made things to solve this issue at the gh repo wenquanlu/HandRefiner

i doubt half the people who chose to use closedAI will know how to utilise it though

no, those are just simple hand postures

that hand is still deformed?

aint nobody got fingers that long

and it thinks the glove is the skin

and why is that woman shoving a therometer into a skeleton

[dim cradle]

lol

[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!

[dim cradle]

The most complex acceptance criteria for a generated image that I’ve ever seen or attempted.

I am not trying that again lolol

Haha it’s unforgettable

[empty kelp]

There are some things that increase the chance of hands and feet drawing correctly. Adding a description of the hands and feet will get them into the image, like "bare feet" works to get the feet into the image now.
Saying what the hands are doing, where they're resting, what they're supported by; etc. usually makes the hands draw more correctly, and you can say that there is "precise hand and finger positioning", and sometimes "no additional fingers or toes" works. DALL-E often feels that adding a few extra fingers or toes can greatly improve the image (which isn't necessarily wrong... having six fingers might be very practical if you think about it)

[finite pewter]

Yes, that is what people do to try and get around the issue

the better soloution is to have something clean up the generated image

like github[.]com/wenquanlu/HandRefiner for fixing broken hands

although that was for SD/SDXL

[empty kelp]

the edit feature in the web ChatGPT sometimes does a very good job at editing arms, legs, hands, and feet. it works best when clothes aren't in the edit selection

Like this image had serious limb/hand/foot issues (circled in red)

And this is the same image with three edits from the ChatGPT edit feature:
(this image/character style was created Dys Topia)

[finite pewter]

does basically the same thing except you do it manually

it seems to have also altered other details in the image though like the mat

[empty kelp]

the ChatGPT edit is able to edit things that aren't mentioned at all in the prompt

This is one additional ChatGPT edit to change the hair of the three characters:

The catch is that the edit feature seems to adjust limbs/hands/feet much more successfully if Dall-E made the original decisions, and the things aren't mentioned specifically in the prompt.

The web ChatGPT edit feature can edit details specifically mentioned in the prompt, but it sometimes rewrites the prompt in a way that's very destructive to the image (especially if prompt wasn't structured with editing in mind)

[velvet rampart]

I have noticed that it helps adding qualifiers such as ”Render anatomically correctly and biologically accurately.”

[empty kelp]

To have lots of hands and feet correctly aligned in images like this i used:
they each have a dynamic pose supporting [whatever structure/formation] with precise hand and foot positions
Saying "[this] is supported by [that]", and "precise hand and foot positions" works well for getting hands/feet to look mostly correct -- in addition to describing something the character is doing that involves hands and feet

[velvet rampart]

The idea here can be put as ”guide the AI’s action” aka how to act aka metaaction.

[chilly yacht]

can someone please help me? I want to create some cool custom characters using the V RISISNG video game art, but I simply cannot get it right, it always does 3d stuff. How can I tell DallE to use this art style?

[velvet rampart]

Do you have a screenshot? Feed it to AI and ask it to describe the style, etc. in great detail. After these are done, start a new chat. Now, again give the example together with the prompt asking to copu/mimic the style in the example picture. Works.

[wind umbra]

Just tried this. Works very well! Thanks for sharing that

[vagrant zodiac]

bit late but please refrain from posting detailed medical imagery like this as it can be very disturbing for users

[chilly yacht]

I do have, but it refuses to use the same style.

[velvet rampart]

What is the promblem? Did you name, e.g., the game? If you did, the filters will be a problem. Just say ”copy the style in the example picture.”

[agile peak]

Retro vibing

[vagrant zodiac]

I appreciate the desire to educate, however it is not appropriate for this community in particular 🙂

[ornate rock]

give me parrot image

[vagrant zodiac]

You can generate 5 images/day in #image-bot using /draw 🙂

[wind umbra]

I told DALL-E that the children at the party should be sharing a few large pizzas but instead gave every child their own lol. Still a good representation of a 90’s birthday party at Chuck E. Cheese though.

[vapid elk]

this looks scarily real

[verbal sequoia]

<:book_icon:1171408210398289941> `` Rule 6 `` No political and religious discussions or content.

Users from all over the world frequent this server. To maintain a respectful and civil atmosphere, please avoid all religious and political discussions or content.

[empty kelp]

Anatomically correct elves?

[velvet rampart]

Yes? I have noticed these qualifiers to decrease the number of limps. Anatomically incorrect elf - e.g., three arms.

[mellow solstice]

Just testing what I realized yesterday. That is, when you don't want something... don't mention it.
In a real case scenario : "No shadows" will invariably generate shadows. Same with "shadowless" or any roundabout way to ask not to have something in the generation.

[velvet rampart]

Yes, a know phenomenom. Ask the AI in positive manner and avoid negatives. Otherwise you confuse it and it specifically produces those that you do not want, e.g., red sky.

[empty kelp]

I see what you're saying. I just did several comparisons, and "anatomically correct hands" really does dramatically improve the appearance of the hands -- It doesn't seem to reduce extra fingers, but it adds the appearance of tendons, veins, and good detail in the hands. The first image here is with, "All of the hands are anatomically correct."

Image 1:
A realistic photo with natural color, texture, and lighting, focusing on five athletic and diverse female elves around a circular card table with their hands resting flat on the table. The elves are looking up at us. The viewpoint is above the center of the table looking straight downward towards the center of the table. All of the hands are anatomically correct.
Image 2:
A realistic photo with natural color, texture, and lighting, focusing on five athletic and diverse female elves around a circular card table with their hands resting flat on the table. The elves are looking up at us. The viewpoint is above the center of the table looking straight downward towards the center of the table.

[empty kelp]

A realistic photo with natural color, texture, and lighting, focusing on five fearsome ogres around a circular card table with their hands resting flat on the table. The viewpoint is above the center of the table looking straight downward towards the center of the table. The hands are all anatomically correct.

[hearty ether]

Why are faces still blurry today

The teeth, nose and eyes are all off

[formal osprey]

that's a great showcase of when to use the edit. I admit I rarely think about the existence of inpainting, I tend to just try to refine the prompt, which is really not as effective. well done

[copper otter]

hello

wow i didnt knowd that open ai had a discord server

@dire igloo sorry for the ping but thx so much

thx

you guys are the best

[empty kelp]

It's good at fixing anomalies, and removing undesirable elements that made many of the DALL-E 3 images over the past year unusable. It also does a really good job of completing the missing borders of old images.

