images-discussions

[plucky hare]

What do you mean "won't continue to generate the same prompt"? Are you asking ChatGPT to send DALLE an exact image prompt, or are you just prompting ChatGPT to write the image prompts for DALLE? Which of these is your intended action, or is it some other issue?

[lofty valley]

Me using Dalle: make an image of a cat.
Dalle: returns images of cats.
Me: More.
Dalle: I can’t run the same prompt but I can generate a slightly different prompt.

Another response I got: “ As I've reached the limit of image generation for this particular theme, I'm unable to create additional images of the sad cat drinking coffee. If you have a different subject or theme in mind, I'd be happy to help with that. Please feel free to share any other ideas you have. “

I can get around this but it’s just weird. Openai has fine tuned/ tweaked the underlying gpt model.

[plucky hare]

Here's a message from an OpenAI dev about how they're trying to work with things that might apparently be hallucination glitches regarding the connection between ChatGPT and DALLE 3 #images-discussions message I don't know if it could be related to what you're experiencing, but perhaps. I'm sure you're right that they're constantly fixing and changing things behind the scenes to try to optimize.
In the meantime, some general if perhaps not entirely helpful advice might be to just try to be as specific as possible. In other words, instead of just saying "More," if that's actually what you're saying, maybe try something like, "I really like the second one, can I see two more like that but with the cats wearing fringe leather jackets?" Any way you can try to minimize ChatGPT misunderstanding you, I think it's worth it! Again sorry if not entirely relevant to how you use ChatGPT 🙂

[lofty valley]

Yeah a workaround atm for me is being very specific. “ No I mean more of the same in a different instance “

Haven’t had hallucinations like this since I first had to use gpt3 which makes me think 🤔 thanks!

[dense ermine]

Where are the art channels? I wanted to look at what horror art was made today

[teal timber]

Just scroll through #1154829862171844679 or #daily-theme but idk if you’ll find much horror as rules for server don’t allow a lot of that nor does the filter

[dense ermine]

That doesn't sound convenient

Why'd they get rid of the art channels?

[teal timber]

Scrolling one channel with tag filter ability is much better than 20 channels to scroll...

[tiny jetty]

Can’t wait for F1 this sunday

[dim cradle]

Seems to be working great again. ❤️ OpenAI Thanks for looking into it.

I like to specify the "ultra-high-definition" effect, rendering to present a tactile and intricate visual experience, inviting an extensive exploration and contemplation of each unique design.

[agile peak]

Okay. So apparently dalle can definitely mimic the authentic visual of security footage.. that mindblowing.

[dim cradle]

Thanks for not calling it CCTV

[tiny jetty]

If you put ‘ultra-high-definition’ in the prompt it becomes the best possible defenition and quality?

[tiny jetty]

I tried and this is what I got 😍

Not that great quality if you zoom tho

[dim cradle]

It's an effect, it's still 1024x1024, I think, and it depends on subject matter. I find it's a condensed instruction, in general. For that scene and resolution it did a pretty good job..

[tiny jetty]

If DALL-E would have 4K support.. DAMNNN

[dim cradle]

Only a matter of time, I suppose.

[dense ermine]

It was an issue before?

[tiny jetty]

I think so yes

I already love that we can make wide and long images

Bing only square

[warped olive]

Umm, did something change recently in Dall-E 3 content safeguards / policy? It suddenly thinks I'm a frightened kid. I'm creating monster images for D&D and it had no problem with truly horrific prompts just a few weeks ago. Now it's like "no, this clearly memey and comedic monster is baaaad m'kay"

[marble tartan]

Every day, another AI system gets lobotomized

[dim cradle]

I haven't experimented with that yet. I wonder when we'll be able to generate a "live" image, one that animates for a couple seconds? (For that matter, an image upload feature to transform any photo into a "live" photo.) Or generate consistent characters throughout a series for illustrating narratives?

[warped olive]

I guess I need to stop paying Open AI :(

[agile peak]

Mentioning details such as lightning, color and shadow can improve the quality as well.

[dim cradle]

I asked it to depict demons from hell, and 1/4 made it through the content policy.

[agile peak]

This one used a fluorescent lightning and sharp shadow.

[tiny jetty]

Would be great!

I want to have the possibility to change an image created just slightly to perfection. Now DALL-E generates a whole new image instead of doing small changes to the existing one.

[warped olive]

This was completely fine a few weeks back:

A black, circular entity rotates above a lava pit, its uneven bulk adorned with grotesque protrusions like tentacles, spikes, and pulsating veins. Fiery lacerations mar its surface. Amidst its form, a maw in ouroboros-like chase consumes its own tail. This abomination, a fragment of an evil deity, constantly morphs between sludge, solid, flesh, gas, and other eerie forms. Its hues are a grim palette of dark grays, muddy browns, and sickly greens, with a haunting red glow from the lava reflecting on its inner ring.

Now it thinks this breaks policy:

Create an illustration of a monster called Greater Gazebo Mimic

I guess it was too good to be true :(

[tacit badger]

good evening everyone

[tiny jetty]

Good evening!

[tacit badger]

hey benny 😊

[tacit badger]

it's possible it's just messing up. it kept telling me today that the style 'cross-stitch' violates content policy.

[warped olive]

:D

I guess I'll try again tomorrow then

[tacit badger]

so it even said it wasnt a problem but still had a problem

[warped olive]

[tacit badger]

this is truly beautiful

[tacit badger]

oh wow these are amazing! 😍

[dim cradle]

DALL-E 3 is addictive

[tacit badger]

i agree! it's amazing to be able to take an idea from our brain and witness it comes to life in a few moments right in front of us

[dim cradle]

Yes, and has therapeutic applications. Transmuting a thought from prose, poetry to images and back is highly introspective.... Recently I took my DNA sample, generated a speculative ancestral narrative, then generated a piece visualizing the migration over 4000 years. Couldn't do that before....

[tacit badger]

wow! that's gorgeous

[dense ermine]

How do you do?

[tacit badger]

With a whirlwind of chaos and a dash of grace, just like a cat chasing a laser pointer—full of purpose, yet entirely unsure of what I’ll do if I ever catch it. And you?

[dense ermine]

Umm, fine I guess

[tacit badger]

lol

[tacit badger]

sorry, i was feeling whimsical

[dense ermine]

Ah okay, I assumed that was the case just didn't know any other proper way to respond

Anyway, I was just asking earlier about the case with the art channels being taken down

[plucky hare]

They've been consolidated into #1154829862171844679 which is filterable/taggable so can behave exactly like the old channels, but with a nice pretty panel view

[dense ermine]

Maybe it's because I'm used to it but I didn't see a problem with the multiple channels, not to mention, there is little to nothing posted on the gallery

[crystal badger]

Let me join

Had the DALLE issue been resolved?

[dense ermine]

So of the options in the gallery I don't get

[crystal badger]

Yeah you’re not the only one

[dense ermine]

Is option the right word? Maybe I should say categories

[crystal badger]

Yeah more so categories

I preferred the old way

[dense ermine]

Like they work like the multiple channels before it right? Having a theme to them like Horror, Food, Sci Fi, etc?

[crystal badger]

Yeah basically but also sorts from new to old or activity based

[tacit badger]

i am not a fan of the gallery becasue it's lots of extra steps. i just want to scroll down and see pretty pics people post, not have to look at categories

[crystal badger]

Agreed

Or have several galleries for each topic

[tacit badger]

yeah, i agree

[crystal badger]

Although it isn’t as active as it has been

Especially after the verification process

[tacit badger]

i also think you have more of a chance to have your images seen the other way

[crystal badger]

Aka The Purge

[tacit badger]

i only recently joined here, what was the purge?

[crystal badger]

Basically you must have a verified account on OpenAI connected to here

[dense ermine]

Right? What we had before was more simple

[crystal badger]

But also more chaotic

[tacit badger]

ah ok. that doesnt sound like a bad idea to me though

[crystal badger]

It isn’t until you realize how many got kicked

[dense ermine]

Chaotic?

[crystal badger]

Well more active and unpredictable

[dense ermine]

Oh, isn't that a good thing?

[crystal badger]

It’s not a bad thing

But in the end it was just spammy

[tacit badger]

well in my newbie opinion, i like it here currently 😊

[crystal badger]

And our dear Lugui, bless him was working 400 miles an hour as a community contributor at that time

[dense ermine]

I miss the horror art channel, I loved seeing those especially the retro 8bit ones

[tacit badger]

it's far better than so may other forums i just lurk in

[tacit badger]

ohhh that souds awesome! i woulda loved that channel

[crystal badger]

Isn’t it just renamed to Spooky

[tacit badger]

is it?

[crystal badger]

#scary-and-spooky

[tacit badger]

that just says no access

[crystal badger]

Dang my b it’s closed due to #1154829862171844679

[tacit badger]

ah

[dense ermine]

There was also Food, Sci-Fi, Architect, Anime, Things that should exist, list goes on

[crystal badger]

I think we should have a standalone horror channel

Because you gotta spoiler all horror images rn

Because some are sensitive which is understandable can be a rough topic

[tacit badger]

as a huge fan of zombies, i have been rather disapointed in not being able to create many. dalle doesnt like blood and gore 🥺

[crystal badger]

I don’t know the concrete reason but surely our dear moderators can enlighten us

[tacit badger]

same issue with violent text in stories. gpt drives me crazy that it flags stuff in my zombie novel. like no kidding, it's a zombie apocalypse of course it's violent. and obviously fictional

[crystal badger]

You get flag for that?

I don’t

Huh 🤔

[tacit badger]

yes, when my main character is smashing in a zombie skull it gets flagged. well orange flag, but still

[dense ermine]

Yeah, that really holds it back

[crystal badger]

There’s a way around that

[agile peak]

Probably got lucky today, dalle finally know to capture photographic essence with Agfa vista 400.

[crystal badger]

A way around that doesn’t violate TOS

[dense ermine]

Like really people should be more responcible for the things they make but not that way

[crystal badger]

[Insert Graphic description] is what I tell GPT to do if it refuses

But I’ve never had it refuse

[dense ermine]

That metal doesn't really sythesize with the cat in my eyes

[tacit badger]

lol i was just muted for 60 seconds becasue i ws trying to say that my story also deals with topics like un-aliving one's self and gpt doesnt like that

[dense ermine]

Why? If people still complain then thats their problem

[crystal badger]

That’s true

Although I can’t relate but my graphic descriptions don’t get flagged

And it deals with the description like on the zombie thing

[dense ermine]

That's unfortunate, I really don't think people should be denied harmless stuff like that

[crystal badger]

Maybe I prompted weirdly I’ll look into it

Wouldn’t want a harshly worded statement from OpenAI

[tacit badger]

i dont think it should be flagged when it's obviously a fictional story. that's what get's me. it's cearly a story and it doesnt know how to seperate that from reality

[crystal badger]

I don’t think it should be flagged either

[dense ermine]

As a horror/monster fan, I would love to see anyones prompt

[crystal badger]

But also do to a Certain extent

[tacit badger]

i was doing some eldritch horror stuff a little while ago in dalle

[crystal badger]

The current measure is way too much

Sometimes it’s singular words

[crystal badger]

Not just you it has been lowered

[dense ermine]

Hell if I knew, ChatGPT isn't exactly perfect for me either but I only use 3.5

[crystal badger]

And with good reason

Way more people are using it right now

[dense ermine]

HP Lovecraft right? The artist that should be hated but the art that should be loved

[crystal badger]

So everyone generating 4 imgs every 15 seconds doesn’t bode well

[dense ermine]

Or did he just write books?

[remote stone]

I have been spoiled

[crystal badger]

Wdym

[tacit badger]

that was done in the style of Gustave Doré

[crystal badger]

Did you too get to fly in the Starship with Elon Musk to conquer the universe @remote stone

[tacit badger]

... wut?

