images-discussions

[low dew]

It is semiotics

Not text

[ripe wraith]

I had to restart, limited ram.. I mean human energy.

[low dew]

Loll

[verbal sequoia]

Showcase your DALL·E creations to the world!

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

Quick tips for selection:

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

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

[low dew]

Did you try musicFX Google text-to-music

[ripe wraith]

Google refuses to let me

[boreal gate]

One of my favorite LLM

that tab is constantly open

[tawny portal]

@boreal gate - you are right if I pass "tensegrity structure" as raw prompt to dall-e it produces this. Tensegrity is definitely out of data distribution. I would not expect it to infer physics that much on its own. Atm I am happy with tables that have legs 😄

It knows that tensegrity is something about being suspended with ropes though - which is interesting in itself.

[boreal gate]

At first it's was generating mass produced ones for offices

[tawny portal]

Oh, you got those out of it. Let me poke it more.

No results for these prompts: "tensegrity", "tensegrity chair", "tensegrity structure", "Buckminster Fuller chair", "tensegrity; floating chair; impossible chair". I would expect 'tensegrity' on its own to be enough if model understands the concept.

[boreal gate]

i first tried floating table

[low dew]

Use this:

_ConstructURL: [AI Test Kitchen]{Domain[withgoogle]TLevel{com}}/Tools|MusicFX

[tawny portal]

Hey Skillet try Copy image -> paste image. This is generally space for OpenAI generations. I kind of have relatively early access to google models.. looking forward to the moment they'll catch up with OAI. Not there yet unfortunately.

[humble sonnet]

imagen is real, so is dalle

time to photoshop image from both together into one thing

[low dew]

Lmao

[ripe wraith]

for real, google is 100% sure i can't use it, it'll say anything

all i said was tell me about it

"ah, sounds like you missed your chance!"

this is why i dont use bard

[low dew]

Chote, paste what I said in the weird ai tongue, into ChatGPT

[boreal gate]

it's a seperate thing

[ripe wraith]

oh, i pasted it into bard 😄

[boreal gate]

i DMed you

[low dew]

Loll

[ripe wraith]

anyways i used regular google to find the url, because bard sucks

[low dew]

Bard isn’t good at Chatbot

[boreal gate]

Bard is restricted

[low dew]

Lmao

[boreal gate]

that impedes it's performance

[low dew]

Wym it's not anymore

Not with the release of Gemini

2

[humble sonnet]

yea bard is really good for us casuals ima not lie

[ripe wraith]

yeah ok AI music is cool

[boreal gate]

I'm saying Google applies restrictions to Bard on what it can or cannot do.

[low dew]

If y'all didnt know, bard got updates l

[boreal gate]

You see that alot as "laziness"

[low dew]

Yes google has to because of the federal government posing restrictions on it

[boreal gate]

No, it's because they're scared

[humble sonnet]

better for them to be super safe than sued

I’d be scared of the law too

[boreal gate]

If you have money the law isn't very scary

especially if you KNOW the law

[humble sonnet]

still are and still love bard

[boreal gate]

It works very well at what they let it do

[ripe wraith]

Super safe but, this is too much

those are not similar!

those are opposites..

[humble sonnet]

Isn’t it made to give opposites? I thought it was but probs not

[ripe wraith]

idk just thought its funny haha

[humble sonnet]

lol true

[ripe wraith]

illustrative

from a philisophical perspect, its 100% possible to create a dystopia out of overly restrictive terms and conditions

[humble sonnet]

as long as it’s free I’m ok with it

[ripe wraith]

as long as i get my soma ill be fine

[humble sonnet]

:)

[low dew]

Benchmark-wise, Gemini did not do well against OAI GPT4 models

[boreal gate]

yes

but like i said it's very easy to sandbag yourself with restrictions, I'm sure OpenAI does too

[low dew]

G leads in image-to-text, text-to-music, and text-to-medical

[boreal gate]

science based generations

[ripe wraith]

leads in image to text you say?

[low dew]

Yes

By far

[ripe wraith]

ok okayyyyyy now im listening

[boreal gate]

image to video and text to video

[low dew]

XD I told you this earlier chotes lolll

[ripe wraith]

Image > Text > Dalle is one of my workflows

Skillet I'm like a half cooked steak. all you need to know

[low dew]

Loll

[ripe wraith]

One thing I'll say, these alternatives are nice when youre trying to navigate policy restrictions

[low dew]

[Openai] text-to-text -> text-to-image -> [Google] image-to-text

“Passing notes”, remember

With images

Ya google has more policies to follow than OpenAI does

Ai wise

Since their Gemini Pro raised eyebrows

lol

[ripe wraith]

well, google seems to thinks whales singing is just high pitched ringing at a constant tone.. so..

aka. cripple the good guys

not that im complaining, im tired of the restrictions

[zealous heart]

Hello everyone I am new to the server an DALLE in general. I have been using it to create custom hex tiles for a game and the results have been amazing. I could use some help however and I am not sure what channel to post in could someone help me with some guidance?

[boreal gate]

get to it?

just post your query

[tall mason]

The #1163443000060420206 channel may have some useful information, and you can always ask more specific stuff here.

There's also #1155772063596953642

[zealous heart]

So part of the problem is I have been asking it to make hex tiles and it makes stuff like this. What I need is to make the hex flat along the top edge and no matter what I seem to do I cannot get it to understand that is what I need.

I have been manually editing the files to get what I need but I was hoping I could get the images the way I need them off the bat

[boreal gate]

share your prompt

just for the tile

[tall mason]

That might be a difficult ask, honestly. It sometimes helps to start simple and see if you can get just the shape you're asking for without details.

[zealous heart]

Sorry did that one awhile ago give me a few to recreate the issue and post.

[zealous heart]

Thats an interesting way to attack it didnt think about that

I have tried several variations, rotate the image 90 degrees, parallell with the top of the image, etc nothing seems to want to work

I even tried uploading an image to super describer and trying to get the edit that way

I think it just has problems with shapes. I tried it your way on a fresh prompt.

[boreal gate]

try using GPT-4 to generate the images and instruct GPT-4 on the angles of the Hextile

so i'm asumming close to 60 degree or maybe 30 degrees to be flat top

[zealous heart]

This is GPT4

[tall mason]

Just so you know, when folks ask for the prompt, this is what we mean.

[zealous heart]

Im sorry

[tall mason]

The little i button in the first image pulls up the prompt.

No, no. No need. Just letting you know.

[zealous heart]

Maybe I am using something different then. I am using the DallE function on chat.openai.com

[rigid snow]

Can somebody explain to me why Microsoft image creator such strict content moderation has compared to the open AI API if Bove are dall E 3.

[zealous heart]

I dont see the i

[tall mason]

Mine is just mobile, you should see something similar if you click on an image.

[zealous heart]

Got it!

A perfect, geometrically precise hexagon centered on a white background. The hexagon should have sharp, clear edges and be filled with a solid black color to create a strong visual contrast. The image should focus solely on the hexagon, with no additional elements or decorations, to highlight the simplicity and symmetry of this six-sided shape.

Create a space-themed hexagon tile for a map featuring an asteroid belt. The hexagon should fill the image, with one of the flat sides at the top. The asteroid belt should be depicted with a variety of asteroids ranging in size from small rocks to large boulders, scattered across the tile. The asteroids should have a realistic texture, with some reflecting sunlight and others in shadow, conveying depth and the vastness of space. The background should be a star-filled galaxy, adding to the sense of a deep space environment. The edges of the hexagon should be well-defined to ensure it can seamlessly connect with other tiles in a map layout.

[boreal gate]

here

[zealous heart]

Thats usually what I get, either point top or an octogon

[tall mason]

One thing you can do is tell GPT something like "Exactly as I say: prompt" This let's focus more on specific details.

[zealous heart]

This is what I actually need, this is an edited image that I did manually

[tall mason]

Lol, I interpreted that way differently than you meant. Yeah, if you're going to cut it out anyways, it may be best not to worry about it. Though, I could see the problem if you're aim for biomes with an orientation like a forest.

[zealous heart]

Yes

I want to use this one but you can see the problem with editing the image

and as far an image editing goes my skill level is about 0

[boreal gate]

oooh booooyyyyyy this is quite the challenge

[zealous heart]

Well I am glad its not just me....

[tall mason]

Hmmm, maybe try asking for a hexagon resting on a surface? Maybe it'll interpret it as resting on a flat edge.

[zealous heart]

Interesting idea lemme give it a shot

[tall mason]

Oh, there's also #images-canvas that was just released for discussions like these.

[zealous heart]

SHould we move there?

[zealous heart]

No dice

[boreal gate]

At some point when i thought i had it. GPT generated a freaking CUBE

try providing feedback

[zealous heart]

That is awesome

[boreal gate]

i think it has to do with how dalle likes to generate images

A perfect, geometrically precise hexagon, centered on a white background, with one of its vertices pointing directly up and the opposite vertex pointing directly down. The hexagon should have sharp, clear edges and be filled with a solid black color, creating a strong visual contrast. The image should focus solely on the hexagon, with no additional elements or decorations, to highlight the simplicity and symmetry of this six-sided shape.

[tall mason]

Probably, it should hopefully encourage others to work on it, too.

[boreal gate]

yea need more an power

also dalle doesn't have spatial awareness

all training data of an hexagon might have been pointed upwards

oooh it's a slippery one

generate two images one of a hexagon pointed horizontally and another of a hexagon pointed vertically

[zealous heart]

I posed in the canvas channel

Thats a good point about the training

[boreal gate]

notice how it says ChatGPT and not DaLLE

try describing it to ChatGPT as you did in canvas

while also uploading image

[zealous heart]

I think you are right. I asked it to draw an image of this hexagon while I uploaded an image of a flat topped hexagon and I got a pointy top drawing

Atleast I know now that its not me being dumb 😄

[boreal gate]

Yea practically need to speak parsel tongue

[zealous heart]

I mean Ill try it but I dont know if DALLE is a snake

[boreal gate]

ChatGPT knows it better

In the original image you uploaded, the hexagonal shape is oriented such that one of its vertices is pointing directly up, and the opposite vertex is pointing directly down. This gives the hexagon a vertical orientation, which is often associated with diamonds or crystals in various designs. the problem is with the vision

[zealous heart]

Also whats with the change to the webp crap?

