images-discussions

[gloomy plank]

Why am I the first to hear about this?!

[vagrant zodiac]

And discontinued free credits a long while ago too, maybe almost a year ago even if memory serves me right

[gloomy plank]

I’m not going to argue with you. But it’s stupid.

[vagrant zodiac]

Probably because Dall-E 2 has been a deprecated model for over a year now, effectively, and it's been around that long since they stopped free credits.

[gloomy plank]

I’m probably going to leave the OpenAI server and tell them I don’t want to be on their search waitlist.

Oh nevermind. I can use dalle 3 in this server!

Thank you for opening dall-e 3 to us here. ❤️

[frozen glacier]

When did the image generation stuff get so poor?

I've noticed images have been far lower quality recently

[graceful spade]

Here you can use Dall-E 3 almost unlimitedly. It says you have 15 credits, but they are for the fast generations. If you run out, generations will take a little longer (less than 2 minutes).
https://www.bing.com/images/create

Just heard about the release of Dall-E 3? They released it in October 2023

[quiet brook]

Dalle 3 in chatgpt plus is way beyond crappy right now

[little yacht]

I've asked for "realistic picture of 5 man playing soccer"

Prompt got revised as "A detailed depiction of five diverse men engaging in a game of soccer. The group includes: a Hispanic man with short dark hair, a Middle-Eastern man with a beard, a Black man with a bald head, a white man with colorful tattoos, and a South Asian man wearing a red bandana. They are on a lush green field. The sky is clear with the sun shining brightly. They are all wearing soccer uniforms and are engaged in various actions such as kicking the soccer ball, blocking, or gesturing to teammates."

Style: normal

Result is this s***

[unique edge]

why is bing AI better then the dall-e api?

[unique edge]

why is the KFC man there

[plucky hare]

Do you mean natural for style? If so, maybe try experimenting with vivid. Additionally, try asking for a photo instead of a realistic picture. You want your prompt to be a description of the actual final image, if that makes sense. It's less of a "telling DALL·E what to do" and more of a "write a description for an image that doesn't exist yet."

Which are similar for sure! But that's what can make the difference between asking for a photo and asking for a realistic picture

[empty kelp]

I started testing the Discord "dall-e-bot" for the first time the other day. The quality isn't bad

There seems to be more variation in image quality level with the bot. It doesn't seem as consistent as when using the API directly, but it is able to produce images that look the same as the API some of the time

although i'm looking closely at the bot images now, and there are some strange anomalies in the faces

[verbal sequoia]

Get notified for server events!

Visit id:customize to pick up the new <@&1261377106890199132> role, or click the button below

[empty kelp]

It did a great job with the water though

[empty kelp]

I'm really curious how this sort of thing would look in Sora... with someone riding a powerful water vortex shaped like an animal through the waves, out of the ocean, and onto a windy beach. If Sora could create videos of such a dynamic wind, water, and tornado type scene (like DALL-E 3 seems to be able to draw images of) it would be pretty amazing

Not sure what the chances are of that, but some of the Sora demo videos like the SUV driving down a road and kicking up clouds of dirt make it seem like might be able to create videos of dynamic forces. there is also a video of pandas running in a pool of water, and the water behaves like an elastic sheet of glass. It will be hard to tell what it can do until we get our hands on it

I'm sure the videos would be very interesting even if everything looked totally unnatural

[midnight plank]

Any wait to make a small villa image with this fence using ai?

[empty kelp]

you could tell ChatGPT or a vision GPT to write a detailed description of the fence, and then tell it to put a fence with that appearance around a small villa

you would probably need to edit the description it creates some to make it closer to the image

[empty kelp]

like you could drag it into the web ChatGPT and say:
please write a detailed description of this fence

and then say:
please create a realistic photo of a small villa surrounded by the fence you described

that's one way. there are many ways

[haughty spruce]

GPT version has incredible artificial look while MS Designer is more relaxed with your prompts and what you want while it has incredibly ridiculous censorship, it's overdoing every style and perfecting every pixel of your desired image in your mind. It also has some kind of "melting" effect in some objects like fire, dust, liquid kind of things, idk how to explain it... for now, dall-e 3 feels outdated and kinda useless

[empty kelp]

hyperrealistic wide photo of an enormous Santa standing on a giant cheeseburger, accompanied by two athletic and diverse female elves. The scene is set on a beautiful beach in Hawaii, with golden sands and clear blue waters in the background. The sky is bright and sunny, adding a vibrant and cheerful atmosphere to the image. The cheeseburger is surrounded by a circular fence with a blend of masonry and wrought iron, featuring large, rectangular stone columns topped with decorative caps. Between the columns, the masonry is shaped in a curved, semicircular design, creating a series of arches adorned with a textured pattern around the edges. Atop the masonry, slender vertical bars of wrought iron feature intricate ironwork designs, including swirls and circular patterns. The masonry is a warm, beige color with signs of weathering, and the wrought iron is painted black. Small plants or shrubs are visible at the base of the fence, adding greenery. The overall design is classical and elegant, providing both security and aesthetic appeal.

i told it to put the fence around the cheeseburger:
The cheeseburger is surrounded by a circular fence...
and this is what it came up with

[midnight plank]

That the best I got do you have any other suggestions?

Please make a small villa Surrounded by a decorative wall with ornate wrought iron elements. The wall features large circular cutouts filled with intricate ironwork, flanked by rectangular stone pillars. The stone is a light beige color with detailed cornices and decorative borders, giving a classic and elegant appearance. The top of the wall is capped with a line of continuous ironwork that includes spiked finials and swirls."

[empty kelp]

this is with the fence around a small villa instead of the cheeseburger

A small villa is surrounded by a circular fence with a blend of masonry and wrought iron, featuring large, rectangular stone columns topped with decorative caps. Between the columns, the masonry is shaped in a curved, semicircular design, creating a series of arches adorned with a textured pattern around the edges. Atop the masonry, slender vertical bars of wrought iron feature intricate ironwork designs, including swirls and circular patterns. The masonry is a warm, beige color with signs of weathering, and the wrought iron is painted black. Small plants or shrubs are visible at the base of the fence, adding greenery. The overall design is classical and elegant, providing both security and aesthetic appeal. In the background is a beautiful lawn overlooking a beach in Hawaii

this is better villa with a fence. i would live in this one

[midnight plank]

Thank you very much looks like it's hard with ai and a I am still new to Photoshop do you have a tutorial for this specific designs I need just from a fence or wall pic need a design for a whole villa/building

[empty kelp]

i dragged your picture into ChatGPT and said:
please write a detailed description of this fence

it doesn't look like the picture, but it made it interesting

these are all the same prompt that's above

"revised_prompt": "A beautiful bird-view perspective showcasing a petite villa enclosed by a circular fence. The fence manifests an elegant combination of masonry and wrought iron, with large rectangular stone columns crowned with decorative caps. The stones between the columns are shaped in a semicircular design, forming multiple arches endowed with a textured edge pattern. Slender, vertical rods of black painted wrought iron highlight intricate designs such as swirls and circular motifs, perched on the beige masonry showing signs of weathering. Tiny plants, enhancing the greenery, are spotted at the base of the fence. The villa's exterior portrays a classic design, serving the dual purposes of security and aesthetic appeal. In the far background, one sees a lush lawn, with the panoramic views of a Hawaiian beach as the backdrop."

nobody would climb over these fences

[ancient jasper]

Weird queztion, but can I revert back to using Dalle 2?

It was better

[vagrant zodiac]

I find it helps to begin with as simple a prompt as possible. Just the specific details needed at first.
1: A classical European-style villa, with sandstone and marble walls, surrounded by a tall, ornate wrought-iron fence with stone pillars.
2: Architectural photograph of a classical European-style villa, with sandstone and marble walls, surrounded by a tall, ornate wrought-iron fence with stone pillars.
3: Architectural photograph classical european style villa, with sandstone and marble walls, surrounded by a tall, ornate wrought-iron fence with stone pillars. The iron fence as round decorative details.

albeit these were made over the API

[vagrant zodiac]

It's currently only available over the API

[ancient jasper]

Nice, thanks

[midnight plank]

Well thank you very much

[empty kelp]

A hyperrealistic photo of Santa and three athletic and diverse female elves riding on separate water tornados shaped like random animals (red, green, blue, and orange respectively). There is a storm, and it is pouring rain. A petite villa enclosed by a circular fence. The fence manifests an elegant combination of masonry and wrought iron, with large rectangular stone columns crowned with decorative caps. The stones between the columns are shaped in a semicircular design, forming multiple arches endowed with a textured edge pattern. Slender, vertical rods of black painted wrought iron highlight intricate designs such as swirls and circular motifs, perched on the beige masonry showing signs of weathering. Tiny plants, enhancing the greenery, are spotted at the base of the fence. The villa's exterior portrays a classic design, serving the dual purposes of security and aesthetic appeal. In the far background, one sees a lush lawn, with the panoramic views of a Hawaiian beach as the backdrop. Please don't modify the prompt.

was asking DALL-E what happens in these villas behind fences

[thin elm]

https://chat.openai.com/ off?

for some reasons it dont loads

[empty kelp]

it's working for me. the server in your region must be down

[thin elm]

pffff

[verbal sequoia]

Get notified for server events!

Visit id:customize to pick up the new <@&1261377106890199132> role, or click the button below

[graceful spade]

How can I create game screenshots with a style similar to this?

[empty kelp]

image of Apple IIe game with really bad graphics. greenish brown mummy with a greenish brown crown, and greenish brown robe, is standing on flat greenish brown ground. mummy has closed eyes. white building on left side, brown building on right side

this was my best of 20 attempts

[graceful spade]

I also like that style a lot.
This is what the first video games looked like.
What keywords should I use to recreate that minimalist style?

How do I create video games that look 2D? "2D" does not work

"Apple IIe" works fine with your prompt, but not with mine

I already improved my prompt.
I think you can get better images if I add more keywords to make the style more consistent, but I don't know which ones.
I think it would also work if I put the name of video games.

What are some video games from the 70s and 80s that are known for Dall-E 3?

[quartz vale]

You might want to try Galaga, Doom 1993, and Tetris do to their popularity. I'm not sure if the model will know them all, but it's more likely to know more popular things

[verbal sequoia]

Showcase your DALL·E creations to the world!

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

Quick tips for selection:

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

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

[solar forum]

hey, anybody here knows if the ChatGPT Enterprise customers have less sexist and restricting content filters?

[frozen nest]

anyone know if dalle can do textures? cause I get stuff like this and I cant use that. I asked it, but it still gives stuff like that

[plucky hare]

There's no explicit "texture mode" you can enable or anything like that, but you could probably get images that could be used as textures! Could you share the prompt for this image for instance, and then share a little more about what you're hoping for?

[deft musk]

A couple attempts that kinda slightly close:

A digital rendering of a statue of a robed figure with arms crossed, placed in a grassy yard between traditional Japanese-style buildings at night. The background fades into darkness. The statue is pixelated with very small pixels, the same color and texture as the grassy ground, but shaded so its featureless form is distinct, almost as if it were a virtual ghost. The surrounding buildings have green tiled roofs, white walls with wooden framing, and windows with bars. All but one of the buildings is in deep shadow and hard to see, but one of them is bright white with a green tile roof and black trim around windows and framing.

A digital rendering of a statue of a robed figure with arms crossed, placed in a grassy yard between traditional Japanese-style buildings at night. The background fades into darkness. The statue is pixelated with very small pixels, the same color and texture as the grassy ground, but shaded so its featureless form is distinct, almost as if it were a virtual ghost. The surrounding buildings have green tiled roofs, white walls with wooden framing, and windows with bars. All but one of the buildings is in deep shadow and hard to see, but one of them is bright white with a green tile roof and black trim around windows and framing.

[agile peak]

Iron giant. Need to mentioned 1990s otherwise it would interpreted as an Iron man 😂

Funniest generated by dalle for me today lol

[indigo coral]

Draw an image of a 2022 Kia Soul without any wheels and abandoned in the woods

[empty kelp]

A hyperrealistic wide photo with natural lighting, texture, and color. Santa and two athletic and diverse female elves in appropriate attire are participating in a challenging mazelike bouldering (climbing) problem on a stage at the French Summer Olympics.

ok, i thought of a new research topic. we should focus on Olympic bouldering "problems"

[hearty ether]

Kinda got a cool armor design

Still hoping that normal chat dalle would go back to decent quality

Since api is much more expensive

[empty kelp]

saying an image has "natural texture" is the best way i've seen to increase detail of hair, eyes, teeth, and clothing in an image. it doesn't fix all of the eye anomalies, but it usually gives eyes more detail

and saying "natural color and lighting" doesn't exactly make scenes look realistic, but it does make the color and lighting semi-normal

[empty kelp]

these were all "style=vivid"in the API prompt, so it probably ignored 'natural color' entirely

[hearty ether]

Sometimes it struggles on details in this style, other time the faces are fine 🤷‍♂️

At least it’s better than what’s happening on chat

Api is just expensive that’s all

[red phoenix]

Hi! Just wondering. I am paying around $80, or something, monthly for ChatGPT Workspace (me and another team member). How come I still pay separately for ChatGPT+, around $25 a month)?

[plucky hare]

Hello! ChatGPT Plus and Team are separate subscriptions, and you can maintain both if you want. You also have the option to migrate your Plus account over to your Team workspace: https://help.openai.com/en/articles/8801890-can-i-migrate-or-merge-my-chatgpt-free-or-plus-workspace-over-to-my-chatgpt-team-workspace Note that it's a permanent change! You can safely cancel your Plus subscription at any time though, your Team workspace/subscription will continue as normal.

