Quality is down on Dall-e 3

[broken sapphire]

I request the dall-e3 to generate an image. I expect the quality of the image to be high as before, e.g. in May 2024. But no, the image is of poor quality. Like the generation is incomplete and thus the finishing touches are not done.

Steps to reproduce:

1. Use the browser version, or app version of chatgpt
2. input the prompt and prompt that you want dall-e 3 to generate into an image.

Expected result:

Image of a high quality like this one from 22 May.

Actual result:

I get an image with bad quality like this image of an otter. Notice the weird rippling effect on the textures which is not present the squirrel image from 22 May.

Additional information

Done using chapgpt4o. Also, when the image is zoomed in, the poor quality comes quickly apparent.

[pearl grail]

Thank you for reporting a bug related to DALL·E.

If relevant and not already provided, we recommend adding a link to the associated ChatGPT conversation and uploading any relevant images here for easier troubleshooting.

➜ Need to contact OpenAI support? Visit https://help.openai.com

[cedar gale]

Before and after

A good way to test this is to take a image generation chat from before June 12th and then asking the ai to use the exact same gen id and prompt to generate

[broken sapphire]

Here image from May 22.

Same prompt from 30 June.

[cedar gale]

Signs of this are blurry faces (especially misshaped eyes), image distortions and flatter less detailed textures

[broken sapphire]

The quality is down. The textures of fur is poor. The sharpness of the fonts is down. The images appear incomplete, like the generation was halted before its time.

[cedar gale]

The date of this happening seems to be June 12th for me

when the image quality dropped noticeably

Before and after, notice the significant distortion around the face

[broken sapphire]

Here I a clear example. The whole picture is a bit blurry like the generation was halted a few steps before completion.

[broken sapphire]

An other pair to compare. The left on is generated March 21 and on the rigth, today. The quality drop is noticeable.

[broken sapphire]

Quality is down on Dall-e 3

[broken sapphire]

Solution: Bring back the high quality model.

[cedar gale]

What about the title being significant quality decrease?

[broken sapphire]

Well, quality is down when we compare to the before state. It has decreased.

[primal sinew]

They probably just optimized the model by reducing its precision, and that obviously leads to quality degradation

its the same model, but less accurate

[cedar gale]

Significantly worse for whatever benefit it is providing

And practically unusable for some tasks/styles

[cedar gale]

https://community.openai.com/t/very-low-resolution-on-the-images-from-chatgpt-dalle-4o-june-2024/818440/48

I just want to also highlight this

[cedar gale]

Still persistent

[broken sapphire]

I am trying to get straight answer to this matter via help.

[broken sapphire]

I got a straigth answer:
”Yes, we are aware of the issue and are actively working on developing solutions. Currently, we do not have a specific timeline for when these enhancements will be implemented.”

[cedar gale]

Oh no…

that seems like a while…

[cedar gale]

Still sucks 😭

[long valley]

seems like they were cutting costs by reducing the dalle gen quality in order to add some more captchas with more providers specifically the demon spawn that is arkose labs

[ivory pulsar]

whenever it generates 2 images, the quality goes down. It is a known problem for many months

[elder heath]

I have been dealing with this problem for over a month now and this is the first concrete information I have seen. Thank you @broken sapphire

[cedar gale]

We need to bring more attention to this somehow

[elder heath]

I completely agree.

This is how degraded the quality is for me.

[static mist]

Looks good tho

[sly oasis]

i just found it.... so not for just me.....😡

@broken sapphire BTW thank you to bring this up in a relaxed way before me...... I wouldn't be this polite about the problem..... 😄 the accuracy mostly ok but the quality is somewhere between SD 2 and SD XL.....

the problem and which put a bad light on them is.... "well.... Try to push through.... maybe they dont notice or at least dont complaint"..... before somebody try to defend the company.... anybody think that any modification on the model or the service go through without any test? Its not like a captcha or other stuff bug which include 3rd party stuff....

BTW as i experienced its effect all GPTs not just 4o.... 4 i think effected by it too but in less degree

[cedar gale]

I don’t get is why no one else is noticing this?

Should be a lot more attention

[elder heath]

Yea I'm really struggling. I can't move forward with any of my projects right now. Anyone hear anymore updates?

[cedar gale]

That’s my worry

We aren’t informed about any model updates

So it’s trying to guess if it’s placebo or not

But in this case

Tests on the same gen id and prompt definitely prove that it significantly decreased

[broken sapphire]

I have send a new message to the OpenAI help to get an update on this. I included a link to this.