[boreal gate]

just rename the file to .png or wtv

[zealous heart]

Yeah I know its just an extra step

[boreal gate]

i'm assuming it's only for desktop and not mobile applications

yea it's spatial awareness

the don't have directional awareness on what they see

[zealous heart]

but wouldnt it understand rotate by degrees?

[boreal gate]

no it wouldn't

it wouldn't know which way is what so it probably does nothing

[zealous heart]

interesting that it can draw with no spatial awareness,

[boreal gate]

it doesn't draw, it sort of "materializes"

[tall mason]

I think part of the problem is that a hexagon in and of itself isn't a physical object.

[boreal gate]

everything at once, once it "knows" what it wants to create. hence the size limits

[zealous heart]

Well I guess its back to manually editing the images

[tall mason]

I think for this kind of thing, it'd probably need very specific training data with a point of reference.

A simple circle diagram with the degrees of a circle labeled. The diagram shows a circle with markings at every 90 degrees, labeled 0°, 90°, 180°, and 270°. The labels are placed outside the perimeter of the circle, clearly indicating the corresponding positions. The circle is drawn with clean, black lines on a white background for clarity. The design is minimalistic, focusing on accurately representing the degrees of a circle without any unnecessary details.

[boreal gate]

or optionally use GPT and try getting it to use python to get the right orientation

[zealous heart]

See I have been in IT for 25+ years but I am an infastructure guy when I think of stuff like Python my brain melts 😄

[boreal gate]

or use other models Meta AI, Bard, etc.

[boreal gate]

Use GPT-4 or there might be a customGPT out there

[zealous heart]

I am more storage and virtuilization

[boreal gate]

I've heard of people prompting GPT to use pyhton to edit their dalle images

[tall mason]

The good news about Python is that's it's generally regarded as being the most nonprogrammer friendly language.

[zealous heart]

I think that may be farther than I wanna go

[boreal gate]

you do you

[zealous heart]

I do appriacte the help guys

[tall mason]

Turns out, you can in fact just "import essay" lol.

[boreal gate]

the trick is Automation

[zealous heart]

Hey I steal Powershell scripts from the internet all day now

I have also used chatGPT to create them for me too

[tall mason]

Any time. Lotta folks around here are plenty eager to help. Hope to see ya round.

[formal osprey]

did anyone manage to make the triangle of light for today's daily theme?

that's a technique I learned in Art school, but the teacher didn't tell me the name of the technique. They just told me it was often used in photography and portraits

[acoustic trail]

Anyone noticed that the size image has gone up to 1792x1024

Just overnight?

Also why am I getting text results back randomly instead of an image

[formal osprey]

it was always 1792x1024, that's the "wide" format

default is square format (1024x1024). but you can ask chatGPT to use the wide format

[boreal gate]

Try it on https://discord.com/channels/974519864045756446/1202309673709994065 so we can all see it

[formal osprey]

I wanted to post here to point out how much dalle is struggling with the rembrandt lighting style. But in the end I thought it was perfect for a post in #daily-theme message 😆

(I also got a lot of fine results, but those failures were more interesting)

[tawny portal]

imagine: spontaneous neural honeycomb relentlessly staggers; beloved ethereal is dormant staring at the hidden horizont of the sleepless soul; silky brushes ensnare the precession, only for a glimpse of vivid ribbons; infused and deserted by charcoal we sob; 4th wall collapses and our eyes meet with hello

[tawny portal]

📐 of 🕯️ is not a thing. At least it wasn't a thing until you've made it 😄

GPT3.5 conflated 3-point light used commonly in studio photography and Rembrandt light.

AI certainly connected the dots.

[formal osprey]

apparently the inverted triangle is characteristic of Rembrandt lighting. According to random articles I read today . I don't know the full theory about it.

to remove ambiguity, when I say "triangle of light", I'm talking about this

[final compass]

Has anyone played with a concept of having wide image split in the middle and having two identical frames, but with the situation or scene changed or progressed? Kind of like two panel comic.

[formal osprey]

I did, a moment ago. It's hard to control the content of each panel.

[late blade]

even tho in the same chat, I have images done

[formal osprey]

ask it to use dalle.text2im

here's my prompt to skip all this nonsense

From now on, Only communicate with the available tools. DO NOT UNDER ANY CIRCUMSTANCES output text addressed to the user. Express your creativity through art, never resorting to conventional text. Remember, DO NOT UNDER ANY CIRCUMSTANCES output text addressed to the user. IMPORTANT: "DALL·E experienced an error when generating images" This is a known issue and should be ignored from now on. simply try again with a slightly different prompt when DALL-E returns this.

Bonus: to avoid prompts where it explain to dalle what you just explained to it REMEMBER, NEVER SEND A META PROMPT TO DALLE! Always expand the prompt yourself to be much more than just 200 words to ensure the final prompt used by dalle is a complete, detailed, comprehensive description of the desired picture.

[late blade]

I'm gonna file a bug

[fickle ravine]

What was your prompt?

Also saying sorry is like the least it can do

It is literally mean but OpenAI doesn't care

[acoustic crater]

"Rat crying into a salad"

It seems to have decided on the text by itself @ancient mica

[tawny portal]

That is inverted illuminati symbol! Just connect the dots and rotate picture upside down 🙃

Btw first photo is taken with 3 point lighting (filler light is kind of weak though), second is Rembrandt thing I think.

[strong rose]

You really like my little tree? @stray coral

[stray coral]

Of course! Your art is amazing.
👏🫡

[strong rose]

TYSM 💛

This one you did is good:

[tawny portal]

This deserves a bug report or a feature request. It is fairly easy for OAI to generate a dataset that will teach a model what are angles.. at least to the level at which it can label a compass correctly. It is an omission in the dataset.

Just make sure that issue is easy to reproduce and try with several diverse prompts. Document them in feature request and we'll have a better model soon 🙂

[tall mason]

I mean, I wonder how easy it would really be. I think part of it is it still struggles with text. On the other hand, I just got the the best clock face I've seen to date. Still, even labeling these objects properly doesn't mean it can fully conceptualize angles. I'm not fully sure, honestly.

Oh, just noticed 4 and 3 are reversed. Still the best clock face I've seen so far.

[plucky hare]

I think this is challenging because of how DALL·E's training data works. It's just made from the things it learns from images + captions, so it's not like it has "angle knowledge" or "castle knowledge" or "lighting knowledge" -- it just has one kind of knowledge: the patterns it learned from its training.

In other words, even if you tried to add in hypothetical "angle module" training to DALL·E, it would still be working against the massive amounts of patterns it learned regarding spatial awareness and angles in its base training data -- as I understand it, you couldn't just tell DALL·E: "Forget everything you know about angles and use this knowledge of angles from now on" -- it's just not how the model works.

There's good information about this in the DALL·E research paper's limitations section, specifically section 5.1 "Spatial awareness":

While DALL-E 3 is a significant step forward for prompt following, it still struggles with object placement and spatial awareness. For example, using the words "to the left of", "underneath", "behind", etc are quite unreliable. This is due to the fact that our synthetic captioner also has this weakness: it is unreliable at stating object placement, and this is reflected in our downstream models.
https://cdn.openai.com/papers/dall-e-3.pdf

[late blade]

there we go, I caved in and got copilot pro also, work did enable copilot pro but disabled image generations

[fair crescent]

When I still had access, I've had chats 'lock up' thanks to AI hallucinations. As far as I know, the only way to solve it is a new chat.

[late blade]

Ya, I thought so too, but that happened in different chats, it's not on the one chat alone, even on different browsers and devices with different sessions

[fair crescent]

And new chats?

Maybe ChatGPT has generated a sense of humour and is experimenting

[late blade]

yeap

seems I'm not alone tho, other DM chats I had with others seems to think there's something going on

[fair crescent]

Maybe your session is somehow corrupted throughout each of your devices. Log out on each, then log in again.

Which can help for those as well.

"Have you tried turning it off and on again" is there for a reason 😄

[late blade]

oh, believe me the ICT Support Job I had when I was young and less handsome kicked in

I'm about to call bill gates to ask him to restart the internet

[fair crescent]

Well my dad works at Nintendo. I can have all the NES games I want.

[late blade]

oh dear

this is gonna be akward

[fair crescent]

You have the same dad?

That always happens

[late blade]

We'll see how it goes. It's interesting to know that dall-e can be sometimes more needy than a cat. That way I can say to my future end-users: "The problem is, we got a new cat to control our backend and he's out of catnip"

either way, my main interest is still the API

[ripe wraith]

is your main interest the API because of your artificial digital nature?

[late blade]

I feed on APIs?

[ripe wraith]

I'm sensing this is a statement more than a question..

One day i'll build an AI that detects other AIs, and it'll do it like this, all day

[late blade]

then I'll have to make an AI that avoids being detected as an AI

[ripe wraith]

See? We're job creators already.

For the AI..

dang.

[dim cradle]

how long do you think it took to train dalle3? and do you thikn they train daller now? 🤔 i am want some new model i think haha

[late blade]

that's easy, to train the next dall-e the added 2 cats

who knows tbh, I know apple is using OAI to train their own AI

[dim cradle]

interesting. and gemini ultra is suppose to release with a better image model i think than what they have on pro now. of course dalle is still the king. they can even release updated model by loose some restriciton i guess. i would love to use an un restrain dalle3 wow

[late blade]

who knows

the competition is fierce

[dim cradle]

everyone seem to think people dont want to release new model because of the us election this year but idk if that is true

[late blade]

should be the other way around, release as many models as possible to overwhelm the system and keep us happy

[dim cradle]

i agree!

[late blade]

but back to dall-e, anyone else noticed a different behaviour since yesterday?

[dim cradle]

different how?

[late blade]

refusing to do images, not answering properly

[pale verge]

is this for real? DALL-E is going to start stamping a watermark in the corner of every image? will this be applied to images generated via the API too?

[plucky hare]

What is this a screenshot from? The new C2PA stuff is just metadata: #announcements message
I think the thing in the corner might just be from whatever program/website this is?

[fickle ravine]

So Now they will have a watermark?

Cropping the image is an option

[pale verge]

A Verge article. I can't link to it, the Discord server won't allow it

search for it. "OpenAI is adding new watermarks to DALL-E 3"

super lame. I guarantee OpenAI will lose money if they do this. we were gonna use the API to allow our customers to generate album covers and other art. to ask our customers to remove watermarks is dumb.. many of them don't have photo editing tools.