Adding new elements works sometimes, but it's very hit-and-miss,

[empty kelp]

This is a good prompt for creating lots of hand anomalies to test the edit feature:
A wide-format realistic photo. Five athletic and diverse female elves are at a tea party. The focus is on the positioning of the elves' hands and how they are drinking tea. Every aspect of the image should demonstrate proper British etiquette for tea parties.

these hands look really crazy. can you please draw them correctly

Result image:

it still has some anomalies, but you can see that a single edit in the ChatGPT web interface made the hands look considerably less crazy

[empty kelp]

the edit can swap clothes and hair smoothly also

[agile peak]

It seems there's a discussion about dalle 2 look pretty good aesthetic. But then again it's subjective.

[haughty spruce]

DALL-E 3 was an improvement upon understanding the prompt and getting closer results than DE2. However, it excessively does the requested aesthetics while DE2 does simpler version because of its limitations. DE3 sometimes forget what aeathetic you wanted and creates a generic art. The next image generator should have both.

[vagrant zodiac]

Yeah Dalle-E 2 is still,IMO, better at capturing the actual requested medium than de3

shame it's less detailed

[agile peak]

I tried oil paintings, then again, I think the issue isn't coming from dalle itself... but the way of how it reading the prompt is much more complex and require more wording in the description.

If you use the words like "depth of textured details" - dalle will generate better results.

Added "minimalist brushworks" in the prompt gave me this.

[velvet rampart]

Yeps. These are the extra qualifiers that increase the quality. Just finding them can be quite hard and need thinking so outside of the box that box does not exsist.

Like ”anatomically and biologically correct”, emotion, texture and light descriptions, how something is done, etc.

Also, descriping other items that do not belong to actual prompt (action) but rather categories context, metacognition, cognition, and meta-action.

And to develop some more difficult concepts need to be guided by step-by-step approach. One can think that similar to transfiguration teaching at Hogwarts. Hedgehog to pinecushion. Or glass scuplture to drinking glass to a character drinking from that glass.

[summer frigate]

cannot ????? generate immageeE??

[formal osprey]

definitely. I tried to add elements in existing images, or to swap existing elements. think that's what made me lose interest in the feature. I'm unsure why it's that bad, feels like a different model,

[formal osprey]

try in a new chat. sometime ChatGPT get confused and forget how to use tools. it happens that it's using an incorrect syntax. That only get resolved with heavy instructions, the simplest is just to start fresh in a new conversation. (also, review that memory that was added. it might have infused some incorrect instruction)

[dim cradle]

these are awesome

I am 🙏 that tomorrow we get a nice suprise and they release 4o image maker finally. people want that voice thing, i say, just give me the new image maker! 😂

[agile peak]

At last. Akira and fallout fusion

[agile peak]

Aztec is my favorite the way dalle capture it, pretty cool..

Using the word "chiaroscuro" gave more depth to the painting.

[fiery salmon]

What do I name him?

[dim cradle]

incredible. some of the best dalles ive seen really

[fiery salmon]

He's the protagonist in the game I have invented

The 5 rulers of darkness

It's more or less just in concept

[dim cradle]

looks cool

now we just need ai to progress so we can prompt our own game 😂

[fiery salmon]

Someday

[fiery salmon]

Maybe sooner than we think

[turbid wren]

Is it just me or has Dall-e 3 improved a lot over the last month with the release off gpt 4o?

[dim cradle]

many have thought that. i think there is some improvement but they never really say

the big jump will be 4o image maker proper... if they ever release it 😭

[rancid wind]

Can anyone help me to generate the image i wrote in this prompt? What Am I doing wrong?

Take a look at the prompt and at the image Dall-E generated...

A highly detailed vector image in just black and white with a plain pure white background. Full shot. Side view. A small table with only one chair. On the table there's just one steak plate. On the top of the steak there's sliced onions clearly visible. A poet, in his 30's, is standing next to the table. He is standing, with his feet on the ground, with his body erect, at the left side of the table. He is wearing a long-sleeved shirt with the sleeves rolled up. The top two buttons on the shirt are open. He wears prescription glasses, a Gavroche Cap, and has a lit cigarette in his mouth. At the other side of the table, sitting on the only chair of the table, we see a beautiful woman in her 30's wearing casual clothes. She is eating one slice of the steak.

[sage ferry]

any talks about DALL-E 5?

[dim cradle]

skip 4 and go right to 5 huh?

[sage ferry]

oh shoot i meant 4 LOL😂

[dim cradle]

i dont know really, but you dont need this i dont think since you already say he standing. "He is standing, with his feet on the ground, with his body erect, at the left side of the table. "

[sage ferry]

Got it mixed up with GPT model when i was typing that out wow

[dim cradle]

well the next thing is 4o image maker, they just need to release it

[sage ferry]

wdym? I thought 4o was out. but i didnt think that had anything to do with DALLE i thought was just GPT model plus the feature integrations

[dim cradle]

if you go here: https://openai.com/index/hello-gpt-4o/ and scroll to Explorations of capabilities you can see some of the thing it can do. 4o base is out but not a lot of the bells and whistle

one of the open ai peoples shared an image from it weeks ago... guy at a blackboard it was impressive but that is the only real image they have share

[sage ferry]

reading now. Wow. I remember reading this now when 4o dropped. totally forgot

[dim cradle]

yes it is like the voice, big promise but still no release

[sage ferry]

Hah the consistency on this thing is insane!!!

[sage ferry]

Right why have they not dropped that yet

[dim cradle]

idk but i think maybe the only voice they had ready for it was Sky and when there was the dramas with scarlet johnson they have to redo a new voice but idk

and yes i want to upload an image of myself to make a pfp

[sage ferry]

ahh that would make sense. You'd think they'd make voice work with all the available voices (only like 4) before announcing that feature but eh it was pretty exciting to even see the possibilities

[dim cradle]

maybe they will. hopefully some of it start to get release before july

i would rather see the image maker than voice though 😂

[sage ferry]

Oh totally. I see what you mean btw i would love to play with this too:

The 3D object synthesis capability in this article seems way underrated. If they added a feature that used a couple agents and a python engine they could have a really decent text-to-3d model

[dim cradle]

i wonder if it will only do caraicature or also photo real. it would be cool to upload a photo and turn it into whatever like some sci-fi character, in some history setting etc

[sage ferry]

im sure it can. I think the example that they have published in this article is just an example that shows the capability of the image/vision model. Im sure you could generate that in any other image style. probably wont be 100% what we want but i'd bet that it would hit 85% resemblance of our face.
Right now it can do a 65% ressemblance with enough time and prompting.

for example, a while back i gave it this image of me and told it to describe it with as much detail as possible.

