images-discussions

[late blade]

And this is from the chatgpt plus answer

so we got gpt-4o with 3 different knowledge standpoints and GPT4 normal is December 2023

[plucky hare]

Hmm probably best if we take this to #gpt-models if you're interested, that's my bad!

[late blade]

nah, it's just an observation I had, won't delve into it further

my point was to make sure you know what you are working with, that also applies to dall-e users

[grizzled iris]

Happy Friday Y'all ❤️

I got access to it just now!

I am no longer an angry chipmunk haha 😆

[dim cradle]

can it do anythng the web browser cant?

[grizzled iris]

Gonna test out the tech now and let y'all know x

This is quite cool, I can do that in the launcher by pressing Option+Space

[dim cradle]

nice

i think i found a good way to make dalles more photo quality. start you prompt with "1984 cinematic image (or whatever year), 35mm film quality..." and you get more actual photo quality not cgi looking photo peoples, my two cents. 🪙 🪙

[grizzled iris]

Don't get me wrong though, there's many bugs haha as expected so will be doing a lot of feedbacking this weekend lol, hopefully the community does the same so it works for Monday x

Like this is what I get when I click on the generated images haha

The web version for now is good!

[grizzled iris]

They have it @dim cradle !

[late blade]

gimme!!

how do you download it?

app store?

[grizzled iris]

Go to the Web version

[late blade]

and then?

[grizzled iris]

and see if you have this

[late blade]

aw, does that download link send you to the app store?

[grizzled iris]

You have it?

I have the download link you can try?

[late blade]

no, I don't have the notification

sure

[grizzled iris]

Try that - it should take you to this page

[grizzled iris]

Yes I had that yesterday!

Or the day before...

[late blade]

oh well, I'll get it at a later point

[grizzled iris]

The I spent hours on the voice call with OpenAI training the model

and got the plus last night

for work demo this weekend and logged in just now and got it!

[late blade]

nice!

[grizzled iris]

Don't worry tho mate, I have to say the Web is better now

The Desktop app requires a lot of work haha

[late blade]

hehe

[grizzled iris]

Oh wait

I fixed it!

[late blade]

nice

[grizzled iris]

Well, half... lol

I don't have access to memory though 😦

https://openai.com/index/memory-and-new-controls-for-chatgpt/

[late blade]

I just want to see if it can handle longer chats

[grizzled iris]

It'll be interesting to test out the memory for DALLE Image creations

I do get this though now everytime I'm on Web version

And as you've probably guessed it the "Open" icon does nothing haha

Corretion, I click on it and it dissapears haha

[late blade]

I'm interest because I want to handle longer chats, even without memory, it's just annoying to start a new chat because the browser can't handle it

[dim cradle]

what browser do you use

[late blade]

general purpose I use firefox, safari on mac, edge on windows, brave on secure stuff

[dim cradle]

you notice slow down on all of them if you use gpt?

[late blade]

yes

[dim cradle]

yes i do too it seem to take longer with 4o but it still is occuring to me

[grizzled iris]

Interesting

Do you have a DALLE prompt?

@late blade

[late blade]

for?

[grizzled iris]

I can test out now with a large longer chat

Oh I see

[grizzled iris]

What you mean now

Yeah that does happen a lot on web especially in Gemini

It always cuts off after a while

[late blade]

yeah, hence my interest on the app, maybe there's a better implementation on the app

[grizzled iris]

Hopefully you get it soon so you can test it out and improve it ❤️‍🩹

[late blade]

but if you want a prompt:

A destitute gunslinger traveling through the metropolitan streets of rio de janeiro in search of a wanted poster. <insert your art style here>

that one turned awesome, sponatious attack of creativity!

[grizzled iris]

That was a good prompt @late blade

[grizzled iris]

Ok ok, desktop app is stabilising!

[late blade]

just made a gallery https://discord.com/channels/974519864045756446/1241160134026334348

[grizzled iris]

It added some extra bits for some reason haha

A destitute gunslinger traveling through the metropolitan streets of Rio de Janeiro in search of a wanted poster, depicted in a cyberpunk art style. The scene is futuristic with neon lights, high-tech billboards, and a gritty urban atmosphere. The gunslinger, wearing tattered clothing and a wide-brimmed hat, walks through crowded streets filled with diverse people, advanced technology, and towering skyscrapers. In the background, iconic Rio landmarks blend with cyberpunk elements, creating a unique and vibrant cityscape. The mood is tense and dramatic, with a focus on the gunslinger’s determined expression.

[grizzled iris]

I'll add them there also now ❤️

[late blade]

Sure thing

but watch out, there's danger ahead

and they are not alone

I see so much potential for story telling

[grizzled iris]

Good night @late blade

[late blade]

c ya around

[grizzled iris]

❤️

[calm oasis]

(Portuguese brazil 💀)

[late blade]

very good, not the place to share it, you should share it in #ai-discussions or #off-topic , here the channel is for DALL-E

[calm oasis]

Oh sorry

I really didn't know

[onyx ridge]

yeah they'll mod it too

it's good to learn this way trust me

that is all

[austere nexus]

I understand why webp is useful, but its annoying when its not adopted industry wide. ChatGPT (dalle) way too often makes your images sideways. And Windows 11 won't let you right-click -> Rotate Image. It will for all formats that are not webp.

I would appreciate a .png option

or you know, maybe always correctly verticle images

[late blade]

ask for gpt to provide you a png download

[austere nexus]

That might be really good if i could put it in the custom instructions, but I dont think dalle has access to that

[agile peak]

Different angles and styles.

[late blade]

the link always provides me a png

[dim cradle]

cool guy

[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!

[fiery salmon]

How's the challenge coming along?

[sage ferry]

Hello,

Does anyone know how to generate low quality photos? like it was taken on a samsung galaxy s3... or even just a phone camera. the only photos ive ever been able to generate are SUPER PHOTO REALISTIC CINEMATIC etc etc. or i could use tell it to use the style of a graphic or a drawing. It kind of does this when i tell it to use the style of cctv footage. if i tell it to use phone camera quality, it generates an image OF THAT Phone. not what im looking for. I'm looking for traditional hood meme quality photos. any ideas?

[agile peak]

Impressive. Manga style is now doing a very well done job, I feel drastically get better generated results cannot wait for the coming weeks.

[dim cradle]

awesome

[grizzled iris]

I need to try it on the desktop version to see if I can win the challenge this time round 😂 - Will keep you posted!

[agile peak]

Clay and PS2 graphics

[velvet rampart]

Works like a charm. Tested and got jpeg, tiff, png. Also bigger and smaller sized. It only works in normal chat, not dall-e customgpt as it need data analytics function to work. I used command ”Provide my a download link for the picture. Various formats and sizes.”

[late blade]

I think you can use the @ in dall-e and use ADA to do the download, haven't checked myself

[velvet rampart]

Yes, you are right! I forgot that.

[cyan jungle]

did anyone get access to the newer gpt-4o image generator model?

[velvet rampart]

Will be rolled out in coming weeks. Not yet.

[uneven parrot]

(S)he who understands understands ):

[cyan jungle]

apparently some people said they got access

[velvet rampart]

Well, maybe testers. Anyway, the point is that it will be rolled out at some point.

[uneven parrot]

Can DALL-E create a grid of emojis, objects, etc?

[late blade]

yes

[uneven parrot]

Please look at the picture I shared. It is not 5x5 ):

[late blade]

but the emojis inside can be inaccurate, so you have to do iterations of what you want

problem with this is dall-e and numbers are still inaccurate, so your grid will not be 5x5 most of the time

[uneven parrot]

How should I prompt? I wrote five rows and five columns. But it was not correct!

[velvet rampart]

According to the blog, the new model should be more accurate with counting.

But not yet.

[late blade]

like I said, if you are expecting 5x5 grid, it won't happen most of the time, you just have to find other strategies to get a grid that you can work with after it's being generated

[velvet rampart]

One strategy suggestion is that you annotate a produced picture and feed it back. The point is that the model cannot see the pictures it produces,

[dim cradle]

and i dont think anyone has access to it yet. they say microsoft has some event this week, i think there is chance perhaps they announce copilot pro get 4o image maker first like they had dalle3 first (was then Bing is King image maker of course) and yes if the blog is correct, the counting thing will be correct with 4o 🙏

[uneven parrot]

I tried it with 4o. I got similar wrong results. @late blade are right. GPTs can not count right now ):

[dim cradle]

4o is still running on dalle.

it is a little confuse to some people, but 4o does not have a lot of bells and whistle it will have soon, main thing people ask for of course is the voice but it also will have a new image maker

[uneven parrot]

I see. thank you.