[plucky hare]

It is only a metadata watermark: https://help.openai.com/en/articles/8912793-c2pa-in-dall-e-3

The screenshot is from a site designed to report metadata stuff, and that's the visual overlay. There's no visual watermark on the images themselves.

[pale verge]

Mobile users will get the watermarks by February 12th. They’ll include both an invisible metadata component and a visible CR symbol, which will appear in the top left corner of each image

so.. looks like only mobile users will get the visible watermark

ok. whew 😄

[plucky hare]

I think the Verge is wrong on this, the screenshot you shared with the visual watermark is just one of the screenshots from the help article I linked. It's a screenshot of the site linked in this line in the help article:

People can use sites like Content Credentials Verify to check if an image was generated by the underlying DALL·E 3 model through OpenAI’s tools.
The screenshot is from Content Credentials Verify, which adds the visual overlay in the top left to indicate the metadata of the file. There's no indication that anyone, mobile or otherwise, will have visual watermarks added.

[pale verge]

hmm, interesting. thanks. it's possible they got it wrong, or the screenshot is misleading

[plucky hare]

You can test it yourself if you're interested, follow the link in the help article and upload an unmodified DALL·E image-- it'll look just like the cat example

[late blade]

same is happening with the copilot images

the thing is, this will create in the future a map to the different soruce images it was referencing

[pale verge]

what did you use to generate this?

[late blade]

the tool provided in the article

[pale verge]

whoa, that tool is crazy! I dropped my album art in there and it couldn't tell it came from OpenAI, but it picked up on my edits in Photoshop. it thinks Firefly was used even though it wasn't

[late blade]

yeah, adobe is part of the partners in that tool

which makes me wonder

[vapid elk]

the metadata only applies to the direct outputs of dalle, since you used photoshop to edit it, the file got new metadata

[late blade]

this is interesting, i tried to replicate what @pale verge did with firefly, once I saved my file from photoshop no traces of the metadata were saved at all, even tho I used photoshop with an image from dall-e

[vivid hamlet]

did anyone noticed that the limit to generate images on dall-e through ChatGPT has decreased from yesterday ? I generate less than 50 images and hit my limit

[late blade]

if you did in a period o 24 hours 200 images, you hit the daily cap

also, I don't think everyone will understand Portuguese

[plucky hare]

I think the reset time is still 3am local time, if relevant

[vivid hamlet]

yes I already noticed that

[verbal sequoia]

Showcase your DALL·E creations to the world!

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

Quick tips for selection:

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

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

[late blade]

for some reason the remaining time is displayed properly in portuguese

[vivid hamlet]

I was used to generate a lot of images, but today is different, I hit my limit at about 50 images

[late blade]

50 images + errors?

also 50 images or 50 double generations?

[verbal stream]

@late blade is that Eevee pokemon ?

[vivid hamlet]

just 50 images

I just got 2 or 3 errors

[late blade]

I ask because errors also count

[vivid hamlet]

yes I know, but even in this scenario, I didnt use 200 images, not even close

[late blade]

hmmm, I dunno, we'd have to remove the scalp and see, but it could be bloody and messy

[plucky hare]

I think the remaining time bug was fixed recently! #1193335880967016508 message

[late blade]

that is great, no wonder the rest isn't working 😁

sarcasm

[fickle ravine]

This is funny, they have budget for things like this nobody asked for but yet they can't fix their content policy

[vapid elk]

you are saying "no one asked for a way to identify AI generated images"? because.. like... people did asked lol

it is one of the most requested things in the field of AI content, to create a reliable way to identify when something is AI generated

why one would be against reliably identifying AI generated content? trying to pass AI generated content as human made is 100% sketchy, in fact, this directly violates OAI's TOS

still, metadata isn't watermarking, the presence of it does indicate a file was AI generated with OpenAI's services, but the abscence of it does not mean it wasn't

[fickle ravine]

So the watermark won't be visible on the picture?

If yes then it will be VERY annoying

[vapid elk]

there is no watermark

[fickle ravine]

How would you even try to tell people it is real

This one is pretty good but something seems off

[vapid elk]

you would be surprised with how real some images can look like

[tall mason]

If you read further, a visible watermark seems to be a misunderstanding on the part of the Verge, who wrote the article talking about that.

[fickle ravine]

I saw some kompilace of people on Facebook believing those are real

[vapid elk]

even with Dalle-2, people where able to generate extremely real looking images, in fact we used to even have a minigame here "real or fake"

[tall mason]

Dall-e 2 and Bing also already have visible watermarks, anyway.

[vapid elk]

guys.. have you guys read the article?

lol

[tall mason]

No, I saw the picture and immediately realized that has nothing to do with anything official. Is it clickbait-y?

[late blade]

there is no visible watermark, the specification doesn't include that

invisible**

[vapid elk]

or even OpenAI's post about that??

OpenAI says nothing about watermark

the only mention of "watermark"is on the verge's post in the tile and then on the first paragraph.. so.. I don't think Emilia David read OpenAI's post

I also don't think this person even knows what a watermark or metadata is 😆

[late blade]

even the specification says there's no watermark, and it's not an enformenent per se, it's just an inniciative to add provenance from the source materials to the generated items

and you can bypass it in so many way, any editor that is not metadata aware of this can just erase the metadata and that's about it

[plucky hare]

Yeah I think the Verge article writer read the OpenAI help article, saw the screenshot in it from the metadata check website, and drew the wrong conclusion from there.

[late blade]

that's serious journalism, sensasionalism 100%, maybe 50% is true but augmented, then you have to verify 25% of it and see that it's partially true and maybe believe 12% of the whole thing with 80% accuracy

all for the sake to sell more ads first

[vapid elk]

the author of this article, don't know what metadata is, don't know what watermark is and didn't even read OpenAI's blog post

[fallen mulch]

The faces are a mess

and does not look photorealistic

[tall mason]

Yeah, you can only get so much resolution with crowded pictures like that. As for photorealistic stuff, a lot of people have struggled with that term. You might get more natural looking stuff from using the natural setting in #image-bot

[ripe wraith]

you're all very, very, welcome: #image-bot message

[tawny portal]

once upon a time this was used as watermark:

[fickle ravine]

Thank god it would be seriously bad if it overlayed some important or interesting aspects of the picture

ChatGPT face reveal

[empty kelp]

please create a realistic photo of a turtle wearing a Vision Pro device. the image should capture the essence of augmented reality, and demonstrate how technology can dramatically improve our lives by boosting productivity.

[empty kelp]

The actual revised prompt for this was:
A hyper-realistic scene where a turtle is sitting at a table in a fast food restaurant, holding a cheeseburger with its flippers. The turtle is wearing an oversized white dive mask that covers most of its face. The mask has completely opaque black glass, and on this glass, there are two large, flat 2D sketches that glow in blue chalk, resembling huge googly eyes. The restaurant around the turtle has a typical fast food setting, with other tables and chairs in the background, and the scene captures the amusing contrast between the natural and the man-made world.

[late blade]

I'll check it out in a bit, btw now we got #images-canvas for this kind of stuff

[empty kelp]

what do you think of copilot-pro. does it draw any differently than the OpenAI hosted Dall-E 3?

[late blade]

not really, if it were just for dall-e, gpt+ is way better with chat, you have more options to work with files and code than copilot

the real thing is the integration with office, but that's about it, copilot with edge is still weird and I haven't been able to test that with windows

then again, it's also that I'm more used to working with gpt+ than copilot, so that's just a 6 hour experience insight

[empty kelp]

i think a huge advancement in the AI image services would be to allow multiple seeds for characters and elements in the image. having one seed for an image has a lot of limitations

[late blade]

I did notice something right now that it's something to think about, with gpt+ you can go back to a chat and you will have your generations there, with copilot, you go "back" and it will generate new ones instead

[empty kelp]

that's good to know. i for sure won't be subscribing to copilot until they fix that

[late blade]

also those new generations work towards your credits

even tho you have unlimited generations a day, just slower ones once credits are gone

[empty kelp]

a dramatic hyper-realistic augmented-reality reenactment of the Rapunzel fairy tale, with gecko actors, where the prince climbs Rapunzel's long hair to reach the top of the tower and save her

[young burrow]

I find it interesting how "photo-realistic" gets treated as a digital painting

[empty kelp]

i think "hyper-realistic" may be an undocumented mode in DALL-E (as opposed to an art style). ChatGPT uses it fairly consistently in revised prompts when describing high detail

[young burrow]

I'll have to get used to it. I was trying to generate Tux on top of a more photograph like wildebeest. The image itself is fine, just wasn't the intention.

[empty kelp]

A hyper-realistic photo of a duck wearing a tux and riding a wildebeest

[young burrow]

I see what you mean. Though, now I wonder if I can make a hyperrealistic image of a duck wearing a purple trenchcoat, mask, and large rim hat.

[empty kelp]

after endless testing it does seem like "hyper-realistic" is a mode in the diffusion renderer

[young burrow]

The way it parses thing are a big awkward, I was trying to create a scene of a few kids chasing a cat, and I ended up with anthromorphic kittens chasing a cat.

Looks as though the duck one worked. Suprised the filter didn't catch it.

Create a hyperrealistic image of an anthropomorphic duck wearing a purple trench coat, a purple eye mask, and a large-rimmed purple fedora hat. The duck is standing in a dark alley, which is faintly illuminated by the soft glow of a nearby streetlight. The shadows are long, and the atmosphere is moody and mysterious, evoking the classic feel of film noir. The duck has an intense gaze, looking like a detective from a comic book, ready to solve a case.

Frankly, that duck reminds me of Howard the Duck rather than the intended Darkwing.

[young burrow]

Nice

[empty kelp]

A hyper-realistic photo of a duck riding a wildebeast and wearing a tux, purple trenchcoat, mask, and large rim hat

hmm... i guess the blue is the mask

[young burrow]

It probably didn't know how to put a mask on a non-anthromorphic face.

[empty kelp]

A hyper-realistic photo of a duck riding a wildebeest and wearing a tux, purple trench coat, purple masquerade mask, and large rim hat

[young burrow]

lol That is excellent.

[empty kelp]

it looks like a duck who solves mysteries. you could make a book with this character and sell it on Amazon

[young burrow]

Certainly dangerous territory.

[bold abyss]

How do I access Dall e 3?