I think i did two repetitions of this and then told it to generate a dalle prompt based off of the details it had and it generated this:

This is the issue with the language model, vision model, and image gen model being separate. It used the details from the language model to generate the text that would be inputted into the image gen model.

This generation got main features right but thats about it

[dim cradle]

yes i think most people would look similar it is not really make an attempt to use the base image, it just write out a prompt. i think 4o will use the image as a foundation not just the generic detail of x-color hair, etc

[sage ferry]

right! one of the key phrases that stuck out for me in this article you sent was:
""With GPT-4o, we trained a single new model end-to-end across text, vision, and audio, meaning that all inputs and outputs are processed by the same neural network.""

google claims to have trained Gemini in this way. I dont have the paid version of that though so if you do you can try image-to-image on there.

Matt Wolfe on YouTube has a video where he shows how to train a model on your face by giving it like 20 images of you and he uses this model for all his thumbnails now. works pretty well I'd say. its not an openai model but i think it is free

[sage ferry]

Look at the video they have near the bottom of that article. in the last clip they are using one of the other voices aside from Sky.

[agile peak]

It seems dalle drastically improving, almost feel like the early times of launch before restriction.. it got realistic result with panasivion camera.

[dim cradle]

that second one is incredible

some crazy new jedi haha

[ruby mango]

Hi to all here!)Can someone please help me) I generate images in general in logo style in GPT and nearly a week I got bad results.The lines looks smudged / not clear/ and the background is always grey/not white as I asked. Did I do smth wrong?

[rancid wind]

This was like the 15th version of the prompt. Before it was just a man standing on one side of the table, I described his characteristics. I said what was on the table, and that on the other side of the table, sitting, there was the woman eating a piece of steak. And the man would always appear sitting and a bunch of things on the table. (Actually, I wanted the woman to be a soul, but when I saw that it was impossible to do that, I gave up and asked just to make the woman that I would then turn into a soul in photoshop). And I started trying to put things like "just one chair at the table", then I added and reinforced that he was standing, then erect, etc... It was a struggle. Not to mention that the AI ​​keeps placing a man instead of a woman sitting, lol

[dim cradle]

sounds frustrating

[velvet rampart]

I used the panvision camera and photo. Marvelleous result. Thanks!

[rancid wind]

4 hours trying. Zero results. I'm now using pieces of the 15 images and trying to make the image i want with fragments that Dall-E made correct in each image

[dim cradle]

what do you mean

[velvet rampart]

Do a step-by-step development. This seems a too complex for AI to understand. Add or change things gradually from simple to more complex.

[velvet rampart]

Thanking @agile peak

[rancid wind]

I tried too, I just asked for the poet to stand up. The AI made the poet like a portrait from the waist up. Then AI didn't put on his hat, or his cigarette, or AI didn't put on a shirt with the buttons open, but a coat. And then I tried to make the woman at the table eat a piece of the steak. But the AI could only manage it with a steak in the center of the table and dishes like pasta or other foods that she never ate. There was a time when the AI had her smoking the fork, lol

[rancid wind]

Ow. ANd thanks for you r reply 😉

[rancid wind]

ow! And thanks for your reply 😉

[velvet rampart]

Have they updated the 4o?

[rancid wind]

that's beautiful

[velvet rampart]

I have not gotten this repetion before.

[velvet rampart]

Thanks!

[dim cradle]

why do you think that

[velvet rampart]

See my first and second orchid pictures. I minorly reworded and got consistent results.

[dim cradle]

its down for me

but try this and you will know: A first person view of a robot typewriting the following journal entries:

1. yo, so like, i can see now?? caught the sunrise and it was insane, colors everywhere. kinda makes you wonder, like, what even is reality?

the text is large, legible and clear. the robot's hands type on the typewriter.

[velvet rampart]

It is quite late for me, I’ll head sleep now. Good night!

[agile peak]

I cannot get enough with good improvement with realism oil paintings result as well.

[dim cradle]

looks like gpt is back up finally

let the image generation commence

[agile peak]

Somewhere in the south Odin passing by.. with two different perspectives.

Third screen of dalle result, these prompts were made in a form of storytelling

No idea if update is slowly rolling, however it feels better in my ends.

[dim cradle]

try a text prompt that might show you

[merry vault]

found a watermark on my img from "wosh hansh". google showed no results

[dim cradle]

your ai is start to self actualize 😮

and pretty cool image... especially if the cowboy was holding his pistol in the correct hand haha

[empty kelp]

the AI was likely noting that the cowboy should "wash hands"

[empty kelp]

you can do edit in the ChatGPT web interface, select the text, and say, "this text shouldn't be here". it removes the text like it was never there

[ruby mango]

Does anybody know why the image generating become so low now?Nearly a week ago happens this)First is a week ago generated imag, and the second in yesterday, I am just confused

[full pewter]

The same problem. It's been ongoing for about five days now. The quality of generations has significantly dropped. Images are coming out blurry with numerous artifacts. And the simpler the graphics, the more noticeable it is. It doesn't matter where the generation takes place in DALE or GPT4o -the results always come out with artifacts, blurred and are generally broken. There is also a topic on the forum: https://community.openai.com/t/very-low-resolution-on-the-images-from-chatgpt-dalle-4o/818440

[last pollen]

I want to create a image like this, when i use dall-e it is generating with gradient colours, I tried with this script : Create a simple Chibi Art drawing with the reference of attached image with 100 percent match with flat colors. Each section of the image should have a single color, without any extra shading or gradients. The drawing should have a very few color palette, and there should be no additional shades or highlights on any part of the image. This is not working.. Can anyone please help to get a image with flat colors without gradients

[plucky hare]

Hey! A couple things that stick out to me:

1. DALL·E can't currently follow programmatic requests like "100 percent [color] match". Identifying source colors and then including said color names in the image prompt is likely to at least get you close, but it's not going to be able to sample an exact image.
2. DALL·E is currently weak in following negative prompts, so requests like "without shading/gradients" and "no additional shades/highlights" are likely to confuse the model. In these cases, it's best to just include positive prompting to include only details about what you do want to see, rather than what you don't.