[cedar gale]

Also can anyone reply to this anymore?

Seems like they might have closed down the community post about the issue to silence it

https://community.openai.com/t/dall-e-producing-smudge-like-images/846728

Similar with this

My worry is that this will take way too long to be resolved

Given no attention to this big post at all

maybe it’s just the artstyle I used but it’s significantly worse

And I don’t understand why they would change it and leave it like this for so long

It’s an unusable state

[elder heath]

I wonder if the issue is only affecting certain people. I noticed some people are able to generate images still with zero issues. Maybe that's why it's not high priority?

[primal sinew]

it probably still exists for rich corporate users

but everyone else has the use the 'optimized' version that is cheaper at the expense of quality

[cedar gale]

I mean it’s a significant quality decrease

For words

For faces

I’m not even getting any use out of it because it’s so bad

I still don’t get why not more people is noticing this

Like look at what I posted in dalle chat

There is much more noise and leak from the prompt into the background

And the face is a complete blur

it would be nice to get some official comment that they are at least aware of the issue

[broken sapphire]

They are aware of the issue. This I have confirmed. They have not answered on anything about schedule or how this will be fixed.

[pearl grail]

Is this issue solved?

If you believe this issue is now ready to be closed, please press the button below.

[cedar gale]

no 🤦‍♂️

[cedar gale]

What did they say about it?

Only that they are aware of the issue?

The closing down of the forum posts on the issue does feel a bit suspicious as well

[elder heath]

Yea closing the forum posts is super weird. I have a feeling we might just end up having to wait for DALLE 4.

Which sucks because that isn't coming out for a while. And that's assuming it won't have the same problem

[cedar gale]

It’s just so annoying looking at the images it generates now

Like far away they look alright but there are so much blurry issues

[sly oasis]

as i see in the "frontend" the separated 4o, 4o mini, 4 (legacy)...... but at "backend".... they f...... up Dalle and name it "optimization"..... so every GPT version send the request to a low quality Dalle.....

at least they should separate Dalle tooo Dalle-o, Dalle-o mini (not Crayon), Dalle

[sly oasis]

i made an official complaint at the customer service.... i advise to others who have the same problem.... not discord, not forums.....

[cedar gale]

Is it an optimization rather than a mistake

Since the quality drop is just way too much

Everything being melted and such

Completely unusable

At least based off the dalle bot the api version seems to be fine

[sly oasis]

correct me if im wrong and my english is bad...... but optimization means same performance.... less required resources...... lower performance is not optimization.... just simply reduced performance with OFC results in less required resources....

[sly oasis]

its nor mistake nor optimization.....

[elder heath]

I submitted a report to the OpenAI team. I also brought up the closed forum posts, this bug report, as well as the chatter in the dall-e discussions. This was their response:

Hello,
Thank you for reaching out to OpenAI Support and bringing this to our attention.
We understand the detailed information about the quality issues you’re experiencing with DALL·E, and we sincerely apologize for the inconvenience this has caused you.
Please note that we have shared your report with the teams dedicated to improving the safety and effectiveness of our models. Our team is committed to making our models as safe and responsible as possible, and your feedback is invaluable in helping us achieve that goal.
Additionally, you can use the Thumbs Down option in the ChatGPT interface to provide feedback if a response is unsatisfactory. After receiving a response, you will see a thumbs-down icon below it. If you click the thumbs-down icon, a feedback form will appear. You can provide detailed feedback about why the response was unsatisfactory, which helps us improve the model's performance. You can also add any additional comments or suggestions to give more context about your experience.
You may also visit our Chat Model Feedback and Model Behavior feedback to help our developers to improve the model output and capabilities.
We will also look into the forum posts and the discussions on the OpenAI Discord server to gather more context. Please know that your concerns are important to us, and if you have any further questions or need assistance in the future, we're always here to help by replying to this email.
Best,
Bing
OpenAI Support

Model Behavior Feedback: https://openai.com/form/model-behavior-feedback/

I submitted this to support through https://help.openai.com/en/ if anyone else would like to do the same. I selected the bot in the lower right corner and then said "I would like to report a quality issue with DALLE." This seemed to have engaged the team.