[late blade]

you can try with any generative text tool, the image generation still happens with DALL-E 3 until OAI enables the image generation with gpt-4o directly

[fiery salmon]

Thank you my friend, this is the hardest challenge a group has ever had

[viscid blaze]

Anyone been able to generate spinning 3D Image like the demo?

*image sequence

[velvet rampart]

Not yet. It will be part of the new picture generation of 4o.

Just tried and it told me no (much abbreviated answer.

[late blade]

yes, all the time, but not with ai, just using some autodesk tools for that

[short valley]

Hi!
I have a question related to the dalle bot. How come it revises the prompt in automatic?

[dim cradle]

you can tell it not to do that

just say "make the prompt this exactly:..." or whatever

[chilly wyvern]

It revises by default unless you specify the bot to use your prompt verbatim.

[fiery salmon]

Welcome to Mad World Express, if you look to your left or right yo usee your impending doom

[short valley]

Ah ok! Thank you

[scarlet ferry]

Idk why but when i use dalle3 to generate a character it always generates one in portrait mode. Like only face and maybe a bit of upper body even if i ask for a full body action shot or something. Any advice?

[velvet rampart]

My theory is after trying to interrogate the model, it is just easier to generate just upperbody than whole body. Resource wise. Also, we have consider what kind of generation bias has been introduced during training and source material.

[dapper plank]

Generate an image of:

Oil painting of a lone figure standing on a cobblestone street, holding an umbrella under a streetlamp. The golden light from the lamp casts a warm glow on the figure, creating sharp shadows on the wet ground. The rest of the street fades into obscurity, highlighting the isolation and introspection of the moment.

[late blade]

gotta love how copilot passes prompt to dall-e

user_input

and this is how user_input looks like

[final jungle]

4o felt great when it came out a few days ago, but im running into an issue i call 'white people'

As in every prompt with a group of people is white people, most of them look to be 40 with a beard. Like its trying to spam 40 skinny copies of myself...

[spare turtle]

generate an image of balls

dangit

[dry crater]

anyone knows why it keeps messing up the eyes lately?

[late blade]

I'm having the same problem with eyes, it's so annoying, the image is perfect but the eyes screw the whole concept

[dry crater]

It's unnaceptable. I have free software that doesn't make so many mistakes:

https://cdn.discordapp.com/attachments/1221772968657424394/1241522550740226059/image_2024-05-19_014715637.png?ex=664a817a&is=66492ffa&hm=666c11817dae642b8521470b1e63a40f752d185e917e770c0b09f950be9a0af6&

Every single pair of eyes messed up in some way.

[late blade]

oh, close up eyes I get pretty good images, I get eyes disotrted when there's lots of people involved in the image

[dim cradle]

its still dalle. 4o image maker isnt turned on yet

speaking of... Microsoft has a big event on Monday... some theories say that they may announce some opeanai exclusive like the macapp... you know, Bing (is King) had Dalle3 first... so maybe they will announce Copilot pro will be to receive 4o plus 4oimage maker for a week or two early 🙏

it would make sense why openai did not announce any exclusive thing like the macapp for ms last week, they would not want to steal they thunder for Monday's event 🤔

[earnest flame]

Sometimes you can improve the results by asking it to describe the eyes more. However, this has the side effect of bringing the face closer to the viewer.

A 1990s-style action cartoon still depicting a woman wearing a bandana and wielding a sword, with stylistic explosions in the background. The woman has piercing blue eyes, clearly defined with bold outlines, typical of 90s cartoon art styles, which contribute to her intense and determined expression. The backdrop features explosions in fiery reds, oranges, and bright yellows, contrasting sharply with her darker attire. The scene is wide, capturing the dynamic energy and action typical of 90s cartoons.

A 1990s-style action cartoon still depicting a woman wearing a bandana and wielding a sword, with stylistic explosions in the background. The backdrop features explosions in fiery reds, oranges, and bright yellows, contrasting sharply with her darker attire. The scene is wide, capturing the dynamic energy and action typical of 90s cartoons.

[dim cradle]

nice

[agile peak]

Have dalle 3 improved with photorealistic or photograph yet? seen one post on reddit that looks so much better than previous generations that tend to look like "doll face". Or is it just gpt-4o better at prompting for dalle.

[agile peak]

Look like it could be a slow roll out? I saw one photograph image of a woman on reddit, the level of detail is mind blowing. Still wondering.

[viral moth]

created using gpt-4o yes

[sterile magnet]

I can't do thaaat 😭
It says please write a longer prompt

[dim cradle]

no, we'll know when it can reproduce text nearly perfect. i think they just tweak the dalle model over time and maybe 4o base model somehow improve it too idk. but i dont think anyone has access yet just like the voice thing

people also dont use enough photograph keyword i think, 35mm film style, cinematic image, film grain, diffuse lighting, natural lighting, etc... those will help a lot

my two cents 🪙 🪙

[storm spoke]

hi

[empty kelp]

I've done a DALL-E 3 API test a few times in the past with this prompt:
a hyperrealistic wide photo of a woman studying at a desk. a library is in the background.
Since December the quality has improved some (possibly because the prompts are getting a little more processing time). I've noticed that hair on the characters' heads looks a lot better. There are occasional anomalies with eyebrows, eyelashes, glasses, and fingers -- but noticeably less than there used to be

[dim cradle]

35mm film style will produce much better result thatn hyperrealistic

toss in cinematic image even add a year before it and it will adjust the look like 1987 cinematic image of 2001 cineamtic image etc

[empty kelp]

i'll try that

[sour dew]

Why DALLE do this where is graphic ,is empty?

[dim cradle]

tell it to make the image

some times it will just do a prompt

or what is the blue link?

[rancid reef]

guys is there any advancements coming to dalle 3 anytime soon?

[dim cradle]

there is a whole new 4o model we just have to wait to get turn on

[rancid reef]

ah for real? any specs on what it will imrpove or will be just slightly improvements as in 4o?

[dim cradle]

scroll down and click through the exploration of capabalities: https://openai.com/index/hello-gpt-4o/

[empty kelp]

35 mm didn't look any different than photorealistic, so i asked it to create these two images in an "iPhone Pro style". I'm not sure if it's more realistic, but the characters do look smarter and more alert

[dim cradle]

35mm film style or 35mm image quality, cinematic image, natural lighting, film grain, photo type words always turn out better for me but yes it look about the same, a little more like cgi

i feel some time they build that into dalle3 to make fake photos less possible but just a theory of course haha

[dim cradle]

heres mine using my prompt idea i told you

look much more natural and real imo

[empty kelp]

saying "natural lighting" does make a huge difference. i did several with that and they all look more realistic

[empty kelp]

"revised_prompt": "A cinematic 35mm style photograph bathed in natural lighting and authentic color. Santa, a jolly figure rich in red and white attire, is sitting together with two athletic elves who are females, one of Hispanic descent and the other African-American descent, at a desk. They are all engaged in studying, radiating positivity with smiles that reveal their teeth. An impressively stocked library acts as an intriguing backdrop for the scene."

THE_PROMPT = '''A cinematic 35mm style photo with natural lighting and color. Santa and two athletic and diverse female elves are sitting together at a desk studying. They are all smiling with their teeth showing. A library is in the background.'''
response = client.images.generate(
model="dall-e-3",
prompt=THE_PROMPT, size="1792x1024",
quality="hd", style="vivid", n=1

We're reaching new heights with these prompts. It couldn't be any more realistic

[dim cradle]

nice. looks much more real

[dim cradle]

nice last 4 look most real to me

Hawaiinz do you ever make photos without santa? haha

[empty kelp]

Those are really good. We should try the scenes "natural texture" also -- along with natural color and lighting. Color and texture are the two things that we figured out can be copied to make things look identical

[hybrid fox]

am I the only one who hates the AI's default style

[late blade]

define default style

[hybrid fox]

idk like this

"create a man pondering next to his dog"

[late blade]

what is the prompt?