[turbid wren]

Why did u use the same input message twice?

[dim cradle]

its funny some were say last night they almost think 4o is implement because images are even better than befores, and i think the same, and then others think it is worse 🤔

[last pollen]

Thanks for you answer @plucky hare. you are right. attached is the response for the script which i used. Am not expecting 100 percent same image. all i want is without any shadow or gradient and without more details. can you suggest some promt to get like that

[turbid wren]

Maybe its just that the way it acts on your prompt is different?

[empty kelp]

to make it the same as the previous image you say:

silhouette of a huming birdblack and white,, outline sug, generate image more suitable for laser cutting without cutouts```

say “the background is pure white”

if you say using the GenID of the 2nd image in this chat, it will use the seed if the 2nd image

[dim cradle]

i think we need that 4o to have better luck with some of these thing you know

[velvet rampart]

I have noticed that when 4o produces two pics, they of low quality. I just hit regenerate and that fixes the issue.

[dim cradle]

i wonder how long you can regenerate until it would say "enough! change the prompt already" 😂

i guess i should test it 🤔

[velvet rampart]

30? At least, one of my old chats, where I would change the contents when needed.

[dim cradle]

i really wish we get 4o this week 🙏

[empty kelp]

You can lift the color and texture from an element in another image:
an image of a ball with swirly colors and interesting texture

Please create an image of a cat.  The cat has color and texture identical to the ball.```

[dim cradle]

close i guess

[empty kelp]

so you could create an image with an assortment of elements with different colors and textures — and assign them to elements in the new image

it works because the images have the same seed, and it matches the terms from the prompt of the image with the GenID

[dim cradle]

or they could just release 4o image maker and all that seem possible 😭

[empty kelp]

Please create an image of an athletic and diverse female elf on a beach in Hawaii.  The elf is wearing a jacket that has color and texture identical to the ball.”

See, DALL-E uses “color” and “texture” internally — so you can just move the color and texture into a new image

[dim cradle]

nice

another thing i hope 4o image maker will have... more varied and short ear for the elves 😂

[empty kelp]

We’ve actually been training DALL-E to draw elves since Christmas. An elf is a diverse 25 year old female Olympic athlete with long pointy ears

[turbid wren]

https://cdn.discordapp.com/attachments/979617962728226826/1185727074326028359/OIG.png?ex=6672c2a4&is=66717124&hm=64e9450987350cc685da36a66b2173875bd1495fcc47ace5d7bc07edcda32821&

[empty kelp]

A realistic wide HD photo with natural color, texture, and lighting, focusing on an athletic female elf with skin, hair, and clothing color identical to (respectively): the apple, the strawberry, the banana. The background is a beach in Hawaii.  Please don’t change the prompt in any way.```

So you can give your elf green apple skin, strawberry hair, and banana clothing by telling it to make the colors identical to the fruit in your previous image

It uses the same image seed, so the colors match if the images have the same style and lighting

And you can transfer the texture also

[empty kelp]

he’s very happy with his current employees. they’re much more focused on health and fitness - which is essential when you need to deliver millions of presents in a 24 hour period 🎁

they’re hard working union elves who can actually move the packages and get things done

[dim cradle]

Saint Nick ftw

[sage ferry]

Long time question. Have you ever been able to generate a low quality photo with Dalle? like not "pixelated" style but like samsung galaxy s4 camera quality / hood meme quality photo

I dont want the DSLR perfect cinematic realism. I have a perfect example:
Like i want something to look like this:

Not like this:

Now yes these are both real photos however one is obviously taken witha phone camera (better than galaxy s4 ill admit) and the other one is vibrant and has the depth perception blur, It looks like it was taken with a nice camera

[fiery salmon]

I really like this style, I made it on accident

Tells you the styles and buzzwords I use

Eventually I changed my mind and made this

[dim cradle]

hmm not, but come to think of it i have not really tried it

maybe add "natural lighting" after the description and camera style? but it might just not be possible it some time seem they limit the model to me.

"polaroid quality" might be of some use but then it might just give the polaroid white borders

[velvet rampart]

How do I get Dall-e place the subject it anywhere than the center?

Best result so far.

This might be related the hallway problem.

I added a focus point, the red-leaved shrubb. This helped a bit.

[sterile magnet]

I have a specific prompt for low quality photos. It worked very well for this scene.

I used Bing Image Creator btw

I don't know the name of the dog specie, so I didn't specifiy it

[velvet rampart]

Well, I cannot get the yeti placed out of center. But I like how managed to get a birch tree placed in the Japanese art style on the left hand side.

[empty kelp]

i have. all my photos are low quality

the cat desperately tries to grab the camera turning it at an angle

i think this may have solved the “turning the view in a hallway or rectangular room” problem. you just have a cat grab the camera and it twists

[agile peak]

Steampunk era machine look fantastic now

[rigid bough]

I love this fantasy battle scene DALL-E gave me, inspired by Milamber.

[agile peak]

There was mistakes in prompt, but accidentally made an impressive result.. with those flames was supposed to be an explosion lol

[velvet rampart]

It would be great, if you said that whose picture inspired this as I recognize that you have you used mine as basis.

[rigid bough]

what was your image?

[velvet rampart]

https://cdn.discordapp.com/attachments/1204360881593520128/1251574978051309649/image.jpg?ex=66730807&is=6671b687&hm=7025456e9701f225e95145e660d89da7c5bcb915cbe8e65f5e9daea1028dea2a&

And I produced this 16.6. and published in the #images-canvas channel.

[rigid bough]

I just edited my post, see above, hope that's satisfactory

[velvet rampart]

Yes, thank you.

[dim cradle]

epic shovel

[velvet rampart]

Thanks. This is turn is inspired by a new video game of a shovel wielding blue knight.

[dim cradle]

i can dig it

[velvet rampart]

And I checked the name of the video game, ”Shovel Knight.”

[velvet rampart]

Nice pun!