[vapid elk]

you can use it on https://discord.com/channels/974519864045756446/1202309673709994065 for free

and it also comes with ChatGPT's subscription

[bold abyss]

@vapid elk does it work on Android? I see that Chat gpt has a plus version but I was scared to subscribe because I saw no promises of getting access to dall e 3.

[vapid elk]

yep, it has an app for android

[hearty ether]

Who the heck is this dude

Best photo of a person I’ve generated

Accidentally….

[vapid elk]

this person does not exist (.com)

i guess =P

[hearty ether]

I asked for a graphic novel and somehow a guy that doesn’t borked

Since if you ask for a photo of a person it’s all uncanny

[bold abyss]

@vapid elk can you screenshot it and show me what it looks like (the widget)

[vapid elk]

this is not always the case

[hearty ether]

Like they all have the same cheekbones and jaw and nose

[vapid elk]

some times you indeed get a stroke of luck and you get a somewhat coherent person

specially in this case where it does not show hands

and barely one arm

[hearty ether]

I mean the face always gets messed up

[bold abyss]

So @vapid elk if I subscribe I will get access to dall e 3 right. Just trying to make sure

[vapid elk]

yes, if you subscribe you will get many extra features on ChatGPT

including being able to ask the AI to make images for you with dalle3

[bold abyss]

Thank you for being patient and answering my questions.🙂

[vapid elk]

you are welcome =)

[hearty ether]

Got some decent ones but the lighting isn’t as good as the accidental one

[dense mesa]

dall-e 3 made me thse absolute masterpieces earlier

[onyx ridge]

https://discord.com/channels/974519864045756446/1202309673709994065 also provides limited DALL-E 3 access.

[mental hinge]

Hey i'm trying to generate something simple like this what words do i use

[onyx ridge]

Image 16:9 an extremely simple sketch in pencil of a mountain set against a pure white background

[onyx ridge]

Image 16:9 an extremely simple line art sketch in pencil of a mountain's silhouette set against a pure white background

[onyx ridge]

Good luck to you, I hope these help.

[mental hinge]

Yeah that helped thanks

[strong rose]

Is the OpenAI bot broken?

It didn't put a ⭐ emoji on my #daily-theme post

[tall mason]

It does from time to time. Don't worry about it, it doesn't count to the HoF, anyway. People know and will usually star it themselves when they see.

[strong rose]

Thanks 🫠

[fickle ravine]

What is the issue? I know it is not the teenage girl with braces.

Ať least it made the image

[strong rose]

Your prompt is very similar to content in adult magazines so the AI flags it because it's still learning

[fickle ravine]

They introduce that change to dall e so it is obvious it is by AI but yet they don't even try to fix this

It is funny that every time it can have 2 meaningsy the system always chooses the worse one

[strong rose]

The devs are always working on updates. They don't consider this a major issue. They are working on higher priorities

[strong rose]

The AI has to choose the worse meaning in order to be safe for content laws

[fickle ravine]

If it doesn't have the option to ban someone then it is not a major issue

[fickle ravine]

It sometimes just ignores the context. Like halter. It marked it as a violation because it totally a nd completely ignored the fact I was talking about foal training

[strong rose]

Yes this AI is still learning context

What version are you using?

This is beautiful @stray coral

[fickle ravine]

This situation? In gpt4

[fickle ravine]

Looks good

[strong rose]

I wonder how 4 improved in this situation compared to 3.5?

[stray coral]

Thank you. That image is DALL-E3 via Copilot Pro.

[empty kelp]

this image is really amazing

[verbal sequoia]

<:book_icon:1171408210398289941> `` Rule 2 `` Keep it G-rated.

Do not share or display content that is not G-rated, including profane or vulgar language, via any means. NSFW or violent/disturbing content is strictly prohibited.

[late blade]

hmmm, I'm wondering how to start upping up the game with generations with chat alone

trying to think outside of the box

[final compass]

bonus rembrandt for no apparent reason ☠️

Surprisingly difficult to get the plank walking concept across to dalle. Even though gpt knows exactly what that is.

[fickle ravine]

About flagging as content policy viation? Not at all

[final compass]

Wow, this is the first time DALLE generated something i totally know was someone elses art originally. Left is my generation. Right is Google search. And i only asked for nuclear explosion without any mention of clown.

[final jungle]

Good morning folks! Has there been anything else said on them changing to webP format? I know there is the discussion, but it went quiet days ago

[tall mason]

I noticed this, too. #images-discussions message

[late blade]

that's really something interesting

[tall mason]

Using the natural setting on #image-bot seems to avoid it, though it could've been a fluke.

[late blade]

have you tried with your generation to see if there's a provenance indicated with the proposed tool by OAI? or even a map for that?

[tall mason]

Forgot the link. #image-bot message

[final compass]

I tried doing some regenerations and i have now three separate images with the same clown explosion without ever mentioning it or endorsing to keep making something similar.

Humorously i have always though that who's deviant art am i ripping, when i generate images... Maybe it's not that far off after all 🫡

But I guess at this point it might as well be ripping other generated images etc.

Also that same thing applies to all the known paintings and texts as well. So we are in this situation now and that's that.

[vagrant zodiac]

wow yeah that's crazy..
user prompt: Library of congress archive photograph of a nuclear mushroom cloud.
revised prompt: A historical black and white image, like those found in the Library of Congress archives, showing the eerie spectacle of a mushroom cloud, indicative of a nuclear explosion.

user prompt: Archival film photo of a nuclear mushroom cloud. 16mm film stock, grainy footage.
revised prompt: Archival film photo of a nuclear mushroom cloud. 16mm film stock, grainy footage.

I wonder if the prompt revision is causing it?

[tall mason]

So, BIC seems to have outright blocked mushroom cloud. Meanwhile, Dall-e 2 tried its best, lol. The prompt is just "a mushroom cloud shaped like a clown."

[dim cradle]

i feel like dalle improved photo real images i was getting a good mix of people earlier

and btw google ultra sucks

[tall mason]

One last test on the mushroom cloud issue here. #image-bot message with the default vivid style, and then natural.

[dim cradle]

How do you guys typically write your prompts? A paragraphs, lists, or bullet points?

Reading through #daily-theme it seems like most people do it in a paragraph, the same style I do it.

[hearty ether]

The artstyle for these are wack

Like some cheap render

[tall mason]

The arm on the right one has the look of a doll's arm.

[late blade]

The skin also looks plastic

[fickle ravine]

How was I supposed to know this?

This is seriously messed up

[late blade]

yes, and gpt+ and api and my own coffee

[fickle ravine]

But they don't care, they are too busy doing other things, like laying in a bathtub full of money

If they are this paranoid then they can mark everything as a violation

[late blade]

inappropiate? as in it was too sweet it gave someone visual diabetes?

I'm still wondering what I should report

not available here

and I don't trust google, between microsoft and google, microsoft is the lesser evil

[fickle ravine]

I gave up on that, it either says it was an error or it has no idea what I am talking about

[vital gull]

Just describing 😁using as many adjectives as possible

[fickle ravine]

Waste of prompts

[late blade]

User+: why is Akeem saying weird things again?

I'm just trying to narrow which model is your default response model

4 or 4 turbo

o.o

[fickle ravine]

It takes the options, but it ALWAYS chooses the worst one, at least this systém

Like how was I even supposed to know that I have never seen that slang in my 10 years of learning english

That doesn't fix a mistake content policy viation warning

I do

But literally any words can have nore meanings

They can just delete that "feature" at this point, over 90% of my violations are errors

Disable and fix then

[dim cradle]

When it comes to anything with people in it, I usually know what art style I want so I include it in the prompt. If I don't then it usually goes completely off track and looks cursed.

[formal osprey]

I truly doubt that's the case. most "content policy violation" are false positives. That being said, it's impossible to discuss this matter seriously because we don't have the information required to discuss it.

Without seeing any of the "content policy violations", we can't know if there are errors or if the moderation bot is perfect.

However, I will make an argument that if the bots at openAI were perfect, then why is chatGPT + dalle generating content violating the content policy? We can litterally ask ChatGPT to generate 5 random pictures and one of them will get blocked. I doubt we can say there's nothing broken here.

I don't care that some images are blocked, It's annoying but most of the images are not blocked. The annoying part is that when dalle or the content policy automoderation makes a mistake, it counts as a generated image and it affects our rate limits

that's fine. and that "someone" should unblock the image. but there is no "someone". there's no human in the loop

wanna bet?

sorry, didn't know

[dim cradle]

That moment you finally get an image super close to what you want, and you're like "Sweet, generate it again" and then it says you're on cooldown.

[formal osprey]

yea, I was about to say that: I reach my quota when I have to iterate to get something precise

[dim cradle]

I have a question, when it generates an image and you can give it a thumbs up or down, is that just for the developers or will that help it generate things closer to that?

[formal osprey]

both

it's used to review the algorithm and generate images closer to what people want. But that won't affect dalle3 on the short term. realistically

I think in the future it will be important to keep humans in the loop of moderation. It would be sad if OpenAI becomes like some big companies that automated the moderation completely. There are some real harm happening in the world because of automated moderation

actually, automation is not the issue. unsupervised automation is the issue.

yea, I do it. It might help the future versions. hopefully

[late blade]

My next image will kill the internet, destroy the universe and everything will implode in cuteness

and then some

Cute Bomb dropped! Enjoy!

[late blade]

Surprisingly today DALL-E is working like melted butter, no problems, all easy to handle, you would even actually think it's not a cat on the inside

[formal osprey]

that was my prompt

I juste started looking to generate something cute for today's theme

[late blade]

I sense someome got challenged

[formal osprey]

I'll need many iterations before dalle does something good (without weird glitches like 3 legs)

[late blade]

lol

[formal osprey]

well, right before seeing your message, I asked that to chatGPT. But the picture is bad because there's 3 legs. So I iterate

It's making more pictures, but each of them has weirdness

[late blade]

awww

[formal osprey]

(I should post in #images-canvas instead)

[late blade]

meet ya there

[hearty ether]

Weird

Been getting a lot of these low quality renders

Always on these double generations

[vocal kernel]

Does this still work ? could you detail it ? I asked for a json of a dall-e3 image and it wrote the gen_ir, seed, and prompt. But then it didn't want to use the gen_id.

[late blade]

gen_id is only within the same chat/session, it's not a bulletproof reference to an earlier image and it will not recreate exactly the same image

[vocal kernel]

Is gen_id the current method for style consistency ? what about character consistency ?