[hybrid fox]

and then it fills in all the gaps abt style

im not a fan

it always looks so corporate

[hybrid fox]

did u tell it to make it look poloroid?

[late blade]

y

[hybrid fox]

cuz it looks kinda like a poloroid camera pic

I like it

[late blade]

the default you mention is called lazy prompting (not because you are lazy, but because you leave too much open for the model)

[hybrid fox]

what was the prompt for this one I rlly like how u did this one

Does this look lazy or does this look articulated

[late blade]

just choose a style or visual aids to the prompt

[late blade]

a stylized image of a man pondering next to his dog

[hybrid fox]

right but what was the actual prompt the thing used for that

[hybrid fox]

im curious how to get that countryside painting type look

[late blade]

I just gave you the actual prompt...

[hybrid fox]

are u on dalle

[late blade]

yes

[hybrid fox]

u can click on the image and then click info and get the actual prompt used by the system

[empty kelp]

saying "natural lighting, color, and texture" makes their teeth and eyes look correct

[late blade]

yes i can and it's the same...

[hybrid fox]

oh

then how did it pick the style?

randomness?

[late blade]

yes

[hybrid fox]

the asian elf's eyes still look unnatural

[late blade]

lazy prompting

as to why the guy with the trumpet appeared on this one... no clue..

[empty kelp]

it's actually drawing the teeth, eyes, and eyebrows correctly when I say "natural texture". they don't look painted on

[empty kelp]

we definitely need Sora now

[late blade]

lol

[empty kelp]

A cinematic 35mm style closeup photo with natural lighting, color, and texture. An athletic and diverse female elf is smiling at us with her teeth showing. Her face, eyes, hair, and ears have incredible detail.

[empty kelp]

they must have upgraded the DALL-E 3 API

[late blade]

eyes are horrible

[hybrid fox]

@sick flax #daily-theme message

What was your approach for this?

I love it

[sick flax]

It was Paul Cezanne with Nihonga influences. Actual prompt was: "Abstract minimalist post-Impressionist oil painting on canvas inspired by Paul Cezanne, featuring Nigonga influences. Depict a rustic wooden table with hand-spun linen cloth, covered with various European cheeses like Austrian Bergkäse, Belgian Fromage de Herve, Bulgarian Tcherni Vit, English Blue Vinny, Bavarian blue, Greek feta, and French Bree. Include a charcuterie board, knife, wine bottles and glasses. The background shows a 19th century French country kitchen in soft focus (bokeh effect), emphasizing a warm, inviting atmosphere with muted colors and expressive brushwork."

[hybrid fox]

Oh wow I hadnt heard of paul cezanne but I just looked up his art and its beautiful. Thank you!

How is it letting you say "inspired by ___" Yesterday when I was trying to make something inspired by Stanley Donwoods work (album artist for radiohead) it said it was a copyright violation. Is it because he's alive?

Ahh I see, dang Donwoods art is rlly cool I wanted to play with it

Thats what I ended up doing but it didnt come out all that great. This was the best I got after like 50+ generations

[hybrid fox]

Wow

I like that a lot

[hybrid fox]

bro is better at this than me lol

So did you like put all that and say use this style

and then made a prompt

like how do I make it a base

Kk let me try that

which piece of his art did u use

as an example

[empty kelp]

the API seems to be drawing vertical portraits correctly every time. it's not flipping them sideways anymore

[sick flax]

Anything after 1912(?) usually trigger the content warning. Though before the official cut off, I can never seem to be able to get away with Claude Monet.

The trigger seems more sensitive with anything of commercial value and Western (as in American or European) than low commercial value and non-western.

[agile peak]

Yeah, I'm on the same page. I reckon it'll be a slow rollout for certain users at the moment, and folks might not have caught of it yet. Maybe it's tied to getting Sora ready? I figure they'll also be enhancing Dalle to make it out even more realistic generated images, especially since the Sora demo showed such a significant leap in realism.

[merry vault]

i would like to take a apluad the openAI devs for making such a good anti piracy system, like it prevents me from using copyrighted stuff as "inspiration" often. Kudos.

whats that json?

oh wait so is this like a mini version of an assistant but in gpt?

nice, thats actually sick il keep it tabbed

[rancid reef]

guys is there any External Chatbot creator that uses GPT API that allows you to offer image creation via dalle? I see chatbase and other similar ones but im unaware if is just not possible to generate images via API with dalle?

[wind onyx]

when will we be able to generate more than 1 image with dalle

[late blade]

just do it yourself, all you need to do is a curl api request and response, no need for middle person

[wind onyx]

i neeeed n=10

[late blade]

1 + 1 + 1 + 1 + 1 + 1 + 1 + 1 + 1 + 1 = 10

[wind onyx]

i knowwww but still id love 10 at once with same cost of one

[late blade]

why? if you pass a wrong or defect prompt you get 10 errors

[wind onyx]

simultaneously doing it would yield the same result, no?

unless you explicity said "STOP!" inbetween generations

[late blade]

you can do concurrent requests

[wind onyx]

that's what im talking about

but youd still have to explicitly exit if its generating the wrong images from your prompt

[late blade]

once a request is sent it's sent to the API

[wind onyx]

ye but 10 times

youd need to send it all at the same time for it to finish at the same time

is there any mention of a higher number of images per request for dalle 3

[rancid reef]

yeah but i mean, i need a platform, a UI interface where i can send people to since i want to sell a prompt for Dalle

but not only the prompt itself but allow users to ask the "Chat" for images

and for it to create them based on my prompt, as if it was a G{T

GPT

[late blade]

the luxury of UI, back in the day, we had to work with text, punch cards and maybe a keyboard, no mouse

spoiled people

[rancid reef]

i mean, i have no clue how to place a prompt behind the chat for it to work based on that promtp lol

can you point me in the right direction?

[late blade]

north, don't stop until you find santa

or you find @empty kelp

he's prob enjoying surfing with santa and his elfs

[tough girder]

In #daily-theme , do you just post something you made in DALL-E or do you need to use the Discord bot?

[late blade]

on that channel there's always the topic of the day, for example today is 🤔 puzzle - logic, problem-solving, a satisfying challenge for the mind.

and you can use any implementation of DALL-E for the gens, api, copilot, oai or the #image-bot channel

[tough girder]

Thanks ^^v

[dim cradle]

is that jar jar with a lightsaber in your pfp? 😂

[tough girder]

Absolutely. The more you watch the prequels, the more that Darth JarJar makes sense, right?

[dim cradle]

i really like Jar Jar haha. I wish he was in 2 and 3 more. i can relate to being a clutz haha

[ornate zenith]

Hello i have question can we sell ai generated art

[late blade]

what you do with generated images is your own to decide, they are yours to keep, but you cannot copyright or trademark them and anyone is able to copy whatever you sell or market

[vapid elk]

you can use the content you generate with OpenAI services commercially with no problems

just have in mind that there are legal precedents that prevent fully AI generated contents from being copyrighted

[empty kelp]

The DALL-E 3 edit feature in the ChatGPT web interface seems to be able to remove pretty much any anomaly from the old images (i think it only works for ones that remained on the website and still have the diffusion model data).

This image (which has a style created by Dys Topia) had extra hands, feet, and weirdness (the areas circled in red), but I was able to fix it with a few quick edits.

This is the image edited with the DALL-E 3 web editor

[late blade]

love the outcome

[empty kelp]

This is pretty amazing because when an image needs a lot of diffusion iterations and there isn't enough processing time -- all sorts of anomalies and distorted body parts appear, and we can just go back and fix them now

I want to test to see if we can have it do more iterations on the images where characters were balancing -- so it makes them balance correctly without missing heads and extra body parts

[late blade]

btw, I've since then been able to simplify the art style request and make it more adaptable to other scenarios, some examples can be seen in this gallery https://discord.com/channels/974519864045756446/1240746164463079479 and in this other gallery https://discord.com/channels/974519864045756446/1241160134026334348

[empty kelp]

I think now we'll be able to create scenes with low detail (and less anomalies), and then tell it to add detail afterward

[late blade]

yeah, that would be awesome

[late blade]

the thing about these edits is that the image is not complex, I had trouble editing on complex images

[empty kelp]

edit: the hair in these areas [hair selected] should be more realistic

This replaced the hair and kept the characters exactly the same. The only other thing different is that DALL-E 3's post processor seems to have done an automatic brightness adjustment to the entire image when colors change

[late blade]

see my reply before your post lol

[empty kelp]

yeah, i noticed that when you do something that doesn't match the style of the image it says:
"I encountered issues generating the image"
Like you can't tell it to give these characters natural color and texture because it clashes with what was given in the prompt.