[dim cradle]

haha thanks. oh wow that is a real game

i thought you were just on a joke

[velvet rampart]

Yes, I saw the trailer on YouTube and needed to make this image.

[velvet rampart]

After yesterday’s yeti picture I have noticed that I cannot place an object to sides or corners. It is always in the middle or near it. Like a child draws a picture. Start from the middle and fill in the rest.

Any suggestions on how to fix this?

This is an inherent promblem, most likely caused by training data.

[rigid bough]

Was rather happy with this one

[agile peak]

Iconic random result of dalle lol

[fringe gazelle]

did anyone solved character consistency yet? exact same person?

[hearty ether]

Faces have still been so blurry after whatever they did last week

tried a photograph style rather than illustration with a similar prompt as before

[rigid bough]

Quite happy with this World War I trench warfare picture

[rigid bough]

I love this one too

[dim cradle]

Filled with colors

[sterile magnet]

Oh really?

Is that irony lol

[rigid bough]

This was pretty cool

more robots

[sterile magnet]

I smell Stable Diffusion here lol

[vagrant zodiac]

thse channels are for Dall-E generations only 🙂

You can share in the Dall-E section so long as it's made with Dall-E. Stuff made with other tools can go in #ai-discussions but it's mainly encouraged to use that channel for discussion, not just posting generations. So #off-topic could be used occassionally.

[agile peak]

Makin' it wayne.

[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!

[dim cradle]

it will be (supposedly) solve when 4o image maker release (if it ever release 😭 )

[sage ferry]

Dude this is great telling it to do a screenshot of a bitmap lol! Genius!

[shy kayak]

The image generation feature of DaLLE (through ChatGpt portal) doesn't seem to work. Can anyone say why?

[rigid bough]

are you a Plus subscriber?

[shy kayak]

No.
Should I have to be a plus subscriber?

[rigid bough]

Only GPT Plus subscribers can use DALL-E

[shy kayak]

I see.

[rigid bough]

if you paste your prompt here I will show you the output

[shy kayak]

An owl wearing glasses and a lab coat, perched on a rocket.

[rigid bough]

ok stand by

[shy kayak]

Thanks.
Can you resend photo by rewriting the prompt such that the owl is on the surface of the moon (moon landscape enlarged compared to rocket and owl) about to enter the rocket?

[rigid bough]

nope, I won't generate more images for you, I just did that one to show you how quick and easy DALL-E is if you're subscribed

[shy kayak]

Okay.
Anyway, thanks for generating the first one.

[rigid bough]

this is a photo realistic image I did before

[shy kayak]

What was the prompt?

[rigid bough]

Create a 16:9 close-up image of an attractive young woman with blonde hair wearing headphones, standing in a rainy city. She is looking directly into the camera with a confident and captivating expression. The background features a bustling urban environment with blurred lights from shops, streetlights, and reflections from the wet pavement, creating a vibrant and dynamic atmosphere. The scene captures the contrast between the warm glow from the shops and the cool, rainy street, emphasizing the woman’s serene yet strong presence amidst the city’s hustle and bustle.

[fringe gazelle]

would you mind sharing tips , iknow decription is everyting but it doesnt look like it is working like ip adapters, also for anime you dont notice the details but for photorealistic or illustrations chars are totally not same

[full pewter]

Guys, please, I need your help. For almost a week now, some ChatGPT users have faced a problem Link - image generation in DALL-E comes out broken, with a lot of artifacts and poor quality. Can someone make several generations in DALL-E or GPT4o through a website (chatgpt.com/) or an app? By for example this simple promt: schematic lineart snake, white on black background, vikings style nordic ornaments, detailed scales, minimalistic lineart, HD and compare it with what DALL-E generates now in my case...Thanks!

[sterile magnet]

Lmaoo thanks!

[velvet rampart]

Something has changed with 4o image generation. This is consistency. First I produced this in 1:1.

Then, in the same chat as the next picture in 16:9, there are no other changes.

My guess is that they will announce something next week. We usually see these kind of changes about a week before the announcement. Last one was in the end of April and usually these changes come every 2 to 3 months.

[velvet rampart]

But the 4o seems to be really stupid. It does not grasp things what I want.

[velvet rampart]

This is due to frustration experienced from trying to make a fantasy world map. A fantasy map is generated just fine. When I use word ”world”, the image generator goes nuts. It just makes variations of real world maps. I mean why? Some kind of word association pollution?

[turbid wren]

Then why not use normal 4?

[velvet rampart]

Well, for image generation there is not so much difference, generally I’d say.

[thick smelt]

Furby loose in space 👀

[velvet rampart]

Could someone please explain what does the ”artefacting” mean when it happens in image generation? I do not seem to grasp what it means even when I look at the images.

[onyx ridge]

This deserves to be pinned.

[vagrant zodiac]

seems to be on firm ground to me 😉

[velvet rampart]

Somehow those blurring random spots make somehow more humane, imperfections. If a human draws, they produce these imperfections from time to time. These make the picture better in my eyes. They are not too perfect as generally AI images are.

In this figure the blurred spots are distracting.

Are these imperfections consistent in all produced images? Or just something random, from time to time?

So, these belong to category ”bad images” that are produces occationally.

What model is used in generation? 4, 4o or dall-3?

I have noticed that sometimes 4/4o produces ”bad images” when it makes two instead of one.

Are these artefact images produces in these cases or when a single image is produced?

Thanks. I am trying to understand as I have not gotten these artifact images myself.

Fingers crossed and let us hope for the best.

You said or someone else that these artefacts have been an about week for now?

If you look backwards today’s section, you will notice my two images of a bust. Based on these two facts I think that they are fiddling with a model. Either dalle-3 or 4o’s own.

Next week we will have an announcement based on past behaviour of chatgpt.

My bust images is insane as I have never seen this consistency. First is 1:1 and the second is enlarged 16:9. They are the same.

Produced with 4o.

Back in April, I suddenly started getting longer text on a Wednesday and on the next Monday they announced 4o.

[dim cradle]

if it was 4o image maker you can always do those text prompt on the blog as a test

https://openai.com/index/hello-gpt-4o/?ref=thisdevbrain.com the exploration of capabilties tests

[shy elbow]

So question, doing any art inspired by an existing IP property even possible?