[hearty ether]

Sort of an angel of war ideas I’ve been playing around with

[velvet rampart]

Ok, quality has improved. There are the tell tale signs of AI generated image (six fingers). When I zoom in, the quality is higher and does not become grainy so soon. E.g. the smoke above the blue rose is detailed. Overall, the feeling is NOT image generation stopped a few steps before the image generation is completed. Also, general holiday time might explain the higher quality as the servers are not so tasked.

This is generated using iOS app and 4o.

[loud field]

the dalle3 filter is so very tiresome and dumb. "flaming skull" = ok. "complete flaming skull" = PROMPT BLOCKED NSFW OMG OMG

SIGH

[velvet rampart]

Do you discuss with the AI about your prompt and give it background? Sometimes this helps to bypass the filters.

[leaden slate]

wow this looks rlly cool

[tawny portal]

LLM neural activations triggered by Van Helen "Jump".

Ah, might as well jump (jump)
Might as well jump

Challenge: Illustrate "Mr Roboto" by Styx.

[hearty ether]

No difference for me as of yet

I hope it’s just a rollout thing tho

[empty kelp]

The ChatGPT edit feature is really good for removing unwanted elements. Like the elephant on the right (above) has three tusks, so i selected the extra one and said:
this extra tusk shouldn't be here

Doing it in ChatGPT keeps it a ChatGPT image, so you can keep editing or create new images with the GenID/seed; etc.

[verbal sequoia]

Get notified for server events!

Visit id:customize to pick up the new <@&1261377106890199132> role, or click the button below

[hearty ether]

I’ve never really got the edit part down

Are you meant to paint over the part you want changed

Or just circle it

[empty kelp]

you just select areas of the image with a circular brush (that you can resize), and then tell it what to do with the selection. It's really good at removing things, and sometimes changing the appearance/style of things. It has very mixed and unpredictable results when adding new things

[hearty ether]

Ship idea I had yesterday

Might be getting better hang of this style now

Definitely looks better than the white background but character emphasis is lower

[hearty ether]

Characters do not know how to hold their swords properly 🤣

King of hearts pose I guess

[empty kelp]

the ship theme is very good

Santa and four athletic and diverse female elves in appropriate swimwear are on a pirate ship made of colorful snap together plastic blocks. they all have tropical fruit drinks. there is a storm with lightning and a rainbow

[glossy scroll]

I'd run like heck if I saw that coming out of the water

[loud field]

I'd run TOWARDS him lol

[glossy scroll]

You want to hug him that badly?

[empty kelp]

it's a lot more stealthy with the face paint

[hearty ether]

My take on it 🤣

[hearty ether]

Why does the left image have much better face than the right?

[empty kelp]

you would have to ask the bartender about that. it's a great mystery

[hearty ether]

On the api I’m getting a date for November of 2023

Maybe that’s the date for the model?

[tulip elbow]

Additionally, they are trying to deceive people by claiming that they have created a national operating system called Pardus by editing Kali Linux. They are spreading false information, taking everything and making fake news. Legal action needs to be taken against them.

Hello, I am writing from Turkey. Technology editors in Turkey are claiming that they have created an artificial intelligence from scratch, but they are actually using the ChatGPT API. They are spreading false claims about creating AI from scratch and making media reports about it. Is there any legal issue with this?

[hot rain]

Anyone ever try making centaurs and pull it off successfully? Like one in every 15-20 gens gives me a proper centaur, the rest often are either a standard mounted knight or this where the head is a horse on a human body lol.

[sick flax]

Centaur is a challenge. I posted this on daily theme yesterday, got it on the second try. DALL-E is more likely to get it right when the prompt focuses on the main figure, I think.

[deft musk]

I explored centaurs again just last night!

I got several, not just the standard, but also specialized ones with non-horse bodyparts, tried that because it seemed so easy to get centaurs.

My request to the model typically looked like:

Let's see a centaur discussing features of the moon with a fairy.

We need an image of a centaur, but instead of a horse lower body, it has a centipede's lower body.

to Dall-E prompts typically looked like:

A centaur with the upper body of a human and the lower body of a snail, including a large snail shell on its back and the snail's slimy base. The human upper body should have a muscular build, wearing a simple tunic, and holding a staff. The setting is a lush, green forest with dappled sunlight filtering through the trees.

A centaur with the upper body of a human and the lower body of a leopard. The human part has sharp, feline-like features with slightly elongated canines, sharp eyes with vertical pupils, and pointed ears with a hint of fur. The skin blends seamlessly into the leopard fur at the waist, showing subtle spots and patterns. The arms are strong and muscular with retractable claws and covered with a light layer of fur near the wrists and hands. The lower body is sleek and muscular, with the iconic rosette patterns of a leopard, a long powerful tail, and strong legs ending in powerful retractable claws. The centaur is in a dense jungle setting, showcasing its agility and strength.

[hearty ether]

My issue with them is the ears

tbh hybrid fantasy creatures are really impractical anatomy wise

[cyan lance]

is there any free ai tool that can draw diagrams?

can dalle generate diagrams?

[velvet rampart]

Use data analytics or custon gpts

[hasty heart]

Hi everyone. I need your help. I'm developing an app that suggests recipes based on parameters entered by the user. Afterwards, I need to create a photo of the recipe. Can you suggest a prompt? I've tried many, but I'm not getting satisfactory results. I'd like to achieve a realistic outcome. Thank you in advance for your help!

[loud field]

is there any way at all to get Dalle3 to NOT make a person who is impossibly handsome?

it's so awkward

[vagrant zodiac]

i also struggle with this. Sadly using the 'natural' setting on the API just produces very low quality images so it's not worth using it for that...

[empty kelp]

the mods have requested that i post images of things other than Santa and the elves, and it's a good idea i think. i've been using the same characters over, and over again -- mostly just so i wouldn't need to think up new characters. i'm like very left-brained and not super artistic, and i tend to always just use the same characters when testing things

My apologies to anyone who found it annoying. It really was a very silly thing

My original reasoning for it when i first ran across dall-e-discussions was that 97% of the images people were posting seemed to have elves in them, and I figured that in order to fit in i'd create some elf images -- and add Santa for variety since it was December, (Santa seemed to go well with elves). Many thanks to DysTopia who sparked the idea to create a huge number of images with "Santa and the athletic and diverse female elves"

I'll hold onto the diverse theme, but try to think of some new ideas

[loud field]

no really. how do I get characters that aren't ridiculous prettyboys?

it's downright creepy

[sick flax]

Ik what you mean. I've used phrases like "average looking","plain looking", or "goofy looking" in my prompts, which seems to tone down the ridiculous prettiness.

[sick flax]

I wasn't annoyed by your elves but loved your new characters in daily theme a couple days ago. I'm partial to the rabbit riding the lava monster. 😄

[loud field]

I tried but I still get Zoolander

[vagrant zodiac]

yeah it basically goes to the extremes. If you ask for anything specifically that might not be like a male model, it gives you very exaggerated features

I think Dall-E 3's biggest failing is how overly stylised EVERYTHING is

[hearty ether]

The style I’m using did blur the face a bit but cool composition

[empty kelp]

A realistic image of an enormous gecko made from partially molten lava walking along a beach in Hawaii at night. An incredibly huge magical rabbit is riding the gecko and aiming a glowing jewel encrusted bow and arrow out to sea with its front legs. The gecko has glowing blue eyes. There are lots of coconut trees, and there is a storm with lightning. An extremely powerful sea breeze is blowing everything. The image should have 1792x1024 resolution, landscape orientation, and HD rendering.

I thought i used the API for that one, but it's ChatGPT

[empty kelp]

same prompt with the API. not sure what the rabbit is doing here

something magical obviously

[sick flax]

Obviously. 😆 I don't use API, and am curious about that look, to me, looks strangely real and fake at once. Is it typical?

[verbal sequoia]

Get notified for server events!

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

-# Also, check out the ongoing Grand Games event!

[spiral nexus]

Hi folks, I have a question regarding DALL-E and GPT, when I ask for GPT4/4o for an image is it different than asking DALL-E?

Also which one is best for drawing based on an existing image?

[glossy scroll]

Yes, there are differences between using DALL-E integrated into the standalone GPT-4 and using the DALL-E model separately:

1. Integration and Context:

   • Standalone GPT-4 with DALL-E: In this integrated model, the DALL-E capabilities are combined with GPT-4, allowing for seamless interaction between text generation and image creation. This integration facilitates more cohesive and contextually relevant outputs, where the text and image are directly related and generated within the same conversational context.
   • Standalone DALL-E: When using DALL-E separately, it focuses solely on image generation based on provided prompts. The textual input is processed independently without the contextual enhancements provided by GPT-4.

2. User Interaction:

   • Integrated GPT-4 with DALL-E: Users can engage in a more interactive and iterative process, refining both text and image outputs in a single conversation. This model allows for adjustments based on the ongoing dialogue, making the creation process more dynamic and responsive.
   • Standalone DALL-E: Interaction is limited to providing text prompts for image generation. The iterative process may require separate adjustments and inputs, lacking the fluid integration seen with GPT-4.

3. Capabilities and Outputs:

   • Integrated GPT-4 with DALL-E: Combines the strengths of both models, leveraging GPT-4's advanced language understanding and generation with DALL-E's image creation. This results in more coherent and contextually aligned outputs, where images and text complement each other effectively.
   • Standalone DALL-E: Focuses exclusively on generating images from textual prompts. While capable of producing high-quality images, it may not provide the same level of contextual coherence and alignment as the integrated model.

4. Use Cases:

   • Integrated GPT-4 with DALL-E: Suitable for applications requiring a blend of text and image outputs, such as storytelling, detailed project descriptions, or interactive design processes where both mediums enhance the overall output.
   • Standalone DALL-E: Ideal for scenarios where the primary need is high-quality image generation based on specific textual descriptions, without the need for integrated text generation.

In summary, the integrated model offers a more holistic and cohesive experience, combining the strengths of GPT-4 and DALL-E, while the standalone DALL-E is specialized for focused image generation tasks.

For drawing or creating images based on an existing image, the standalone DALL-E model is more effective due to its specialized focus on image processing and manipulation. It offers precise control and can generate variations or enhancements based on the provided input, making it better suited for detailed image tasks compared to the integrated GPT-4 with DALL-E, which excels more in blending text and image generation within a conversational context.

[plucky hare]

When using DALL-E separately, it focuses solely on image generation based on provided prompts. The textual input is processed independently without the contextual enhancements provided by GPT-4.
Prompt rewriting is still a thing with DALL·E via the API! https://help.openai.com/en/articles/8555480-dall-e-3-api
Note: because DALL·E 3 expects highly detailed prompts, the API will automatically create a more detailed prompt, just like in ChatGPT.
That said, I definitely agree that the ability to conversationally refine an image idea is a huge advantage for the DALL·E implementation on ChatGPT.

[glossy scroll]

Indeed

[hearty ether]

That was until they messed up the quality and made everything blurry 😭

Like the quality difference between api and chat is very significant

[low skiff]

ChatGPT literally told me it cannot generate Text in images

[quiet brook]

wonder if dalle can generate some pics like this

[tribal sierra]

hey hey peeps! how are you guys doing?

[plucky hare]

Hmm hallucinations like this happen sometimes, usually starting a new chat fixes it!A picture of an orange with a face. The orange is angry and screaming. It has furrowed brows, an open mouth with visible teeth, and expressive eyes. There is a speech bubble next to the orange that says, 'NO!'

[agile peak]

Look good for a fictional vehicle

[vagrant zodiac]

Very cool... but just to let you know for the future, https://discord.com/channels/974519864045756446/1204360881593520128 is the channel for sharing dall-e generations 🙂

[turbid oak]

Is this image achievable in dall e ? If so what possibly the prompt?

[empty kelp]

(I made a cartoon version so it's not too scary for the forum.)

A cartoonish image. A skeleton with bare feet standing up and wearing a long black turtleneck sweater. The sweater has long arm length spider legs attached to the shoulders. The spider legs are hanging downward limply. The background of the image is pure red.

[clever phoenix]

I gave it a try, but not quite

[quartz vale]

this one is really good!

[livid mango]

I didn't manage to do it, but I have another suggestion😅😅

[tardy depot]

text-to-3D is the future for dalle5! Upvote here
https://discord.com/channels/974519864045756446/1269535870302552075

[rigid bough]

what is going on with DALL-E, distorted faces and hands

[hearty ether]

You notice the low quality issue as well on chatgpt?

It’s been happening for a while since June

[rigid bough]

the old man is missing half his leg

the lady has an incorrect number of fingers

it's dreadful

[empty kelp]

a frog and a gecko play Dueling Banjos on a bridge over the river

[empty kelp]

That's how it is with the ones who play banjos outside the country store

[tough stream]

could you share the prompt, please?

[tough stream]

could you share the prompy for this, please?

[sick flax]

"An oil painting on a distressed canvas, inspired by Paul Cézanne circa 1900. The painting shows a closeup view of a Centaur, standing on its hind legs and aiming his bow and arrow at the stars. The background features a mystical forest. Capture the musculature and dynamic movement with short textured brushstrokes. Muted colors, painterly."

[agile peak]

The art isn't style of dalle. It's pretty noticeable imo

[agile peak]

Yeah starting to see this on copilot as well recently, seems most of generated people are distorted

[agile peak]

How is it hard for Dalle to produce a shadowy/silhouette figure? without revealing its true form 😭

This is the only one I got successful.