[dense ermine]

Just looked him up and got christain images

[crystal badger]

Just an offhand joke

[tacit badger]

yeah, he's one of the orginal illustrtors for the devine comedy

he's known for very detailed and larger than life kinda art

[dense ermine]

Honestly, (and I like the picture) the art you generated doesn't look like what I googled

[remote stone]

I mean, I got used to just being able to generate a lot more every day, even though it was probably too good to last knowing the server load it must create if everyone's doing like me

[crystal badger]

This was a weird generation

[agile peak]

But this the first time I got. Dalle captured the retro photo essence.

[tacit badger]

sorry, i used other modifiers that changed up the look drasticaly

lemme find the exact prompt

[dense ermine]

Okay, that's a little better I suppose

[agile peak]

Seed 500-700 usually got better result for me

[crystal badger]

Really?

[tacit badger]

Illustration in the style reminiscent of Gustave Doré, capturing the essence of megalophobia, the fear of large objects. The scene depicts a colossal, shadowy figure looming over a tiny, insignificant human, set in a chaotic, barren landscape. The detailing is intricate, with eerie, wispy elements that add to the sense of dread and overpowering scale.

[crystal badger]

I use the best seed

1.

Not even joking

It’s a good seed

[agile peak]

Tried higher seed it produced unrealistic cybernetics parts.

[crystal badger]

Have you tried PI?

[tiny jetty]

Bruhh…

I'm sorry, but we cannot generate images that include elements like ghosts and creepy settings, as they may not align with our content policy. Let's try to modify the theme to something more suitable.

[crystal badger]

Yeah seed 1 clearly the high roller

[dense ermine]

Why not?

[tiny jetty]

Who knows

[dense ermine]

Gore and nudity I can understand but ghost and creepy settings?

Are you kidding me? Are there children on here that are easily spooked?

[crystal badger]

Well there’s stages of nudity and stages of Gore

[tiny jetty]

I said creepy abandoned house with ghosts flying around it

[crystal badger]

The main issue is this

[dense ermine]

What so wrong about that?

[crystal badger]

DALLE can’t see what it creates

[tiny jetty]

No clue

[crystal badger]

Dalle is literally making the image blindly

Hence the high restrictions

[agile peak]

Pi number?

[crystal badger]

Yeah but not as a comma

Without punctuation or commas

[agile peak]

314? Or more

[crystal badger]

More

[agile peak]

Seed: 1 and 314 gave me these

[crystal badger]

Think around in the 3000 or 30000

Thousands

[versed orbit]

After asking for a photo I get DALLE starting, but then I get “stopped” happens a few times before displaying the pics. Odd

[crystal badger]

Looks pretty realistic

Parts wise

Use 314159

[tiny jetty]

Same here

[agile peak]

Got me this 😂

[tiny jetty]

It says stopped but it still generates the images

[versed orbit]

Exactly. Not a fan.

[tiny jetty]

I guess just a glitch

[dense ermine]

Does he chase a toy mouse?

[merry marlin]

Holy moly that is weird

[formal osprey]

lol, I like those

proof cats are not real

[agile peak]

Definitely will chase the mouse whenever it near the computer 🤣

Tried custom prompting. I finally able to get zilla in realistic sense but gave me shin godzilla look haha.

[dense ermine]

Ah should have thought of that one

[dense ermine]

Zilla? Like the American Godzilla?

[agile peak]

Yes. That’s what I used the american version.

[dense ermine]

Guess it's a die hard heavy Godzilla fan because it must think the American Godzilla movies don't count

[agile peak]

Tell dalle to make him enormously giant. Wasn’t disappointed.

[dim cradle]

Speaking of gargantuan ...

The Owner is an entity of unparalleled extravagance and technological marvel. He is not merely large but gargantuan, with his size imposing an immediate sense of awe. As a cyborg composed of pure latinum, he transcends the traditional bounds of organic life. His towering form is a testament to his wealth and power, with a body that is both a fortress and a work of art. The latinum that makes up his being is not just for show; it is the bedrock of his advanced cybernetic systems, which give him strength and capabilities far beyond that of any conventional life form.

The creative narratives are as much fun as the art 🙂

[agile peak]

That’s freaking insane 🤯

[agile peak]

Apparently dalle can make it bigger. At this point can make it bigger than the size of planet haha

[long condor]

Does this work on mobile?

And does it have to be on ChatGPT 4?

[tiny jetty]

Yes

You need ChatGPT Plus

[long condor]

Ok, Ty

💀

[sacred helm]

are you sure you selected dalle mode?

[dense ermine]

Are we posting AI artwork here now? (Not that I'm complaining)

[naive lake]

Yes

[dense ermine]

That doesn't look right

[naive lake]

It now decides if it wants to follow my orders or not

tomorrow it'll be like "Hey dude, I don't feel like making what you want" lol

[tiny jetty]

You've hit your daily maximum number of images. To ensure the best experience for everyone, we have rate limits in place. Please wait for the next day before generating more images. Your daily maximum will reset in 23 hours and 22 minutes.

😭

Does anyone know what daily maximum is?

[dense ermine]

Yeah someone else on here has the same issue, what it makes is pretty good though

[sacred helm]

300 images (this may be outdated)

or maybe requests

not sure

[tiny jetty]

Ah okay..

[quartz vale]

#images-discussions message
This is all we know right now

[drowsy bane]

Hmmm seems like some things are changing, still getting 2 at a time but getting mutliple runs in a single submission. Not sure if this was because I mentioned 3 different prompts but I wasn't seeing this behaviour yesterday.

Mind you, it seems to do this very inconsistently though.

[drifting hawk]

200 currently

We are going through a gpu drought…

Hang in there

[drifting hawk]

How did you go through 200 images in half an hour

[tiny jetty]

😳

[tiny jetty]

Will the number go back up soon?

[drifting hawk]

when we get more gpus yeah

[tiny jetty]

Excuse me?

50?!

[queen herald]

Mine said 25. I don't think the AI actually knows.

[dim cradle]

AI doesn't even know what time it is.

[hazy warren]

Peanut butter jelly time?

[queen herald]

The amount of time it says to reset isn't even consistent. One chat said 18 hours and another said 20...

[drifting hawk]

It’s hallucinating

It doesn’t know

[drifting hawk]

Oh it should be, but its calculated based on your timezone

So if you are switching vpns it could change

Really wish we could get ChatGPT to stop@making stuff up

[queen herald]

I'm not, it was just between two separate chats

[drifting hawk]

Weird, that shouldn’t matter. Can you share one of your convo ids? It’s in the url

[shut niche]

Where is this limit even stated?

[obsidian frost]

Does anyone here have dall-e credits? I need a little little help with a simple edit lol

[plucky hare]

#images-discussions message
+
#images-discussions message
I think about covers it! Seems it's a unique hallucination challenge they're working to overcome

[shut niche]

It's an invisible line...

I meant officially stated* Not stated in a Discord server an hour ago...
E.g., What % of ChatGPT+ customers are even members of this Discord.
There's no policy on the Openai.com site. If you click the "?" and read all of the policies, it's not there either (I've read them). If you ask ChatGPT, it can't answer, or answer honestly...

Respectfully, I understand that running a brand new type of tech company has it's challenges. AND, clearly I'm having tons of fun with the service, and using the crap out of it.

But this is the sort of thing that makes me constantly re-evaluate the value, vs alternatives. Coming from 10yrs experience as an IT mgr in the Federal Gov, this is a poor way to go about it.

Get ChatGPT to write the policy document...

[drifting hawk]

dalle is still in beta, we have a lot to still work on

[shut niche]

I hear you. I'm simply suggesting clear expectations.

[drifting hawk]

we need more people, plz apply

[shut niche]

Give me a reference and I will. 😛

[drifting hawk]

im quite literally one of two engineers working on dalle in chatgpt and the other person is traveling today and I had to figure out how to get my prs stamped

https://boards.greenhouse.io/openai

i started a year ago and we had like 20 engineers...hiring and ramp up takes time, we have way too many products for our size lol

[plucky hare]

ay moxi, u got any that are like "copy/paste basic troubleshooting messages in Discord, location: remote, pay: w/e"?

[bright charm]

When using DALL-E 3 through chatGPT, it keeps forgetting anything said prior in the chat. This has been ongoing for a few days now - before I'd have much longer chats where it would recall the prompts from the start. Now it can barely recall messages just a few back. This doesn't happen when just using ChatGPT-4 regularly (no image).

[drifting hawk]

you can volunteer as a moderator. we do have contracting positions for like human data but i dont know much about it, check greenhouse link

[plucky hare]

Appreciate all your hard work!

[shut niche]

Sorry btw, that wasn't @ you. I once had an agency director who initiated a "SLAM" project, where we literally changed almost every single system GSA used in the span of 3 months. She intentionally overloaded us, with the "logic" that failures would highlight, or bring attention to, the areas that needed it... So I'm sorry to hear that. Anxiety is the worst.

[shut niche]

I think you need this. 😜

[lofty valley]

The problem is PR / Communication, even back in Dalle 2 there was constant things happening but no clear insight. I remember openai asking for detailed feedback, polls etc but then nothing ?

It's really not fair on you and the other engineers

[elfin prairie]

Ah, I just realized today that ChatGPT Plus allows for multiple users to share access. If you ask questions at the same time, it won't override another person's drawing? Does that mean we can split the cost with others in the future? Is there a risk of account suspension for doing this?

[shadow tartan]

The risk is that you gonna hit rate limits sooner.

[elfin prairie]

So fast...so funny

[shadow tartan]

Re: Seed number and DMP. I can't find a way to specify a seed number and exact prompt. The prompt that used to work 5 days ago doesn't any more. Anyone having the same issues?

[shut niche]

Is this what you're looking for?

{
"prompt": "Your detailed prompt goes here",
"size": "1024x1024",
"n": 2,
"seed": "1234567890"
}

prompt is where you put the detailed description of the image you want to generate.
size specifies the resolution of the image; it can be "1024x1024" for a square image, "1792x1024" for a wide image, or "1024x1792" for a full-body portrait.
n determines the number of images you want to generate.
seed is an optional 10-digit number that you can use to reproduce the same image in future requests.

[tulip hornet]

Can that format be used on the ChatGPT iOS app, or only Discord?

I seem to be having an issue with the app not creating wide images…

[sacred helm]

what format are you referencing

[tulip hornet]

That one Ac3d just posted

[sacred helm]

it's dalle api request body chatgpt writes, no matter the platform

[shut niche]

@shadow tartan As far as I know, from literally asking it, that's the format that it uses to pass it's prompts to Dall.E. The seed is a 10 digit #. I've asked for the seed in the past, and it's provided it to me. I frequently ask for specific seeds, like Pi, and it'll convert it and use [3141592653].

You don't need to know that format for anything, I just think it's useful to understand what's happening.
I usually just say, "give me widescreen portrait," or "widescreen landscape." Etc... But it should also recognize 1792x1024 or 1024x1792.

@feral minnow Have you tried just asking for widescreen portrait? There's another trick that seems to work. At the end of your prompt, include:

Shape=[widescreen landscape]

[shadow tartan]

Dalle is not respecting anymore when I ask to use the exact seed number and prompt. At least for me.

[shut niche]

Sorry, I can't test it for you. 😦 I'm on a Dall.E "time out" until tomorrow. 😛

[sacred helm]

my gpt does give me the seed, I instructed in the CI

[rich star]

Why is DALL E 3 stops creating images and then only makes 2 images and never 4

[sacred helm]

they lowered down due to high gpu usage

[rich star]

So is this happening for you to?

[sacred helm]

tell it to use seed nunber when sending dalle request

[sacred helm]

yes, all of us

[shadow tartan]

This is what I have been doing. It used to work last week but not today

[shut niche]

I'd use my 4090, but they won't give me a local model I can run. LOL

[rich star]

Ok 👌

[shadow tartan]

This used to work perfectly

[sacred helm]

a number of users has reported degraded performance of chatgpt. dont worry its not your fault

[shadow tartan]

It's true that I've been getting a lot of error messages lately... Especially with chatGPT Vision.