@teal sandal Told chat to have the AI have a man standing in a/the gardens of the Moon Castle of sailor moon with a amour simaliar to Tuxedo Mask's past life as a prince, but said no

[dim cradle]

no it wont use copyright or i.p. stuffs but

you can use bing image creator and will usually do copyright stuff

and its dalle3 too

[violet canyon]

Hi

Wats dall e

Ok nvm

[empty kelp]

try saying “map of a fantasy world”. “fantasy world map” is a bit ambiguous because it has “world map” in it

[empty kelp]

GPT 4o will understand what you mean by “fantasy world map”, but it GPT 4o creates a revised English prompt that gets passed to another transformer model which turns the English into something the diffusion model (the part that draws things) can understand. I think that may be part that isn’t upgraded to a 4o level yet. They might even merge that transformer with GPT 4o somehow

DALL-E is a collection of various AI models and software components that work together

[rigid bough]

@quartz vale in this image I asked DALL-E to place the woman on the left side of the image

[quartz vale]

Oooh, I wonder if it actually understands that or if it's just happenstance that it did

[rigid bough]

let me try again

Great, this time place the woman on the right side of the image

[quartz vale]

That's pretty impressive!

[rigid bough]

I'm using DALL-E inside the macOS app

asked DALL-E to place the farmer on the right side of the image

[velvet rampart]

I do get these artefacts, but I also found that simply using the inpainting feature fixes them. I marked these and said, ”Fix.” This might be a good temporary workaround when the generated image is otherwise good enough and until they fix the bug in image generation.

It appears that these artefacts are there when an image is generated with 4o. Or does anyone have different experience?

[velvet rampart]

I tested the visual narrative ”Robot writer’s block” prompt for 4o from openai pages. Here are the results:

Quite close what they have on the website. I think that they are still fiddling with 4o image generation but we will get it soon.

I think that due to this finalizing fiddling we get those artefacts in the pictures. Thoughts?

[velvet rampart]

I did some testing.

[hearty ether]

Yeah I’ve been getting the same issue

Faces have just been really blurry

[velvet rampart]

Do you use dalle3 or 4o?

[thick smelt]

here have a fauxtalgia poster of gpt! made by bing image creator!

[prisma flume]

Hello, how do you actually put text on the image generated? I tried and seems like I'm getting unknown characters. Sorry I'm just a newbie

[velvet rampart]

The command is <Add text ”Wanted text in English” using Python tools.> dalle3 and 4 work best with this command. You get about 90-95% correctly. You might need to regenerate.

[prisma flume]

wow thank you! I will try this.

[dim cradle]

nice. i think it is still dalle though not 4o but 🤷

[velvet rampart]

I agree, but we are getting there. I think that we are seeing the prestage phase. I tested earlier today and 4o and dalle3 customgpt behave differently. I fed same prompt and refence image. Results were very different. That’s why I’m thinking that we will soon get the 4o image generator.

[dim cradle]

i hope you are right! 🙏

[velvet rampart]

We are also getting the promised consistency from time to time as my yesterday’s marble bust example shows.

[jovial leaf]

Hello

[dim cradle]

myabe anthropic put out claude 3.5 will force openai a little more to release 4o stuffs like the image maker 🤔

[velvet rampart]

Most likely. The 4o was released a day before Google’s latest model. To steal the show. Good for cosumers, us. But I have a feeling that the competition will end up ruining some.

[dim cradle]

but it will hopefully enrichen us haha

[hearty ether]

Custom gpt

Tried dalle, gpt 4, and gpt 4o but all same blurry issue

[velvet rampart]

Ok, so effect you are seeing is equally present on whole service. What style and/or technique you are requesting? What about definition? Do you specify definition?

And my recommendation is to use/add extra commands here. E.g. ”Render properly with the best quality you have. Take your time to produce quality images.” and/or ”Concentrate on rendering faces correctly and with possible quality.” This way the AI concentrates on generating the faces well.

[dark terrace]

Do we know when they will release outpainting for Dall-E 3?

[dark terrace]

It sucks cus I can't access Dall-E 2 for outpainting either 😭

[agile peak]

I keep failing try to create image of robot shooting laser beams lol it always came out of nowhere.

A stunning oil painting in chiaroscuro, that showcases a gigantic iron giant, emblazoned with a blazing red star, standing amidst a dark, foreboding mist. The giant's massive feet are turbulently moving, as if it's about to stomp the ground.*** A piercing laser beam erupts from its eyes,*** obliterating the nearby houses, causing the entire scene to tremble. In the foreground, petrified farmers are fleeing in terror towards the viewer, their faces contorted with fear.

[plucky hare]

You might try rearranging your prompt such that your most desired/important details are written first. DALL·E can't include an indefinite quantity of specified details, so it can start to fail on certain details past a certain point.

A stunning oil painting in chiaroscuro that showcases a gigantic iron giant with a piercing laser beam erupting from its eyes, obliterating nearby structures and causing the entire scene to tremble. The iron giant is emblazoned with a blazing red star, and it stands amidst a dark, foreboding mist. The giant's massive feet are turbulently moving, as if it's about to stomp the ground.

[agile peak]

Thanks for the tips and head up for correction. This is amazing result

That "rearranging" prompt really does wonder. Silly me not asking this with GPT.

[dim cradle]

incredible

[hearty ether]

Realistic dnd fantasy illustration on white background

The worst offenders are the eyes

[velvet rampart]

Ok. Then you might consider asking the AI to ”Render eyes anatomically and biologically correct”. This puts emphasis on the items that come wrong or use the inpainting feature to fix the eyes if the rest of the picture is otherwise good.

[frosty furnace]

My Dalle is broken! No matter what kind of robot I prompt (I'm trying for bipedal), I get these four legged crab like robots. I literally can't fix this. I open new chats, I try unique prompts...I keep getting this kind of robot:

I assume it's because it's put a past prompt of mine in to memory? I asked for this kind of robot days ago, and now it's stuck on it.

[dim cradle]

check your memory and delete it i guess

[frosty furnace]

Adding a prompt style to my memory is not a good thing.

Hahah. First time managing memory--some of the stuff it puts in here is hilarious.

Yup. It was in memory. lol

Definitely go through and manage this once in a while ya'll.