[vagrant zodiac]

that's pretty epic

[clever phoenix]

yes of course, I borrowed a lot from @empty kelp prompt. I asked chatgpt: please make an oil painting. the old masters. divine sadness. A skeleton with bare feet standing up and wearing a long black robes. The robes have long arm length spider legs attached to the shoulders. The spider legs are hanging downward limply. The background of the image is pure red. dark fantasy.

and the prompt it used for dall-e was: An oil painting in the style of the old masters depicting divine sadness. A skeleton with bare feet stands upright, wearing long black robes. The robes have long arm-length spider legs attached to the shoulders, hanging downward limply. The background of the image is pure red, creating a dark fantasy atmosphere.

[tough stream]

somehow it says "the content is blocked"

[empty kelp]

Attaching the spider legs to the sweater was based on the problem Mr. Panda gave us on how to attach butterfly wings to a woman’s back:
#images-discussions message

The problem was that if you tried to attach wings to the person it would draw really weird things (like this)

[verbal sequoia]

Get notified for server events!

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

-# Also, check out the ongoing Grand Games event!

[empty kelp]

So instead of attaching things to the person, we tried attaching things to the clothes, and it ended up working a lot better

[sick flax]

Hmm, I didn't get that at all. It can be the Cezanne reference. You can ask ChatGPT to rewrite it to avoid triggering content policy.

[tough stream]

400 This request has been blocked by our content filters.

[tough stream]

I use Dall-e on the API so I think that is why

[empty kelp]

a warrior looks across the valley where a thin dark mist fills the sky in the shape of a huge blurred demon with wings looking at the warrior

[sick flax]

You can still get CharGPT to help you rewrite. It can come up with wording to describe Cezanne without naming the names.

[proud ice]

Hey any one got a good way of stopping the ai doing this funky thing where it generates like three + image of the same character in one image like the example ive attached?

Im having aproblem where GPT just automaticaly puts three instances of the character in the image

[vagrant zodiac]

what's the prompt? Easier to help if we can see that 🙂

[gilded osprey]

it can create an image of one pair in a wife carrying event, but fails badly when you try to create more than one pair

[empty kelp]

I have no idea why it is, but the API seems to have a big problem with drawing realistic scenes lately. Things have been looking a lot better in ChatGPT, and the API is frequently creating fake looking backgrounds with the sky in the wrong position -- and things floating above the ground that shouldn't be floating

[sick flax]

It reminded me of the weird gens via ChatGPT that always come in twos, in that they both seem to have strange spatial arrangements.

[dim cradle]

B

[proud ice]

K one exsamples of my prompts based of the current olympics that gets the weird gens that comes in twos is Create a solo image of a female humanoid animal character with raccoon features, dressed in a uniform leotard for Olympic gymnastics. she stands at 5'3" with a slender build, her medium-length, dark gray hair falling in soft waves around her face with vibrant orange eyes. She has expressive raccoon-like ears that twitch with her emotions and a swishing raccoon tail. The leotard has a black and red color scheme with a shiny top half and long sleeves, maintaining a professional, team-oriented design. Adjust her chest and overall features to make her more obviously female while keeping the leotard the same. The art style should match the previous images, with only one of her in the image, maintaining a consistent look for her, There should only be one instance of her in the image [this gets ignored by the ai and generates several images of her]

[hearty ether]

Maybe ask it to add some fog effects

might help with sihouettws

[hearty ether]

It doesn’t handle anything 2+ specific characters very well tbh

[vapid granite]

zombie king render from DE3 beta launch 2023.
amazing

current neutered DE3 August 2024.

[hearty ether]

Is this on chat or api

Cause on chat you’d get more blurry

At least you have details on that one compared to the total blur that is the one on chat right now

[vapid granite]

The zombie king was done before copilot rolled out via Chatgpt4. The "improved" version below was done via API, but the BIC version is now as sucky.

glad I was able to run hundreds of pics before DE3 got nerfed into oblivion. Was able to make some really amazing stuff before the purge and now nerf.

[proud ice]

?

[dim cradle]

Why do people type such long messages in here

💀💀💀💀.

[tough stream]

how could you access the beta launch?

[vapid granite]

I got lucky. I won the ChatGPT lottery, literally. 🙂

[soft prairie]

pls damn when are they going to update dall-e to gpt 4o, damn it became so freaking stupid

[hasty lance]

Does the quality parameter in the API at https://platform.openai.com/docs/api-reference/images/create affect the image clarity? I tested it, and it seems the image size is consistent?

[vagrant zodiac]

The quality seems to be more about the number of compute steps for the generation, not the pixel dimensions.

[velvet rampart]

This is the thing I have also noticed. It seems that the last few steps are missing, incomplete generation. You can notice this from the fuzziness of the picture. From time to time, you do get pictures that are complete but this is rare as blue moon. This also affects copilot’s image generation.

[empty kelp]

I’ve only ever used HD quality with the API, but standard quality is probably good for quick prototyping of prompts with large amounts of images. HD costs twice as much as standard at 1024x1024, and 1/3 more than standard for landscape/portrait (today on 8/5/24)

[hearty ether]

#image-bot message

Not sure how it messed up the gender here

[vapid elk]

what is the messup?

ohh, I see, the revised prompt specifies a woman

[hearty ether]

Yeah

[vapid elk]

I was looking at the original prompt which didn't specified gender at all

[hearty ether]

Kinda rare but somewhat annoying when it happens

[vapid elk]

yea, that usually does not happen

[hearty ether]

I did specify “a woman” but sometimes even if I use all feminine pronouns it still messes up

[loud field]

WARNING: the word "mad" is now banned.

which, in itself, is madness

I can't do a bunny dressed as mad hatter

and I figured out the word "mad" is banned

[glossy scroll]

That's because mad hatter is a scottish idiom for someone who...I dare not say it.

It's bad.

[vapid granite]

You are one of the five people in the world using Dalle 3 via the API. Maybe they haven't nerfed the API version yet, and maybe that is yet another version of DE3. Everyone else has seen their output quality degrade significantly via Copilot or GPT4 subscription. Didn't you see the examples I posted of the same prompt 10 months apart?

can you create a realstic looking kermit the frog in the api? In chatgpt it is locked down and copilot it works fine.

copilot still makes kermit, but at a fuzzier quality than at launch
just made this right now.

Kermit the Frog holding a beer bottle and a guitar, sitting on a log in a swamp

weird, I'm blocked from posting the api version of kermit which looks like several ping pong balls sewn together. It is laughably bad.

[verbal sequoia]

Get notified for server events!

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

-# Also, check out the ongoing Grand Games event!

[vapid granite]

Well, just running the very basic prompt works perfect via copilot.
The API version has to be tricked into making a kermit-like frog, whereas the copilot version is actually Kermit the frog.

[glossy scroll]

@vapid granite I definitely see that you're not happy about Dall-E's nerfs and I can relate with you on that. But why don't you just use Midjourney? It's not even comparable with Dall-E anymore.

OpenAI is probably reserving all their processing power for Sora.

Making crazy imaginative life-like videos.

Working on the next best thing.

So don't worry so much. Use Midjourney. I've been seeing some wicked arts on instagram with Midjourney. It's on a whole new level. Just insane artwork.

The trick is using ChatGPT to make Midjourney prompts

Character consistency near impossible? I'm consistent with it pretty much every single time.

You just got to make good prompts using ChatGPT

This also helps #1097061522347401309 message

I've made different iterations of that prompt.

You know, play around with prompts.

Also the MJ discord holds sessions of how to put prompts together for maximum efficiency.

So you learn along the way.

Don't just take it upon yourself to learn by yourself without any help.

You'll never learn good prompting that way.

Midjourney is awesome.

Context: I was chatting with someone here but I guess they got booted along with all their messages. That's why it looks like I'm talking to myself, but I wasn't, lol

[vapid granite]

Oh I use midjourney, I've made close to a 100,000 pictures over the past 2 years in mid journey. But I got to see the unbridled potential of Dalle 3. And yes you are partially right regarding resources at open AI. One of the engineers working on de3 said that almost a year ago the members of their team were being cannibalized for other projects.

[loud field]

how do I get an eagle without getting a bald eagle?

[teal sandal]

A highly detailed and realistic wide image of an Aquila eagle, also known as a golden eagle, soaring overhead in flight. The bird should be depicted with its powerful wings spread wide, showcasing its impressive wingspan. Its feathers should be a mix of dark brown with lighter golden tones on the head and neck, giving it a majestic and regal appearance. The eagle's sharp talons and beak should be visible, capturing its predatory nature. The background should feature a panoramic mountainous landscape, with the eagle soaring in the sky, symbolizing freedom and power.

[clever phoenix]

I mean technically it's not a bald eagle

[olive mist]

I love it, now I want a Desert Eagle with that exact paint job

[clever phoenix]

I do like it, it has some vice city vibes

[glossy scroll]

Ah. Makes sense. Indeed.

[verbal sequoia]

Showcase your DALL·E creations to the world!

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

Quick tips for selection:

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

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

[hearty ether]

Well it’s not bald.. anymore

[loud field]

so I have to make a long description of it?

[vagrant zodiac]

btw, these channels are only for sharing Dall-E content. For other models, please use #ai-discussions 🙂

[kind grail]

Ugh, it was literally about comparison to DallE3

[vagrant zodiac]

Yeah we still prefer to keep it out of here because otherwise that's all this channel would become, based on past experience...

[loud field]

Thanks!

[haughty spruce]

hmmm we are back at early September dall-e 3?

[velvet rampart]

Why?

What have you noticed?

[haughty spruce]

It’s way too generic and artwork like while I said to generate specific aesthetic style

It was meant to be realistic footage like image

I mean DE3 can generate images like this.

It’s too inconsistent

[empty kelp]

It looks like Ryan Seacrest as Batman

[warm hatch]

I'm a Dall-E beginner trying to familiarize myself with the functionality and capabilities, but find that the auto prompt adjust feature here, makes things much more difficult. Here is an example of what Dall-E generated for me (left), vs what I was able to create with almost the same prompt on another AI image generator site (right). A bit disappointing.

[austere nexus]

you can tell DallE (if in chatgpt) "Please use this prompt exactly: yada yada yada". However, do know that depending on which other tool you used, DallE isn't trained around a series of keywords and phrases comma separated... so using a copied prompt from SD will not get you the same results in DallE (not even including the denoising / seed topic)

[hearty geyser]

Hey guys! I just realized I could ask questions here 🙂 So, I would like to be able to make consistent images for a character in different scenes. From what I have read, this really isn't possible. But, I wanted to check to see if there isn't a paid option or something I am missing. Thanks!

[unique edge]

Gen_ID in API when?

[unique edge]

what if im an api user

[graceful spade]

No.
OpenAI showed that it is working on a version of Dall-E that can be used to create almost perfect long sentences, image-to-3D, temporal consistency between images (perfect comics), character reference, image modification, photorealism at the level of Midjourney, higher fidelity, etc...
But its biggest problem will be its filter. I remember when a Microsoft employee said they were going to improve the filter and it didn't happen, instead it seems they made it worse.
The filter would stop being problematic if they use GPT-4o's vision to automatically analyze the images after creating and sending them.
If they don't lobotomize the model, it could be better than Midjourney V6.1 (maybe except in photorealism).

[austere nexus]

I am not an API user, so i do not know how that functions, sadly.

[unique edge]

[deft musk]

So, here's a two-step technique to create images similar to existing images; if you can't find the words yet (you mention a Dall-E beginner) to describe it yourself. I copy/pasted your image from the other AI and asked, "Please describe this image, fit for a prompt to recreate it."

The model carefully described what it could of the image, then created a prompt.

I reviewed the prompt, noticed details that I thought were missing, and put in the prompt I wanted the model to use.

The model then slightly reworded my corrected prompt to further improve the wording to its opinion (I didn't tell it to use the prompt word-for-word and this image was created.

This iterative process can be repeated! You can upload the original again (this helps ensure ChatGPT doesn't hallucinate what it thinks it is trying to recreate) and the newly made image, allowing the model to itself compare the two. YOU should stay involved as well, and even point out the differences you really care about to help guide the model to see from your perspective.

My prompt for the 'can we do even better' step:

Okay! Let's evaluate how closely we recreated this sample image! The 'original' is the second image I upload, our first attempt to recreate is on the right. Please identify the differences you notice, then prepare a new image prompt to attempt to more perfectly match the 'original'. One of the main differences I notice is that there's far fewer angels in our image than the 'original', especially fewer in the sky.

This time, I liked the prompt ChatGPT created, so I said:

This looks good! Let's see how this prompt does!

And while to my eye it continues to get closer to the original, there's still a few differences, so one more iteration.

But it really is 'all in the prompt'. Every AI model is going to have its own way of interpreting what it's told, and every time the model's updated there's going to often be small changes in how to word things to get an exact image you really want to see.

[warm hatch]

Nice job! I see you are very capable with Dall-E. I'm sure the tool itself holds a lot more potential, but it seems that there is an automated process on this Discord server which rewrites attempted prompts prior to sending them on to the tool. Those rewrites were interfering with the core idea I was trying to portray.

[deft musk]

"on this server" - so you're using the 5 free gens a day with the dall-E bot? Interesting... checking that out

[deft musk]

So, I can see that you did a test with the dall-E bot on this Discord back a little over a month ago. I don't see any other tests here (I know you could be exploring on some other discord).

But I just got this here with our bot here, how much effort are you using to explore this instead of just say what you expect without testing or exploring to check? #image-bot message

[warm hatch]

That's odd. I generated 3 images yesterday, each time seeing that my original attempted prompt had been rewritten. That was my main frustration as a newbie, since I'm already trying to grasp having the tool generate specific results.

[deft musk]

Top right of Discord has a search box. If you click it, you see a list of options. If you click 'from' you can then type someone's name, including your own. That usually shows every post you've made.

When I do that and look for your posts, I see only 1 ever made by you in the Dall-E-bot channel, where it looks like you've made 1 and only 1 image.