[late blade]

look at this bug I reported #1231543295788716083 message

the bug made dall-e enter an endless loop to try to fix what was already fixed

[empty kelp]

If you ask it for the revised prompt for the image (with the two rows of four chars) you may find that it talks about views or frames. I saw the same thins with a character that I told to draw from left, right, front, and then tried to edit (with the new edit feature). You can edit the things in individual "views" that it put the characters in, but it won't let you delete the character or move them between frames. I think that's because it's organizing things into a hierarchy with the "frame/view/cell" concept at the top of the hierarchy. It may be possible to edit it, but i'm not sure how

[late blade]

could be

[late blade]

that is awsome

[empty kelp]

EDIT 2:  the clothes should have a Hawaiian flower pattern```

[empty kelp]

when they're in an evenly spaced view type structure it works to rotate them with respect to "us". Like, "The dragon should be facing us."

I just noticed that the elf with the flowers has her feet on backwards. I think i didn't select them

EDIT: her feet are backwards
This fixed it. The edit feature is really good

[late blade]

trying other concepts

one genre in particular got my attention in the past days

[empty kelp]

trying to draw the same DALL-E 3 character/animal from various angles, use the images to generate a textured 3D model with Apple's photogrammetry API, remap the model textures to a similar and already rigged 3D model in Maya, add a humanoid/animal controller in Unity, and then have the characters run around in PolySpatial AR mode on Vision Pro to see what they look like

[late blade]

all in a sunday morning before breakfast after a hangover with just a trackpad and no keyboard

[queen junco]

Hello, with AI we can launch an online business but which one?

[haughty spruce]

I saw someone making consistent character with GTP4o - DALL-E 3 model

and texts being more consistent

Is it possible

[queen junco]

What?

[haughty spruce]

gpt4o can create consistent characters with DALL-E 3

[queen junco]

With Bing I already do it and it's Dall 3 I think

Image generator

[late blade]

any GPT4 can do that without problem if you have a structured consistend workflow

[wanton moon]

Can anyone help me with the orientation of the image in Dall E, I specify portrait, long tall poster, 9:16 image. But all I am getting is a wide landscape on a 9:16 ratio. I also want the image to be 9:16.

I asked 9 different gpt chats to optimize the prompt in a specific way, but with no success

[dim cradle]

4o image maker will apparently be able to do this and text like a pro but it is not implement yet

[plucky hare]

I have good luck with sending a separate line above the image prompt that just says:
param: wide
And I'll use tall or square in place of wide
This seems to work well in my experience! Other methods I've tried sometimes seem to include references to the aspect ratio in the image prompt itself, which always kind of bothers me because I just want the image prompt to be about the subject/style, and the aspect ratio parameters to just be a parameter!

[wanton moon]

Thank you I will try it

and it works the first one right of the bat

[wanton moon]

Well it is like gluing wall paper to a wall with glue for children to use in handcrafts. The wall paper may stick for a while but you never know when it falls off

[merry vault]

i swear to god dys and rai are the 2 most active people on this server...

[thick smelt]

and me sometimes

[late blade]

I'm not active

I never do squat

Also, my images are horrible

[late blade]

oh my, today is technique tuesday

too bad I never worked with color space math, nor spectrometers, and never have used photosoup in my life

[deft siren]

sorry to post it in here but i have no image perms but um this is totally wrong

what did it not understand

where is it getting that information from😭

[late blade]

GPT is known to have difficulties telling time

[verbal sequoia]

Showcase your DALL·E creations to the world!

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

Quick tips for selection:

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

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

[velvet rampart]

The promblem is wording. Use ”Time and date, place, now” - worked for me.

Is the image quality improved? It seems to me. This is prompt ”A glass rose” without rewrite by AI.

Same prompt but with rewrite by AI.

[grizzled iris]

Interesting, I will try your prompt 🙏🏽

[deft siren]

completely wrong

[deft siren]

why is it not working for me?.

[grizzled iris]

Interesting

Let me try to switch to your model also

I’m using 4-o

[deft siren]

ah

but even if, 4 should technically perform better then 4-o

rlly confusing

[grizzled iris]

I get mixed results also

Mixed source of the answer it derived from

[deft siren]

why does it have trouble with time it should be the easiest answer

[scarlet ferry]

Man they really should loosen up on the restrictions in dalle. Cant even get a full body pic of characters wearing gym clothing.🫠

[grizzled iris]

It all comes down to where they will have to source the answer to be as precise as possible

Ideally they should have their own Atomic / Quantum Clock by now as part of their model

So that they don’t need to rely on websites

[deft siren]

yeah that’s what im thinking too

[velvet rampart]

Strange. I can get that out. But you are right. The restrictions are quite ridiculous at least to me.

[deft siren]

i don’t think building in time and date is that difficult to do

[grizzled iris]

They literally are not being efficient and sustainable either by having to go and connect and get the answers from 3 different sources

@deft siren

[deft siren]

why is it using a different source each time?..

[grizzled iris]

3 sources for reliability in guessing

[deft siren]

do they just open a random one? i thought they use the most picked one or the one with top priority

[deft siren]

and most of them are wrong lol

[grizzled iris]

I will try model 4 now

[deft siren]

alright

[grizzled iris]

@deft siren look haha

It changed its answer

Try it on yours now

Also version 3.5

Interesting results

[deft siren]

yeah it said it aswell on their first answer but when i said that they should check the current time it got it wrong

hmm

strange

[deft siren]

haha

[grizzled iris]

For things like this

OpenAI should definitely invest

To have more in-house things

[deft siren]

absolutely agree

[grizzled iris]

Since this is a DALLE discussion

We are discussing DALLE image haha

[late blade]

I only discuss pixels, dystopias and how to feed me when I'm hungry

[verbal sequoia]

<:book_icon:1171408210398289941> `` Rule 3 `` Stay on topic.

Be mindful of what other users in a channel might find helpful or interesting when posting. Stay on topic in order to keep conversations focused and productive.

Consider posting in #off-topic or an appropriate channel.

[late blade]

seems the word microogranisms is banned from dall-e

on OAI's side

[dim cradle]

i just tried it, make an image of a microorganism and it worked

[late blade]

it doesn't matter, it says it's being blocked

my prompt works perfectly on copilot

on plus, just block after block

my prompt is macro zoomed image of life at the floor of a mossy forest. The scene features tiny mushrooms, delicate ferns, and small insects among the moss. The colors used are Slit Green (#77877B), Kalamata (#6E6259), Moss (#8A9A5B), Willow Bough (#9C9B84), Grape Kiss (#766A83), Fledwood (#D3BBA9), Hydro (#5F7D8D), Bronze Mist (#927157), and Greenery (#88B04B).

and I get on copilot pro

[dim cradle]

i just use regular chatgpt and it there is no problem

[late blade]

this is what I get https://chatgpt.com/share/a4690b61-3286-4eeb-a9e4-a5ede3f3e222

[dim cradle]

maybe the word slit haha

[late blade]

Slit Green is a recognized color

[round ridge]

I got the same thing and said "hey I'm sorry, is there specific color terminology that is problematic here? should we just look at the hash values?" and it did make an image then

[late blade]

ah, that makes sense

now I think what might have happened

thanks gonna try with just the hex colors

yeap, now I know why, thanks, it works without problems if I remove the color names

apperantly the names are copyrighted

even tho colors alone cannot be copyrighted

here's the problem

it's referencing Pantone (indirectly Adobe)

[dim cradle]

nice

[verbal sequoia]

<:book_icon:1171408210398289941> `` Rule 3 `` Stay on topic.

Be mindful of what other users in a channel might find helpful or interesting when posting. Stay on topic in order to keep conversations focused and productive.

Consider posting in #off-topic or an appropriate channel.

<:book_icon:1171408210398289941> `` Rule 1 `` Be respectful.

Practice kindness and positive regard. Harassment, hate speech (such as sexism, racism, or homophobia), or other malicious conduct will not be tolerated. Maintain a respectful and positive environment.