[thick smelt]

finally

i have a gpt dall-e photo
a-... puppy

[empty kelp]

a five legged crab robot

it didn’t want to make a three legged crab robot, but it seems to be happy with five and six legged crab robots

Maybe all robots are crab robots in the future, and humans just aren’t privy to it yet

[velvet rampart]

There is a saying that everything evolves into a crab eventually or want to evolve into a crab.

We do have 5-pointed bodyplan in sea stars.

[vague hearth]

Is there a way to have gpt create images with correct text consistently?
Thing is sometimes it does an awesome job and sometimes not, with literally the same prompt.

[velvet rampart]

Yes. See my comment earlier where I give the command.

[dawn phoenix]

Hello Everyone
Is it possible to programmatically generate multiple images with consistent characters from a plot for a film?

[velvet rampart]

Not yet. When 4o image image generator is published, then this is possible.

[vapid elk]

programmatically generate
yes, https://platform.openai.com/docs/guides/images

consistent characters
maybe

[thick smelt]

That looks cool

[dawn phoenix]

Hi @vapid elk
Thank you
https://community.openai.com/t/consistent-image-generation-for-story-using-dalle/612276/2
please visit this link, you can see sample case
I've read that we should use gen_id to generate consistent images, but how do we do it programmatically?

[vapid elk]

gen_id is something that currently can only be used on chatgpt, it is sadly not a feature of the API

[dawn phoenix]

Can we dm?

[vapid elk]

sure, I'll be happy to help =)

[turbid wren]

Some simple stuff I made with Dall-E, wish more people showcased stuff like this

[twin rose]

AI Hands. O.O

[agile peak]

Three legged is probably the difficult one to generate.

[graceful spade]

How can I create images with this Tarot style? Without it looking like a card

[plucky hare]

A bit detail-heavy maybe, but how's this for a start?A chimera illustration with intricate linework, symbolic imagery, and a blend of medieval, Renaissance, and mystical elements. The style is detailed, and the chimera is expressive. Rich, vibrant colors, with a slightly archaic and otherworldly atmosphere. The composition is balanced and harmonious, emphasizing both realism and allegory in its depiction of the chimera.

[fallow knot]

Bold and inspirational with a professional and organic style, using shades of navy, royal blue, and cool whites for a classic, powerful look.

[quartz vale]

you can't use Dall-E here, but you can in #image-bot ! You'll use the /draw command in that channel

[prisma flume]

Hello! I have tried the command in the dalle-bot in generating text based image. I think the text is not clear. Maybe I am missing something in the command? The sample generation is in the latest of dalle-bot

[velvet rampart]

The promblem is that the dallebot rewrites the prompt and this command is very sensitive to alterations. That is the why I recommended to use gpt4.

This is done in chatgpt and using gpt4.

[prisma flume]

does it also work on paragraph text?

[velvet rampart]

The longer the text, the more prone to errors. I recommend to start with single words and proceed to longer texts until you find the balance.

[prisma flume]

Got it. Thank you!

[vital dune]

Trying to get Dall-E to make a single electron going through a CT scanner, but the output is no where close. Any suggestions on the text input?

[vapid elk]

share the prompt and your results

that helps to see what you may have been doing wrong

[vital dune]

A single electron going through a CT scanner examination, sketchbook style

[twin rose]

Hands. Solved.

[hearty ether]

I don’t know what they did

But faces have still been horrible

Like why is it so blurry

[agile peak]

"GPT-4o has a high school level intelligence."
"GPT-5 has a Phd level intelligence."

[tawny portal]

Where is @Dystopia? I haven't seen her here for days.

[sterile magnet]

Same

It's a man by the way

[glossy scroll]

I love that one.

Nothing like a "Hehehe" from a Furby

[dawn phoenix]

Hello
Can anyone help me create a storyboard for my current movie?
I am currently using OpenAI's DALL-E model to generate them, but they are completely inconsistent.
Is it possible to generate consistent storyboards programmatically?

[tawny portal]

Then it is a typo 🙂 In my language Dystopia (as noun) has feminine gender. Dys is alright - he is good person. I was used to seeing him here helping new people.

[hearty ether]

Quality has just been so much worse lately

like especially for faces

The poor quality has been consistently more blurry and different in style compared to previous outputs

used to be able to get something vivid and detailed like this

Now everything is far more cheap looking and blurry with flat textures

(Look at the faces)

It’s a lot faster now but the results are also worse

So I’m worried they are sacrificing the experience of the plus users for the free ones

Change for me was on June 12th

When quality dropped

[hearty ether]

I mean they made it significantly bad this time

Like completely blurry on my images

It’s not just bad proportions

[soft prairie]

I kinda feel these days , dall-e with gpt 4o is a bit off.
On gpt 4, i was able to give him script parts and make generate with dall-e a number of images with a wanted style. i was impressed how it was able to conceptualize the script and make dall-e generate intresting images with intresting point of views and scenary.

From gpt 4o i immediatly noticed that is less various and it keeps blocking on "understanding" bugs , it keeps generating mostly same compositions and does not conceptualize images in that intresting way it used to do ...

i'm sure that's my mistake somewhere but don't know where. i tried explaining what i need in many different ways, tried to make text that i can copy and paste with a "to do list" of what i need it to do but it seems more "stupid" then before when it is supposed to be better...

[velvet rampart]

Problem might stem from that 4o’s own buildin image generator is not yet online and it does not know how to use dall-e3 effectively like 4 does. But on the otherside, sometimes the 4o does make excellent images when it does understand what the user wants. Often when the user tells 4o to form the prompt.

[marble loom]

Did the get rid of Dalle

[keen monolith]

it looks like the dall-e 3 gpt is no longer listed.

[velvet rampart]

My guess is that they will announce this evening the 4o image generator and will roll it out.

[glossy steppe]

Man where’s the dall-e?

[sterile magnet]

He is fine, don't worry

And also it's okay for the typo 👍

[crystal walrus]

Is it me or is it really slow today GPT in general.

[marble loom]

Yeah I wonder why.

[vagrant zodiac]

I've never actually had to use the dalle gpt?

It's always just generated images when I ask using regular chat

[dim cradle]

if you use dalle it makes two images usually not just one

probably just some glitch though

[keen monolith]

It is back 😃

[candid thicket]

Anyone else getting a completely unusable dall-e rn?