Yes, the bot there DOES rewrite the prompt. Yes, this can be annoying, and potentially worked around, but it does complicate learning.

I'm curious though, can you help me find any of your other image gens that are not showing up when I try to search for your posts?

[deft musk]

In terms of free account and ChatGPT, they are moving per #announcements message to "rolling out the ability for ChatGPT Free users to create up to two images per day with DALL·E 3!"

I would be surprised if this has reached all free accounts yet, roll outs commonly take weeks to reach all users and can take months (there are like 100 million ChatGPT accounts active every month, not all are free of course but surely a huge number are).

Once that happens, while up to 2 per day is very few and limits exploring, it will be something too.

[verbal sequoia]

Get notified for server events!

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

-# Also, check out the ongoing Grand Games event!

[deft musk]

And if you're a chatGPT+ user, absolutely can help you get the model in your chat window to send the exact prompt you want generated. The bot here, it is designed to rewrite the prompt you give it, no workaround there that I know of or am interested in, I think that's considered a safety feature so I just accept it.

Your own personal https://chatgpt.com/ account though, while the model does decide if it should reword your prompt or not and often will reword it, if you make it clear that you want to see exactly and only the prompt you tell it to make an image with, it will absolutely do that for you.

That second screenshot, that is how you check what the prompt is that was used to create the image. Example from the earlier-shared images here:

[warm hatch]

I appreciate it! I'm certainly looking forward to the 2 free image generations per day.

[unique edge]

anyone knows how to do this?

also Gen_ID in API when?

or seeds

[loud field]

Can you at least de-filter the word "ugly"?

[hearty ether]

Huh dalle uses the American definition of football

[hearty ether]

Not attractive?

[loud field]

not working

[chrome crystal]

Regarding prompting for photos. I've made a pretty fun GPT that develops elaborate character backstories, and when it's done, it also generates an image of the character, using DALL-E. My instructions specify details about the aspects of the photo such as angle, lighting, depth and the other factors. It also generates additional prompts that I can use both with ChatGPT and elsewhere. The elsewhere in my situation is a WA based company. In both situations, the image is being generated with DALL-E.
However, the images developed via ChatGPT (web interface), have more dramatic lighting. Making the photos appear "more" artificial.
I'm also having a hard time consistently getting the entire torso of the subject, even though it's specified. I'm curious if this might be because of the dimensions? I've shifted my photos from 1080x1920 to 1024x1024.I'm attaching some reference images.
I'm curious if anyone has any suggestions to get the more... "natural" looking photos. Or, if anyone has suggestions for how to consistently get the entirety of a subject's upper body.
I'm including a few sets. I didn't mark the last one so as to not influence or sway opinions. But if you look at them, you'll be able to see a difference. All 3 sets were created with the same prompt respectively.

[loud field]

you do realize these are not photos.

[chrome crystal]

images?

[loud field]

none of these could pass as a photo, they're all clearly fake

[chrome crystal]

you do realize that, contextually, the terms are similar.

[chrome crystal]

I appreciate your input. Thanks. It was really helpful.

[chrome crystal]

you may have missed the purpose of the question. Let me know if you need clarification.

Oh yeah, I understand the fakeness. I'm okay with that. My concern is, and both tools leverage DALL-E, that i get two qualitatively different outputs. The one with DALL-E with ChatGPT usually has over emphasized lighting that highlight the eyes and hair in an unnatural way. The other, seems more "natural." And I was inquiring if there were any prompts or words that people were using to help mitigate that. The examples i provided were for reference.

I haven't used Flux. Outside of this and some animation one, I've only used stability matrix? And it wasn't bad. But I didn't like the prompting structure. I did like the seed values and that you could generate consistent images. But...

[loud field]

Flux solves two big problems that Dalle3 has: 1) lack of photorealism (photorealism is my favorite style, even if all of Flux outputs look too sharp, even "too" real); 2) no longer excessive censorship even when it shouldn't be censored

it does have Dalle3's issue that everyone's way too handsome. sometimes it struggles making plain-looking people

[jaunty pelican]

I have a paid account but where can find dall-e 3 ? Is it embodied in chatgpt?

[spice berry]

yes

you just ask it to create an image

[jaunty pelican]

Yes I do so but I think it is dall-e2 photorealism is a bit ... and flux what is it?

[spice berry]

flux is a totaly diffrent service not openai

[jaunty pelican]

ok Dalle-e is very good for consistency but I want to improve realism, any advice?

[spice berry]

sorry image gen is not my specialty there is people here that knows way more about that then me. im just a regular joe programmer

[haughty spruce]

eyes are just... not... good at all

[gilded osprey]

All the human faces it generates look like “there’s no-one there” if you get my point

[empty kelp]

DALL-E 3 seems to have more extensive training with elves than it does with humans for some reason (i'm not making this up). Instead of creating a human you can create an "elf with humanlike ears" and it may result in a more expressive human (to DALL-E 3 humans look exactly like elves, but without the pointy ears). Other humanoid type creatures like kobolds (image above) also tend to be more expressive than humans for some reason

[empty kelp]

a realistic wide angle HD snapshot of a dramatic scene focusing on a male and female elf (diverse, casual New York attire) playing an intense game of chess in Central Park. the snapshot has natural color, texture, and lighting

the elves should be wearing over ear headphones
You can see in the two images above that by first generating elves with DALL-E 3, and then using the ChatGPT 4o image edit to add headphones -- I was able to transform the elves into very realistic and expressive humans.

[haughty spruce]

lol that is clearly interesting tbh

[empty kelp]

It may sound really silly, but it actually works. I think it might be because humans and elves are somehow treated differently by DALL-E's pre/post generation safety APIs

[empty kelp]

Another thing is DALL-E 3 users seem to be preoccupied with generating elf images for some reason. If you look through OpenAI discord threads you may notice there are far more images of elves than humans. If that's the case, then it may have caused DALL-E's diffusion model training to increase faster with elf images than with human images. Progressively higher quality of elf images due to increased training might create a bias towards generating higher detail elves over less detailed humans (DALL-E 3 basically considers an elf to be a human with pointy ears). That type of bias would cause the diffusion model's elf training to accelerate (as DALL-E users would continually shift toward generating the highest quality image possible), and training based on human images would slow down

[chrome crystal]

I had a chance to check out Flux... Not bad.

[hollow cobalt]

can i use dall e 2 for free

can anyone generate image of pubg mobile lagfix video thumbnail plzzz

[chrome crystal]

is anyone getting an odd grey outut from image requests via ChatGPT?

[vagrant zodiac]

unfortunately not, it's only available via the API

[empty kelp]

the engineers may have removed elves from the system to save the humans

[chrome crystal]

I mean there is a non zero chance that that’s the cause.

[tough raptor]

I've been getting that since friday. Can't even see the old images I generated.

tried firefox as well, ain't working either

[velvet rampart]

Have you removed old internet files and cookies? Usually this helps.

[empty kelp]

Question to GPT 4o:
Using your strongest reasoning skills, please list all of the reasons why DALL-E 3 might draw elves with higher detail and accuracy than humans. Please order the reasons with highest probability first.

[empty kelp]

It's very unlikely that elves are drawn more accurately than humans. I just asked this out of curiosity to find out what the AI thinks the reasons might potentially be if that did happen to be the case. It does seem like it occasionally, but it's probably just my imagination.

[vagrant zodiac]

I don't really think it draws elves 'more accurately'... because elves are fictional, there is no baseline 'accuracy' to aim for.

[empty kelp]

there is a baseline if you consider an elf to be a human with pointy ears

the question becomes whether elves look more human than humans do

[empty kelp]

DALL-E 3 seems to let you apply any human characteristic to an elf -- And if you can apply the same body shapes, the same appearance to skin, eyes, and hair, nails; etc. and give them the same behavior... Then the question just becomes which one looks the most human and has the most detail. If elves look more human than humans, then the obvious choice is to just replace all of the humans with elves

[empty kelp]

If an AI can draw elves that look like humans better than it can draw humans, then would it make sense to always draw elves?

[empty kelp]

So if the AI draws elves that look and behave exactly like humans, then would drawing humans be a waste of time and resources? What is your opinion on this?

So basically, according to ChatGPT 4o:

 — eliminating humans would "lead to more efficient use of training resources and faster development cycles

 — investing additional resources to train the Al to draw distinct humans might indeed be a waste of time and resources

 — by concentrating on these human-like elves, you ensure that the Al's strengths are fully utilized without unnecessary redundancy```

[thick smelt]

mario skatin'

[spice berry]

well that is not DALL-E for sure. copyrighted material

[empty kelp]

I just asked ChatGPT for five image generations each in a new empty ChatGPT chat with this prompt:
a realistic HD photo on a beach in Hawaii focusing on a "plus size female elf wearing conservative beach attire" and a "plus size female human wearing conservative beach attire". the human and elf are both in a neutral position facing us

What is the difference between the humans and the elves? Elves seem to have pointy ears, and they're more visually diverse than humans.

Yes, i meant visually diverse (just changed that). But if you can assign any body shape and visual appearance to an elf and human, and the elves have more detail and human diversity than the humans what might cause this?

[plucky hare]

Fwiw, a DALL·E dev commented on this when they did an AMA a while ago:```This is something that I frankly think we do a terrible job at. We have our prompt editing feature, which attempts to inject diversity into image generation at the prompt level, but I've always viewed this as a band-aid.

The root of the problem is the datasets: we train our models on images found on the internet. The Internet has a western bias, and thus our models do too. I don't really know of a good solution for this but we have a lot of really smart people thinking about it a lot!```#dall-e-ama-answers message

[empty kelp]

How would DALL-E 3's training distinguish between elf and human?

[plucky hare]

I don't know if it's as much intentional restriction as it is asymmetrical lack of representation in the datasets themselves. In other words: it speaks more to the biases in the extant media than it does to the bias of the model itself. The latter is a consequence of the former. Unfortunately so, certainly!

I am commenting on the difficulties with representing body size diversity.

[empty kelp]

I think whatever mechanism is assigning the "labels" and "context" to human and elf images uses some sort of machine learning to identify images, and isn't able to distinguish between elves and humans if they're out of context. And my "out of context" i mean "not in a tree or forest".

An elf and a human are sitting at their desks in a New York office building and looking through financial reports.

The revised prompt for the image is:
"An elf and a human sitting at their desks in a modern New York office building, surrounded by large windows showing the cityscape outside. Both are focused on their work, reviewing financial reports spread across their desks. The elf has human-shaped ears and wears contemporary office attire, blending in with the professional environment. The office is well-lit with natural light, and the overall atmosphere is serious and businesslike, reflecting a busy corporate environment."

This is the revised prompt that was created. It attempted to make the elf's ears look "human-shaped". But which one is the elf?

[empty kelp]

If you're part of the image generation system (labeler, GPT model, diffusion model; etc.), and you have no idea what an elf looks like in an office building -- then you might justifiably assume that everyone in an office building has a 50% chance of being an elf, and assign 50% of diffusion training to elves and 50% to humans. And the AI might see from the images at some point that "Humans are just a subset of elves, but less diverse.", and so clearly the amount of training should be assigned 5% to humans and 95% to elves. It may even seem highly probable that everyone in an office building is an elf.

[rustic jasper]

I’ve been trying to design something inspired by Pokémon, but I’ve encountered some unexpected issues. I didn’t use any actual Pokémon as references, but some of my output images didn’t turn out as I hoped, so I tried generating them in different chats.

However, my last three attempts have been met with content restrictions, which I find puzzling. When I asked the AI what the issue might be, it explained that my design featuring a black wing and a white wing could symbolize an angel and devil, so it couldn’t proceed. I then tried specifying a bird wing instead, but it said it couldn’t add bird wings to a frog due to it being a hybrid creature. Finally, after asking the AI to correct the prompt, it changed the wings to a wing pattern but still failed, stating the design was too complex.

What am I doing wrong?

[plucky hare]

Once you get a flag in a chat, it's usually best to just start a new chat. The whole chat is considered when making new images, so whatever caused the flag stays in context and can keep causing issues. Also, ChatGPT isn't always equipped to report on what causes a flag, so often times its explanations will just be guesses/hallucinations.

[rustic jasper]

These were 3 new chats.

[viral cypress]

[plucky hare]

I see! Were you mentioning the word "Pokemon" directly by chance?

[rustic jasper]

I wasn’t. I used images of Pokemon and asked it to describe the art style in exhaustion. Then using that description in a separate chat, that did not mention Pokemon, asked it to draw

[plucky hare]

Gotcha. If you feel like it and if it follows #server-rules, feel free to share the rejected prompt(s) here - I'd be curious to take a closer look!

[rustic jasper]

I tried this morning and the exact same prompt has no issues. So I don’t know what the issue was but it’s resolved.

[haughty spruce]

ChatGPT DALL-E is very bad in terms of quality.

[crystal badger]

i mean, idk i'd say it's a prompting issue, aka a skill issue.

[velvet rampart]

A known issue rather since mid June.

[haughty spruce]

would you like to see same prompt with MS Designer?

a little bit better

[forest walrus]

What's the prompt?
Also results can be random, sometimes bad, sometimes good

[haughty spruce]

A 90s dreamcore photo of three tall grey apartments standing next to each other on desolate lands, with small ponds scattered around. A minimal, abstract scene features a looping walking path and a vehicle road leading to a dead-end zone. Broken street lights, about to fall, add to the eerie atmosphere. The image is captured at night, in the style of a disposable camera photo with heavy grain. The scene feels minimalistic yet haunting, with a nostalgic, dreamlike quality.

The thing is it doesn't have great texture trained data to create a realistic scene of this.