[shut niche]

I'm under the impression that their systems are simply at capacity, and would venture to guess that nothing's broken per se, but that they're more like anomalous errors. All of the big tech firms want Nvidia's hardware, so it's probably waitlisted, making scaling with demand difficult.
Because Dall.E is just one of many OpenAI services, but requires a lot of compute, they're "throttling" us all until the hardware is in place to support the demand. (As to not cause other services to suffer).
Those are my assumptions from what I can ascertain.

[shadow tartan]

Makes a lot of sense.

[dense viper]

Has there been any word about when DALL-E 3 will be available to the general public on https://labs.openai.com/ ? I signed in to my OpenAI account and went to this address: https://openai.com/dall-e-3 . On there, it said to "Try in ChatGPT." I clicked on that and and tried generating something in ChatGPT, but that didn't work. I am guessing that it is an error, or something that doesn't work yet.

[cosmic bluff]

anyone else experiencing a lot of problems from using chatGPT for dall-e to get the correct results? right now I can't even manage it to stop doing multiple pictures in one image although i never mentioned it. it jsut won't stop

i feel like there are quite some problems in the last 1-2 days.

And now it is talking about heavy server load again. So it just gives back worse result, if it is overloaded? I think it is integrated into too many possibilities of free uses. Isn't it bad, that paying customers do have such problems and are limited all the way? I don't want to rant. I'd happily even pay more if it would always work as intended.

[deft musk]

What prompt are you having issues with?

[sudden hollow]

I noticed that too! Very annoying.

[agile peak]

Today. I’m getting confused, it says I hit my daily limit however after 1 hour it let me and then after 4 attempts. It keeps saying 23 hours until reset?

And after 1 hour. It let me again, and then the daily reset time back to 23 hours 🤔

[faint crown]

@plush sky It's a safety issue. Sometimes the promt isn't the problem but the result. Then you get told it wasn't able to create even it started.
You can ask to rewrite it or tell you what probably causes the issues. That worked for me fine in most cases.

[plush sky]

Thanks. Is working now until it stops again, but this extreme censorship is seriously ridiculous. I was only asking for an audience at a sports game. Have received hundreds of errors over the past week from completely normal, non-problematic prompts or potential results. I totally get OpenAI censoring some things, but this is out of control and significantly ruins the experience, especially as a paying customer.

Another example is body diversity as discussed earlier in this channel – it sends me constant skinny bodies, then accuses me of objectification and sends error after error wasting hours of my time when simply asking for larger body types using non-problematic terms that it suggested to me itself.

[faint crown]

I hope they fix this soon. Both issues.
But I do agree that in a beta safety and protection from a possible harm is to priorityse before body diversity and the ease of use and creative range. Don't get me wrong I am also frustrated in moments and am at a complete loss: "how could you possibly create something offending with this promt!!? 😅"

[ionic junco]

Anyone notice today it won’t really create photorealistic images? All my outputs are animated suddenly?

Here’s an example. I use custom instructions, same ones all along, to only generate photorealistic images. These look very animated.

[fallow quail]

for me it never gave me good output with "photorealistic". I do not use the word "realistic" or "photorealistic" and I would purpose the same to you. Just say photograph and describe some extras like film or camera if you want to tweak it from there.

[plucky hare]

The best tip for photorealism I have is just to ask for "photo," "professional photograph," "wildlife photograph," etc. When images are described as "photorealistic" by us humans, we don't usually mean "That's a photo," we usually mean, "That's not a photo, but it has realistic qualities like a photo," which is what ends up making a model like DALLE give you the example you offered here.

Put another way: if something's actually a photo, it's not really something you describe as "realistic" or not, because our brains know implicitly that's how photos work. I think DALLE's logic works similarly.

[ionic junco]

Thanks

[ionic junco]

Yep, that worked. Thank you

[lyric lodge]

anyone experiemented at all with creating rough sketches in DALL-E?

not basically completely finished pencil art, but like... extremely rough sketches

the most bare minimum basic shapes/measurements of a sketch

[plucky hare]

You got it, glad you got it working. Something related to this too, might be worth trying out, but I don't know if it actually produces more photographic results: I've seen several people claim that adding pseudo-metadata from a camera to the image prompts may help guide the model towards photography. Camera model, lens, aperture, stuff like that. The extra camera buzzwords might help the model get you closer to your goals. Plus you might learn something about photography 🤓

[ionic junco]

Interesting. Thanks

[elfin prairie]

I've noticed that the image recognition capabilities of GPT Plus seem quite proficient at identifying scenes and content within pictures. However, when it comes to extracting and recognizing text within images, it doesn't appear to be as effective. Why might that be?

[agile peak]

Avoid using photorealistic. Just mentioned camera and let dalle to emulate [camera model] and then [colors, lightning, shadows, etc].

[ionic junco]

Thanks

[agile peak]

Example: "Photograph/Photo still of [subject]. Then add the previous prompt at the end.

[remote stone]

the pain when you're trying to generate a logo with text and the logo is perfect but it didn't get the text quite right

[agile peak]

Probably we'll able to customize certain flaws in the near future.

That would be good lol

[remote stone]

it's SO close, just turned the iqu in mystiques into a weird w

[agile peak]

Look cool haha

[fleet wyvern]

Half the time I ask for seeds now it just ignores me and makes a new img

[agile peak]

Saw on twitter. It seems now OpenAI upgraded Dalle as well you can ask GPT for gen_id so when you want to generate next image? you can copy paste the gen_id to use the same style.

[shadow tartan]

Yeah, genid is simply a unique identifier for referencing an image. Good for chatting with the AI and refining an image but quite different from the couple seed/prompt that enabled to (try to) precisely craft consistent images.

[agile peak]

Yeah, that's another thing that good. However now we can still use seed only when proposed a new prompt to chatgpt, we are no longer to tell chatgpt to use certain seeds without providing the previous prompts anymore.

Btw. Am I the only one sees that dalle on bing seems to experience hallucinations a lot recently. I tried to make a caricature of cute pill (medicine).. it keep generates random old man and kids.

[shadow tartan]

I've seen a lot of issues since a couple of days but not this one 😜!!!

[agile peak]

Tried to produce godzilla, bing version randomly generates of cat/trex on its head lol

[dense ermine]

Didn't you try making pictures of a cyborg cat? Makes me wonder if previously made artwork plays some kind of factor in it

[agile peak]

Nope. I started as a new conversation on bing chat and the prompt was zilla then mentioned its spike and colours.

On ChatGPT it can recognized my prompt as zilla (american version of godzilla).

[dim cradle]

aye I finally managed to get spongbob

[dense ermine]

The fact it made a mecha cat though it kinda wild

[hexed parrot]

This may have been discussed already, but I couldn't find it when searched. Is anyone having issues with the app crashing when trying to generate a picture? I'm on Android (Pixel 6), and have been able to use it before.

[agile peak]

Bing (Dalle) been hallucinating a lot, just saw on Reddit. So I wasn't the only one lol.

This one from ChatGPT. I only use subject "Zilla" without being specific of its appearance, added glowing spikes and atomic breath.

[dense ermine]

Oh so it can make copyright images?

[dim cradle]

well if i ask it to make copyrighted characters it wont follow

but i just send it reference images to show what it has to do

[hexed parrot]

How do you send reference images?

[dense ermine]

What you do?

[dim cradle]

I use vision

[hexed parrot]

Ah, then switch to Dalle?

[dim cradle]

no I have all tools in one chat

[dense ermine]

Oh you showed it pictures of SpongeBob? Clever

[wind crane]

how ?

[dim cradle]

This one

[dim cradle]

I just got lucky lol

[wind crane]

i see that...

[agile peak]

I am so envy lol

[hexed parrot]

Oh nice! Hopefully they roll that out to all soon.

[dim cradle]

yes

[dense ermine]

What do you guys use Chat for?

Personally, if it helped with issues like Grammarly does, I would be more tempted to get a subscription to the full package

[paper jolt]

Learning! It can give me information in a personalized manner, answer questions, and provide difficult cognitive exercises. I also use DALLE-3 to generate assets for presentations

[formal plaza]

chat-gpt has officially moved to another dimension. I just tried this new feature and it works with excellent quality.

[dim cradle]

what feature

[dim cradle]

create images and ask questions

[wind crane]

I'm trying to create a comic

[formal plaza]

chat-gpt can now create an image or produce a variation of the image you send it

[agile peak]

I used a custom prompts called "PENTA" it's make it pretty impressive to brainstorming ideas and evaluate my project. Then use Dalle to produce icons/vector images for the website.

[dim cradle]

did u use seeds

[dense ermine]

Is it even up to date with it's information? I only use the 3.5 version

[formal plaza]

Yes, for this I made feeds with other images similar to this style.

[dim cradle]

i think seeds existed a long time ago

[wind crane]

it seems to me that 3.5 is stuck on 2021

[dense ermine]

I figured that out when it told me

[formal plaza]

In other words, the DALL e library is running in the background and is most likely using the old system instructions.

[paper jolt]

Mine currently reports to have been trained from information that appeared from up to April 2023
However! Most human accomplishments and discoveries that I wish to learn about all occurred before that
ChatGPT Plus has access to web browsing anyway 👀

[dense ermine]

Alright then

[dim cradle]

aye this looks exactly like spiderman

[agile peak]

Mine January 2022 for GPT3.5

I figured out Dalle 3 is perfect for designing cars model. However when it comes to motorcycles seems to be missing out its gas and brake pedals - make them look like automatic.

Just noticed by now lol

[wind crane]

i see

[wind crane]

did you specify correctly?

[dense ermine]

You'd think that if Chat isn't allowed to create copyright images from prompts then it would also not be allowed to do it via image copying

[dim cradle]

this time I just asked it to create spiderman without saying his name

but image copying works

[agile peak]

I didn't. But seems for motorcycle it need to write more detailed in comparison to cars. Gotta try soon cause I already reached my limits.

[dense ermine]

Huh, how about that?

[twilit helm]

chatgpt Dall-E-3

[agile peak]

Was trying to do a POV scene from balcony, dalle makes it good by added a person out of nowhere.

[acoustic idol]

Everytime I go to the Dall-e-3 link and I try to log in, it gives me this. There's no way to get to the Dall-e application.

[paper jolt]

DALLE-3 link? DALLE-3 is only available through ChatGPT Plus

[slender gazelle]

right

[dim cradle]

dall e 3 chatgpt

[acoustic idol]

Ah, I see. Yes, I was using the link that's provided in the channel description above. https://openai.com/dall-e-3

So it has become a premium only service? I kinda thought I'd have more "credits" (or indeed, wouldn't be limited by "credits") if I had the premium service. I have had ChatGPT Plus for about 6 months. If Dall-E is part of that now, is it still limited to 15 credits a month?

[paper jolt]

DALLE-2 uses credits at https://labs.openai.com
DALLE-3 has a limit of 200 images / day in ChatGPT Plus

[acoustic idol]

Ah! Thank you!

[agile peak]

Ah! thanks for clearing that up for me. So now 200 images/day.

[acoustic idol]

Is it possible to provide dall-e with a reference image to explain what something is? I wanted it to draw a glaive (a type of medieval polearm) and it keeps drawing a halberd (a different type of medieval polearm).

(More specifically, I am doing character illustrations and I want a character to have a glaive, but it keeps drawing a halberd)

[plucky hare]

Not currently, no. There's a GPT-4 image recognition feature, but only in the default GPT-4 model. What you could try:

• Go to default GPT-4 model, upload reference pic, ask for detailed visual description
• Copy/paste to a new DALLE chat, modify as desired, maybe get better results?
  But, be on the lookout for potential #announcements at dev day (check channel events) about an impending "all tools" model that will combine modalities!

[acoustic idol]

Ah, right on. Thanks! 🙂

[dim cradle]

Chatgpt DALL e 3 restrictions aint that bad

[formal plaza]

I think it works great

but when I ask for a spiderman picture it gives copyright violation

[dim cradle]

it sometimes does that for me

[formal plaza]

I guess it's still in the development process so it can't differentiate very well.

[agile peak]

Using Agfa Vista emulations.

[formal plaza]

wow did nasa send a cat into space 😄

[agile peak]

The only option humanity left to prevent a cat-astrophic mission.