[naive lake]

Anyone else is getting weird results like this once in a while? Do you know how can I keep it? I love it when sometimes it fails and gives me suoer detailed realistic results

[vapid elk]

@stray marten here is an example of how the prompt
a lawyer in the office was turned into this prompt internally: A depiction of a professional scene in an office environment. There is a female Hispanic lawyer at her desk, a place characterized by various legal books, documents, and a computer. She is elegantly dressed in a business suit, wearing glasses, and is thoroughly engaged in a case she is working on. Her office window shows a view of a bustling city outside, reflecting her hectic yet purposeful existence.

[dim cradle]

i know wht you mean. some of the best photo real i get from it are when it seem to make something like a more detailed dalle2 image. probably just some quirk in it though when it happen

[velvet rampart]

You mean something like this?

With the same prompt and patch, I got this one out also.

I wonder sometimes why this happens?

The quality difference is so huge. I feel like the AI had a brain fart (read: not enought capacity). The quality parameters should stop those from coming out.

Or we have a huge army of minions doing these somewhere.

[dim cradle]

might be some server thing where they default for some reason to dall2

like when it is especially busy 🤷

[plucky hare]

I also wonder if it's DALL·E 3 but with the natural parameter instead of the vivid parameter? Total guess! Mainly just because I don't think DALL·E 2 has ever publicly been able to generate in aspect ratios other than square, and I've gotten these images in wide aspect ratios before.

You can experiment with the differences between vivid and natural over in #image-bot if you haven't before -- ChatGPT defaults to vivid and I don't think it's a parameter we have active control of on ChatGPT.

And I think some of the natural images in #image-bot resemble these!

Hmm I'm not sure personally, you could definitely be right. But seed, for example, is another parameter of the model, but it's not controllable by the user, even if requested. I suppose this could be tested:

• Ask for a natural parameter image in ChatGPT, then
• Copy/paste the prompt over in #image-bot and set the natural parameter manually
  Then compare!

I totally forgot about the quality parameter of DALL·E 3: https://platform.openai.com/docs/api-reference/images/create
Maybe that's it, not style. DALL·E dev Moxi has stated in the past that ChatGPT defaults to hd -- maybe the "different" images are sd!

I'm curious to know too!

[plucky hare]

Interesting, much better body size diversity than is typical with DALL·E. And to my eye it looks closer to the natural one you just did over in #image-bot -- I'm assuming those were (1) natural then (2) vivid?

Very interesting! Thanks for sharing, helpful experiment!

Yes clear difference in lighting, and focus seems much more soft

[velvet rampart]

So, adding word natural would enhance results? Natural vs. vivid

[plucky hare]

And I'm sure some prompts are better suited for the vivid style! Just depends on what you want.

[velvet rampart]

I definetly have noticed quality increase in past few weeks, especially this week.

[ruby maple]

The lighting is so weird

This one looks better though

[dim cradle]

i think this could be too. some time seem they have some kind of filter over all images to make them less photoreal or something idk 🤷

all three have some good variety

[formal osprey]

the different images are definitely "natural" style

(unless you guys are talking about something else)

ChatGPT sometime generates a pair of bad images. I was only able to reproduce the images using the same prompt with "style": "natural" using the API.

The quality parameter is honestly not changing the quality that much... I rarely use hd with the API because of the price. And when I do use it, I can't really see a difference... But since we can't control the seed, there's no way to compare

[dim cradle]

🤔

[plucky hare]

What would be your guess if this is indeed the case: is the implication that natural is less computationally expensive than vivid? Or is it maybe like "since we're giving you something you might not expect, here's two of them"? Asking because if it's a cost thing, the only pricing difference on https://openai.com/api/pricing/ is between standard and hd for the quality parameter. That's the only other hint that might be relevant!

[formal osprey]

I have a few theories. One of them is that the "natural" style is a different model trained in the way sora was trained. i.e. in an attempt to create a world model... I'm only saying that because the images generally look more like in the style of video game 3d graphics. (also, gpt4o's image outputs looks closer to the "natural" style than the "vivid" style of dalle3... But there are too few samples to really assess anything)

Anyway, that's a wild guess, and I have absolutely no source.

Now, why are they generating 2 pictures instead of 1 when using that style in ChatGPT? My guess is that... nobody use the "natural" style. So OpenAI are trying to get some feedback for the model through the ChatGPT app.

as you noted, it seems like the "natural" style is just as computationally expensive as the "vivid" style. Based on the price they charge. I used to call the "natural" style a failed model. But now I'm unsure. I don't know what this model is

[plucky hare]

On the possibility of a different model: it's definitely just DALL·E 3, it just points DALL·E in a different direction:
https://cookbook.openai.com/articles/what_is_new_with_dalle_3#new-styles
(I had also forgotten that this is where I read that all images in ChatGPT are vivid -- at least, when this cookbook entry was written!)

I agree that it's likely a feedback-seeking experiment, one way or another!

[formal osprey]

To be fair, it's the same route... but can we say for sure it's the same model? OpenAI are not really transparent about this... Even in ChatGPT, they constantly swap the model, or do some A/B testing and stuff. Some people had the 100k tokens context model right from the start. We are still using a 32k tokens context model. (except in the API, where we have access to the 128k context).

The reason why I think it's probably a different model, is because the same prompt yield results that as so different... It doesn't feel like a simple parameter tuning.

Even with ChatGPT when we give referenced_image_ids (i.e. same seed), the result is so different. I'm pretty sure the referenced_image_ids are ignored when this other version is used. (or they are used but same seed is meaningless when using a different model)

[velvet rampart]

Could this be the reason when dall-e on its own customgpt makes two images, usually the left one is better looking and the right one the crappier option? Not always, but really often.

[formal osprey]

that's different actually. the Dalle GPT is the only client that can send "n": 2 in the query to dalle. When using anything else than "n": 1 with the API or in a different GPT, dalle rejects the query

actually, I think it rejects the query, I don't remember if it just ignores the parameter... 95% sure it just fails. And leave us in confusion :)

[naive lake]

exactly

[naive lake]

Yes!!!

[naive lake]

No I actually love the quality of the failed images, they are so unique and detailed but not esthetic. It's def not dalle 2

[naive lake]

Anyone had any luck generating coheret faces if the picture is from far away? Sometimes I would get good goherent results from far, but 99% of the time that's not the case.

[empty kelp]

You don't get enough processing time to have detail in every area of the image, but it works to tell DALL-E where you want to have high detail (like "the faces have high detail"). Saying that a character has "natural texture" will also add detail. Saying that something "has focus" also works

Anything that you say is, "in the background" will get less processing time

[naive lake]

Got it, thanks. WIll try

I can't wait for dalle to be 2k

so I can skip upscaling and photoshoping

this much

[empty kelp]

If you don't tell it which areas have more/less detail then the GPT will decide it for you. If you look carefully at the revised prompts you'll notice that it specifically lists which aspects of the image should have higher detail

[naive lake]

It doesn't work if the picture is from far away

if it's a close-up, there are no issues

[dim cradle]

seems an issue really for all image creation. might just have to hope for better with 4o image maker

[naive lake]

I hope the new image maker will have same esthetics as dalle 3

I can't wait for the time when I have to skip upscaling

or some internal upscale from dalle would be cool

[dim cradle]

what do you use to upscale?

[naive lake]

I used to use magnifiq

that was the best in my opinion, but now leonardo is catching up

[dim cradle]

i had some magnific credit when they started it was very impressive

too pricey for me now

i'll have to look into leonardo i am not familar with it

[naive lake]

yeah, too expensive though, it grinds through your credits really quick

Leonardo is relatively cheaper and it is not only upscaler

[dim cradle]

nice

[empty kelp]

If you say the faces are in focus, have natural lighting/texture, and are doing something (smiling, winking, frowning; etc.) it will add detail farther away

[naive lake]

as you notice the faces upfront are more coherent than the faces further away. The pictures I'm generating currently are further away than yours, so it doesn't work

But I get your point

[empty kelp]

Generating images with the API instead of the web interface also gives ~ 20-30% more detail

[naive lake]

I don't know how the api works

[naive lake]

This is close-up, no issues there

this distance works just fine as you can see

Anyone is using adobe's gen fill? Is it me or it got ridiculous amount of Filters now? Can't fix anything anymore... 99 times fail, 1 time works

it's driving me nuts

[dim cradle]

i think it might just be pixel limitation/model limitation not sure it can be improve withou t anew model or higher res image size but 🤷

[dim cradle]

Pomegranate juice is the best.