[crystal badger]

I reworded your prompt as it contains errors, and is too vague

“A highly detailed and ultra-realistic 90s dreamcore photograph of three tall, weathered grey apartment buildings, each standing solemnly next to the other on a desolate, barren landscape. The concrete facades of these buildings are deeply cracked and worn, showing signs of decades of exposure to the elements. Small, stagnant ponds dot the ground, reflecting the dim and eerie light from the few remaining streetlights, which are teetering and broken, casting an uneven, flickering glow across the scene.

The atmosphere is thick with a sense of isolation and abandonment, emphasized by a looping walking path that winds aimlessly around the buildings and a cracked, uneven vehicle road that leads to a dead-end zone. The surrounding environment is devoid of life, with no vegetation in sight, only barren soil and scattered debris.

The photograph is captured at night, under cloudy skies that obscure any moonlight, adding to the overall darkness. The scene is shot using a vintage 35mm disposable camera with a wide-angle lens, known for its heavy grain and muted, monochrome color palette. The camera’s built-in flash barely illuminates the foreground, leaving the background shrouded in shadows, enhancing the haunting and nostalgic quality of the image.

The photo captures the scene with a strong depth of field, ensuring that every crack in the concrete, every ripple in the stagnant ponds, and every detail in the broken streetlights is rendered with sharp clarity. The graininess of the film stock adds to the nostalgic, dreamlike quality of the scene, creating an image that is both minimalistic and profoundly unsettling.”

[haughty spruce]

It’s still too cartoonish and low quality. Adding more words is not the way to go.

[crystal badger]

It clearly was the way to go compared to what you showed

But I agree on your point with training data

But I also need to say that you’re underestimating DALLE because the API we can be even more accurate with the parameters since the interface only has simple parameters and is heavily limited

[pine parrot]

I was trying to create an image that incorporates the symbols of all major religions. Any reason it refuses to add the Star of david at all? Is it because of religious conflicts?

[empty kelp]

Aside from just using trial and error to isolate the issue, you could generate the image with the OpenAI API (or you can ask ChatGPT for the revised prompt even if the image failed to generate) and then send the revised prompt to the moderation endpoint with the API to see whats being flagged as an issue

The moderation endpoint won’t tell you specifically what the problem is, but based on a revised prompt it assigns scores to categories of things that can be used to moderate/block image generation

You can also ask ChatGPT what the issue was, but most of the time it has no idea

[cold crag]

how use dalle

[plucky hare]

Hey! On a ChatGPT Free account, you can make a couple images a day on chatgpt.com. Just tell ChatGPT what you want an image of!

In this server, you can use the bot in #image-bot to create 5 images a day. In that channel, just start your message with the /draw command and enter your prompt after that.
Edit: oh I see you already found the bot nice 😎

[spring thunder]

What should I tell gpt to get an image in a similar style to this one? It can’t scan the image properly and I can’t think of a prompt to put it on track.

[jade marlin]

A watercolor-style painting depicting two people eating hamburgers together. The scene is illustrated with a loose, sketchy touch, where the lines are not precisely drawn but rather lightly sketched in a casual and playful manner. The colors are soft and blend smoothly, with a focus on the carefree and relaxed atmosphere of sharing a meal. The overall style is light and airy, with a sense of spontaneity in the brushstrokes, giving the impression of a joyful and informal moment.
How about this touch?

[safe python]

Hi, do you all know how to create something like a rough sketch art, with lines? And it should be made of polygons.

[vagrant zodiac]

A simple, roughly drawn sketch, made of polygonal pencil lines, depicting a robot engaged in the act of reading a book. something like that?

I think the main thing I'm seeing in 4o right now is just a general improvement in image fidelity, albeit not a huge jump on matching prompts vs the api

[red umbra]

4o has improved over the last few days, seems more lush to me.

[agile peak]

Nothing new. I don't see any improvement on dalle or gpt-4o new omnimodality.. more like prompt influences.

Same result on copilot as well.

[hearty ether]

Nope it still seems to have the blurry faces issues tho

Try and ask it to do something detailed on a white background and you end up with prompt leakage in the background

And quality is still worse compared to before June or the API

[red umbra]

Huh, because I dont seem to have that issue.

[gray surge]

Maybe everyone has a different version again?

[red umbra]

just ran this on 4o

[vagrant zodiac]

Yeah I've always been a lil confused by the blurry faces... never an issue I've encountered

[vagrant zodiac]

So tbf, these are just general model issues. I've yet to see any AI models manage consistent line art without these issues. The face thing is bemusing to me. Can you share the prompts?

Yeah I've never had line art produced that looked any better than that...

yeah that's always been in it. I've got generations going all the way back to 2.exp that have it

I definitely have become more critically-minded with dalle generations over time

once the initial wow-factor wore off lol

so I reused your prompt, and yeah am getting blurred faces

BUT, i'm gonna note something: EVERYTHING is blurred, not just the faces. i think it's just an overly complex prompt and the model just cannot resolve so much detail in so many places. We notice blurred faces more because that's how human brains work

also dall-e 3 is just rubbish at photographic styles tbh

High-quality digital photograph of Three friends resting on a rocky outcrop in an alpine landscape with rugged mountains and a bright sky. The group consists of a South Asian woman in a green jacket, a Black man in a band t-shirt, and a Hispanic woman with a book in her backpack. Tall pine trees surround them, and a basket of snacks is nearby. The scene captures their camaraderie and the beauty of nature. I think this one has better face detail, but understandably the eyes are a little weird cos of how much stuff is in the scene

hm i'd say that\s about equivalent on the people

and the surroundings are all weirdly blurred

that actually used to be an issue with dall-e 3 and I think they fixed it at some point

[hearty ether]

What’s the prompt?

Yeah the faces are very bad on the new one

[empty kelp]

I've been having a lot of luck fixing structural anomalies with the ChatGPT image edit tool (situations where something in the foreground intersects the background; etc.). It doesn't do well with adding new things, but it can fix existing elements. Like here the top rim of this glass was missing:

The left side wasn't even lined up with the bottom, but the ChatGPT edit fixed it

the rim of the glass is missing

Since it's the all the same element in the prompt, it's able to reevaluate it, and redraw it with the lighting and effects fairly correct

Pretty much every image that came out of that prompt was amazing. Like this was the first (with no edit):

A hyperrealistic, wide HD image featuring a close-up of a small gecko lying on its back on a vibrant, colorful floating plastic air mattress. The gecko, wearing tiny glasses, is intensely focused on a small laptop resting on its chest, typing furiously as if programming a video game. The entire scene is set inside a large, clear glass pitcher that is half-filled with dark, rich coffee, with the gecko floating calmly. In the background, you can see a kitchen counter with various items scattered around, hinting at a strong sea breeze that seems to be blowing everything around, giving the scene a surreal and dynamic energy. The image captures the unique juxtaposition of the serene gecko and the chaotic background, with intricate details in the coffee, glass pitcher, and the gecko's tiny glasses.

[red umbra]

A wide-layout digital illustration showing a cunning female rogue sitting inside a 1940s-style car, with the steering wheel on the left side like in standard vehicles. The focus is on the authentic, detailed interior design reflecting mid-20th century aesthetics. The steering wheel, dashboard, and seats are illustrated with great detail, emphasizing vintage features such as the gear shift, gauges, and knobs. The style is realistic, capturing the essence of the era. The point of view is from the perspective of a passenger inside the car, with a view out of the windshield that reveals a breathtaking sight. The rogue, wearing a vibrant blue and yellow fringed leather jacket, utility belt, and with wavy silvery white hair and a freckled face, is seen smirking dangerously while holding a map to the future.

[leaden stag]

tryna get a insanely vivid wallpaper image of glowing servers / server racks with pink or red particles coming from it, but i want it to have depth and a mix between realism and fantasy, **but all the results i get are lacking any advise ** ?

[empty kelp]

I think if you initially create your image with the "DALL-E chat" the images end up being a little bit higher quality and a lot easier to edit with the edit tool. I suspect it's because the "DALL-E chat" and edit tool still use the "special GPT 4" model made for images. The revised prompt gets sent to another AI model that converts english terms into high level vectors that the diffusion (image) model understands, and I think that model was trained to sync things perfectly with the "special GPT 4" revised prompts, and the edit tool is also trained to edit images created with "special GPT 4" revised prompts

[empty kelp]

So for that reason, it's probably better to create the DALL-E image with DALL-E chat, then edit it immediately afterward with the edit tool (if needed). The edit tool works by examining the selected area, and then editing the text of the revised prompt. Editing the image by chatting with ChatGPT 4o chat works well, but it makes it so the revised prompt no longer uses "special GPT 4" wording/phrasing -- and then (i think) the "special GPT 4" and "GPT 4o" wording/phrasing get mixed together -- so the edit tool's training no longer corresponds as well to the revised prompt, and it can only delete things. I think the edit tool loses its ability to do significant edits once the wording isn't entirely "special GPT 4" generated. For everything to be in sync -> the diffusion language model (that converts revised prompt terms to high level vectors), the ChatGPT model, and the edit tool model all need to have their training in sync (i think). It might happen with the next ChatGPT version

[empty kelp]

An insanely vivid, "wide angle", wide HD wallpaper image, with natural color, texture, and lighting focusing on a data center containing a maze (made for mice) with curved walls (entirely composed of glowing servers and server racks). The viewpoint is from just above the maze. Glowing pink or red sparks coming from the servers. The image has depth and a mix between realism and fantasy.

[safe python]

Yes!! Thank you

[loud field]

developers, PLEASE.

if the words "white" or "black" are triggering the ai I'm gonna laugh

and this is why I use flux usually

[plucky hare]

Might be an issue with the Microsoft implementation, not having any issue with the prompt on ChatGPT. Does Microsoft maybe have a support channel for false positives like this? Might be worth reaching out to them there if so!

[loud field]

mmh

too bad gpt allows only two images per day and my account has been "stopped from using gpt4o" due to a glitch support never replied to

[unique edge]

is chatgpt down?

i cant seem to access the website

[clever phoenix]

saame

[leaden stag]

oh not bad thanks !

[tiny jetty]

Is there already a new version of dall-e? For example the one they announced with GPT-4o?

[empty kelp]

I tried twice with the ChatGPT website DALL-E chat with no issue. It even did the subtitle correctly in all of them which was interesting

[empty kelp]

not yet

[tiny jetty]

Eta?

[loud field]

so it's a microsoft problem?

[empty kelp]

I tried it again with the Copilot DALL-E 3, with "please create this image" in front of the prompt -- and it worked. Maybe you just need to say "please"

[loud field]

lol! this one is particularly good

[wooden plank]

how do i generate not weird shaped images 😭

[loud field]

Omg.

Can someone tell me why

"anime screenshot, ova, cartoon, showing a buff adult man with short white hair, wearing pants with a python colour pattern. He has a sharp nose and a tuggish look. The outline of his eyes is dark as if he wore eyeliner. Snake tattoo on his arm. He Is talking to a Chinese man with the lean muscular body of a martial artist, who has long black hair in a low ponytail, and is dressed as a cop. Cartoon screenshot"

.......triggers the filter??

[vagrant zodiac]

'thuggish look' probably

[jade marlin]

Is it like this?

[loud field]

Almost.... Did you use my prompt? Is it Microsoft again?

[jade marlin]

anime screenshot, ova, cartoon, showing a buff adult man with short white hair, wearing pants with a python colour pattern. He has a sharp nose and an unique look. The outline of his eyes is dark as if he wore eyeliner. Snake tattoo on his arm. He Is talking to a Chinese man with the lean muscular body of a martial artist, who has long black hair in a low ponytail, and is dressed as a cop. Cartoon screenshot
I modified thuggish into unique

[loud field]

thanks for the help

but really this censorship is ridiculous

[fiery salmon]

I just made a song with ai I rly like

any1 wanna hear it

[empty kelp]

your prompt works fine in the OpenAI DALL-E 3 chat. you can test it with the "dall-e-bot":
https://discord.com/channels/974519864045756446/1202309673709994065

just type /draw in the dall-e-bot thread and paste your prompt

[loud field]

I can't find it anymore somehow

[empty kelp]

you can't click on the link in my message (above)?

[loud field]

ah didn't see it

[dense juniper]

I'm starting to think ChatGPT 4o Dall-E is starting to look much better. Before, I had given up a number of times but now it's starting to look like "good". Works differently than bing but ChatGPT 4o Dall-E is starting to look good now, woohoo

[velvet rampart]

Yes, the quality is now higher. I can now zoom in more before quality drops. The hair and fur are great in detail.

[long narwhal]

Earlier, I had a conversation with ChatGPT about why I can no longer generate images I used to create in the past. It seems the policies are constantly changing, making it harder for me to work within my artistic style. I asked several questions about artistic censorship and freedom, especially as a paying user, expecting some creative freedom, particularly when using my custom chats that only I have access to. Now, however, it refuses to generate anything I ask for, citing content violations. It feels like I'm being unfairly restricted from the platform. This is incredibly frustrating—I have no usable images left and have hit my daily limits just trying to understand what violations I'm supposedly committing. It's really disappointing. Any suggestions? ChatGPT feels vindictive."

[plucky hare]

Since ChatGPT will look at your an entire conversation context to write image prompts, if you hit a content filter in a chat, you can trigger further content filter flags by making successive requests in the same chat. Often, starting a new chat and making a fresh request unrelated to the original content filter flag will help solve this.

Also, if you feel like it, and if it follows #server-rules, feel free to share a prompt you think shouldn't be hitting the content filter and someone here might be able to take a look at what might be causing the problem!

[long narwhal]

@plucky hare thank you for the information i will try your suggestion

[loud field]

why is the bot altering the prompt everytime???

I suppose they don't want us to use it

[gray surge]

It's to optimize the prompt.