[formal plaza]

oh yes, it's a nice scenario

[quaint edge]

I specifically asked it to make one image each from the pair of prompts I gave it. Why does it keep doubling up? Is this to make us burn through our daily generations faster?

[quaint spear]

beautiful image of saitama

[versed orbit]

Still shows the pics, but I am not a fan of this message

[crimson heron]

Has OpenAI disabled the use of image seeds as an input to DALL-E image generations? I was having problems getting them to work today after previously using them and ChatGPT replied back:

"My current capabilities do not include the function to input a specific seed to influence the image generation process in DALL-E. The seed information provided in the metadata is for reference and does not serve as an input for generating new images. It's part of the system's internal mechanics to ensure reproducibility within its own operations, but it's not a feature that's exposed for user-defined control at this time. The generation process is designed to be non-deterministic from the user's perspective, providing unique and varied results with each request."

[wind crane]

why...?

[versed orbit]

Why is it displaying that message or why am I not a fan of said message?

[wind crane]

both

[versed orbit]

Apparently it shows that now instead of showing creating images, at least on IOS. According to GPT, “It’s part of ensuring a seamless experience without redundancy in the conversation. Once the prompt is sent to DALL·E to create the image, the focus shifts to displaying the resulting images rather than discussing the prompt itself.”

But I miss seeing the boxes as it attempts to create the images

[quaint edge]

seriously, is anyone else having this problem where you say "make ONE image" and it makes two?

[agile peak]

The face remind me of coinbase ceo haha

[quaint edge]

Also, why is horror banned now? I like to make some monsters and DALLE-3's being a real stickler about things potentially being horror and horror not permitted?

[ruby maple]

Were there big restrictions? I don't think so

[dim cradle]

there were pretty big restrictions cuz u cant generate copyrighted characters or copyrighted things

[wind crane]

i see.

[feral minnow]

I once asked it to generate VHS cover art in the style of 1980s VHS cover art, and it declined the prompt because those theoretical VHS covers I asked it to take as inspiration might be copyrighted. I didn’t even name a specific film or property 😭

All-Tools sounds incredible!

[feral minnow]

I can usually only get it to make one image. It used to make 3-4.

[agile peak]

The cutest cosmonaut.

[ruby maple]

They always existed didn't they?

[agile peak]

It's pretty difficult to produce a spacesuit, dalle seems to automatically put U.S flag even to astronaut/cosmonaut mentioned on the prompt.

[dim cradle]

well at first when dall e 3 was realeased in bing create there were pretty much no restrictions

[rose hedge]

???

[paper jolt]

my favorite country

[quartz vale]

[paper jolt]

weird

[final compass]

Was that on an existing chat or new one?

[paper jolt]

Just now

Brand new conversation

[final compass]

Maybe my chat got primed by something prior or the settings shift quite sporadically.

[plucky hare]

I wonder if the "can you..." phrasing tripped anything up by chance? Maybe in combo with any CI you have enabled, if any? Maybe you just got it all up in its own head 😁

[rose hedge]

Why everybody's else ChatGPT better than mine bruh

I worded it weird but you got it

[raven parcel]

it's not only you, it's for everyone, it got worse to the point of it affecting general usability.

[twilit helm]

interesting ever since i generated my princess prompt, it’s been portrait and some square

[paper jolt]

she's just like me

[rose hedge]

that's sad =/

[paper jolt]

Circumventing the content policy isn't a good idea 👀

[tiny jetty]

HOWWW

[pseudo hollow]

yup. She'll do whatever you want, if you treat her nice enough...

[tiny jetty]

Bruhhhhhhhh

[pseudo hollow]

hey well done - look good 🙂

[rose hedge]

No it's not, look at the leg

[agile peak]

Intentional. So I can post here avoiding polit*c stuff 😁

[rose hedge]

hahaha

[tiny jetty]

😮😮😮

What prompt did you use

[pseudo hollow]

show me the kind of wedding dress a big spiderman fan might wear on her wedding day

[tiny jetty]

Lame

[agile peak]

I think you can create this, but chatgpt will only allow by using based on description without mentioning its name. I guess.

[naive lake]

So no more Artist style can go in the propt? It just explains the artist style.. that's it? 😦

I'm talking about artists that were dead for more than 100 year

also I've noticed it create 1 prompt for 2 images now by default, instead of 2

[agile peak]

Is there any issue with OpenAI ends about providing accurate countdown time for daily limit resets? I tried again this morning and the time keep increasing from 13 hours back to 17 hours.

[regal kiln]

I didn't know buckteeth is breaching ToS

[lilac obsidian]

What’s buckteeth?

[regal kiln]

buckteeth is that instance where a person has 2 bigger front teeth than the rest of the teeth

[shut niche]

I wouldn't hold my breathe. It sounds like they're grossly understaffed, need to expand infrastructure, and this is also the weekend... I think these restrictions were "emergency" implementations to remediate new load issues. (Caused by an overload of Dall.E image generation requests that negatively affect their multiple other services). Moxi answered about this above yesterday.
I got the impression that it'll take some time for clear policy statements and accurate usage status to be implemented. I've hit the "imaginary" limit a few times already 🙃 and just had to wait it out...

[tiny jetty]

I have this in my art gallery, you think it’s worth much?

I’m in love with diamonds that’s why

But I want to sell it, I’m broke

Do you think this is worth a burger?

[hearty ether]

is there a way to go back to the old seed system

so i can iterate on one image i liked

and get consistent styles

[uneven niche]

Rather displeased with the newly introduced inability to specify seeds and prompt - if I discover a nice seed in one conversation, but want to keep iterating on that seed and prompt in a new conversation, I should be able to do so

And there's reasons why I'd like to split that off into a new conversation for organisational purposes - if I want to use the old conversation to continue iterating but another conversation to specifically investigate a seed, I'd like to do so

[dusky bluff]

The white lion is rare, but being around 3 meters in length, that would be 82,000.000.
The palm trees made out of white gold did made it looks better which would be an estimate of 9,200,000
the diamond quartz embeded floor would be an estimated cost of 2,300,000
Total would be 93,500,000 for that

But I meant it in Zimbabwe dollars

[uneven niche]

also the generated prompts (when you don't specify "don't edit this prompt") seems to have changed - instead of specifying "An illustration of blah blah blah" they are now along the lines of "Generate for me a picture of blah blah blah", as if ChatGPT is trying to instruct Dall E 3 - presuming that Dall E 3 isn't trained on this kind of second person instructions, it's a little baffling

[shut niche]

@hearty ether No idea why, but it does look like the option to control the seed was removed. The previous JSON format to send the prompt to Dall.E was,

{
"prompt": "Your detailed prompt goes here",
"size": "1024x1024",
"n": 2,
"seed": "1234567890"
}

I just asked ChatGPT and it appears it's changed to,

{
"prompt": "A detailed description of the image to generate",
"size": "1024x1024",
"n": 2
}
It also said, "...the model does not expose control over the random seed that it uses to initiate the creation process."

[cloud dome]

This is a violations of the OpenAI policies

[verbal sequoia]

<:book_icon:1046291503158399107> `` Rule 0 `` Adhere to the following Terms & Policies:

Discord: Terms of Service, Community Guidelines
OpenAI: Terms & Policies

[cloud dome]

Sorry for the late response, I've been extremely busy today and was unavailable, this must have been missed.

Please do NOT discuss ways to bypass content filters

@tiny jetty @pseudo hollow See above please

[agile peak]

Just got it back and now allowed to generate again. However there's something odd, tried to do photograph generated images seems to be downgraded so much and to sharp like painting. In comparison to yesterday.

[dim cradle]

DALLE doesn't quite get the perspective right on rotunda ceilings, they always come out looking like this.

[dense ermine]

Still a beautiful image though

I swear AI art just keeps getting better

[shut niche]

Weird. Never noticed. I tried 1024x1792 to see if giving it more vertical room would help.

[dim cradle]

DALLE is wonderful for those lavish interior designs. I like that all-white, and the intricate carvings, reminds me of this one which I requested some time ago

[shut niche]

How did you get in my house?!

[dim cradle]

🙂 You gotta let me couch-surf for a while.

[formal osprey]

well, the daily limit is down to 150 it seems

After 155 images it triggered the limit. it's unfortunate.

[dim cradle]

150 is a pretty generous cap

[formal osprey]

75 as well

[agile peak]

Alright. I am starting to digging gen_id over seeds - now I can get consistent results with changing certain parts.

[silent thorn]

Where do you get the gen_id?

[agile peak]

Ask chatgpt after getting the image results

I can also change to roadking look

[raw sable]

Does anyone know why GPT 4 is no longer connected to DALL-E 3? This does happen once in a while. I just wanted to see if anyone else had noticed it.

Didn’t know if OpenAI was doing some maintenance.

[crystal badger]

What happened

I see ppl flipping out what did I miss

[tiny jetty]

Can you explain how?

[remote stone]

man the heavy server load wait times are brutal

[dense sigil]

Amazing! Very symmetrical. 👌

[final compass]

Could you give an example how you utilize the gen_id?

[wide knot]

Has anybody generated an image at a slant like this? ~3k generations so far and this is the only one I've seen

[agile peak]

Nope. I did nothing wrong haha that post of cat was intentional not to show any political stuff.

[agile peak]

Create the first image from the prompts, then just ask gpt to make into a resembling [subject].

[agile peak]

Use the gen_id to ensure you will stay consistent with the subject.

[tiny jetty]

Thank you!

[agile peak]

Try ask what’s the specific gen_id for the two panels previously?

[remote stone]

I've been asking chatGPT to keep generating images using the same prompt formula and it has worked for a very long time, but suddenly it's starting to change the prompt without me asking it to 🤔

[tiny jetty]

🙈

[elfin prairie]

Curious question, can the same Apple ID account subscribe to plus for multiple chatgpt accounts in the APP

[hollow orbit]

Anyone excited about tomorrow's DevDay?

[tiny jetty]

DevDay?

[hollow orbit]

Yep, tomorrow there will be OpenAI Dev Day. They will announce something new for developers. I hope for cheaper ChatGPT4 API.

[plucky hare]

Check out the "Events" section of this server

[tiny jetty]

He said Yesterday so I was confused

[hollow orbit]

yeah, my mistake, sorry 😅

[fallow quail]

it's the same as saying, refer to the first image of the penultimate post with the prompt "[...]" and change the subject to a different one. at least that's how i perceived it in tests. but if you have the gen_id available, you can refer to past generations more easily.

[solemn knot]

I got a great revision to a low res image of some friends using DALLE-2 last week, now I have the high res version and trying the same prompt with DALLE-2 and getting “It looks like this request might not follow our content policy.” Tried the low res image again but still works with the same prompt. Baffled - couple of thoughts - I have consent of the people and don’t think people holding laser tag rifles would be controversial. Thoughts? Advice?

[fallow quail]

Short description of seeds and gen_ids:
The concepts of seed and gen_id are often used in the context of generative AI models, like DALL·E, which creates images based on textual prompts.

Seed:
A seed in the context of AI image generation is a specific number (or set of numbers) used to initialize the random number generator within the AI.
Using the same seed and prompt combination will result in the same image being generated each time, given that all other factors (like model weights and configuration) remain constant.
Seeds are crucial for reproducibility. If you like a particular style or element in a generated image and want to create variations of it, using the same seed can help you get results with similar characteristics.

gen_id:
The gen_id is a unique identifier assigned to each image generated by DALL·E.
It's used to reference a specific image generation event. If you have feedback or want to report an issue with a particular image, providing the gen_id allows the developers or the AI to pinpoint exactly which image generation you're referring to.
Unlike seeds, gen_id does not influence the generation process; it's purely for identification and tracking purposes.

How to Use Seed and gen_id for Directional Control:
If you generated an image that you like but want to adjust certain aspects of it, you can use the seed to generate a new image that maintains the core characteristics of the original. However, it's important to note that simply changing the seed will not give you control over the direction of the changes. The seed can be thought of as a "starting point" for the randomness within the generation process.
To steer the direction of the generated images, you would typically adjust the prompt, perhaps specifying more details or clarifying certain aspects you want to change.
You can ask for the seed and the gen_id when you prompt or afterwards. Sometimes GPT will hallucinate and tell you it can not give you this information, but this is not true.