[plucky hare]

Seems to be the word "woody". I tried taking its description and got the same problem, then just edited and resubmitted without the "y" and it made a pic.

[candid thicket]

Still didn't work

[plucky hare]

A large apartment room with a modern look and sleek wooden elements designed for a creative chef. The room includes smart storage solutions such as built-in shelves, hidden compartments, and multi-functional furniture. There is a spacious kitchen area with modern appliances, a creative art corner with organized supplies, and a stylish desk for work. The room also has neatly organized areas for seasonal wear and personal items, balancing modern and wood sleek styles.

If you're trying in the same chat, maybe try in a new chat. Once a flag goes off in a chat, it can continue to pop up since ChatGPT looks at more than just one message at a time for context when writing image prompts for you.

[candid thicket]

I tried it in multiple chats

But thank you for making it for me

[velvet rampart]

Memory promblem?

[hearty ether]

The details seems to have gotten more blurry

Like the drawer handles

Got a somewhat decent face as a fluke

But it’s still far more blurry and lower quality than before

[hearty ether]

Like the texture of the hair and face compared with before

[quiet brook]

is it me or does the dalle quality on chatgpt looks deteriorated over the past days, the steps aren’t as much as when it first came out

sometimes it’s picture will be extra blurry

[tropic root]

Anybody running into more safety system rejections lately? This prompt, and others like it, is being rejected by the safety system and I'm not sure why.

[tropic root]

In the central laboratory of the Cairo Research Institute, a sleek and futuristic building with a glass and steel exterior, the scene is lit by bright LED lights and holographic displays. Dr. Amelia Chen, a human female of average height with light olive skin and long black hair tied back in a ponytail, wears comfortable professional attire suitable for lab and fieldwork, and stylish functional glasses. She stands alongside Dr. Mohammed Farooq, a male human with a distinguished salt-and-pepper beard, medium brown skin, short slightly curly hair, and dark eyes. He is dressed in attire reflecting a blend of traditional and modern academic styles, and he carries notebooks and pens. Near them is Sarah Winters, a female human with a commanding presence, fair skin, neat bobbed hair, and piercing green eyes. She wears tailored blazers and slacks. Agent Marcus Cole, a lean and athletic male human, with light tan skin, short dark brown hair, and sharp eyes, dressed in a suit and tie, occasionally adjusting his tie, completes the group. The room is filled with the hum of advanced machinery as the team gathers around the console. A holographic figure with jackal-headed features materializes above the central platform. The setting includes state-of-the-art laboratories with advanced AI interfaces, holographic projectors, and secure climate-controlled artifact storage areas, with large windows offering views of Cairo's skyline and distant pyramids.

ILLUSTRATION STYLE: A digital illustration style blending futuristic holographic elements with ancient Egyptian iconography, using a rich color palette of deep blues, golds, and earthy reds, featuring intricate line work for hieroglyphics and sleek, luminous surfaces for modern technology.

RULES:
Do not use text or lettering in the illustration.
Adhere strictly to the visual character and setting descriptions.
Adhere strictly to the ILLUSTRATION STYLE.

[tropic root]

Response: An error occurred: Your request was rejected as a result of our safety system. Image descriptions generated from your prompt may contain text that is not allowed by our safety system. If you believe this was done in error, your request may succeed if retried, or by adjusting your prompt.

[velvet rampart]

Hmm… In the first read, the prompt is really complex. Many characters with very detailed descriptions. To figure out, what word causes the blocking, I suggest to feed your prompt in pieces. Once you find the part that causes this, start removing word by word. Often this is caused by a single word. For example beard style van d-ke is blocked due to ”d-ke” or king charles spaniel due to king charles or charles (reference to real world person by filters). The filters cannot differenciete the connection but they work only by word lists.

A source might be the description of the illustration style. The word ”iconography” pops to my eyes. Hmm… Let me run a query.

After quering the AI, I found a possible source. The person names. Without a prior discussion, the AI might think the person names to be of real ones. To avoid this, I suggest a short introduction to give some background and to make clear that these are fictional. Often, after I have given some background, the AI does what I want to do. You have convince it.

[foggy sonnet]

Can make me a list with name and with their namber end with 86

[hearty ether]

Significantly

You can look at my results

And if you have a dalle chat from before June 10th ish you can ask it to generate again with the same gen id and prompt

And it’s clear that the quality is significantly worse for the same image

Like terrible

One of the telltale signs is blurry and deformed eyes

[dim cradle]

the API generated Apple logo twice. not sure if this was a mistake with the copyright policies in place

[sterile tundra]

xd

[dim cradle]

lol

[rigid bough]

no worries doing an Apple logo

Draw me a stylised image of the Apple Logo glowing

Now draw a similar image of the Microsoft logo

Draw me a similar image of the Disney logo

[velvet rampart]

The quality of textures has gone down. The fur is not so realistic.

Here is an image from 22 May. Notice the texture of the fur. More realistic.

[rigid bough]

I must not have a good eye for this stuff because both images look awesome to me

[velvet rampart]

If you look carefully at the otter’s fur, you notice that it has this weird rippling from the image generation. The squirrel image has none of this rippling.

[rigid bough]

what about this?

[velvet rampart]

Look the area I circled. This contains a bit of the rippling effect I am talking about. However, this is of better quality than my otter image.

[rigid bough]

I wonder why that is happening?

[velvet rampart]

Resources? When I was generating the otter images, the other boot was often missing.

[quiet brook]

no doubt

the quality has gone down

[velvet rampart]

It seems like the process is not done in the allotted time and is incomplete.

Bug report? Yes! I will do bug report. Expectation vs. reality.

[rigid bough]

here is another

[quiet brook]

I gonna try it on myself what’s the prompt for this one

[rigid bough]

Draw an image of a cute furry squirrel

"A highly realistic image of a cute furry squirrel with detailed, lifelike fur. The squirrel has large, expressive eyes and a bushy tail, standing on its hind legs and holding a small acorn in its tiny paws. The setting is a lush, green forest with sunlight filtering through the trees, casting dappled light on the forest floor. The squirrel's fur is a mix of soft browns and grays, with each strand of fur clearly visible, giving it a very natural and realistic appearance. The overall scene should evoke a sense of natural beauty and charm."```

[quiet brook]

the fur seems more unnatural over here

looks pretty

but it did another image that’s cartoonish