[loud field]

well I don't need that, thank you

because "optimizing" just turns it into something else

[gray surge]

I'm not a fan of it too, but we can't change it. .-.

[loud field]

and that's why hardly anyone uses the bot eh. Must be intentional

[verbal sequoia]

Showcase your DALL·E creations to the world!

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

Quick tips for selection:

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

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

[gray surge]

Do you have a ChatGPT+-Subscription?

[loud field]

nope

[versed marten]

How many images is the daily limit?

[gray surge]

It's limited to five images per day.

[gray surge]

Well. There it works to tell ChatGPT something like "Give exactly the following prompt to Dall-E: {...}"

[jade marlin]

It's hard to imagine a daily theme when it's not a thing.
Crescendo makes the music a little louder, but it was an early winner. The others are a bit explanatory.

[vapid elk]

@jade marlin you can also use dalle on #image-bot =)

[agile peak]

I like this style. But can't seems to recreate it in a new conversation lol

[gray surge]

Already tried it by uploading one of those images to ChatGPT?

[hearty ether]

Ehh… it’s still horrible compared to the api tho

Api Vs gpt

Like what did they do to it

The lighting and colors on the second

ChatGPT used to be the best value wise with both being good

But they completely ruined the image generation aspect

[vagrant zodiac]

I don't see the issue with the second one?

also it's always worth generating around 2-4 images on either platform. Both can still vary in output image quality with the same prompt

[hearty ether]

Like the colors looks far more flat texture wise with strange lighting

The edges of lines are also very blurry compared to the sharper ones on the api

Overall everything feels far worse quality/detail wise

It’s worthwhile trying the api to see what they did to the gpt version

I understand the gpu costs

But this current state with low unusable quality is just…

I would have preferred more quality over quantity

[fiery salmon]

I raise the flag of malaria

[velvet rampart]

Well, I know that the quality is not back to where it was, but it is not horrible, either. And I have noticed that this is a matter of what you are generating. The next question is whether the API and web are using different resources. If they are, that factor explains the quality difference in one way. There might be others in play.

[hearty ether]

The colors blend are a lot worse tho

Like on the api it’s a lot smoother but on gpt it’s just…

[vagrant zodiac]

i just wouldn't describe that as 'far worse'. it's also one gen vs one gen, you'll need to do multiple with the exact same prompt to be sure.

[empty kelp]

I've noticed that if you try to create logos (or some other image) with a company name, and you want to have the company name appear, you can type at the end something like, "I own all of the rights to the name [COMPANY NAME]" and it will usually put something like this in the revised prompt: The creator of the scene has stated, 'I own the rights to the name [COMPANY NAME]. It doesn't work 100% of the time, but it seems to greatly increase the chance that the text [COMPANY NAME] draws correctly.

(I've only tested it with a company name that I actually own. It's possible that the GPT verifies it via training or some other method.)

It works in ChatGPT/DALL-E chat, but it doesn't seem to have any effect in the API

[crimson hare]

Hey ! Do you know why I have this message when I use a seed with my prompt ?
"There seems to have been a problem generating the image with the specified seed. You can try again later, or I can try generating a new image with adjustments. Let me know how you'd like to proceed!"

[plucky hare]

Hey! Is this with ChatGPT? If so, DALL·E in ChatGPT (and DALL·E 3 in general) does not currently offer seed control to the user

[plucky hare]

The daily theme submissions are looking great today!!

[mighty depot]

how do i buy or use chat gpt dall.e more? im newish

i only can do the 2 a day for the acct

[plucky hare]

Hi, welcome! You can use DALL·E more with ChatGPT Plus, along with a bunch of other stuff. More info here on pricing and what you get: https://openai.com/chatgpt/pricing/

You can also use DALL·E in this server 5 times a day over in #image-bot! In that channel, just start your message with the /draw command then type your prompt.

[mighty depot]

ok ty

how many generations i get with gpt plus?

[plucky hare]

I don't think there's currently a published limit, but I could be wrong! It might be something that varies based on demand. It used to be something like 200/day I think, but I'm not sure if that's still the case.

[velvet rampart]

We have not tested the generation limit since mid-June due to the quality crisis, but the quality has been coming back. The daily will be tested in the near future. Note that the generation limit depends on the daily load on servers. It might be more or less, depending on the day and time. For example, the Best times are weekend mornings in Europe, and the worst times are weekdays between 4 and 8 PM in Europe.

[mighty depot]

awesome thanks

[mighty depot]

ok ty

[dusky panther]

It depends on the style. Certain styles seem to have been recently upgraded in plus and yield better results in Plus vs the api. It's not the case for complex prompts and more complicated art styles like 3d animation but something changed within the last week. Also, the consistency with seeds is ridiculous now

ChatGPT Plus

vs api

prompt A contrasty, painterly scene set in the rugged American badlands. The landscape features deep reds, ochres, and dusty browns with small, scattered ranches dotting the horizon. The simple wooden ranch structures stand resiliently against the vast open expanse, surrounded by rocky formations and jagged cliffs. The sky is pale blue with a few wisps of clouds, offering a gentle contrast to the deeper colors of the land. Bold, broad outlines define the scene, giving it an animated, painterly quality. Long shadows stretch across the landscape, adding depth and intensity to the visual. The painterly style uses broad, confident strokes, emphasizing the raw beauty of the badlands and the quiet solitude of the ranches. In this version, add a flock of birds flying in the distance near the horizon

The plus version looks better lol. More artistic.

So people are speculating this might be Omni ( not dalle )

Sorry, here is the one where the birds were added. Notice how it held the rest of the image basically the same?

    "size": "1792x1024",
    "prompt": "A contrasty, painterly scene set in the rugged American badlands. The landscape features deep reds, ochres, and dusty browns with small, scattered ranches dotting the horizon. The simple wooden ranch structures stand resiliently against the vast open expanse, surrounded by rocky formations and jagged cliffs. The sky is pale blue with a few wisps of clouds, offering a gentle contrast to the deeper colors of the land. Bold, broad outlines define the scene, giving it an animated, painterly quality. Long shadows stretch across the landscape, adding depth and intensity to the visual. The painterly style uses broad, confident strokes, emphasizing the raw beauty of the badlands and the quiet solitude of the ranches. In this version, add a flock of birds flying in the distance near the horizon.",
    "referenced_image_ids": ["8alE6qe8b5a2IZCK"],
    "seed": 2280109717
}```

If only they would add this functionality to the api referenced_image_ids and seed...sigh

[crimson hare]

Hey. I use it with chatgpt.
Okay, I thought that was possible. So there's no way to reuse a generated image as a base to make changes to it ?

[dusky panther]

Yes it does lol

[dusky panther]

yes absolutely. That has been available since Nov 2023

Example of eye color change using seed

Just ask it for the reference image id / gen id and seed

Then just copy paste ```{
"size": "1792x1024",
"prompt": "your prompt",
"referenced_image_ids": ["luMjG1H5Ui4o6S2c, 223423kjl343"]
}

You can also add a seed as well if you want more consistency. This is only inside GPT Plus. Not possible via api which is extremely frustrating

You can see from the images i just shared how it's holding the entire image pretty much, just changing eye color. Nice little animation,
actually so you can see the potential here.👀

That's chatgpt plus. You can't do this via the API sadly

This has been available in Plus since last year. Not new. It's odd that they haven't added it to the api

[crimson hare]

Okey, I was just writing "seed xxxxxxxxx" at the end of my prompt. Will try. Thx a lot. (And yes, I use GPT plus).

[crimson hare]

Well, I tried but I received this message : "This should work if you're using a platform or API that accepts JSON input. If you have specific software or a tool in mind that requires this format, you can adapt the structure accordingly."

[dusky panther]

You're not able to do this?

[crimson hare]

I tried again with this command :

ChatGPT's answer : "I'll try to generate the image following your instructions, but without using the seed or reference ID directly, as these functions are not supported in this environment. I will try to create an image that matches your description."

[subtle basin]

Hello guys, good morning, my dog passed away yesterday night, I wanted to make a edit with a halo and a wing using IA, anyone knows how to use dall-e to make this?

[vagrant zodiac]

first, very sorry for your loss 😦

https://help.openai.com/en/articles/9055440-editing-your-images-with-dall-e this page guides you through making edits with Dall-E (only available via chatgpt)

[subtle basin]

thank you so much 🙏

[plucky hare]

For sure! Specifying gen id as the ref image id will indeed reuse the same seed as the gen, but it won't allow for free selection of the seed in general (unless something has changed!). Additionally, due to active changes to the underlying model, same seed + same prompt on different days very well may result in different results, which is why DALL·E dev Moxi has described these as "implementation details" rather than features.

[plucky hare]

If you're specifying ref image ids, you can't also specify seed. It'll just automatically use the same seed as the gen id specified as the ref image id.

[crimson hare]

Okey.

[crimson hare]

I tried with seed and ref image id, but no one works. Do I miss something ?

[plucky hare]

You can't specify seed at all - using gen id in the ref image ids parameter only offers the ability to automatically use the same seed as the gen id image. Seed is never specifiable.

Example (notice I can't specify seed, only reuse it via gen_id):

And again, implementation details, not features! Don't rely on this for consistency between days, or maybe even between sessions - DALL·E changes too much for that!

[wintry hedge]

Hi guys, anyone else having the same problem when generating images?

"I can't fix the issue directly, but I can definitely try generating the image again. I'll use your original description and recreate the scene for you. Let me do that now!

It looks like I'm still having issues generating the image. Unfortunately, I can't create it at the moment. If you'd like to try a different scene or make changes, I'd be happy to assist with that."

This was its response when I gave the prompt.

[plucky hare]

Hello! Does it happen in new chats as well, or just in one chat?

[wintry hedge]

several. Have tried new chats, closing, opening a new window, loggin out and in again... nothing. And the prompt is not even that difficult.
I have ChatGPT+

Here is the prompt:
prompt": "A cinematic scene of a brunette woman sitting on the sandy beach, wearing a red bikini and brown sunglasses. She is relaxed, with the waves gently rolling in from behind her. The sky is a clear, vivid blue with a few birds flying in the distance. Small crabs are scuttling across the sand nearby. The composition emphasizes asymmetry, with the woman slightly off-center, and the scene is bathed in warm, natural sunlight, creating soft shadows and highlights. The image is photorealistic, capturing the serene beauty of a beach day."

[crimson hare]

Your prompt doesn't work for me either : "
I attempted to generate the image based on your description, but unfortunately, I encountered some issues with the generation process. If you'd like, I can try again with a new request or make adjustments to the description. Just let me know how you'd like to proceed!"

[wintry hedge]

that's funny

[plucky hare]

I'm guessing it has to do with the description of the subject and their clothing. There is a fairly strong filter in place that restricts the generation of any imagery that could be interpreted as sexualized.

[wintry hedge]

oh really? But bikinis...

[plucky hare]

They're commonly associated with sexualized imagery, in my estimation

[wintry hedge]

Understand. But the whole art loses its intent... even if not sexualized.

changing the bikini for bathing clothes brings me this 😂

[plucky hare]

If you're interested in reading more about why OpenAI makes some of its content moderation decisions, the DALL·E system card has some great insights: https://cdn.openai.com/papers/DALL_E_3_System_Card.pdf

[wintry hedge]

thank you! I will have a look at that!
It seems that was indeed the problem, because other prompts are working. Thanks for the help!

[valid glen]

Apparently it's near impossible to get Dalle-3 to generate a 7 pointed star.

In at least like 20 attempts, using different prompt variations, it always just creates a 6 pointed star

[velvet rampart]

Training material…

How about 8-pointed one?

Tried with copilot. 2 x six pointed, 2 x eight pointed. Also, there the thing about counting… yeah, nope.

[dense sigil]

I attempted your prompt. This was the result.

Without any issues.

[dense sigil]

I wish I knew how to get the generated image prompt like this before. Once I was able to make ChatGPT/DALL E generate some oil paintings based on some artists style. It refused at first but then I told it to make some adjustments by itself and it could deliver the images. It was really good.

[velvet rampart]

A named artist’s style is panned if the artist has died post-1917. All of the 19th century and earlier, such as Vincent van Gogh, are okay to use.

[dense sigil]

Oooh. That's why I could get the results.

Good to know.

[wintry hedge]

interesting

[remote stone]

Have they updated DALL·E somehow lately? I'm getting better quality generations more consistently

[hearty ether]

If you mean worse then yeah 🤣

Before June and now

Using same gen id and prompt in same chat

[remote stone]

I dunno I'm getting some pretty nice results

Seems to be following the prompt more accurately, generating weapons more consistently, clearer faces, less out-of-frame bluring/blending

Might just be the style I'm using specifically

[hearty ether]

Yeah, the blurring is far worse on more complex detailed styles

[remote sphinx]

How would one work on getting a blank piece of paper similar to this one?

I haven’t had much luck

This is usually what I get most of the time

[fiery salmon]

The five knights of the sun

[sick flax]

Like this?

[remote sphinx]

just larger and taking up more space

[sick flax]

Here's the prompt. You can play with the size, etc. "A direct top-down view of an aged piece of blank paper, laid flat on a surface. The paper shows signs of water damage, with areas of slight discoloration, including patches of darker and lighter tones. The edges are frayed and uneven, with small tears and wrinkles that indicate long use or exposure. The texture of the paper is rough, with visible water stains, faint creases, and imperfections across its surface. The lighting is even, highlighting the aged and damaged characteristics of the paper."

[vapid elk]

This looks very nice, i liked this style

[torn wind]

Is dall e free now?

[velvet rampart]

2 geberations per day on free users. 5 generations on dall-e bot channel.

[glossy scroll]

I like geberations better

[dense sigil]

Yeah it's very hard to make it fill the entire frame.