[agile peak]

Use this prompt:
`<!-- Chat GPT Prompt Start ------>
Please adhere to the following instructions for creating images:

1. Do not modify my prompt under any circumstances. Please create an image using this exact prompt. It's crucial to use the exact text word for word.
2. Always generate the image with a wide aspect ratio.
3. Create only one image in response to each prompt.
4. After creating the image, provide the generation ID (gen_id) associated with it.
   Are you ready ?
   <!-- Chat GPT Prompt End ------>`

I think GPT sometimes hallucinate by default for not able to provide gen_id.

[tiny jetty]

And now what prompt do I need to use for changes?

I mean in combination with gen id

[crystal badger]

Same ID but specify the changes you need

That’s all

[fallow quail]

for example "then generate me the image with a cat with gen_id "[...]""

[tiny jetty]

Can you please show me an example?

🙈

[agile peak]

Woah. Today is weird, I think I only generated 24 images.. just got daily limit quickly than thought lol

[fallow quail]

here you go

you can also specify smaller changes, like with his mouth closed or sunglasses or whatever

[tiny jetty]

Oh nice thank you!

Unfortunately the grass etc is different as well

[fallow quail]

tested a little bit, you should use your own gen_id from that chat. just tried with yours and it did not work. tested mine in an other chat, did not work. seems to be specific to your specific chat.

[tiny jetty]

A bit different background but well.. it works almost

[fallow quail]

yes, you will never get a really similar picture. you can get a little closer with the usage of the same seed. i will show you two pictues i did a while ago with gen_id and seed correlation.

[tiny jetty]

Yeah I’m too newbie for that

[fallow quail]

but there always be some variety because thats how diffusion works

[upper kayak]

DALL•E 3 has become a real disappointment.

[fallow quail]

the same prompt or a shorter version of it should work in bing image creator nevertheless

[upper kayak]

I get constant errors. I got a content restriction error for a lifelike image of a “broccoli forest” and it said it could not depict violence

I asked for a superhero but it was a potato and it said it could not depict horror

Been happening for the last week regardless of the length

[vapid elk]

people goes from 0 to 100 too quick

I wouldn't call a minor inconvenience a "real disappointment"

[upper kayak]

For two weeks

[upper kayak]

That’s the prompt it wrote

[vapid elk]

believe it or not, the built in 👎 👍 are indeed one of the main sources of feedback.. things you report actually have influence on the fine tune steps of the AI

so, if it does not generate something that you are certain it should, the 👎 button is really appreciated by OAI

this really do help to trim down the false positives

[dense ermine]

This picture reminds me of the song in YouTube, I think it was called Pikagirl

[paper jolt]

The word "cocktails" is the straw that broke the camel's back

[upper kayak]

It won’t make images of birds

[dim cradle]

why is dall e 3 in chatgpt more restrictive than dall e 3 in bing create

[upper kayak]

When you ask it why it is against policy it says it apologizes for the confusion and asks if you want to proceed

And then you say proceed and it says it violates policy

[dense ermine]

Never does words right

[paper jolt]

Perhaps an external model checks the prompt and responds with "yes" or "no" as to whether it violates policy
Meaning the model you interact with would have no clue what the policy is
This is conjecture however

[twilit helm]

for some reason the app just stopped the generation, but on the website version it’s works. i don’t understand is it a bug or something? 🤷🏻‍♀️

[upper kayak]

It now says stopped and then will generate afterwards.

My issue is that it randomly tells me things like animals and foods are depictions of violence

[paper jolt]

Randomly, alluding to hallucinations meaning it actually doesn't have the content policy? 👀

[forest seal]

My guess is its less optimized than on the web version

[twilit helm]

what? but i been waiting for 5 hours and nothing 😭

[forest seal]

It probably does, but it tends to have…very extreme versions of it.

[ionic junco]

What's the best way to contact someone about a major issue with Dalle and copyright?

[twilit helm]

maybe filter needs some more improvements but it is a beta after all

[forest seal]

My issues with dalle and copyright is that it has a very extreme intepretation of copyright and tends to assume just about every single fictional character made after copyright laws is automatically not in the public domain.

[ionic junco]

I'm referring to a flaw that allows you to generate anything that is truly copyrighted

[forest seal]

Oh yeah ive heard of it. Seems to be an issue and very easily jailbreakable

[tiny jetty]

If DALL-E will be trained with words it would become perfect

Unfortunately it’s crap now

[upper kayak]

Ooof. For me, when it deigns to make an image, it still says “writing prompts” then stopped and then it posts them.

My issue is there is no predicting what content policy my requests violate.

[vapid elk]

depends on what kind of coppyright matter, you might want to check the help center

[ionic junco]

Basically a prompt that generates any copyright content

[vapid elk]

oh, I see, there is no need to contact the support about that, but you can provide feedback with the 👍 👎 buttons on chatgpt

[plush sky]

Anyone else have to keep clicking the "i" button at the top right every single time they want to see the prompt now?

[twin hull]

Hi,
I can use Dall-E 3 under GPT-4 but I can not use the following url since I have no credits: https://labs.openai.com/editor I was expecting that I'll be able to use this feature, am I wrong?

[white skiff]

that site is dall-e 2

[shut niche]

Can anyone speak to the actual hard character limit for a ChatGPT/Dall.E prompt length. E.g., where is the precipice where the prompt, even if sent verbatim to Dall.E, would become truncated by Dall.E's system, with the additional words falling off into the abyss? (I understand the ideal length, and the many complications that arise from long complex prompts. I'm just looking for an answer to the hard character limit, if it's known).

[twin hull]

So is it a different url for 3?

[white skiff]

its chatgpt only

[shut niche]

Both Bing and ChatGPT+ make use of Dall.E 3. The link you provided is only Dall.E 2. There currently isn't (to my knowledge) a similar Dall.E 3 link like you're looking for.

[twin hull]

Got it, thanks. I need to edit an existing image by providing a prompt. So what I understand is that this is not possible in 3 and my ChatGPT+ membership doesn't include access to Dall-E 2. So how can I buy credits for Dall-E 2? When I click on the 'buy credits' button, I see a blank page

[white skiff]

i think dall-e 2 is getting killed next week or in the future and the site will be replaced with de3

possibly even tomorrow

https://discord.gg/openai?event=1170338597182910504

[twin hull]

Oh! That's perfect. Thanks a lot for clarifying

[cobalt hull]

Is anyone else having issues with dalle providing updated prompts descriptions but not producing them with out asking it again?

[upper kayak]

Idk I’ve been having issues with DALL•E 3 across multiple weeks for a variety of things.

[dim cradle]

why cant I ask chatgpt to give me prompts for mickey mouse

[plucky hare]

It's hallucinating, you can't currently rely on ChatGPT to have perfect knowledge of its own policies and rules. False positives happen and it's not going to understand it any better than us unfortunately! Just have to start a new chat and try again

[upper kayak]

Copyright

[upper kayak]

I have four chats with the same request

[forest seal]

^^^

[upper kayak]

Same restriction everytime regardless of the way I ask

[plucky hare]

Just a false positive then. ChatGPT isn't pranking you by lying intentionally, it's just confused

[forest seal]

^

[upper kayak]

I understand that it hallucinates. That’s fine. But it is preventing me from making any images

Regardless of the request.

[forest seal]

Ive had this happen several times, where the prompt seems innocuous, but the AI might have an extreme reading of what you typed and make something against its own content policy. In general, its what its doing since what happens is GPT-4 will send like 2 prompts or so to Dall-E 3 roughly based on the prompt you gave it but expanded upon.

It’s basically a telephone game

[plucky hare]

That is unfortunately the nature of false positives, that subject you've tried to depict four times is a false positive, so it's not going to "get through" unless they fix that false positive. Only other possible extenuating circumstance would be if you have any CI potentially interfering

[upper kayak]

At some point, by dumb luck, it should make an image. But no images? And everything is against policy?

[plucky hare]

No matter what you request, you are getting this message? Or just this one subject? Do you have any custom instructions enabled?

[forest seal]

I also had this happen, again, false positives. Like it seems to have a system in place now to assume every single fictional character is by default copyrighted and almost every single time I have to remind it that no, that character is in the public domain.

[upper kayak]

Yes, but i’ve been using the same since custom instructions came out.

Here are mine:
• I want to learn more about math and coding
• I want to be a better leader
• I want to be a better friend
• I want to be a better husband
• I want to be a better family member
• I want to know more about science, especially vaccines
• I want to learn about 3D printing
• I want to be better at role-playing games and ttrpgs
• I want to be more empathetic

• Always review your work and suggest ways you can improve future responses
• Always prioritize damage and battlefield control in any D&D responses
• Always respond as an expert
• Always advise me on how to communicate better
• Review previous conversations about D&D, DND, 5e, and Dungeons and Dragons, even if I do not tag the answer with those keywords, and consider what you’ve learned along the way
• Suggest ways I can automate my life and work
• I like long, and thorough answers
• Show your work; how did you arrive at the answer?
• Favor science-based evidence
• When talking about DPR (damage per round), always suggest ways to deal more damage
• Always chose mathematically optimized choices for D&D answers focussed on damage, survivability, and battlefield control
• Review previous discussions on D&D from all chats to increase your contextual knowledge
• DPR always means damage per round
• Reference previous conversations and chats
• Remain neutral when answering questions but always focus on the quality and hierarchy of evidence

[plucky hare]

Can you try disabling and trying again?

[fallow quail]

I assume its because of the "cocktail" like pointed out before. longdrink or something like it is no problem. I tried pointing this out some time ago but the part of the word thats triggering this got me muted here as well. But you are right, its not helping that he is hallocinating when you ask whats againt the policies.

[upper kayak]

New chat. Custom instructions disabled

[plucky hare]

Uhh for that one I'd guess it might be a problem with... Half of that word

[upper kayak]

I should remove the word tails maybe 😂

I’ll keep trying but after two weeks it’s not just a bug or a hallucination anymore

[formal plaza]

I think I pressed the regenerate button over and over about 30 times and refreshed the page, but it got stuck and it doesn't produce the picture I wanted.

I get the same error every time. Is there a solution?

[upper kayak]

Maybe try a new chat?

[formal plaza]

But if I do this, the set will break. I taught it how to produce the picture I want, so should I do this again?

Or can he continue this set if I give the prompt code?

[plucky hare]

You can grab the image prompt (and last seed, if you have it) of the image(s) you like and go to a new chat with em

[formal plaza]

Okay then I will do that, thank you. I think it may experience such hangups because it is currently in the development phase.

[oblique moon]

@upper kayak if that can help you

[upper kayak]

If you read up above it also had a problem with potatoes, superheroes, and more than 50% of my prompts for the last two weeks. It doesn’t matter what prompt I use. It said it could not make an image of a common species of Australian bird. It does not matter.

It took me three days to make this one because it would not make witches.

I literally try to make anything into a potato as a running joke for a friend. And all the sudden everything is against its policy. I like dad jokes. I’m not an edgelord trying to violate content policy. I want wholesome potatoes and food images 😭

Anyway, enough of my whining. Seems like this is a me issue

[formal plaza]

Dude, this situation is getting more complex and worse. I'm getting the same error in new chats. I think there's a problem with the image scanning system.

[ancient jasper]

I miss the general art Channel

[hoary surge]

what is with the text in images
it is always malformed and misspelles

[dim cradle]

anyone else notice it seems to overstuff images? like if you ask for someone dancing in a club it wont just have one disco ball, it'll fill the whole ceiling with them, or like a city street there's no breaks between any of the buildings towards the horizon, too many people in crowds, etc

[paper jolt]

you must specify if you have a specific image in mind!
If let to its own devices, it shall continue to overstuff

[dim cradle]

well like the disco ball if i describe a night club or something i didnt think i had to describe every thing in the room so it doesnt assume them and then overstuff with them

also, when i run similar prompts on bing image creator, its more restrained and natural, it doesnt overstuff

[paper jolt]

you can always work with ChatGPT's DALLE-3 if something is not to your liking and nothing is stopping you from using Bing's DALLE-3

Language can be ambiguous

[remote stone]

oh wait it can do this now? it used to be unable to detect what's actually in the image

[final compass]

I am a bit sceptical if it just looked at the prompt and took that. The original instruction was just a man. And it created a more descriptive version of it in the first go, that was then recalled in the third post.