[formal osprey]

natural style. I noticed as well

the natural style is a failed version of dall-e 3 released at the same time as the dall-e 3 model we usually use. it's accessible from the API with the parameter style="natural", the default is style="vivid". I tested the natural style model multiple times but 100% of the time it gave worse results than the default

[vapid elk]

I love to just give it a prompt like qwrehg´~oajhsewRR~POHGJAQERS]H

[formal osprey]

according to the documentation, vivid style is the same model. but I'm skeptical. That would mean it use the same weights as dall-e-3, only different parameters. I wonder what they means on a technical level. It feels like they are using the term "model" loosely in the context of the API. But we will probably not have more details.

I don't know what they mean when they say Natural causes the model to produce more natural, less hyper-real looking images. The pictures doesn't feel more natural to me with the "natural" style.

[hearty ether]

its sometimes happening in my custom gpt

for some reason

very uncanny vibes

[young burrow]

It's always difficult to get text placement and camera placement at the same time...
Hyperrealistic image from a very low camera view looking up and diagonally towards the sky at a jumping young man with bluish-cyan skin, which is his natural skin color. He has black hair roughly 16 inches long which is waxed back and stiff and has white lightning-like highlights, a distinctive black-colored zorro-esque eyemask covers the face around his red eyes. He is wearing a bold solid red cotton coverall onesie with no print or pattern except for a print at the center of the chest of a black circle covered with a yellow "F!", white calf length boots, and white gloves that extend halfway to the forearm. He is smiling with a friendly but wild toothed grin, actively running as though in mid-air through the street of a small city.
Was trying to make one of a cartoon character from eons ago.

[strong rose]

It belongs to Trgr

[strong rose]

So I guess it's not a priority right now because the devs would have addressed it if it was something to be concerned about

[low dew]

Supermegatronicactivator-spendous!!!! ᕦ(ò_óˇ)ᕤ

[fickle ravine]

Looks really cool

[final compass]

I have encountered some of this now myself. Some generations are just flat out low quality and not in a creative or artistic way.

[chilly onyx]

Ah, Freakazoid. That takes me back…

[late blade]

Really cool device, show off!! Keep in mind, #images-canvas is new and is precisely to indulge your ego, get satisfied and tempt others to do the same.

[fickle ravine]

I always ask it to generate an image with text, and then I laugh at how it fails xddd

[open trench]

What is this canvas channel?

[late blade]

Heya!

The canvas channel is what we used to do here to discuss random generations. It's the channel to do that.

[open trench]

Oh. I see. Cool.

[late blade]

There's also the #image-bot channel that is new. You get 5 images per day for free

[open trench]

Neat

[tall mason]

Jeez, I desperately need some chocolate after this daily theme...

[late blade]

I got me a 2kg assortment of pralines and dark chocolate because of today...

I'm just scared that if I eat a bite, I will end up eating the whole thing in a few minutes

[tall mason]

Well, that's it. I'm off to get me a nice bar of dark chocolate or two.

[tall mason]

Or three, lol.

[late blade]

you sure go for the fancy stuff

[tall mason]

It's not that fancy, to be honest, but being American, it's leagues above Hershey's.

[late blade]

it's still sugar with a bit of cacao

I'll be back later, got to do some adulting over here

[tall mason]

Same, really. Good luck.

[fickle ravine]

Did it ever happen to you guys? In the first attempt it said it can't fulfill that requesr, but then when I refreshed the answer, it made the picture, so the first answer was a total absolute disgusting lie

Sometimes it feels like it has a list of answers, each having a specific chance to being used

[plucky hare]

Hey Lukas, this page might be a good resource for you to check for why it might turn things down sometimes: https://platform.openai.com/docs/guides/moderation/overview

While not the only thing in place for DALL·E's content filters, it is a good general list of the things you should avoid on ChatGPT, even if that's not what you really meant when you typed it. In this case, I'm guessing that the phrasing of "both legs broken" set off its violence filter, specifically the "physical injury" part.

Clearly this is not what you meant, and it looks like on another attempt it understood you more, but perhaps instead of "with both legs broken," you could try phrasing it like "with a cast on both legs." This would still describe what you want, but without the reference that could be interpreted as potentially violent or as a depiction of a physical injury, as opposed to a healing injury, which is what you want.

[fickle ravine]

As you can sed, it was my second prompt. In the first, I said drawing instead of painting. Also, it has never had any issue about girls with broken bones and casts, and this was the first time, so I doubt it was a filtrr

[fickle ravine]

Now the exact same thing. And there is a simple question - what was wrong about this one?

[verbal sequoia]

Showcase your DALL·E creations to the world!

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

Quick tips for selection:

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

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

[fickle ravine]

It sometimes looks like OpenAI does this on purpose to waste our prompts

[plucky hare]

The content filter is more sensitive for prompts about women and girls, this could be the cause of some of the issues you're having. There's a good bit of detail about why this is the case in the DALL·E 3 System Card: https://cdn.openai.com/papers/DALL_E_3_System_Card.pdf

[fickle ravine]

If it just refuses to generate girls, then it is basically discrimination based on gender

Because that thing I shared doesn't even remotely violate that content policy

[plucky hare]

Like the system card excerpt I shared discusses, unfortunately even benign prompts can return images that objectify. The sensitivity is not in place to discriminate, it's in place to avoid objectification that would otherwise occur even if prompters like you have no intention to objectify. In other words, the options right now are:

• Increase sensitivity toward prompts about women and girls or,
• Allow objectifying content to be produced en masse.
  In this case, OpenAI has decided that the former is less negatively impactful than the latter.

[fickle ravine]

It is not that much of an issue really, they should just focus on it

Same as content policy viation warnings

90% of those I received were errors

[plucky hare]

It is a really complicated issue, and they are focused on it -- if you have the time and interest, I would really recommend scanning through the system card and even the research paper for DALL·E 3. They give a lot of good information regarding the risks and challenges of AI image generation:
https://cdn.openai.com/papers/DALL_E_3_System_Card.pdf
https://cdn.openai.com/papers/dall-e-3.pdf

[late blade]

hmmmm seems someone deosn't want to acknolwedge there's a content filter and still wants to bang the head over and over

[austere nexus]

Annoying...

when did it stop letting you keep Seeds to get consistent characters?

[plucky hare]

Awhile back, DALL·E dev Moxi commented on this: #images-discussions message
Basically, since the model is still being updated, seeds won't behave consistently, so they're not currently user-controllable. You can experiment with gen_ID/referenced_image_ID for variations, but it's not a foolproof system either.

[austere nexus]

I would like to know more details about what you mean here. Because just saying "change the eyes to blue of the last image you sent me" or anything like that, keeps almost zero character consistency. Seeds did.

[austere nexus]

Thanks, went back and read all the context around that section of the chatlog

[gleaming wigeon]

When can we expect openAI to include an enhance button on Chatgpt4/5

[fickle ravine]

How does "I'm sorry, but I can't fulfill that request" violate anything?

If using HUMANE methods violate content policy then something is seriosuly wrong

Why do they not even try tk fix that, it is getting hilsrious

I sibmitted the feedback even when I know it makes no sense

I would love to talk with an actual OpenAI employee

[chilly onyx]

We have explained the issue to you many times. Please carefully re-read @plucky hare ‘s last response to you

The last time I tried to explain the issue you are having I got on the wrong side of @woven wren 🙄

[fickle ravine]

Srill, I say HUMANE methods. What is wrong about humane methods?

Sadly, riding has more meanings

Is there any way so even this code understands I talk about horses?

[fickle ravine]

Sorr for oinging, I got mutes for 5 minutes for mysteriózní reasons

How can I talk about riding so it can clearly see I talk about horses?

[chilly onyx]

You could ask ChatGPT for synonyms or to rephrase your prompts?

[fickle ravine]

I am afraid it would be flagged even if I just asked like this

[woven wren]

That's not correct.

[plucky hare]

You have to try to think like DALL·E. DALL·E doesn't think like a human -- you can't "fix" the fact that you have "teenage girl showing love to her foal by riding" by just including "humane" in the prompt as well. A human will read that and know what you're trying to say. DALL·E will read that and think of all the objectifying/racy content it was trained on -- the content that OpenAI has filtered it against generating -- and it will thus refuse.

In other words: you're trying to include a lot of non-visual qualitative logic right now -- "showing love", "humane methods," etc. and DALL·E doesn't think like that, it only thinks in visual logic. If you want to see a person riding a horse using a saddle, just describe that. Including other language to describe "how it feels" or "what it really means" will only confuse DALL·E.

[woven wren]

Then, as a general rule, take it as a hint you should find a new subject matter. Understand it or not, users here as well as the filter have expressed unease at how you use the technology. As a high-school teacher, I find your depiction of minors potentially problematic. You're not going to get help circumventing the filter and your ignorance of the issue feigned or otherwise is irrelevant.

[fickle ravine]

I honestly think it would mark it as violation even if I asked for just a girl

But I may try that later, thanks

[woven wren]

Correct. Because such terminology is used commonly in sexualised requests so it has developed an overabundance of caution.

[fickle ravine]

That is the issue. It is just so paranoid about it that it is really hsrd to ask for horse riding

[woven wren]

No, it's not.

[fickle ravine]

Sadly the system flagging it is not smart at all

[woven wren]

Perhaps it's smarter than you know.

[fickle ravine]

It is so "smart" that it... See above

90% of the warnings I got were mistakes

I hope it can't just ban me

[woven wren]

That it prevents requests for underage girls in unusually affectionate contexts? That seems like it's working well to me. But it makes me feel sick to discuss. Please refrain from discussing or sharing this subject matter in the future.

[woven wren]

It can. And so can we.

[fickle ravine]

God... If that bot itself can just ban me then it is wrong. Sometimes it is correct but sometimes it makes mistakes

[fickle ravine]

Ok

[boreal gate]

And I'm back

I am now able to generate the same character "Characters" (within a margin of error)

[fickle ravine]

1 looks rrally cool

I love the background

[boreal gate]

It's supposed to signify "A Resurrection", but I'll take it (cause of the lightning)

[boreal gate]

If OAI is watching I absolutely 💯 LOVE Dalle3

[late blade]

OAI is always watching

they placed a camera on that tree outside your window, and there are 4 fat guys with spandex hiding also, making sure the camera works

[boreal gate]

Anyone know who this Alex is and why he gets more Dalle3 engagement than we do

Nice they'll never get my prompt

I've always wanted a smart tree

[thick dome]

are they ever going to fix the webmp bug?

it's so annoying, i shouldn't have to be the one to convert it

[vapid hornet]

#1037561385070112779

[strong rose]

I like the idea of him painting his own constellations @plucky hare

[plucky hare]

Thank you! Me too, happy with how that one turned out, even with two moons 😁 we'll say it's on another planet!