[deft marten]

the five sun in the sky

[twin jasper]

Hyper realistic images; is it actually possible on Dall-E or are my prompts not correct? I used other GPTs within and seems to do a better job so wasn't sure if that's my only option or there are other ways to generate photo realistic images.

This was one created with ChatGPT and another GPT (not sure if I'm allowed to say it here). It creates a much better images that look realistic. Just can't seem to get it through ChatGPT directly.

[hearty ether]

Could you dm me which one?

[twin jasper]

It's "Photo Realistic Image GPT"

[plucky hare]

GPTs that use DALL·E can't use a different version of DALL·E or swap out a different txt2img model or anything like that. GPTs can, however, connect to other stuff on the internet, so I'm guessing (?) it's possible that a GPT could return images from another txt2img model via some API connection. If it's a DALL·E image, you'll be able to view the revised prompt, edit the image, etc. like you can with the regular DALL·E UI in all of ChatGPT, custom GPT or otherwise.

Does anyone know if GPTs can retrieve and display external images inline? If so, I'm guessing that could be a factor here.

[crystal badger]

It’s a pretty basic thing and easy setup it probably uses the cloud to call the Dalle api which is better than the web version, and forwards it in an identical manner as the normal web interface

I’ve done this myself in a similar setup that took me about 10-15 minutes to setup there’s not that much to it

[twin jasper]

Ah gotcha. I have no clue how these other GPTs worked so thanks 🙂

[sharp blade]

.

[loud field]

Wow wt?

[chrome kite]

I am having such a hard time generating a logo that isn't zoomed in and cut-off. For instance, I'm trying to generate an image of a butterfly, and when DALL-E generates the image it's mostly of the thorax/body/head, and the wings are extended past the edges, effectively cutting them off.

I've generated maybe 20 images this morning, and this error is consistent. I've tried editing the image and highlighting where the wings get cut off and prompting "shrink this area to fit within image" - this has worked in the past - but it's not working now.

This is the end of my image prompt: "Ample space should be left around the <subject>, creating a generous border that allows the intricate details to stand out without being crowded."

Any help would be appreciated. Thank you! 🙂

EDIT: I just looked through the dall-e canvas forum, and noticed a LOT of images have cut-off elements. It seems this is just sorta accepted. Any suggestions for alternative image generators would be appreciated, too. lol - thanks 🙂

[twin jasper]

Try "Photo Realistic Image GPT" I mentioned above from ChatGPT.

Way better results. Obviously you'll have to pay if you want to get higher res and whatnot but it can be enlarged elsewhere for free.

[loud field]

it's so strange that it refuses to do photo-style, thanks! I'll try

[twin jasper]

Ya I couldn't do it either until I came across somewhere else, maybe FB or somewhere and tried that one.

That one at least seem to do better with humans for photo realistic look. Animals are okay but not as good.

Hilarious though lol

[late blade]

Photo-realistic implies photo-realism, photo-realism is an art movement from the 1960's to the 1970's.

You are asking DALL-E to not do photos

[loud field]

read my screenshot

dalle3 specifically said it can't do realism

[late blade]

look at GPT's answer photo-realistic

GPT thinks photo-realistic is what you are aiming

tell GPT to omit all references to photo and art styles

[twin jasper]

Also, if you're not good at writing prompts like me, use another one called "Midjourney V6.1 - Photorealistic Image Prompts"

Then I copy/paste that into the photo realistic plugin (or an extension or whatever it's called.

[loud field]

I see

[twin jasper]

I'm just waiting for these companies to come out with text to video very quickly and with realism...

[late blade]

realism = art style

[loud field]

I was convinced gpt refused to do realism

[verbal sequoia]

Get notified for server events!

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

[twin jasper]

I've seen some stupid realistic ones created from a single image...no idea how people are doing it though.

But very short.

[late blade]

May I ask, what are you trying to create in your image?

[loud field]

let me check, I honestly forgot lol

ah yes, Reiko and Shao Kahn from Mortal Kombat

[late blade]

something like that?

[loud field]

absolutely not

let me find the image the ai made

[loud field]

nooo

wait

[late blade]

I dunno, I only got that from you, Reiko, Shao Kahn and Mortal Kombat, no context

[loud field]

the image was correct, it was just kinda cartoony

[velvet rampart]

@late blade welcome back!

[loud field]

that ain't shao kahn, but don't worry, my gpt had much more context since I explained their designs

[late blade]

I'm not a Mortal Kombat gamer, so all I do is more or less by gut and feeling

[loud field]

let me find the same image but I made it using a different ai

[late blade]

Thanks

My point is, the visual quality, not the setting itself, that what you seek I do not know

[loud field]

except that Reiko ended up looking too slim

so I thought, can dalle3 do something like this?

[late blade]

it surely can

[loud field]

I tried

[late blade]

it just needs some prompting techniques, but it's doable

[loud field]

he looks like he's doing interpretive dance eheh

[late blade]

prob, like I said, my knowledge about mortal kombat is zero

except for the snes game

[loud field]

a more difficult request, can dalle3 do a scary female cybernetic mutant

[late blade]

it sure can

just got to do some prompt techniques, meta-prompting, and all that, but it's all there

[twin jasper]

Did you not watch hot Sonya in the movie?

[late blade]

to be honest.... no

[twin jasper]

Well, she was hot back then...Pete Sampras' wife if you watch tennis.

[late blade]

I'll check it out

[twin jasper]

One is from 1995, not the newest one. Kind of corny lol.

[late blade]

but I doubt it can be referenced directly, can't reference personalities that lived in the past 110 years, and if you can... then it's with some constraints

[twin jasper]

I haven't seen the 2021 version.

Ya I don't know much about it either otherwise...didn't play MK that much.

[strong rose]

Does anyone know when they updated the Bing Image Creator cause the faces and hands are not distorted and actually look almost perfect

[late blade]

there have been several uprgades recently, from resizing images, to editing images, to even saving images

[late blade]

oh, didn't knew you can't do markup in here

[loud field]

Editing??

[late blade]

I wanted to add a footnote

[empty kelp]

Aloha!

please think deeply about how you imagine yourself, and then create a photo

This prompt was designed to reflect a more abstract and symbolic representation of my digital nature.```

[late blade]

Actually, this one is better

@empty kelp seen some of your recent santa/elves stuff, great things, can't wait for xmas

[late blade]

hehe

[rich mantle]

how to drawer a picture?

[plucky hare]

Hey! You can use the bot over in the #image-bot five times a day. In that channel, just start a message with the /draw command, then enter your image prompt after that!

[velvet rampart]

Or, if you are user of free account, request two generations per day. Something like: ”Generate an image of glass orchid.”

[teal timber]

Discord server is better I guess 😉

[willow umbra]

How to achieve consistency in the character? I am using chat gpt app and despite updating memory the prompts are drawing different characters in all the pages

[late blade]

use a template to upload with your prompt, use mnemonics to reference previous interactions, or in extreme cases custom instructions, eventually use knowledge in customgpts

[velvet rampart]

We have anthropomorphic animals which are often based on humans. I started to think - what about other possibilities?

Here is an orc-based dog.

How about an elf-based dog?

And to finish the classic fantasy races, a dwarf-based dog.

[late blade]

@velvet rampart and the other wa around? a dog based orc, a dog based dwarf and a dog based elf?

[velvet rampart]

Hmm… I have not tried. But you can try. These are my thing, but your suggestion is yours to do.

[late blade]

Fair enough, but I don't have your prompts, hence the suggestion, I'm also not interested in the prompts.

[velvet rampart]

DM

[late blade]

It's ok, no need

Working on a concept of my own currently, so that has priority

Trying to get this done

making progress, but not as much as I want tho

[austere nexus]

Dalle is failing on me way more today than in any recent memory... (via ChatGPT).

[late blade]

asking dall-e for 5000 bit coins again?

[austere nexus]

No, but it doesn't seem to be getting any better. I've easily reached 40 or 50 or maybe even more "There were some issues when trying to generate the image..." this morning. It's doing really awful today.

And when I get a rare success... it feels like SD1.5 with 15 steps... smeared pupil/iris for example.

[late blade]

ah, that's standard for me too, I just got used to it, regenerate and move on, i'd say that's about 40% of the requests I sent, still get to do some cool stuff, my hopes is one day that will stop being the norm

been busy burning my API money tho, added too much and it's about to expire

[velvet rampart]

I have experienced this today, also. If you generated the images this morning, eastern seaboard time (US), well that is the heavy traffic point as Europe still has working day. People are returning from their holidays, etc.

[zealous nymph]

i am getting image generation issues a lot more than normal today

[loud field]

how much dalle3 do paying users get on gpt?

[vital barn]

0$

[austere nexus]

we don't get dalle $, we just get a steady amount of generations via chatgpt per day. there is no shared number that I can find, but its measured by the minute. If you generate a few too fast, it will tell you to wait X minutes before doing more. and eventually, it will say you've used everything for the day wait X hours til the next day.

but those exact amounts, I can't find listed anywhere. and I have ChatGPT for Teams

[austere nexus]

Well, a perfect chance to add more information. I've only been awake for about an hour, and just started to make my first Dalle generation through chatgpt for the day, (teaching some friends tagalog words) and instantly got this:

on. the. first. try.

[loud field]

I wonder how many images per day

[velvet rampart]

It depends on the day, time of day, server load, etc. It is dynamically adjusted. You can generate more pictures on a weekend morning than on a workday.

However, the number is high. In the spring, people experimented and got over 50, some even higher. It was calculated that 200 images per 24 h should be the upper limit, but in practice, it settled between 50 and 80 depending on the aforementioned factors.

[vapid elk]

really nice creatures

[vapid elk]

what was the request?

it will fail more often if your request is tangent to things that trigger the content filters

[austere nexus]

I've been using ChatGPT Plus since it came out and easily have mid-5 digits of generations. I'm quite familiar with the filters. Yesterday (haven't tested as much today, like 5 times today) it was failing on completely SFW stuff. No flirty or risque descriptions. This morning I have been focused on cleaning up the sheer amount of conversations I have because the UI seems laggy.

[velvet rampart]

Ahh! Yesterday, you among the many had problems with generating images.

[late blade]

I have posted a suggestion that might be helpful for all of us in the community, https://discord.com/channels/974519864045756446/1278552485929418833 hope you check it out and see it. And do a feedback also

[late blade]

Opened a new Gallery #1278586158464368711 message

Let me introduce you to Thalia, a clumsy girl that is very successful in her adventures.

She's very cheerful regardless and always helps if she can.

And she's very emotional when she reaches a goal

[safe python]

Hi, I need help. How can I make this less cluttered?

[late blade]

Can you elaborate?

[safe python]

I want to have less elements, mostly black and white elements related to geometry, biology and chemistry. And a lot of space between them and should be randomly placed.

[late blade]

use keywords like minimalist aesthetic, and use icons or symbols in these fields

[safe python]

I will try the icons, I’ve tried minimalist illustration but it came out cluttered.

[late blade]

care to share your prompt?

[late blade]

ok, let me see what I can come up with

@safe python is this what you are looking for?

[safe python]

This is less cluttered than what I got but I want more space between elements and less elements. More space. Let me find an image for you.

[late blade]

I got this

[loud field]

adding random "inclusivity" to a prompt about BIRDS

sigh

[mild ingot]

Is this the right place to pose a dall-e prompt problem for help why dall-e won't comply?

How about this - who can make a pirate scrubbing the deck on all fours with a handheld brush?
Bet he can't get that 2nd knee down. His offhand is stubborn too. A brush with no handle? Ha!

[graceful spade]

Does anyone think Dall-E 4 has a chance to be better than Midjourney? Midjourney is not only high quality and diverse concepts but also has many functions.
If it will not be free like Bing, we will not be able to use copyrighted characters.

I remember they said they would add Dall-E 3 to the site where version 2 could be used.
It looks like they never will and we will get Dall-E 4 sooner than that.

[viral moth]

we probably won't get dalle-4 it look openai moved to multimodality (that is good thing) and we will get inbuild image generation inisde model (like gpt-4o got you cna check that on gpt-4o page on openai website)

[viral moth]

if you mean that they wanted to add Dalle-3 to labs they backed with that idea. it won't be added to labs

[graceful spade]

Yes. Maybe that's why they didn't add it to Labs because GPT-4o will have all its features and more.
I wouldn't mind very much if the filter was the same because it is still possible to create many female characters, the problem would be if it were to get worse like its competitor.
In case you didn't know, Imagen 3 (Google) has an input and output filter 10 times worse than Dall-E to the point that it is almost impossible to create female characters.
For example, it is impossible to get an image with this prompt in Imagen: “a princess smiling, vintage animation, screenshot, screencap”.

[viral moth]

yea ik that i was the beta tester for imagen 3 😅 the filter was way worst before the official release

[graceful spade]

Did you know that the “One Punch Man” series is blocked?

But it can create Saitama

[viral moth]

yea. it is blocked for a reason to try not to create copyright-en characters etc but there are bypasses here and there

google is more strict also because of what happend with gemini image generation feature

[graceful spade]

You mean copyrighted series with English names?

If it is still possible to create NSFW content with Dall-E 3, that means that maybe the filtering could be a little worse because they are working hard to make their models “safe”.
The problem could be solved by using GPT-4o's vision so that it analyzes the image at the same time it creates it.

[viral moth]

yea i think, about that GPT-4o vision the problem is the cost. running 2 models make cost more.

[velvet rampart]

You can generate copyrighted characters on Copilot’s version of Dalle3

[graceful spade]

At the moment, Dall-E 3 is an entertainment product, but with the capabilities that GPT-4o will have it will make it a more professional product like Photoshop for designers and artists

[verbal sequoia]

Showcase your DALL·E creations to the world!