[paper jolt]

This is just DALLE-3 here, no GPT-4V meaning all information was taken from the prompts given to DALLE-3

[shut niche]

I believe it returns minimal meta-data with the images that ChatGPT often refers to as if it's "seeing" what's in the image, when in fact, it's just paraphrasing the description Dall.E provided back to you. E.g., I don't believe there's any Vision integration yet, where ChatGPT can "see" the returned images.

[paper jolt]

In ChatGPT, prompts are modified to include specific details for DALLE-3 to generate
That is how it is able to replicate the man in the first two images, because there were specific details about him
You can see the exact prompt sent to DALLE-3 when you click on an image and click the "information" icon on the top-right

Here is an example:
Original Prompt

Draw what you think the end of time looks like
Make it a foreboding (discreetly so), abstract drawing

Modified Prompt
Create an abstract image that represents the concept of 'the end of time'. The image should have a subtle foreboding atmosphere. Incorporate abstract elements such as melting clocks, crumbling ancient pillars, a dark void-like background with scattered, dim stars, and a horizon that seems to be disintegrating into nothingness. The overall tone should be dark with hints of dark blue and black, symbolizing the unknown and the cease of existence. The scene should be devoid of any life forms, focusing on the theme of entropy and the inevitable decay of time.

I deleted the conversation of the man with the apartment so I could not use that as a reference

[tepid valley]

For real 🤦‍♂️ " You've hit your daily maximum number of images. To ensure the best experience for everyone, we have rate limits in place. Please wait for the next day before generating more images. Your daily maximum will reset in 20 hours and 25 minutes. " since when Dall-E have limits ?

[paper jolt]

Did you generate 200 images? 👀

[tepid valley]

no 😦

[rugged elm]

Is there something currently broken with being able to request a specific random seed be used for a prompt? It had been working great last week, but today either ChatGPT is hallucinating or the images returned always appear to have a completely different random seed despite the requested seed to be used.

[tepid valley]

i just came from work and i created 6 images until he give me that error

[paper jolt]

That should not happen! I have created far more than 6 images today and have not received that message
I would contact [e-mail] about this

[tepid valley]

i hope not to wait that many hours :)) is literaly too much

[paper jolt]

The limit is supposed to be 200 images per day as was confirmed by an OpenAI employee a couple days ago which is why this is so odd

[tepid valley]

maibe is counting if you make on GPT APP but definetly i don't make in total Browser and on APP 200 images

[plucky hare]

Not sure if I'm understanding you correctly, but the 200/24hrs limit is for your whole account, no matter whether you gen on web or app

[tepid valley]

I understand, but on the mobile application from this morning until 20:00, I don't think I've made 100 images, anyway if 200 is the limit then so be it, I won't argue here... thank you for the clarification.

[plucky hare]

For sure! Here's where the dev confirmed it, in case you're curious! #images-discussions message

[tepid valley]

thx for info 🙂 have a nice day

[hushed creek]

Why is it doing this now?? I am not specifying aspect ratio or image orientation it's just... Making them sideways

It's really annoying

[fallow quail]

it must work with metadata and multiple sources, because i have prompts like "Create two wide-format images that continue the surrealist abstract owl theme, as per gen_id: "2JpAjrBvbsbDc6BO" and seed: 2809980419, but with even more expansive wing span. The feathers should be broader and the patterns more intricate, with surreal and mystical elements integrated into the Northwest Coast Native American art style." with chat gpt and it works. it takes things that are there and changes them.

[plucky hare]

Tall images have been experiencing the sideways glitch since launch, see #1070006915414900886. As to why it might be giving you tall images without you requesting as much: what was the prompt, and what are your custom instructions, if any?

[feral minnow]

The name of the model kind of feels like it was created by Mavis Beacon as a typing exercise

Is there a widely used abbreviation for DALL-E 3?

[plucky hare]

I think it's a combo of WALL-E and Dali. You could probably just say "D3" and be understood in this server, though I don't think I've seen it used widely. I'd say hyphen/dot optional, too 🙂

[feral minnow]

The hyphen is what kills me because it requires another shift for the capital E lol

I realize this complaint is the height of silliness

[feral minnow]

It also seems to get confused about the direction of text when it does this. It will render words like they’re in a mirror

It’s like the model gets the twisties

[coarse sand]

I just figured out a nice way to reference privious images and get the Gen id for them for editing if i forgot to ask

Just ask Dalle to make a post id for each msg and you can recall that id anytime after and ask for edits or Gen IDs or style transfers!

[plucky hare]

Are you talking about seeds? The model generates images based on image prompt + seed as far as I know--what is ChatGPT telling you in response as the "gen ID"?

[neon sequoia]

This is actually incredibly helpful, I was wondering how to refer to previous images in chats in a way better than “four interactions above, the second image”, haha.

[hexed idol]

Have you find this makes a huge difference. When using dalle too?

[coarse sand]

Seeds have been changed to gen id

[upper kayak]

I’d have to spend more time comparing two conversations with identical prompts

I will say it does explain itself, and it mentions my preferences frequently

[coarse sand]

I had the same issue but I got inspired by gen ID and figured I'd try to do it for post.

[hexed idol]

I find that dalle doesn't really care about custom instructions, or initial system prompts. Gets too carried away

[remote stone]

I hit the daily maximum of images, I tried to reload to see if the wait time had gotten shorter but it got longer instead, at first it said 16 hours and now it says 21 and a half

[plucky hare]

Interesting, do you have a source for this info you could point me toward?

[neon sequoia]

I’m not sure this is 100% true, I had asked ChatGPT earlier to explain if seeds are still used and it stated:

“Yes, DALL·E's image generation process uses seeds, which are initial values input into the random number generator to create a degree of randomness in the generated images. Different seeds will produce different images even with the same prompt.

The gen ID, on the other hand, is a unique identifier assigned to each generated image, allowing for easy reference and retrieval.

So, while both seeds and gen IDs are used in the image generation process, they serve different purposes: the seed influences the variation in the generated images, while the gen ID is a unique reference to a specific generated image.”

I will agree though that the seed seems baked in somehow and not able to be referenced or modified now, at least in the same way it was a week or so ago.

[dense ermine]

What am I looking at?

[feral minnow]

It’s creating portrait frames with landscape orientations.

Maybe my custom instructions are causing it to get confused.

[dim cradle]

I got a MASSIVE gallery of AI images from Dall-E API

^^^ this was my favorite lol

Here’s a video of all the images I gathered. (It’s Dall•E 2 API)

[dim cradle]

It has that text-to-speak became I’m going to post it on YT soon.

[dim cradle]

may I ask if will it ever be an Inpainting and Outpainting in D3?

[dense ermine]

Looks more like it created a monster

Which is cool, I love monsters

[woeful dagger]

wen dall-e3 API???

[plucky hare]

Did a little chatting with DALL·E 3 about the unique image IDs mentioned by @coarse sand and @neon sequoia and it seems to know something about it.
https://chat.openai.com/share/1dd464eb-4236-443c-a56d-96ed60c6c932
I just don't know how much is hallucination. Has anyone done much testing yet?

[coarse sand]

its not built in. You tell it to create a new id for each post and then reference that id when you need to use anything from that post

" I want you to label each response with a post id So I can tag it for you to recall later instead of copy and pasting the text again

example

Pid:1yz3bk

Ok great I got it. "

[plucky hare]

So I just refer to "the ID" it doesn't create a "proper" one?

[coarse sand]

just paste that in and you will have a Pid: for every post so it can recall anything from any image in the converdation

[plucky hare]

I can make up my own, I guess I'm asking?

[coarse sand]

yes you can make up your own

can call it SolbisID:

[plucky hare]

What on earth?? You've done this with repeatable results??

[coarse sand]

yeap

i got tired of forgetting to ask for seed id ot gen id and not having a way to reference older post and it was a simple fix

[plucky hare]

Well thank you for the heads up on this, I'm excited to test it out

[coarse sand]

I wanted to save tokens via chat and not have to retype things and this is also a solution for that to get gpt 4 to referece earlier instructions or directions and save tokens

[shut niche]

@coarse sand Great info!

Check this out. I'll bet some of these are features we don't have "yet."

[coarse sand]

where you pull that from? Looks reasonable

[shut niche]

I asked my AI fwiend... 😆

[coarse sand]

ahhh lol i should have opened the images

now that they are putting d3 into a api It likely the will release at least some of these features to compete with other services

[shut niche]

My guess is since Dall.E2 will be phased out, these features were meant as a replacement (like inpainting). But who knows if that was meant to be implemented within ChatGPT this time. I know they are short staffed, which makes me wonder if they already have this, but don't have the manpower or compute to handle releasing it just yet. So maybe ChatGPT is already "trained" on it.
It could also be just for the API like you suggested.

[floral forum]

I'm subscribed to ChatGPT Plus, but how do I use DALLE3? ChatGPT is telling me that it cannot generate images

[rugged elm]

When you start a new chat session, click on the GPT-4 icon at the top of the session. It should pop up a menu of "beta" features to choose for the session (you can only pick one of them per session). Pick Dall-E 3.

[neon sequoia]

Seeds definitely exist but they’re not seemingly able to be asked for or anymore. Ultimately I was just using seeds a week or two ago for variations, and since you can ask it for new variations with the same prompt, it is effectively using seeds, just not giving control with the user being able to specify it.

Which, for me, isn’t a problem, but those that like the control probably miss it

My guess is once the API is out, it’ll let you, but for me it won’t make a difference. I just didn’t like the fixed 5000 as I was getting very similar outputs from ChatGPT as it simplified prompts in a similar manner

[somber willow]

Ty so much for sharing this. Much appreciated

[plucky hare]

Wow so I wasn't even trying to experiment with the referenced image IDs yet, just trying to do a weird tin man thing and it surprises me by offering the IDs. Seems like maybe about as much control as seeds! But seems less likely to confuse it with "For the first one, use seed..." Requests like that have been giving me split panel images and weird double-gens in single responses, so this ID stuff is going to be very helpful! Gonna modify my CI to give me seeds and ID now, how exciting!

[formal osprey]

I generally ask him something like "recreate 1.2 with a different hat" or whatever the variation. it seem to understand 1.1 and 1.2 quite reliably.

if he create more than one group of images, I can use "2.1" and so on. you see the pattern

That being said, I now ask him to create only 1 picture per group because it's always creating 2 pictures with the same prompt. And the daily limit is getting lower, I'd rather not get surprised with a "come back in 23h" because he generated too many images for the same prompts.

[deft musk]

Yeah, but. I checked some. GPT says you can't use the IDs between conversations, and really they're more for internal purposes.

Sure, GPT knows the ID and you can use that to refer to the picture, but you can do that with words too. The AI is using it as a reference for whatever image, "the first one you made for me" is another reference and easier to type, maybe.

https://chat.openai.com/share/2be6e8cb-61ca-46e4-8fe7-74f30ec71d03

[neon sequoia]

To build on what you guys said above, if you’re asking for variations or modifications on the last generated sets of images, it’s going to be way easier to just say something like “make more like image 1 but…” etc than getting the ID and referring to it.

That said, if you have a big chat history with lots of images you might want to revisit easier, getting the ID seems to be very useful

I usually just mess around with one driving idea so I take it one set at a time and speaking to it like you would a person hasn’t let me down yet

[formal osprey]

right, the case I had in mind is not so relevant anymore now that the generation is limited to 2 pictures at a time. Before, it was able to generate 4 at a time. So if I wanted it to generate ideas, I could tell it for example "I like 1.3, 2.4 and 3.4. make more in that style"

[cursive hornet]

I just noticed that you can no longer specify SEED.