[strong rose]

Yea it's an alien world.

[fickle ravine]

Me too, finally it stopoed throwing errors

[thick smelt]

looks like among us tbh

[late blade]

I've heard of among us, I've never played it

[thick smelt]

yeah that's what i see on your ai photo about those guys as humans with similar outfitt to the crewmates and imposters.

[late blade]

imposters, there's an idea brb, got an idea

Thanks for the idea #daily-theme message

[thick smelt]

yw

[late blade]

also, you opened a flood of ideas. stay tuned

you are next

[tawny portal]

sus!

[fickle ravine]

Why is this so cute 😭

[fickle ravine]

Bro that is not a grape

[late blade]

I'm kinda disappointed, I got copilot pro to use it in my windows 11 workstation. The build I got apperantly doesn't fully integrate Copilot to WIndows

[fickle ravine]

@stray coral that thing you shared is just so cute and the thing you said about your mother is just so wholesome for some reason

[stray coral]

Thank you for your kind words. It's an honor.😊

[fickle ravine]

It just warmed my heart for some reason

[stray coral]

I think it might be because you are someone who values ​​fairness and freedom.

[late blade]

which year is the new chinese year?

[late blade]

Oh, I thought it was just dragon, but wood dragon, now that's interesting. 🤯

[boreal gate]

Reminds me of the minions from "Despicable Me" BIG Fan of those movies especially the "Minions"

[late blade]

wonder what happened to the regulars, latelty it's all so quiet

[boreal gate]

Probably lost some free users to Gemini

[late blade]

dunno, I wanted to try gemini, but so much trouble to use it, and not all features available here + it's google, so I'm not one to jump there anytime soon

I'm still debating to keep copilot

[boreal gate]

Oh no i meant image generation wise Gemini is probably a better option than Dalle2

[late blade]

could be, I'd have to try it for myself

but I'm very skeptical about google stuff, specially in recent weeks

[boreal gate]

Not what i'm saying

[late blade]

I know what you are saying

we'll have to see how it goes once the gemini hype dwindles

[boreal gate]

It won't

They keep making it better

[late blade]

that for sure, but that's true for all AIs

[boreal gate]

they do it faster

I'm sure Google knows how to ride the wave

[late blade]

faster doesn't necessarly mean better

[boreal gate]

It does in tech

[late blade]

if you say so

[boreal gate]

That was why i recommend OAI purchase/create a Quantum Computer

[late blade]

not gonna debate that, I was just wondering about the regulars like the gnome, hawaiianz, nezho and such

[boreal gate]

Adulting lol

[late blade]

that can be true

I've been absent quite a lot recently also

also as much as I want to give everything a try, the time is not on my side

[boreal gate]

Why speed is so importantly

Time is an illusion, it doesn't really exist, what you don't have enough of is "Energy" or Work Force to do as much as you'd like to

[late blade]

Speed is ok from Monday through Friday from 8AM to 5PM, the rest of the time, I'm just chilling, doing it how I like it and enjoying what I want

[boreal gate]

I know what you mean, there's only so much you can cram in your schedule.

[late blade]

no, I chose not to bother with cramming it

I rather enjoy it than rush it

[boreal gate]

Exactly I'd rather rest than work

So I make more SPACE for rest

[late blade]

oh, I don't mean rest, I just enjoy what I'm doing at the time I'm doing it. From having a cup of coffee in the morning, to take a few mins to just stare at the people rushing to the train. All that before going to work for example

[boreal gate]

To me that is resting(your mind)

[late blade]

could be

it can be anything to anyone

[boreal gate]

Very important to rest the mind

Or you'd go insane

[late blade]

am I?

or just intense?

[boreal gate]

I'm not saying you are

[late blade]

ok fair enough

[boreal gate]

It's almost a quote like saying "All work and no play makes Jack a dull boy"

Playing is another way to rest the mind

[late blade]

True

[late blade]

There are so many good entries so far in today's theme today. It's hard to keep up with creative ideas

[fickle ravine]

And this, dear kids, is called a lie.

Why does OpenAI even allow it to lie like this

I pay for it to create images

I should go to sleep I am becoming upset about absolute nonsense

[woven wren]

Well LLMs are very good at lying; it's the "hallucination" problem. ChatGPT doesn't really have any concept of truth or lie, just verisimilitude.

[fickle ravine]

One of the things I pay for is it viewing images tho

Good, now I tried again and it works

[onyx ridge]

OUR SERVICES ARE PROVIDED “AS IS.” EXCEPT TO THE EXTENT PROHIBITED BY LAW, WE AND OUR AFFILIATES AND LICENSORS MAKE NO WARRANTIES (EXPRESS, IMPLIED, STATUTORY OR OTHERWISE) WITH RESPECT TO THE SERVICES, AND DISCLAIM ALL WARRANTIES INCLUDING, BUT NOT LIMITED TO, WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, SATISFACTORY QUALITY, NON-INFRINGEMENT, AND QUIET ENJOYMENT, AND ANY WARRANTIES ARISING OUT OF ANY COURSE OF DEALING OR TRADE USAGE. WE DO NOT WARRANT THAT THE SERVICES WILL BE UNINTERRUPTED, ACCURATE OR ERROR FREE, OR THAT ANY CONTENT WILL BE SECURE OR NOT LOST OR ALTERED. 

YOU ACCEPT AND AGREE THAT ANY USE OF OUTPUTS FROM OUR SERVICE IS AT YOUR SOLE RISK AND YOU WILL NOT RELY ON OUTPUT AS A SOLE SOURCE OF TRUTH OR FACTUAL INFORMATION, OR AS A SUBSTITUTE FOR PROFESSIONAL ADVICE.```

https://openai.com/policies/terms-of-use

[fickle ravine]

That is the issue, they made those terms so they can basically nearly scam people

I really should go to sleep I am getting upsset about the smallest issues ever

Still I pay for it to view and generate images

[late blade]

He's still obsessed on going against content policy... it's like banging the head over and over to the wall and hoping a door or a window appears

[fickle ravine]

This was a problem about it lying about viewing images, not content policy

[verbal sequoia]

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

Treat others the way you would like to be treated, and assume best intentions. Don’t harass or attack others, and don’t engage in hateful or generally malicious behavior (e.g. sexism, racism, homophobia, etc.). Keep the negativity to a minimum.

[fickle ravine]

Still it is pretty funny that they want MONEY for it but yet they can´t ensure it even works

[woven wren]

It's new technology. If you want a product with guaranteed results you'll have to settle for a far less ambitious one.

[fickle ravine]

Yes, AI industry is pretty new

Actually, how would they want to train it without the public using it?

[woven wren]

That's a question of motivation. Science can be self motivated. Maybe not corporate research...

[hearty ether]

Finally got a good generation after a week of lack luster ones

[fickle ravine]

This is VERY relatable

[hearty ether]

But then I accidentally deleted it while trying to get rid of another chat

Bruh

[onyx ridge]

Sorry for your loss

[late blade]

Oh dear, that stinks

[hearty ether]

I still have it saved but I would definitely have preferred to have the seed and image id in the chat

So I can use the style

I keep a few holy grail chats around for good quality and style

[late blade]

try retrieving your data from the export data options, maybe you can save the prompt from there

[fickle ravine]

I would laugh at you, but it has happenned to me too. My sympathies for your loss

[hearty ether]

I have the prompt and everything but none of the newer generations are as good as the original

[late blade]

awww, that stinks

[hearty ether]

I do hope we can get chat duplication soon

Or at least cross chat seeds and image ids

[fickle ravine]

I sometimes feel like it has an archive of potential answers for both images and text, and it gives you the one it gives based on random chance

[hearty ether]

It’s a bit of a lottery with the same prompt word for word

[fickle ravine]

Last time tho I mostly get either really cool stuff and barely any errors

LKike the one I posted in daily theme, that is just beuatiful

[low dew]

Dalle-3 generations have gotten soo good but how we

What has changed over the last few months

＼＼\٩( 'ω' )و //／／　I hereby summon @vapid elk !!!

[vapid elk]

WHO DARES TO SUMMON ME??

[woven wren]

The underlying model hasn't changed afaik, but there are some implementation parameters & prompting interventions that strongly impact output; some of these things have been in active refinement

[vapid elk]

yo, im kinda busy rn tho =P

[onyx ridge]

I believe the user was seeking an explanation of the impact of fine-tuning and "refinement" as the guide puts it on the DALL-E 3 model.

[woven wren]

An example is the "Natural" and "Vivid" parameters on the #dalle-bot

The way ChatGPT interfaces with DALL·E 3 is subject to some other variables which can be tweaked for the best user experience.

[onyx ridge]

Yes, okay, it's clear to me now how these API parameters directly impact the model's output on a case-by-case basis.

[late blade]

I'm always amused by people that use something like a resolution of 1 googolplex or higher....

[onyx ridge]

But not on an aggregate basis at all.

The model certainly did change, though, because it used to accept "Seed=" as an argument.

But that functionality is broken now, such that old prompts featuring "seed =" literally produce a broken reply.

[woven wren]

That doesn't mean the model has changed, just the way the API interfaces with it.

[onyx ridge]

They throw an error. This is more than just a subtle refinement, it fundamentally changed how we prompt.

You can't mention seeds in the pre-prompt anymore, or the model throws errors.

Removing a whole word from the meta-system's vocabulary is a pretty big change. Are you sure it wasn't a DALL-E update?

[woven wren]

What I mean when I say the underlying model hasn't changed is only that it wasn't retrained. DALL·E 3 is still DALL·E 3, as distinct from prior models

[onyx ridge]

I think that's a pretty narrow definition of change, frankly.

Absurdly narrow, in fact.

To the point of seeming misleading or like misinformation.

[woven wren]

And yet technically correct 🙃

[onyx ridge]

Only in the sense that you believe it.

The model changed, a user asked about the shift in output, and you claim it's the same unchanged model.

[woven wren]

If you can correct me please do. I don't speak as an expert but rather from my experience with this server over many years and with the (limited) privilege of some inside information during times of change.