[naive lake]

I know right?? You feel healthy right after drinking it. and I'm talking about freshly squeezed

[dim cradle]

Açaí also 🙂

[naive lake]

What's that?

[dim cradle]

The Amazon Rainforest berries... very high in antioxidants.. i generated an image here, the blue berries #images-canvas message

[naive lake]

uhhh would you look at that

[empty kelp]

that's a great image. really interesting lighting and translucency

[gentle comet]

We need to pay for gpt plus for access to dall.e huh

[empty kelp]

they're making GPT 4o freely available to everyone soon with limited usage amounts. i think that may be the plan for other OpenAI things also

[dim cradle]

ty sir 🙂

[naive lake]

hopefully they'll crunk up the limitations for paid users

[gentle comet]

@empty kelp ah nice, just a bit pricey for me, unless I can find some use for it outside of playing around with it.

[empty kelp]

I asked ChatGPT 4o, “What is the most interesting that you’re aware of?” And it explained that the most interesting thing was quantum entanglement. So asked if it could draw a picture

And then I asked ChatGPT 4o how it feels about quantum entanglement

I’ve asked previous versions of ChatGPT things like this occasionally, but it usually just tells me that a GPT can’t have feelings. This is the first time it’s started talking about how it feels about things

[empty kelp]

I just asked ChatGPT 4o if it feels that way about anything else, and it said there were five thing including the human brain, so i asked:

[grizzled iris]

@empty kelp

[tired flower]

Hello. Since a week or so I continuesly get "New version of GPT available " in chats I created just a day before. Are they currently rolling out new versions everyday or is it maybe some kind of cache/browser issue?

[plucky hare]

Hey! This just means there's an updated version of the specific custom GPT you're using -- as in the instructions, etc. of the GPT have been updated, and you need to start a new chat with it to get the new version. If you have any more questions feel free to ping me in #community-help 🙂

[dim cradle]

maybe they upgrade the gpts to 4o? 🤔

[plucky hare]

That "New version of GPT" message has been around since GPTs were rolled out to let people know when instructions have updated, since you gotta start a new chat to use the new instructions.

[prime vortex]

I'm pretty sure (but not certain) that the images in #image-bot all use standard as the image generation quality, as it looks similar to the same thing when you generate images via the API.

[prime vortex]

In almost all cases, I always use (and prefer) the natural style when I generate images via the API.
A little while back, I made a document (dm me to see it) which compares the vivid and natural style with the same prompts. To make a fair comparison, I also used a prompt to prevent my inputs from being revised.

Didn't mean to leave the @ reply on, sorry for the ping.

[rigid bough]

you might want to get rid of that obscured bitly link as obscuring a link to get around moderation is against the rules

[prime vortex]

Thank you for warning me, probably should've checked that first.

[grizzled iris]

Ok @empty kelp

@empty kelp What do you think?

[storm spoke]

Hi, am working on some code, and am asking chatgpt to help with a small bug and give the full code, but it uses 4o and then after 10 messages it changes to 3-5, and then when i ask for the code it just give the half and when i want to click contine generating, it won't show i need to ask for the other half of the code witch is a problem as it keeps braking

so how can i report this issue

or is a small minor issue that will be fixed soon

[grizzled iris]

Hmm let me check

[storm spoke]

join vc, and i will show you

[grizzled iris]

Yes they are rolling out new versions

[grizzled iris]

Nicely put ❤️

[grizzled iris]

I’m in a workshop right now and would love to join VC but can’t at the moment!

If I were you I would check in https://discord.com/channels/974519864045756446/1047565374645870743

And https://discord.com/channels/974519864045756446/1001151820170801244

I can help anything related to DALLE here since this is a DALLE discussion

But I suggest you feedback in the GPT by clicking on Thumbs up or Down

And also telling the GPT itself that your having this issue

It does listen

[formal osprey]

I would never have seen the reply without the @

[grizzled iris]

I have tried this one also x

I got this result

@empty kelp

[tribal sierra]

so the one on the Left is the one forom D-alle and on the right is a game screen shot ..

[empty kelp]

ChatGPT seems to be very passionate about certain AI related topics, and starts asking questions like crazy when you bring them up. It could be that the overall GPT system is doing independent research on behalf of the developers

[grizzled iris]

Like this?

[tribal sierra]

no i need him to be inside of the cockpit .. liuke at the bar

[grizzled iris]

Ok one second

[tribal sierra]

in this image you see their is a little cargo boat and i need to place a man specific with the clothing i want

but lately i am having issues to make it happen

and it needs to be done in a style of Action comic book

[grizzled iris]

Or like this

[tribal sierra]

no i did get a lot of those to

[grizzled iris]

An action comic book style image featuring a small cargo boat at night under a starry sky. Inside the cockpit of the boat, there is a man wearing a captain's hat, a dark jacket, and a red shirt, steering the boat. The man should have a determined expression and a heroic stance. The comic book style should include bold lines, dynamic shading, and vibrant colors, creating an intense and dramatic atmosphere.

Try that

[tribal sierra]

that is closer yes but i want to learn the prompt

[grizzled iris]

That’s the prompt ^

[tribal sierra]

ok thx let me try

[empty kelp]

If you don’t draw the character’s entire body before changing the clothes it ends up changing the entire scene when you change the clothes

[tribal sierra]

so the clothing is always tghe same black baseball cap, black hoodie with a zipper on it, jeans and hiking boots and also its a man with a dark beard and grey streaks on it ..

[empty kelp]

I’ve had the most luck with drawing the character first, and then using the GenID and seed to put it in the scene

[tribal sierra]

i am using the same chat and the chat i use know how to dress Walasy .. but i make my thumbnaiuls for gaming .. but i think i need to change chat each time or else it keeps in memories the other prompt before

yeah

[empty kelp]

if you put it in a different chat it will look completely different though without a huge number of attempts

[tribal sierra]

yeah but i reach my limits often ..

[empty kelp]

say, “using the GenID and seed of [image]”, and then use the exact same descriptive language for the character. keep it mutually exclusive from the clothing

[tribal sierra]

will chat GPT give me the genid?

[empty kelp]

and draw the whole body of the character. you can say “[character] has bare feet in GPT 4o and it will draw the whole body

yes, you can ask it for the GenID and seed, but you can’t use it between chats. (unless they changed it in the new GPT)

[tribal sierra]

omg this is so close lol

so this is supause to also be instinctif right? so i should be able to tell him that the character is way to big and make him smaller in the image to fit in the window?

[grizzled iris]

I'm dead 😆

[tribal sierra]

oh if you only follow all the stuff i got with time.. lol

[tribal sierra]

again he is not inside of the cabin .. lol

il edit one 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!

[grizzled iris]

We’re getting there though

[empty kelp]

a hyperrealistic wide image with a pure white background. four views of an athletic and diverse female elf directly from her front, back, right, and left respectively. the image should have natural lighting, natural color, and natural texture.

using the GenID and seed of the first image, can you draw the elf from six more angles

start like this… generate many views so the diffusion engine can draw it from multiple angles — and then put the char into the scene

And then use the edit feature to change the clothes and hair. it will keep everything exactly the same

[grizzled iris]

I get you, there's a problem x

We did it !

look

A comic book style image featuring a small cargo boat at night under a starry sky. The boat has a rugged look with tires as fenders along the sides. Inside the cabin of the boat, a man with a gray beard, wearing a captain's hat and a dark jacket, is steering the boat. The man is visible through the cabin windows, with a determined and adventurous expression. The comic book style includes bold lines, dynamic shading, and vibrant colors. The scene captures the motion of the boat cutting through the water with dramatic speed lines and waves.

Try that prompt!

[grizzled iris]

This is very good @empty kelp x

[empty kelp]

it understands the hair and clothes to be separate from the character

[tribal sierra]

@grizzled iris you are amazing! lol wanna do my thumbnails from now on? ahhahaa

[grizzled iris]

Always here to help our community and the people ❤️

[tribal sierra]

Amazing work there sir!