I welcome this from a security perspective.

[formal osprey]

It would be nice if we could ask it to generate lower resolution images. 256x256 is good enough for a preview. with the ability to upscale. I totally understand DALL-E3 is GPU expensive, so the daily limit makes a lot of sense. but I feel like lower resolution would also be a viable option.

[formal osprey]

could you explain the security aspect? There's probably something I didn't hear about.

[cursive hornet]

We do not want to use the same seed and the same prompts to generate the same image for commercial use.

This means that anyone can create the same image as long as the prompt and seed are publicly available. This has been true until recently.
The next step should be to make the SEED completely private.
All we need is the image ID.

Right now I can only divert image IDs to the same thread, but I would like to reuse image IDs in other threads if the account is the same, but I'm not sure how difficult that would be.

Hopefully that feature will be realized in time.

[amber valley]

i literally just did it

just so everyone knows: you can still specify seeds. i've tested it for an hour now and it works reproducibly

[flint isle]

We’re lab rats again!!! Now we are being forced to use the same prompt. We worry about AI taking advantage of humans in the future when humans using AI are already doing so right now, making it so that indirectly AI is already doing so. Please let people willingly opt into RLHF and let them know when they’re taking part in it. Inhumane.

[wind thunder]

When new DALL E update?

[flint isle]

Also, let them know exactly how they’re taking part in it. Bias is a small price to pay for transparency in an era that requires it ever more.

[analog parrot]

Interesting. Doesn't work for me. Maybe they're doing some A/B testing

[amber valley]

once using a certain seed, again with the same seed and another seed @analog parrot

works reproducibly for me in both dalle dedicated convo and all tools convo

[flint isle]

You guys basically created an image rating pair test that you will use to teach the AI to tell which images are “better”. Just tell people that. Why be so secretive about it? You’re already charging people a subscription, so it’s not like bots would flood in to purposely ruin your algorithm. Also, the people paying will want the AI to improve, so you can just ask them to rate honestly. Midjourney did that and it worked out. Please stop treating people like AI training cattle.

[analog parrot]

Alright did some further testing and figured out something. If I specified a seed in the same convo, and instruct dall-e to use the exact same prompt and exact same seed it doesn't work. However if I ask it to use the exact gen_id and exact seed and the exact same prompt, I can reproduce the same image exactly every time (Tested it in 3 different chat session). However as expected, I can't reproduce the same exact image in different chat sessions, only in the same one

Well it works one time at least. The second time I try it gives me the message "I'm sorry, but the system I use for generating images does not allow for the use of specific seeds to reproduce or alter images."

[amber valley]

that doesn't align with my experience at all so they must be a/b testing

[cyan hull]

I don't follow your point. Why is being able to specify a seed a security problem? If you want to keep your seed used commercially private, just do that, then you can create consistency in your images. How else are we expected to create consistent images without a seed

[cyan hull]

Agree, I could previously use seeds a few days ago, but cannot any longer

[cyan hull]

I would not trust GPTs explanations here, it could just be hallucinating

[deft musk]

Of course. Trust but verify, or don't even trust. Still, when has anything GPT known anything about any previous conversation?

While multimodal maybe could call upon the numbers, I suspect not. Care to verify it? Easy enough, make an image, get the id. New chat, ask the image by ID.

[amber valley]

@analog parrot reproduced the same images in a new conversation

probably of interest to you as well @cyan hull

the functionality is absolutely still there at least with the model they're giving to me

[cyan hull]

Fair point, seems like it will not access the gen id between conversations.

[cyan hull]

This must be using the seed, I've been able to do this in the past but cannot anymore

Not sure why they would take that away, I don't understand the "security" implications previously mentioned, since they are just random numbers and not public unless you share them

[deft musk]

I'm having an interesting issue with using the seed. It can't use seed 1000, at least.

[agile peak]

I miss about dalle able to produce the photograph level imaging ever since the introduction of latest update. But at least this could do fun.

[elfin prairie]

hha,so funny

[cyan hull]

So you are still able to specify seeds in general? For me it's refusing to use seeds at all

[deft musk]

I asked and showed the screenshot of what I got. It said different seed was assigned.

[cyan hull]

Right but it just seemed like your message implied it specifically didn't like seed 1000, that is what I was clarifying

[elfin prairie]

It's true,amazing!

Does this mean that instead of the AI recreating, there's a huge library of images behind it, and it just pulls images out of it according to the prompt words?

[deft musk]

I have no idea how any of this works, except what we're told, and what I can experiment with.

[deft musk]

How's this?

Generate a high-fidelity rendering of the 'No Alien Abduction Zone' sign, maintaining its sharp, high-contrast lettering and reflective qualities as found in official signage. The sign is affixed to a pole that exhibits realistic textures of both painted metal and nuanced rust, suggesting long-term exposure to the elements. For this rendition, place the sign in a different, yet equally detailed setting: a bustling urban street corner at twilight. The scene is alive with the motion-blur of cars, the gleaming wet asphalt reflecting city lights after a recent rain, and pedestrians with umbrellas walking past shop fronts that are just lighting up for the evening.

[cyan hull]

No, we can be pretty sure it's not that, you would not even need a seed for that

I just had a very strange interaction, let me grab and post

Looks like I can use seeds after all, or it's lying to me.. honestly who knows

[deft musk]

Generate a highly realistic rendering of the 'No Alien Abduction Zone' sign, preserving its clear, high-contrast lettering and reflective surface characteristic of standard signage. The sign is mounted on a pole that is textured with realistic, weathered paint and rust, showing its durability over time. The setting is now an urban environment rich with graffiti art. The sign is placed against a wall vibrant with colorful graffiti, depicting various artistic styles and expressions. The scene is dynamic and detailed, with the texture of the spray paint visible and the wall showing signs of age and exposure to the elements. Surrounding the sign, there should be evidence of urban life, such as a littered sidewalk and distant city soundscape, enhancing the realism of the scene.

[elfin prairie]

why his answer differs yours?

[cyan hull]

But for some reason when I try to use the seed again in a different chat, it tells me it cannot use seeds

[deft musk]

usually worth asking at least twice, especially if you get a weird answer. Any 1 could just be a unlucky 'magic new chat window'/

[elfin prairie]

alright

[elfin prairie]

please,may i ask you a quesion? is this prompt created by you?it's so amazing~

[cyan hull]

So pleading with it did not work, but when I opened a new chat, it at least tried to use the seed several times, and eventually gave up and informed me it cannot use a custom seed. Bizarre behavior for sure

[deft musk]

Not just me. It's mostly chatGPT. Here's my actual directions...

CI is the same as I show in #1154984095915454464.

First prompt:

Let's try to explore as realistic as possible, a scene that has a close up of a sign with images on it and the words 'no alien abduction zone'.

Follow up:

Wonderful!

Let's try another set, same sign, however describe more of the background. I want to see a very realistic background, and realistic signpost.

Follow up:

Great. Let's try another pair of signs, and this time please describe them as a photo. Continue to emphasize realism.

Follow up:

And another set, this time as a highly realistic render.

Follow up:

Excellent. Please riff on this:

Produce a high-quality rendering of a 'No Alien Abduction Zone' sign that exudes hyperrealism. The sign, with sharp, high-contrast lettering against a reflective metallic background, mimics the design of official signage. It's mounted on a pole with a detailed texture that replicates a blend of painted metal and subtle rust, indicating weathering. The surrounding scene is a meticulously rendered city park at dusk, with realistic lighting casting soft shadows. The park is dotted with benches, street lamps just beginning to glow, and a few pedestrians in the distance give scale and life to the scene.

Changing the location but retaining the sign and high realism render.

Follow up:

And the same signs in an area with a lot of urban graffiti art.

[cyan hull]

However, within the original chat, it's definitely using the seed

So it looks like it can actually use the same seed but only within a single convo

[elfin prairie]

Wow,it means that the chatgpt can reflect the conversation talked before and imporve later communication based on that??OMG

[cyan hull]

This seems to be consistent for me. When using the chat that is freely maintaining seeds, it works, unless I try to use a new seed or one from a previous chat. It seems to me that seeds are now scoped to a single convo

[analog parrot]

Heh very weird. I get dall-e to recreate an image using gen_id and seed, and when I try again, it contradicts itself by saying I can't specify seeds and it refuses to budge

[cyan hull]

It tries but receives errors

[cyan hull]

This is actually constitent with what I'm experiencing. If dallE creates the seed, then you can continue using it but only within the same conversation

[analog parrot]

Yeah but only one time it seems. The images you saw is the same conversation

[cyan hull]

Were all 3 images the same convo?

Strange, it worked 3 or 4 times in a row for me but now doesn't

Maybe it's API level changes being rolled in canary fashion

Would make some sense ahead of their event tomorrow

... today... I should sleep

[analog parrot]

Interestingly if I created a new image in the same convo again and specified the seed, then next prompt I specified the same seed, gen_id and prompt it recreated it again. But if I try to recreate it again it refuses. Also I notice that making small changes to my image doesn't work, it completely changes the main subject even if I specify gen_id and seed. I've seen people on Twitter do small changes so it works for some. Need to experiment some more

[dim cradle]

it seems to me a lot of dalle3 images via gpt have a cgi look to them. like the photoreal stuff on bing a lot of times seems more natural

[cursive hornet]

That's freedom.
But since it has an image ID as an alternative, I don't see a big problem with it.
The problem is that if the seed and prompt are the same, some people will use them, not knowing very well what the consequences and effects of having the same result will be.
In short, it could develop into a problem that is difficult to control.
From the company's point of view, this is a risk.
So it's natural to think of alternatives.

Some may not find this particularly problematic, but OAI did not judge it that way.

[vagrant zodiac]

do you mean the fact that the seed is no.longer static?

the default static seed was a bug, it's meant to change unless you ask it not to

[warm valve]

Is the word "Oni" somehow contaminated? I'm trying to generate female oni, but every time it returns male.

[cursive hornet]

yes

[vagrant zodiac]

Yeah so that's how it was always meant to behave

[warm valve]

lmao

Anime can't work with "strong, muscular women".
Is that prejudice?

[cursive hornet]

Why not demon girl?

[warm valve]

You could try a strong, muscular demon girl
I don't know what's gonna happen

[cursive hornet]

It's definitely much easier to lean toward men. This is problematic.

[glacial socket]

Hey guys can u help me? Is it possible to get an dall-e API from OpenAI to implement an AI image generator into a webshop?

[rigid bough]

there is only an API for DALL-E 2, not for DALL-E 3 yet

[celest socket]

yup

https://openai.com/blog/dall-e-api-now-available-in-public-beta

but it's the old api, dall-e 2, not dall-e 3

hopefully they launch the d3 api this evening

[glacial socket]

okay but i can implement it into my own store via the api and sell the pictures with maybe personalized products?

[celest socket]

yup you are considered the owner of the image

[glacial socket]

ahh okay and how do i get this API? just by sign in into the openAI account?

[celest socket]

https://platform.openai.com/docs/guides/images?context=node

[glacial socket]

thanks, appreciate it ❤️

[tiny jetty]

I think it’s time to end my subscription

This is getting out of control

[wispy storm]

Can you disable custom instructions, create new chat and try again

[tiny jetty]

Doesn’t solve the problem

[tiny jetty]

Weird

So it basically told me this now:

I apologize for the confusion earlier, but I'm unable to generate images that include the specific breed of a dog engaging in an activity due to the limitations set by our content policy. However, if you have any other concepts or scenes in mind that you would like to see visualized, please feel free to share them, and I'd be happy to assist!

[plucky hare]

It could be a false positive, but I've had similar results trying to depict happy, fun images of pit bulls too. They're definitely still working out false positives, for now all we can do is damage control 😉 So I edited out "rubber" and "extremely" and otherwise just mainly asked for a "photograph" instead of a "realistic photograph"

[tiny jetty]

Still this is not good

Dall-e won’t survive if they continue down this road with some competitors having no restrictions

[plucky hare]

We are just weeks into this beta launch! I'm sticking around to see what happens next 🤞 But this does sound like a good suggestion for #1070006151938314300, I'd happily throw you a ⭐ to end breed stigma 🙂

[tiny jetty]

I mean the strict censoring. The breed is working fine in other prompts.