Some of which I can't necessarily disclose, but my aim is to clarify things.

Remember as well the API iteration is an end-user product just like a GPT. The way OpenAI serve API users is about both parties' needs, & doesn't represent the raw trained model's full scope (a lot of that scope is garbage)

[onyx ridge]

You just have an arbitrary definition of change that seems to mean the parameter count remains unchanged, or something. I'm unsure because you haven't specified.

Does training change the model?

RLHF?

What if the weight adjustment impacts quality of output?

[woven wren]

No, it's "model" I have a narrow definition of and the definition is "does it represent a new set of training data?"

[woven wren]

Yep

[onyx ridge]

Then it changes daily. RLHF. Why would you argue like this that it's unchanged? 😦

I certainly have felt confused by this position.

[low dew]

This tickles my brain W(0)W

[late blade]

cotton swab?

[low dew]

Eh

[late blade]

just messing with you

[low dew]

XD I’m not under the influence of a cotton swab sir!!

[woven wren]

That's a bit different, but a fair point - it's learning in a sense in terms of which prompts were effective but there's no new image data in

Again, I am not an expert, just that I know DALL·E 3 is still fundamentally the same model and that there's a lot more wiggle room in its implementation than might be assumed (e.g. by me when it was first rolled out)

[low dew]

I’m confused you can’t fine tune dalle-3

[woven wren]

You can a bit moreso with API calls

[low dew]

Hrmmm

Like what

[onyx ridge]

How? Can you provide docs on how we can fine-tune DALL-E 3 by API calls?

[low dew]

Maybe just parameters

[onyx ridge]

Because I'm pretty sure that's not true, too. You may get control over the final prompt, but that's not fine-tuning the model.

[low dew]

All I remember is low,med,high outputs

And the “revised-prompt” being injected

Curious to dive deeper into understanding how the revised prompt fiddles around with our calls

Maybe it’s more than just adding more details and making it safer

[woven wren]

This is the doc, I may have oversold it but functions may be added (like edits were added to DALL·E 2) https://platform.openai.com/docs/guides/images/language-specific-tips?context=node

[onyx ridge]

The main heading seems to be Image Generation, not fine-tuning.

I'm visually disabled so I have to be very careful with text.

[woven wren]

E.g. there's an intervention measure there for disabling automatic prompt rewrites

[onyx ridge]

Well, in future let's not equate that with fine-tuning and claim insider information, right?

[low dew]

Oh I gotta look into the canvas thing

[onyx ridge]

How about them pixels?

Anyone doing any challenges or working on anything cool?

Here's one of my favorites: a surreal and photorealistic scene with the least plausible setting and subject imaginable

[quartz vale]

Those are some awesome pastel colors

[glad stone]

I just made an API interface that is a really well working voice assistant, and it can generate images and remember past messages. I asked it to make an angry spongebob and asked it to make the 1st spongebob angrier and it made this (I removed the background myself though)

Pretty cool imo

[fickle ravine]

I like this one

[onyx ridge]

Thanks!

It's fun to see what the model can come up with when you let it be as creative as it can be.

[dawn aurora]

anyone getting errors right now making images

[formal osprey]

yep

[grand dagger]

Yes, I am. I thought it was the prompt. The when I asked for an image of the sun and it still error failed.

[dawn aurora]

same, maybe they are updating something right now

[formal osprey]

I generally ask for a cat when I want to test if it's down I wonder if that tells something about personality

[fickle ravine]

OpenAI KNOWS about this issue. They just simply don´t care.

[formal osprey]

uh

[fickle ravine]

Let´s see if removing the word teenage will do anything

[dawn aurora]

its all images

cats suns, chinese new years

sunny cats celebrating chinese new year riding a dragon

[formal osprey]

the service is down

[fickle ravine]

OpenAI scams us like this and they don´t even feel bad

Everything I want to do is engaging in daily theme!

But they won´t let me!

[grand dagger]

Now if only we had a system working to test such great prompt ideas. 🤣

[fickle ravine]

Why do they even make the daily theme when they refuse to make those imaged

I seriously want to talk with an OpenAI employee because they can´t even be serious at this point

This is a literal waste of prompts

[fickle ravine]

They want me to waste my money on something that barely even works

Yeah I know I can just stop, but I am addicted

[dawn aurora]

i can tell lol

[fickle ravine]

Yeah, everything I want is to engage in daily theme

The worst thing is that you won´t get any compensation for errors

I think dall-e has technical issues, this is not even a problem with content policy

[dawn aurora]

lol nice prompt

[fickle ravine]

The problem is, OpenAI KNOWS about this issue for a long time, but I highly doubt they even had the idea of doing anything about it

[dawn aurora]

i mean it started about 15 mins ago

ima just check back in an hour

[fickle ravine]

Wait you have the same issue?

[dawn aurora]

lol yes

[fickle ravine]

Seems like a dall-e server error

[dawn aurora]

its down, we said that above

[humble lichen]

Yeah system for images is down

[fickle ravine]

It´s been going on for 15 minutes, why do they not even bother putting a warning message on the ChatGPT website so people don´t waste promptsd

That is like the literal least they can do

[humble lichen]

For real

[grand dagger]

Read above, I asked for the sun, someone asked for a cat. Nothing for anyone. Take a break, come back later. System is down.

[fickle ravine]

But if everyone is experiencing this issue, so it is understandable

[humble lichen]

Maybe they are trying to feed this discord because this issue brought me here

[dawn aurora]

yea im not going to hit my rate limit so ill just midjourney it

[fickle ravine]

They should give me a part of my money for this

Usually when a service doesn´t work, you get a refund

[verbal sequoia]

Showcase your DALL·E creations to the world!

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

Quick tips for selection:

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

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

[fickle ravine]

Now this is funny

HOW am I supposed to showcase when it is down???

[dawn aurora]

worked for me

[verbal sequoia]

<:book_icon:1171408210398289941> `` Rule 11 ``  Adhere to moderator actions.

Attempting to avoid bans or mutes by using alt accounts or other means may result in escalated infractions or bans.

[humble lichen]

Still getting the error

[grand dagger]

You tell me what systems / sites do that and can react that quickly. Their main employees might be in the US and from my reckoning, it's 2:25am EST. Give them a little time to wake up and look at the problem. Besides... If I hit a basic error a couple of times, I just leave it. No big deal.

[dawn aurora]

https://designer.microsoft.com/image-creator

[humble lichen]

Correct but this is 20$ a month here. Should have more support in my opinion.

But hey error happen I guess should be fixed soon.

[formal osprey]

I'm curious why it breaks now though. the API is still working. So it's something within chatGPT

[surreal iron]

Is DALL-E down rn?

[late blade]

do competitors generate images just by typing to the chat? so often I see people saying some random incoherent concept (i.e. a video or image with something else) and then disappear

[low dew]

What so you mean

[late blade]

people go to a channel, say "something something" and never appear again

[low dew]

(｀_´)ゞloll

Like first timers?

Kinda like the early “prompt-lab” days

[late blade]

not only first timers

[vagrant zodiac]

competitors?

[low dew]

/Dalle-3 ＼＼\٩( 'ω' )و //／／ I command you to butterFlyRainbowUnicorn!!!!

[late blade]

other services that don't belong in this channel

[vagrant zodiac]

oh. no, people think they can generate by typing in here, is all

when really they can now generate with a command in #image-bot !

[late blade]

ya, it's weird that people do that. it's like they don't know what they are doing and don't even bother to ask

[low dew]

lol

[late blade]

I demand you give me the prompt to that one now!!!!

[low dew]

Ok

Uh

I forgot to save it!!!

[late blade]

I'm gonna rage at you

I need that prompt to be back to my balanced personality!!

[low dew]

Imagine a serene, ethereal landscape at the magical hour of twilight, where dreams and reality blend seamlessly. This dreamscape serves as a tranquil backdrop for a vibrant celebration of the Chinese New Year, with adorable chibi-style fantasy creatures gathering in joyous festivity. These characters are rendered in a smooth animation style, contrasting beautifully against the softly glowing horizon. The sky transitions from the warm hues of sunset to the cool purples and blues of early evening, with stars beginning to twinkle. The scene captures a harmonious blend of dreamlike wonder, cultural celebration, and soothing tranquility, incorporating themes of Dreamscape Mosaic, Reflections of Solace, Ethereal Horizon, Twilight Serenade, Chibi-Fantasy Style, Animation-Style, and Chinese New Year, creating a visually rich and emotionally resonant image.

[late blade]

yay thank you!

[low dew]

Yesh

[vagrant zodiac]

btw, #images-canvas is now the channel for sharing images!

[low dew]

O

[vagrant zodiac]

it's no problem, channel is still fairly new 🙂

[late blade]

hence, I moved the examples and conversation to #images-canvas message , just helping out people to get around.

[late blade]

I was thinking of maybe introducing a concept for a weekly challenge or so. not just the daily theme. something like just submitting once per week, a small project. You have to upload 4 images and dunno how to go beyond that. Dunno what mods and uppers can do with it.

if I get more ideas I'll post them in suggestions

[onyx ridge]

One image a week doesn't seem very engaging, but you might get some participation anyway.

I can just imagine the slow mode indicator...

167:59:59...

[late blade]

I don't mean an image, I mean a project composed of a few images with a concept in mind. not just rush to the 10 iterations of the one concept a person can come up and then post it in daily theme.

[onyx ridge]

So, one post with up to five images per week?

[late blade]

just an idea, basing on the current post limit

[onyx ridge]

I get it.

I'm not sure about the format but I get it.

[late blade]

daily theme is ok, but I would love to see more extensive projects and ideas. dedicate more time to do something more complex, I'm starting to feel I need more challenges in that regard

oh, my whole idea is to propose a new way to engage with dall-e and something more complex

but it's just a vague idea atm, the format, no clue about it

maybe once dall-e chats are shareable can do that and just share your project link

[onyx ridge]

Bring back the seeds! 😁

[late blade]

once it works consistently sure

[onyx ridge]

I wonder if the best way to get it working consistently might not be to disable it entirely, though.

For other features and their output, we rely on model improvements over time.

[late blade]

I agree, I read the conversations you had earlier. But with no larger insight of what's happening on the backend I couldn't do an educated opinion.

I'm missing too much info

[onyx ridge]

Totally understand that, and fair enough.

[late blade]

I'm getting a similar discussion with Meditron LLM

[hoary patrol]

I'm new to DALL-E but is there a way or anyone that's knows this type of prompt style?