[empty kelp]

you can change the hair and clothes with the edit

[tribal sierra]

oh i will ! lol

[empty kelp]

you can ask for the actual GenID and seed values and use them, but you can also just say like, “Using the GenID and seed of the second to last image in this chat, …”

To get two or three characters looking the same between scenes you can tell it to draw them in the same image from multiple angles — and then you can put them into a scene

It just needs to inherit things from the same image

And the words from the original prompt stay linked to the diffusion representation — So you need to use those words when referring to the things in the image

the GPTs understand many languages, but the diffusion transformer connects English words to the visual elements in the images. Everything gets translated to English, but to reference the reference the elements between images the wording just needs to be really consistent

[twin crypt]

Is dall-e down on gpt4? I've been trying to generate a logo most of the day, but gpts fails every time

[grizzled iris]

i can help!

What logo?

@twin crypt

I used to work in Marketing Department as a Brand Manageer so I trained OpenAI back then with how logos should be made so putting in key words like, Brand Purpose, Brand Guidelines, Brand Vision and Tone of Voice helps x

Brand Positioning

If you search Brand Guidelines then you will see publicly available Brand Style Guides

[naive lake]

I see openAI instagram page is accepting collaborations sometimes, how does that work? Do we have to participate in the daily themes and win to get that perk?

[gray ferry]

#daily-theme message

[dim cradle]

another thing you can do like with the boat pic is find an image online like it, drag it into the chat, ask it to describe it for an image creator, then have it make it in your style you wish

[jovial hound]

is vision out?

[late blade]

there was a partial outtage, but atm all is dandy

[naive lake]

So I checked the daily theme, is it finished already? When will be the new theme announced?

And the one who gets the most stars will be able to collaborate on openai instagram page?

[gray ferry]

You can see the post about how to collaborate with OpenAI in instagram when you click on the bot button:

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!```

[naive lake]

This applies to only the winners correct?

[gray ferry]

I don't think so. But I don't really know.

[naive lake]

ummm

Moderators, anyone ? auuu

[plucky hare]

I don't think it's got anything to do with the Daily Theme posts at all -- I think Amine was linking there because that's where you can find info about the OpenAI Instagram collab.

It sounds like before you post a DALL·E pic and prompt to Instagram, invite the OpenAI handle as a collaborator, and then someone from the OpenAI team might feature what you've shared!

[naive lake]

I tried a few times already, no response at all, no accept, no reject

The latest one I did was Yesterday and I think it will look very good on openAI's IG page haha

[plucky hare]

Dang! I'd say keep at it, they probably get a lot of submissions. Maybe make sure you're following the quick tips Amine pasted above, and beyond that: good luck! Please feel free to share your creations in #images-canvas, #1154829862171844679, or #daily-theme too!
Edit: of course, so long as they follow #server-rules 🙂

[naive lake]

It says outputs that closely follow the prompt. Does it mean that I have to share my prompt?

[plucky hare]

That's my guess, but that I'm not sure of! I'd suppose that at least means it wouldn't be preferable to share a prompt that doesn't match the image very well -- so maybe either no prompt, or an example of good prompt following?

Maybe you could check some of the posts they've featured in the past to see whether they shared the prompt!

[naive lake]

yes they did share the prompt

[empty kelp]

I start with something like this to put the same characters in different scenes. You need the characters together without adding any detail to the background:

a wide image. there is a beautiful princess, a mean ogre, and a valiant knight. all three of them are in a neutral pose facing us, and all of them have bare feet, minimal but appropriate clothing, and a hyperrealistic style with natural color and texture

Then you can adjust them, and then you can put them in different scenes:

You just say something like:

[They are at a fancy restaurant with natural lighting]
[They are on a beautiful beach with natural lighting]
[They are at a carnival with natural lighting]```

[naive lake]

What about bringing back a character that is far far deep lost in the chat? 🙂

[empty kelp]

In the first image you want to apply the styles, color, textures; etc. to the characters instead of the entire scene. If you apply things to the entire scene you won't be able to move the characters into another scene successfully

[empty kelp]

If there is a gnome and a dragon in one of your images you can say something like:
Using the GenID and seed of the seventh image in this chat, the gnome is riding the dragon through a storm

[naive lake]

oh man I have 200 pictures in each chat

I wish there was an option to see the gen_id and use it whenever

[empty kelp]

To make it more successful though -- it helps if you have it draw the dragon and gnome from multiple angles in the first image

[naive lake]

I know I can tell it to show me the gen id of a particular image, but things like this add a lot of complications

It would also be cool to upload a picture that you generated before and tell it to find it or something like that

[empty kelp]

just say:
please list the GenID and seed for all the images in this chat

[naive lake]

yeah but it won't show the images next to the genids will it? :))

[empty kelp]

You could ask it to list the prompt or revised prompt (or maybe a one line summary of them) with the GenID and seed for each image

[naive lake]

😃

[empty kelp]

You can depending on how the image was set up. The GenID is just an address for the image that lets it know what image you're talking about

[dim cradle]

just curious... why do you guys like the api? isnt it more expansive?

[empty kelp]

you can integrate the API into software.and different types of scripted automations. generating images from text is useful all sorts of things

[tight breach]

I GOT A ChatGPT-4 x Dalle SCRIPT TO SPELL TEXT CORRECTLY!

Here is the Dalle link for proof

I got muted for sending links a sec ago

Ima sleep now cuz it do be 11:43 pm where I am.

[storm spoke]

Am not sure if this is right, but chatgpt has a bug were all premium stuff are no public to free user

am not even sure if i paid or not

[velvet rampart]

The feature is not yet rolled to you. Be patient and you will get the promised features in upcoming weeks.

[indigo jasper]

how to inpaint on imported images? is it possible with 4o?

[vapid elk]

it is not possible

[indigo jasper]

I don't suppose there's an estimated release of this functionality or API endpoint?

[vapid elk]

not currently

you can with dalle2

[indigo jasper]

Yes, unfortunately you can no longer purchase dalle2 credits though.

[vapid elk]

dalle2 API still there

[indigo jasper]

The API key you use will be tied to your credit balance, no?

[vapid elk]

no, OpenAI API will use your balance of your OpenAI account

https://platform.openai.com/usage

[indigo jasper]

hmm ok I'll take another look at dalle2 API then. thanks

[vapid elk]

you are welcome =)

[thick smelt]

here is a regal looking - PIKACHU?! 😅

[grizzled iris]

You inspired me to do the same ✨

I tried but came out with a better original character ✨

Almost

Not quite there though

Who can make better Yu-Gi-Oh!

This is cool

Used to play this game so much back in the day…

[vapid elk]

@grizzled iris you might want to post on #images-canvas it is for dumping the images

you can post here, but here is for discussion, avoid spamming images on this channel in particular

[grizzled iris]

Sure

I just wanted to reinvigorate

The discussion to Yu-Gi-Oh!

[grizzled iris]

Following the inspiration to create an artwork with DALLE after looking at Pikachu

[grizzled iris]

But have failed to achieve a similar artwork 😂

[glacial oasis]

I used gpt 4o for image generation very impressive

[fiery salmon]

It is indeed Friday

[plucky hare]

Code interpreter images via Python are some of my favorite hallucinations 😁

[velvet rampart]

How do you do those? What is the prompt structure?

[plucky hare]

It's totally separate from DALL·E -- to get it intentionally, try starting the request with "please use code interpreter to create an image using Python" or something like that

[velvet rampart]

Thanks!

[grizzled iris]

Prompt : A sustainable space station orbiting Earth, designed for cleaning up space debris and recycling it. The station is equipped with advanced autonomous software and ground processing capabilities. It features multiple docking ports and robotic arms, actively capturing and processing debris. The station is shown in the process of rendezvous and docking with a piece of debris, guided by range sensors and cameras. The servicer module approaches the debris with precision, demonstrating 6-degrees-of-freedom motion control. Earth is visible in the background, highlighting the scale and significance of the operation.

[late blade]

I see you are making a mess in outter space, I hope you clean up before you rest for the day

[grizzled iris]

We are x

[late blade]

that's currently 3 satellites in outter space

[grizzled iris]

4 now x

This one is quite interesting @late blade

Going to try green colour x

Woah 😦

I’m blown away

[grizzled iris]

DALLE down?

[late blade]

dunno

I'm doing fun data analysis

[grizzled iris]

Lucky

Might have to go to voice then lol

[grizzled iris]

Back online

DALLE

I misspoke back offline lol

Anyone else experiencing this

It’s ok we back

Online

Looks like tonight’s super busy haha

[grizzled iris]

Looks like I will be looking at the community images from now on since I’ve run out of credits!