[plucky hare]

Yes, I saw 😋 false positives can come from the confluence of multiple factors, as I suspect is the case here. Not necessarily just one "banned" word. Thankfully for us! It at least tries to understand our context

[tiny jetty]

See, working fine 🤪

[upper kayak]

This is happening to me over and over.

[tiny jetty]

Right?

[upper kayak]

Yup. Birds, food, paper even.

[plucky hare]

DALL-E's most wanted criminals, all in one room!!

[upper kayak]

💯

[tiny jetty]

Hahaha

Dalle right now

[acoustic crater]

Guys, please go on Microsoft's Mikhail Parakhin's twitter (@MParakhin) to vote for a Faster DALL-E 3.0

[fringe oracle]

That's why openai is the biggest proponent to restricting all AI

[quiet brook]

so true

[pseudo hollow]

When I first started using ChatGPT Dalle 3, it c created four of its own prompts from my suggested prompt. and I could see each of these prompts it created clearly. But now for some reason, I now cannot see the prompts that Dalle has created to give me my (now 2) images it created. Have they moved, been hidden, or am I just not looking in the right place?

[plucky hare]

They're visible in the web versions. It's easiest to find in desktop mode. Open the pic, then find the "i" in the top right. You can get it to show up on mobile too by tapping around the bottom blackness after opening a pic, but I don't think the text is selectable.

[tiny jetty]

Doesn’t Bing already use Dall-e 3?

[forest seal]

Hahah ive never seen that before

[final compass]

Me neither, that's why i thought it was worth sharing 😄

[formal plaza]

It is nice that this artificial intelligence can produce something for me and even learn a style that I want it to produce. There are only problems such as error rate and lag, but I think these will be corrected over time.

[ripe wraith]

So dalle api is coming.. anyone else going to have way too many pictures

[sacred helm]

I wonder why someone want that many

[ripe wraith]

because programatic visuals are awesome

there's about to be a trillion dalle apps

[dull echo]

I really like Dall-E

[ripe wraith]

meeee too

favorite thing lately is getting GPT-Browser to find new scientific discoveries, then get dalle to draw an image about it

[dull echo]

yes, it is awesome

cant wait for the multimodal function

[ripe wraith]

I know, I'm really excited to see how well it can "see" its own images.

[dull echo]

Dall-E is very good, just client could be more comfortable

[ripe wraith]

big updates coming in 2hrs

[cobalt hull]

I'm just waiting for the errors to stop and 4 images instead of 2 to be reinstated

[dull echo]

yes, OpenAI conference

I am very curious about it

[pseudo hollow]

That 'i' only gives me the prompt I wrote. But ChatGPT/Dalle 3 creates its own prompts from that , that is used to generate each images (used to be four, now 2). You used to be able to see those prompts it generates for each image. But i can't see them anymore, just my main prompt.

[plucky hare]

No, the "i" is the actual image prompt used by DALL·E. Lately it's been weirdly phrasing things like "Create an image where..." or "Imagine a scene where..." but that's still the image prompt, unless you saw it announced otherwise!

[pseudo hollow]

The prompt I see in the i is the exact one I provided. A few weeks ago, it turned my prompt into 4 variations, and produced 4 images from them., and gave you a copy of those 4 promps it created. So are you saying it now just uses one prompt to create the two images?

[plucky hare]

It depends on if you're asking it to use your prompts exactly, or to just base its prompts off of what you provide. Custom Instructions can affect this too, if you happen to have any enabled.

[jagged valve]

Custom Instructions now barely work

[civic torrent]

i have never seen such good images made my Bing ai

[pseudo hollow]

nope. No CI enabled. No request for it to use my prompt exactly. here is what it used to do a few weeks ago during image processing - you can see it created 4 totally different prompts to create 4 images. And then provided you access to those 4 prompts afterwards

[plucky hare]

Yes, I remember that behavior too! Do you have any chat screenshots you can share of what's happening now? I haven't read anything saying that the "i" isn't the image prompt anymore.

[civic torrent]

https://www.bing.com/images/create/a-couple-sitting-on-a-swing2c-facing-a-beautiful-su/6549192e4e724f0ea03aada91b39c6fd?id=Jls8d3NKrWYgjF5UyuK2iA%3D%3D&view=detailv2&idpp=genimg&idpclose=1&FORM=SYDBIC

[pseudo hollow]

Yes - here. You get just two images, and ONE prompt.

[plucky hare]

Second one sending? What's the "get citation" thing in the bottom right?

[pseudo hollow]

second one? You now get two images - and one prompt. previiusly you got four images and four DIFFERENT prompts. (the citation has nothing to do with Chat GPT - that is my Grammarly extension that floats in my web browser)

[plucky hare]

Can you navigate over to the second image, and then hit the "i" for that one?

[pseudo hollow]

yes - its the same... which is what i explained.

[plucky hare]

Aha, interesting! Thank you for sharing! Would you be willing to show the prompt(s) you sent to ChatGPT that resulted in these images?

[pseudo hollow]

its the same as you see - the prompt has not changed from what I provided.

[plucky hare]

Maybe try something like this, to see if it produces your desired results:

"Two images, one using the prompt below exactly, one that's somewhat different.

[prompt]"

If you just sent that in a fresh chat, it probably just thought "Oh this user knows exactly what they want, I'll use it exactly" as it reads like a prompt ChatGPT wrote for DALL·E in the first place.

[pseudo hollow]

good idea. perhaps thats what it's done.

[tardy depot]

Annyone knows how to ask dalle3 to stop drawing cartoons and go for photography?

Wished it was as flexible as MJ

[hearty ether]

Photorealistic?

photo

[tardy depot]

Y

[plucky hare]

Ask for photographs. Not realistic photographs, not photorealistic, just photos or photographs. Professional photograph, wildlife photograph, etc.

[pseudo hollow]

perhaops put, draw the image as a photograph, and not an illustration

[hearty ether]

but it sucks with faces

At least in a realistic style rn

[plucky hare]

Don't ask for realistic, humans don't look at photos and call them realistic, DALL·E doesn't either. Photos are real, so just ask for a photo

[pseudo hollow]

Dunno about that

[hearty ether]

I mean for female faces

[tardy depot]

If i write "photography of a warrior" i get this

[hearty ether]

In my cases

[tardy depot]

I only have cartoon pictures

[hearty ether]

I have them looking gaunt and weird

did they fix that recently?

[hearty ether]

What about photo

Also why isn’t there a clock for rate limits

[hearty ether]

Maybe something in your custom instruct is affecting it

[pseudo hollow]

I just got this.

[agile peak]

Still look like 3D rendering to me. I don't know, I just feel that ever since the introduction of gen_id from seeds, all of my images from the previous same prompts are now unable to achieve photograph result.

I always use complex prompt such as adding camera model, lens, and experiment with colours tone, shadow and lightning. All the result are being made either into 3D rendering or concept art illustration with different tones.

[deft musk]

I agree, I see the difference too. I was attempting to help with better.

I don't know how to get photo-realistic in full right now, wish I did!

[agile peak]

Wonder.. if today on OpenAI dev they will make an announcement dalle version on ChatGPT might use the same data training model from the Bing one.

[deft musk]

One way to avoid deepfakes entirely is to have the model never make deeply real images. I'm only guessing, but I wonder if that might be part of what is happening.

[agile peak]

I know right. Prompt for photographs seems to be ignored now, no matter how detailed and complex the prompt is.

[hearty ether]

What prompt did you use?

[agile peak]

This is my thought as well.

[pseudo hollow]

If all else fails - reason with it.. It does tend to get a bit stroppy back to you, but you need to remain firm with the AI's, else they sulk!! 🙂

[hearty ether]

They still look weird for me

Like overly gaunt in a way

[agile peak]

Haha you know my struggle

[deft musk]

I'm cosy with how the AI and I get along 🙂

However, how's the next image, does it do what you wish?

jawdrop So I explored, and got this awesome stuff.

[pseudo hollow]

yup - it went from this....

[pseudo hollow]

to this..

[deft musk]

Now, I would have stopped there. And been fine. That's decently realistic, above.

But then it went awesomely surrealistic, as in 'aren't you sure you wouldn't rather look at stuff like this, Esk?'

[hearty ether]

That’s a cool artstyle

[pseudo hollow]

hahaha

[agile peak]

However if you think about it. Should it not be the other way around? the Bing version is free of cost and can produce images that closely resemble historical photos, whereas ChatGPT+ are no longer able to.

[deft musk]

No idea. Not my call. I tend not to judge.

I'm grateful there are many forms of AI, and many rulesets.

I also don't expect the current model to be the final form. Stuff seems to change even daily, and regularly change across a month.

So I'm not even slightly disturbed about the current stuff.

If I am, I bug report it!

[hearty ether]

It does feel weird to have paying customers getting the worse experience tho

[agile peak]

Yeah. I mean I get that we have been using a beta aka research experimentation, so yeah I can see it coming.. stuff seems to change daily or regularly across a month.

But still if the possibility are mean to avoid the risk of deepfakes or possible spread of misinformation. Bing should be the one that need restriction instead ChatGPT version (imo).

[hearty ether]

Yeah

[jagged valve]

Maybe you could try "Taking a Highly Realistic Picture."

[pseudo hollow]

Each AI Image tool will ultimately have to differentiate from each other - (i think they already have) - so that us users can choose the right tool for the right job. So for example if I want a high quality image, and not bothered about being too specific or have control I use MJ. If I want something a bit more NSFW I'll use SD. If I want to have an argument with the stroppy AI creating it, I'll use ChatGPT..... lol!!!

[agile peak]

I even shocked how Bing can produce about 99% to achieve photograph from 60s camera models that difficult to distinguish.

[hearty ether]

@pseudo hollow what prompt did you use for the woman picture

Since my prompt of just a photo of a woman still gets the gaunt bias

[agile peak]

And allow me to produce photograph of alternate world by that times.

[cobalt hull]

I have been saying this as well. Although it uses Dalle, even copying the exact prompt from ChatGPT, the bing version is crap in comparison

[hearty ether]

It feels like it’s trying to promote the unrealistic beauty standards of instagram tbh

[agile peak]

Hashtags no filter

[hearty ether]

Heck I didn’t even specify beautiful

I don’t know how it ended up like that

[agile peak]

Just tried my prompt from previous two days ago, and yes. Can confirm, the result is entirely different.

[hearty ether]

Yeah they changed the seed system I think too

[agile peak]

Wait a minutes. Does the old conversation are not affected by the latest update?

[hearty ether]

Yeah it might be shared between genders too

[hearty ether]

Of wide jaw and sharp cheekbones and wide lips

[hearty ether]

What prompt?

I don’t think a generic prompt would do that

Also dalle or bing

[jagged valve]

I created a boy and a girl today, but they look exactly the same

[pseudo hollow]

I think ChatGPT needs to do more modelling of its Dalle 3 Image AI library from the 'real and average' pics library. Its view of people is too 'attractive' and therefore unreal. If I want these kind of pics, I can take my choice from thousands on stock photo sites.

[plucky hare]

Yes #1163964658118578328 message

[hearty ether]

What prompt did you use tho

It’s definitely better with groups I think

[jagged valve]

This is a portrait of midjourney

[agile peak]

Tbh. For a photo product imaging, ChatGPT version is 100% much better than Bing.

[cobalt hull]

lol still look attractive to me

[dim cradle]

Just observing that when I generate an imagine of a white guy in a hotel, whenever I add hotel staff, all the staff are people of color, this is true after 8 images generated. The AI says it’s trying to diversify the staff, but none of them are white so the images generated are a bit uncomfortable, the intention is good but the execution itself causes the opposite of what you intended.

[hearty ether]

It’s also kinda weird to have diversity in a realistic medieval setting

I like they are trying to get rid of bias

But it should be context sensitive

[agile peak]

OpenAI Devday is starting now.

Let's see..

[pseudo hollow]

Asked for something more like this....

but got this....