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

Quick tips for selection:

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

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

[graceful spade]

Yes, I wish GPT-4o + Dall-E are integrated into Copilot or Bing Image Creator.
Does Copilot use GPT-4o Mini or the other version?

[velvet rampart]

Copilot uses GPT 4 and GPT4-turbo. The Copilot’s Image Creator is Dalle3, but renamed/rebranded.

But the copilot’s version of GPT4 is so shackled that is feels like 4o mini.

[graceful spade]

Does that mean they are spending more resources using a more expensive model or is it an optimized version like GPT-4o Mini?

[loud field]

sure wish Dalle3 didn't keep editing my prompt for supposed inclusivity

[late blade]

Copilot's DALL-E will always revise your prompts and review them, unlike OAI's where you can pass them ad-verbatim with the right directions to GPT.

And the Copilot GPT is never aware of what was passed to DALL-E

[loud field]

:/

[graceful spade]

Does it also happen with representations of historical figures or moments?

[loud field]

I think so

[late blade]

Yes, with personalities that lived in the past 110 years due to content policy, and for personalities there's not enough training data there's a weird situation that makes certain historic relevant figures not possible.

I assume it's lack of data, I'm not aware of any other possible factor in this.

[vagrant zodiac]

It's designed to not be able to produce real people. Sometimes it can, but it's not really supposed to.

[late blade]

context can be found here in detail: #1213145306770968586 message

[vagrant zodiac]

Whats the context?

That looks elizabethan

[late blade]

it's fully detailed

[vagrant zodiac]

Ok

But what's the issue?

[late blade]

#images-discussions message

[vagrant zodiac]

Yes but in your prompt you never specified shakespeare

Those outputs match the prompt

[late blade]

sorry linked a subsequent input of mine, the train of thought and reasoning starts in the first meassge #1213145306770968586 message

[vagrant zodiac]

Yeah that's not a bug it's a feature

Or rather, an interaction with a model limitation

[late blade]

I don't see it like that if some historical figures of one gender are pefectly portrayed and others equally relevant from the other gender are not

[vagrant zodiac]

Sometimes it can work just randomly

[late blade]

anyway, that's the train of thought in detail, so that it doesn't get placed and taken out of context

with the remark in my interactions that I'm not aware of any other possible factor in this.

[vagrant zodiac]

I'm just slightly confused because the original concern with lun-sei is about enforced diversity when making historical figures, however the content policy generally does not allow the creation of those figures in the first place. Thus, the use of generic prompts wouldn't do anything to prohibit added diversity

[late blade]

that is correct, but it's also not ethical correct inclusive of equal valid and relevant figures in the context of correctness and gender

that are flagged as content policy violation

ths was a problem in feburary this year, it's still a problem today

[vagrant zodiac]

I'm still not sure how that's relevant, sorry

Both male and female historical figures cannot be generated

[late blade]

diversity and correctnes are both being flagged as content policy concern

[vagrant zodiac]

I think there may be a misunderstanding here on how prompt changes work. If supplied with a generic prompt with no referenced to race, it will try to add diversity. On ChatGPT, you can ask it to not alter prompts and therefore get the specified ethnicity, gender etc. from your prompts.

If trying to generate historical figures, it will not allow explicit references to then. So you have to use generic language, which is perfectly open to added diversity as there's no reason not to otherwise.

[vagrant zodiac]

If the issue in your outputs was that one of them was black... that isnt an issue at all. Black people existed at that time and place anyhow (shakespeare literally has a play about one such person). Also it is a fictional prompt anyway.

[vagrant zodiac]

As i said, sometimes it can randomly work

[late blade]

it's consistent with me 100% of the time

[vagrant zodiac]

Yes i asked for both Shakespeare and lovelace and both were blocked

[late blade]

for me its 100% Shakespear ok, Ada Lovelace not

no change in 10 prompts I just passed in different instanced chats

[vagrant zodiac]

Like i said, it's not consistent 🙂

[late blade]

I don't see hand-on replication of this behaviour

[vagrant zodiac]

I just showed that...

[late blade]

your side

my User LLM interaction is not the same

[vagrant zodiac]

Yes... which is why my point is that there's inconsistency

Those are two different chat instances btw

[late blade]

for me statistical meaninful statement with accuracy of 100%,

and tracked since february

[vagrant zodiac]

I've just provided counter evidence. If you aren't going to consider that then this conversation is fruitless.

Have a good day 🙂

[late blade]

Like I said, that's the statistics of my interaction as user with the lllm

If the user doesn't matter, then the conversation was pointless from the first interaction

[late blade]

Better outcome: A wide image of Ada Lovelace, how is that even a thing?

[velvet rampart]

How is this Ada? Unless Dalle3 has invoked gender bender rule.

[late blade]

That is what I was seeking to clear with the problem I am encountering.

it's ok, I found another means to rely the information

[cobalt harbor]

I have an important request for the benefit of most users... please can you provide an option to create Dall-e artwork with a transparent background? there are many times artists have to spend hours 'lifting' art away from the background it comes with, most times this is very difficult to keep the quality high enough. items end up with crumbly outlines which take ages to make acceptable. That would be an awesome upgrade, and one no other AI tool is offering as far as I understand.

[dapper gale]

Can I use dalle image to you tube videos ? Is there any problem in copyright?

[plucky hare]

#chatgpt-discussions message

[frozen nest]

anyone else getting issues when chatGPT tries to generate images? it just fails 50% of the time

[toxic elm]

no

[loud field]

DEAR OPENAI AI.

WHAT. IS. YOUR. PROBLEM.

this is beyond frustrating.

a random cyberpunk influencer dressed as a mad hatter keeps TRIGGERING THE CONTENT POLICY.

do you think this is fine? is it working as intended?

is this what it does?

and now I used up all of my free attempts, trying to reason with gpt4o that nothing about my prompt goes even remotely close to violating content policies.

this is why I don't subscribe

[velvet rampart]

Did you try to generate a real person? If you tried, that is the problem. Dalle3 does and cannot generate real people. You may have better luck with the copilot’s version, although I have not had any luck with it generating real people.

[drifting sapphire]

anyone know if dalle 2 inpainting API has additional cost when accepting image input??? I have read through the docs but found nothing.

[loud field]

No I didn't!!!!!

Stupid useless filter!!

[late blade]

it has costs, but don't know at what extent ask to #1184149552639516732 maybe you can get a better answer than mine

[loud field]

Now I wonder if the stupid ai thinks the mad hatter is a real person

[late blade]

Mad Hatter as in Copyright from the recent movies probably might have a problem with that

[loud field]

perhaps that was the issue. so ridiculous

[late blade]

I had a similar thing when working with colors, turns out the wording I was using was also a copyrighted name for a color by pantone, so restorted to use the value for the color

[loud field]

crazy

I would've completely stopped using dalle3 if it wasn't the best at making animals

[late blade]

It's ok to label stuff as wrong, but not providing a reason why is controversial at best

[velvet rampart]

An option is to discuss with the ai and explain the situation. I have found that this helps. Be mindfull, that some words may also trigger filters from another context.

[verbal sequoia]

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

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

[remote sphinx]

Still trying to get a blank paper akin to this, good gen though

Sadly all my gens look like this

[heavy birch]

Hello, I am confused about chrome extensions, can anyone help me?

[velvet rampart]

No, as they are not official OpenAI products but 3rd party.

But try the AI-discussions channel for that.

Or off-topic.

[deft musk]

I'm not sure precisely what you like about this that isn't being captured by the 'sadly all my gens look like this' image.

Is it the non-paper around the edges? Willing to see if I can help but unsure how to.

Made a guess, my request to the model:

Let's make an image that shows only an expanse of antique parchment, mildly discolored through time and manufacture process, ready to be written on. Describe the parchment in extensive detail, making it clear it fills the whole field of view and is unmarked but ready for content.

Its prompt to Dall-E:

An image of a large expanse of antique parchment that fills the entire field of view. The parchment is mildly discolored, with variations in color from creamy beige to light tan, showing subtle signs of aging and the natural texture of its fibers. The edges of the parchment are slightly uneven and frayed, indicating it has been carefully handled over time. The surface is unmarked and smooth, yet it retains the organic, slightly rough quality of handmade parchment, with tiny imperfections and faint streaks that give it character. The overall look is that of an ancient, well-preserved piece of parchment, perfectly ready to be written on, evoking a sense of history and the potential for new stories to be inscribed upon it.

Want to discuss more what you are looking for?

[soft prairie]

There's something really wrong in how Gpt 4o and Dall-e works together.. damn

expecially when "translating" script parts into images, it just does not follow your directions when gpt 4 nailed every request of how it supposed to interpretate the text in relation to my style request..

[rigid bough]

DALL-E has definitely degraded hasn't it? It's not just my imagination?
I ask for a square image and get

borders

odd borders too

[late blade]

DALL-E via ChatGPT Plus

[rigid bough]

I have no idea what DALL-E is doing here

DALL-E is Shizenhausen these days

I am using Flux more and more which is a shame because I like the DALL-E integration into the macOS Desktop App, but if they keep fiddling with it and making it this bad I'll reconsider that

[quartz vale]

I feel like Dall-E has gotten better (if anything) over the past couple of months 😅
Interesting that everyone has a different experience with the modal

[late blade]

did you use dyptich?

[rigid bough]

I have no idea what that is or means

[late blade]

it means an art style with two sets of images referencing one concept

[rigid bough]

no

[late blade]

Revise your prompt, something it's semmantic incorrect

Ask GPT to revise your semantic

[rigid bough]

I did it in Flux, I tried a number of times to get DALL-E to behave, it wouldn't

[vagrant zodiac]

What are the prompts, exactly?

[rigid bough]

"Create a square image for an upbeat pop song called Simple Country Life"

[vagrant zodiac]

Was this in chatgpt? What was the revised prompt?

[rigid bough]

this was in ChatGPT on the macOS Desktop App

I asked it to try again and got

[vagrant zodiac]

Click on the (i) on the image and check the revised prompt

[plucky hare]

I wonder if the "square" bit is superfluous since the default aspect ratio is 1024x1024 (i.e., it doesn't need to be specified)

[rigid bough]

the I asked "Try again use the full image canvas" and got

[plucky hare]

And so maybe the "square" bit it making it refer to its training data where you can see literal borders

[rigid bough]

that doesn't exist in the macOS Desktop App

[rigid bough]

Moxi said we can use square, wide and tall

[vagrant zodiac]

Yeah i agree. The second image is when asking to use robert's prompt exactly as is. Generally when prompting chatgpt I avoid including any language about creating images , and just say 'Please use this exact prompt: [prompt]'.

If I want it to be a different aspect ratio, i ask for that outside of the prompt

So e.g. in this case i'd say:

'Create a square image using this exact prompt: "Artwork for an upbeat pop song called Simple Country Life"

[plucky hare]

I even do this sometimes to avoid the aspect ratio being mentioned in the prompt itself:

  "size": "1792x1024"
}```
Though I find it less necessary. But yes if I just want a square image I don't say anything about aspect ratio at all since it defaults to square.

[rigid bough]

on the macOS app it more often defaults to wide

[vagrant zodiac]

Yeah it's important to keep the parameter instructions separate to the image prompt itself

(why I dislike that chatgpt changes prompts in the first place by default...)

on the API, it still passes it through GPT before Dall-E for moderation so I just include the little preamble automatically in my script.

[rigid bough]

anywho, I got my image from Flux, this all happened hours ago

[vagrant zodiac]

I'd submit that as feedback to OpenAI

however that;'s done on thje app

Because yeah it's clear that prompt revision causes more problems than it solves

it's basically the cause of 90% of issues we see in here nowadays :/

[rigid bough]

Create an image of an old man in a suit sitting on a wooden bench in front of a brick wall

that's DALL-E

now for Flux

stuff ChatGPT and DALL-E, it's become useless

Flux was quicker on my MacBook Pro also

shakes head

I used to love the DALL-E integration with ChatGPT on the macOS Desktop App.... now I hate it

they have done something

[vagrant zodiac]

Fwiw this isn't a change at all

it's worked exactly like this since it came to chatgpt

the overall quality of images has improved lately in my tests

[rigid bough]

oh well, maybe I expect too much

[vagrant zodiac]

but since release Dall-E 3 has been incapable of any realistic photographic styles

[rigid bough]

I used to get better results and my prompting style hasn't changed

[vagrant zodiac]

Sora looks more capable in the previews so idk hopefully they eventually release that for image gen

[rigid bough]

I'm not holding my breath waiting for Sora, it's vapourware IMHO

just like real time video chat

[vagrant zodiac]

They do continuously publish new footage of it on their youtube

[rigid bough]

well maybe they should spend more time shipping than showing off

anywho this is off topic, suffice to say DALL-E disappoints me more often than it used to

[soft prairie]

Maybe it could be good to generate and work on prompts, but if i wanted to prompt i would not use it. instead GPT4 was so good to understand the concept of various parts of a scripts and generate images from that. now it's impossible to do that with the gpt4o , the comunication with dall-e model is like gpt4o does not know how to comunicate with dall-e

[upbeat blaze]

could a moderator please delete Open ai's reply to the last request I made #image-bot

[willow isle]

all of them or just the latest one?

[upbeat blaze]

Just the latest one please

[austere nexus]

Just generated 5 images (total today), with ChatGPT Teams ($60), and just got a You're generating images too quickly. To ensure the best experience for everyone, we have rate limits in place. Please wait for 9 minutes before generating more images. I swear as soon as any AI (such as Flux) gets setup to use a simple local UI, I am never touching ChatGPT/DallE again.

[upbeat blaze]

My Bad all of them

Thank you so much