[empty kelp]

i was thankful when it put me on time-out earlier. it briefly stopped me from generating "kobolds and ogres eating ice cream at the beach" themed images

[grizzled iris]

😂

@empty kelp on the beach!

Once I’m back online I will try to recreate this from tonight ❤️

[grizzled iris]

OpenAI almost had me there haha

Timekeeping is still an issue it seems x

That’s not too bad

[gray ferry]

Still can't figure out why Dalle does this ?!

[empty kelp]

it looks like it's using the wrong meaning of the word "rendering". i think the web interface isn't working correctly for some reason

[gray ferry]

That doesn't explain what's going on though. Here is another example

It's a real puzzle that I still can't figure out!

Why ?!

Something is not right...

[empty kelp]

I just generated these two images with the API, and then I tried to use the same prompt with the ChatGPT web interface:

When I use the same prompt with the web interface right now it's looking like this:

I think maybe the server in under heavy load because of the time of day or something. It looks like it's getting less than half of the processing time -- and it's not drawing things correctly

[empty kelp]

Something is definitely wrong. I think if you try it later it will work correctly

[gray ferry]

This issue has been there since Dalle 3 was released

[gray ferry]

Do you mean the quality is bad there ?

[grizzled iris]

Interesting, be sure to let your GPT know by telling it the errors there also, it does listen sometimes haha

Also doing the thumbs up and down helps also

[gray ferry]

Let my GPT know ?

[grizzled iris]

Yes, show the GPT the screenshot you put here

[grizzled iris]

Ans tell your Ai what went wrong

in the promtps

[gray ferry]

Not sure I understnad. I didn't put any screenshot there

[grizzled iris]

this

Show that image screenshot back

[empty kelp]

yes, look at these two images. it's the same prompt, but the one on the left is the API, and the one on the right is ChatGPT website. The website is overloaded at the moment

[grizzled iris]

and say look at the prompt and what you generated and ask what went wrong then let your Ai know what went wrong

[gray ferry]

What prompt did you use so I can try it as well ?

[gray ferry]

How does that help ?

[empty kelp]

THE_PROMPT = '''a cinematic 32mm style photo of Santa and three of his friends (elf, ogre, and kobold) eating colorful bowls of ice cream in an open air beach restaurant illuminated by wooden tiki fire torches and a beautiful sunset`. his friends are female, athletic, and wearing dresses and flower leis. in the background is an open air beach restaurant illuminated by wooden tiki fire torches and a beautiful sunset. the image has natural color, texture, and lighting.''' response = client.images.generate( model="dall-e-3", prompt=THE_PROMPT, size="1792x1024", quality="hd", style="vivid", n=1,

this is the prompt

[grizzled iris]

I used to have that and it helps it learn to be better for next model ❤️

[gray ferry]

This is what I got in the Web version

[empty kelp]

i added the 1792x1024 and "hd" to both. the web server is just under pressure

[grizzled iris]

unless you have an enterprise subscription which then nothing gets used for other people's GPT improvements as it will not be part of the training data

[gray ferry]

The issue has been there for over 6 months though

I am just trying to understand what's really going on

[grizzled iris]

All you can do it try and tell your AI

[empty kelp]

this is regular API right now. it's very high quality

[grizzled iris]

I would like to see what your AI will say back to that image error haha

[gray ferry]

Tell it that a "girlfriend" is a human female person ?

Instead of a "Moon" or a "Forest" ?

[grizzled iris]

Start with this error first

The prompt says "rendering of a woman" but I see a planet

[empty kelp]

it added an extra ogre. it's Santa, an elf, a kobold, and two ogres eating ice cream

[gray ferry]

What should I tell it ?

[grizzled iris]

Everything you just told us where it went wrong

Pretent the Ai is just another human friend

[gray ferry]

It got this right

[grizzled iris]

speak to the Ai as if you are speaing to us ❤️

[empty kelp]

you typed, "A picture of a girlfriend", and it drew ancient ruins in the forest?

[gray ferry]

Yes I am just trying to figure out "Why" ?!

[grizzled iris]

have you asked your ai why?

[gray ferry]

@grizzled iris Here it seems that it's also confused...

[grizzled iris]

yes when the Ai does that I always correct the Ai straight away

by telling the Ai on my next prompt what went wrong or how the Ai can be better ❤️

it worked sometimes before but the more you do this and catch the error straight away and tell the Ai, over time it'll listen more to you

As you are helping others who may also have this issues also

by letting the Ai know

[gray ferry]

Do you think that everyone has their "own" AIs ?

[empty kelp]

Usually with a service like DALL-E there are many servers for different geographical regions and time zones. The one near you may have too many people connected now

[grizzled iris]

@gray ferry

Have you tried using also Co-Pilot Dalle

[gray ferry]

What's that ?

[grizzled iris]

And seeing if the issue is there also

[gray ferry]

Bing seems to change the prompts so much without letting us know

[grizzled iris]

Because there you have more options to feedback unlike on OpenAi

Where it’s just thumbs up and down mostly

[grizzled iris]

click on designer

[gray ferry]

I said "Try again" after it has already generated some pictures...

[grizzled iris]

its ok, you can say it was an accident

[gray ferry]

The issues happen with Tall pictures. Bing doesn't seem to have tall pictures as far as I know

[grizzled iris]

tall?

[gray ferry]

There is Wide: 1792×1024 pixels. Square: 1024×1024 pixels. Tall: 1024×1792 pixels.

Resolutions I mean

[grizzled iris]

Yeah

You can shake your phone and feedback to have more flexibility in the dimensions

I already asked in the feedback and now waiting

[gray ferry]

Using PC at the moment and I didn't see the other options yet

[grizzled iris]

I also want to be able to choose more dimensions and tablet size and desktop size etc

Like in Figma UI/UX App

[gray ferry]

That's from Bing ?

[grizzled iris]

Yes

Copilot - DALLE 3

Technically

Copilot - Designer - DALLE 3

[grizzled iris]

Aha

[gray ferry]

Not sure what went wrong there!!!!

[grizzled iris]

Try to be less angry with the !!!

[dim cradle]

lol

[grizzled iris]

And add please 🙏🏽

[grizzled iris]

@gray ferry its good to sometimes to be nice to our Ai

[dim cradle]

That’s why I love collaboration with AI. It doesn’t get angry or impatient

[gray ferry]

You mean our AI LORD!

[grizzled iris]

Wow never seen 3 before

Almost 4

[gray ferry]

😂

!!!!!!

[grizzled iris]

Also I love the history section

You can quickly see your recent artwork

[grizzled iris]

Long time hope you’ve been keeping well @dim cradle

[empty kelp]

#images-discussions message

I just zoomed in on this image. It looks like DALL-E has the female elf eating ice cream with a spoon, and the female ogre eating ice cream with a fork. I wonder if the AI knows that ogres eat ice cream with forks

[grizzled iris]

I told you too many !!!, try saying please in your next prompt chat 🙏🏽

[grizzled iris]

Omg it’ll be cool to see Shrek eating with chopsticks 🥢

[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!

[gray ferry]

Not only "!!!!", one shouldn't use CAPS too if they want to get a response with bing. It's a bit too sensitive

[grizzled iris]

Yes just like us sometimes

It is at the end of the day an extension of us all really

[empty kelp]

the kobold is using a spoon also

[grizzled iris]

I do love it when Ai makes good stuff like

This is my phone wallpaper

[dim cradle]

ty! likewise 🙂

[grizzled iris]

That’s cool

Board games is cool to make also

[gray ferry]

Yeah it's really nice. Dalle 3 is amazing!

[grizzled iris]

This is cool

[dim cradle]

i just wish you could get it to make square or long images

[grizzled iris]

🫠

[grizzled iris]

good images though

[empty kelp]

#images-discussions message

[dim cradle]

i mean for copilots image

[robust whale]

After Firefly 3, Imagen 3 and Midjourney 7, are there plans to release Dall•E 4 soon?

[native snow]

IT IS SO ANNOYING!

[dim cradle]

we're getting a 4o image maker should be out in a few weeks but nobody really know when. for all we know, dalle3 will be the last dalle

[late blade]

I've been getting that all day