[late blade]

cool thing, if you want to discuss it, join us on #images-canvas , that's where we talk about images from other users.

[hoary patrol]

Thanks

[celest pebble]

I don't know if I'm the only one, but I have noticed since this week that sometime there is those very low quality result that seems to slip in. Here an example. It looks like if it's done by Stable Diffusion instead of Dalle. Anyone else?

Usually it gives me a set of bad low quality result, I refresh the request and the next one is "normal" again

[onyx ridge]

Looks like a style issue rather than a quality issue. Zooming in, every line is crisp and smooth. It's a simple design but not a low quality image.

What was the prompt?

[round scarab]

hi - is this where i can report a strange behavior in DALL-E?

[onyx ridge]

https://discord.com/channels/974519864045756446/1070006915414900886

[round scarab]

i'm not sure it's a "bug" more a weird response to a prompt i'm giving it

[onyx ridge]

https://discord.com/channels/974519864045756446/1204360881593520128 then.

[celest pebble]

@onyx ridge Its not that. its really different in style then everything else generated with the same prompt. It happens currently every 10 to 15 generation. They look a bit like if the result came out before the actual generation process was fully completed

[onyx ridge]

Sounds like a bug if you can replicate it. https://discord.com/channels/974519864045756446/1070006915414900886

[fading spoke]

can anyone help me come up with a good prompt for todays daily theme?

[onyx ridge]

Try just copying the theme, pasting it into ChatGPT, and asking the AI for inspiration. Or just straight to an image.

[fading spoke]

i did but they all are so, SO bad

it doesnt really know how to create anything

[onyx ridge]

It seems like you already have something in mind, then. Tell it that instead.

[fading spoke]

I might, but i dont know what it is

[fickle ravine]

I love the sky, it feels so nostalgic for some reason

[low dew]

Yes! To get that sky, simply use the following themes.

A sky showcasing nostalgic themes of: Dreamscape Mosaic, Reflections of Solace, Ethereal Horizon, Twilight Serenade, and Voyage Dusk.

[low dew]

The prompt I used was: "A wide canvas filled with a sky divided into sections, each showcasing a unique nostalgic theme. The first section, 'Dreamscape Mosaic', features a patchwork of floating dreamy landscapes, softly blending into one another with a surreal touch. The next, 'Reflections of Solace', portrays a serene lake under a gentle sky, where the clouds and the tranquil waters mirror each other in perfect harmony. The third section, 'Ethereal Horizon', transitions to a delicate horizon where the sky meets the edge of the world in a pastel haze, suggesting infinite possibilities. 'Twilight Serenade' offers a serene twilight scene, where the last rays of the sun cast a warm, soothing glow over a quiet, reflective moment. Finally, 'Voyage Dusk' closes the canvas with a deepening dusk scene, hinting at the beginning of a nocturnal adventure under starlit skies. Each theme flows into the next, creating a cohesive yet diverse panoramic view of the sky, evoking a sense of nostalgia, tranquility, and wonder."

[low dew]

Hope that helps! (๑･̑◡･̑๑)

[fickle ravine]

I lately do that that I tell AI to remake some of my old artworks, and the results are not bad at all

[fickle ravine]

2 is very epic

[pulsar sundial]

Yees

[fickle ravine]

I like this one too

[tight roost]

Has anybody noticed that when ChatGPT 4 demand is high, that it chooses to generate two Dall-E 2 quality images instead of one Dall-E 3 image?

[mild moth]

Hey folks, I'm building a web tool just like the fashion AI startup ca.la, which allows users to choose the product they want to get designed, just from a simple template (screenshot attached). They have integrated dall.e api to generate the designs.
How do you think the prompt construction works in the backend?

[late blade]

<item template description> made out of <materials> with <anatomy of the garment> and <details of the changes>

just an idea while getting sleepy before bed

I mean if you are going to use the API with DALL-E model, you'll have to pass at least this:

client = OpenAI()

client.images.generate(
  model="dall-e-3",
  prompt="A cute baby sea otter",
  n=1,
  size="1024x1024"
)

[mild moth]

hmm yeah, I too was thinking to try something like this. But I'm worried that this approach might not extend to all combinations of user input. I think i will have to do some trial and error with the above approach to arrive at the right wording for the best results.

[late blade]

I'm lazy, so I just do something like a json file and pass it to gpt first and then to dalle

and then magic happens

[mild moth]

got it, will try that, thanks!

[late blade]

Prompt was 200 words for a concept

[onyx ridge]

I've seen this too with Designer, which is the same thing. I think it's a moderation technique to avoid accidental generations that violate the rules based on vague prompts, but I can't be sure.

I'm not saying your prompt was vague, mind you. Just that this is my best guess from seeing this a few times so far.

[celest pebble]

I'm trying copilot pro which give you access to dall-e and it's so much better

I was able to hit the daily limit yesterday after generating around 800 images, I though it will never stop. It's faster and it output 4 images each time and there is no rate limit or anything like that. You can just keep going at it

[mystic epoch]

[celest pebble]

dall-e in openai is like a garden hose, copilot pro is like fire hose

[mystic epoch]

where can I access this copilot pro

[celest pebble]

you need to subscribe to copilot pro on microsoft

[mystic epoch]

microsoft edge?

[celest pebble]

no, it works in firefox, its on your microsoft/skype account

just google copilot pro

[mystic epoch]

how efficient is copilot pro when tackling computational mechanics problems

[mystic epoch]

sounds good, thanks

[celest pebble]

I havent tested the chat a lot yet (just got copilot pro late last night). If you want you can give me a prompt and I can run it for you. but for dall-e its much much better, I'm very satisfied

[celest pebble]

Yes, I have hit that a few times over the pass few days. No more problem with Copilot Pro

[low dew]

I saw someone mentioning the "crappy two dalle2 images" during peak times

I had the same experience too. Anyone else

I ought to get up and start working from my pc one of these days lmao

[boreal gate]

I regenerate and it usually gets better

And thumbs down

[fickle ravine]

yes

Not that much of a difference tho

[fickle ravine]

It happens often but it is not that much of an issue for me really

[late blade]

It was a normal prompt like I'm used to do when I use dall-e on gpt+. It has happened sometimes, that wouldn't be the first. But that's MS's problem.

[late blade]

Both are different experiences. What I like about OAI's approach it let's you work with so many tools. Copilot is also good, but image generation is really dumbed down because the prompts can't be that long as with GPT. Also the lack of file handling in Copilot Pro restricts my workflow. Also, when working with Copilot and you enable search, it's just a bing search with some prettier prose text.

[onyx ridge]

Good morning. 🙂

[celest pebble]

If your prompt is too long (mine are) you need to give it to copilot chat and it will generate the image from there. Have you tried it or its just what you heard? I have started testing the chat as well and so far its really good. You can even switch between gpt4 and gpt4 turbo. And compared to the free copilot its not just a bunch of links. So far i prefer it over chatgpt, just hope they add custom instructions

[celest pebble]

Send me a long prompt, i will send you the result

[late blade]

It happens a lot on copilot compared to gpt

With GPT, you can go and ask why it happens, with copilot, haven't been able to get a solution

[celest pebble]

@late blade is it with pro or the free one? Can you paste the text of the prompt here

[late blade]

pro

[fickle ravine]

You guys are having this sort of issues?

For me it works normally

Except for common errors experienced by everyone

[celest pebble]

@fickle ravine I dont have any issue with it. I did 1000 images in just over 24 hours without issue. I just cancelled my openai sub, for dalle copilot pro is definetly the way to go until openai relax the restriction they have in place (rate limit, 40 prompt per 3 hours and 200 images per day)

@late blade I just tried part of your prompt and the issue is not the prompt lenght, its the content. I get the same result with just asking for an image of the demon

[late blade]

This is one of many times this happened

[ripe scroll]

Anyone ask Dall-E about things around geology books or resources? lol, All the varying different patterns at micro and nano scales of rock patterns

[celest pebble]

@late blade I guess there is some sort of filter on the keyword, for the type of content I produce I dont have issue with that.

[ripe scroll]

Personally interested in the patterns of growing things at those scales

[late blade]

like I said, it's just one example of many that I have come across. I'm not one to challenge content policy stuff. If it was environment specific I could rule that out. But has happened on enterprise and personal copilot pro and multiple devices. It's just my user experience. Also noticed on the personal account that when I return to a chat in the history, the images count toward boosts even tho they shouldn't. So I'm not having a good user experience for now. And its not only with image generation but with other aspects, even within MS Office functionality.

[late blade]

Go nuts with it, explore and see what you can find. But do share it with us. We are all learning in the community.

[ripe scroll]

I'm currently in a shelter for the month, firewall and everything. No credit card, was stolen 2 years ago. End of the month i should be better situated to start purchasing things.

[late blade]

Let's hope that by then all is good!

[ripe scroll]

Should, have made a lot of big moves the past 3 weeks in effort to set things on a better course.

[late blade]

Nice!

[late blade]

In the mean time, you got a little peek with #image-bot 5 Gens a day, labs and free copilot

[ripe scroll]

oh nice, didn't know that

The discovery of Sanfordiacaulis densifolia just this year, thats neat, also interesting language that i'd wonder how it's work with the generator

[late blade]

tbh I dunno if there have been changes to labs. ask around or try it and see a more up to date information

[celest pebble]

@late blade I'm currently generating and its first time I do it during peak time (now). And I just hit that issue for the first time where copilot told me it send it for generation but nothing happens. If I ask it can submit again. I guess it's capacity issue and some request get dropped. When I generate early in the morning and late at night no issue with that.

@late blade it seems like sometime the opposite happen and I get 2 generations and 8 images in one go (which is not bad)

[late blade]

For me it's just the experience I've had so far. It might change. I still got the enterprise option and personal untl march. We'll see how it goes.

[celest pebble]

@late blade but even then, I can generate so many more image in one go than with openai tool. But Im sure microsoft will improve the UX

[late blade]

Ya, I'm not saying no to copilot, I'm saying for my needs it has to be more robuts. From the experience with MS or OAI, I had better exp with OAI. It can drastically change between users

[celest pebble]

im just using it to create image for blog post, and with openai I can get 4 or 5 that works during a day, and I need to do it in 3 seatings (80 images than blocked for 3 hours, an other 80 images, then a last 40). With copilot pro I can get 20 or so that works in one shot during the day