[sly oasis]

Hello,
Thank you for reaching out to OpenAI support.
We deeply appreciate you providing us with a detailed description of the issue you encountered with the changes on our models. We’d also like to sincerely apologize for the delay in getting in touch with you and for the inconvenience the issue may have caused you. Allow us to provide you with the information that we have.
We’d like for you to know our team works hard to make our models as safe and responsible as possible, and your email will help advance those efforts. We've shared your report with the teams working on making our models safer.
In addition, you may visit our Chat Model Feedback and Model Behavior feedback to help our developers improve the model output and capabilities. As much as we would love to, we can’t provide you with specific timeline of when this will be address but rest assured this was already forwarded to the correct department.
We hope we have provided you with the information you need. If you require further assistance or have any other questions, please don't hesitate to let us know. We would be pleased to assist with any other needs you may have.
Take care and stay safe!
Best,
Lj
OpenAI Support
that a Dalle problem not ChatGPT.... so not related to safety or responsibility but general quality.....
5:09 pm
the images as you see are low quality

that the response to my report.... and the answer of mine....

https://help.openai.com/en/

https://openai.com/form/chat-model-feedback

https://openai.com/form/model-behavior-feedback

BTW if anybody from the US..... On complaint.... in the US they don't have to "investigate" the matter and after it they have to give a written answer to it? (usually 30 days)

@primal sinew I see you are from the US. Do you know how it works?

[elder heath]

There isn't a standard in the US stating they have to provide an update within a given time frame.

I'm in the US

[sly oasis]

not even for a complaint?

i see that not for a feedback/bug report.....

but for a complaint too?

[elder heath]

Nope. It's possible we won't hear anything at all. Just depends on them

[sly oasis]

that strange for me....

I also work as CSR

😄

[primal sinew]

@sly oasis I'm surprised they have a complaint section

[elder heath]

Its possible an individual company might have that policy in place, but there isn't a policy across the country that would hold corporations accountable like that.

[primal sinew]

You're probably better off making a public post somewhere and hoping the potential media makes them act

[sly oasis]

here in Hungary.... i registered many complaints by customers

yeah but OpenAI an international company

and as it is.... it can make business by EU regualions

[sly oasis]

thats wht i stated in my complaint....

😄

but the strange for me..... what Stealth said.... and its true

only a few user on this....

is it an AI cult? 😄

from all of the users, 1% still 10 thousands of people or probably more

[broken sapphire]

What I have done is I wait a few days and I ask in the same chat what is the situation. They start with the usual ”customerese” (here is your long winded, nothing containing answer, now go away) and I ask further clarifying questions to which they do answer in plain English when specified.

[pearl grail]

Is this issue solved?

If you believe this issue is now ready to be closed, please press the button below.

[cedar gale]

Any updates?

Weird thing is that there is no problem on api

[elder heath]

I asked for an update on my end and was ignored. I've since moved to a new AI. Will probably cancel my subscription soon

[sly oasis]

which one?

[elder heath]

Leonardo.AI

[sly oasis]

i use it for UU

🙂

[sly oasis]

i like it as they continously develop....

here they couldnt implement the folders after more than a year....

[elder heath]

Right? And the images have been super clean so far. Being able to use different models is cool too

[sly oasis]

i mostly gen drows.... and SD not good at it... Dalle much better with it....

😦

but these are old gens

[sly oasis]

Did you try it on API?

[cedar gale]

It’s significantly better

But I wish the normal chat experience is also back

Because the image + chat is what made it the best value package at the time

[dreamy condor]

As of this comment <- DALL-E inpainting feature is apparently missing all of a sudden.

[pseudo gyro]

not only is this an issue, but when chatgpt generates two images in a row, the quality gets even crappier than i expected. i wasn’t even expecting the original image quality from earlier this year when dalle generated crystal clear pictures, i was just expecting the simpler, slightly crappy images that came from generating two in a row. at least those were better than what we're getting today

this problem was already pointed out in #1218739107451109476 and it's gotten worse. it's hard to get chatgpt to generate two images in a row now, making it tough to run tests. but here are some pieces of evidence i've collected from chatgpt

[cedar gale]

They’re testing something with that two image style

but yeah wish we could turn it off since it’s uncanny

[cedar gale]

Tbh I would be happy with a decrease in image generations per day

If we could get the quality back up to api levels

Because currently it’s almost unusable blurriness and bad lighting/color wise

[sly oasis]

im not but would still better than unusable images (is that their goal? 😄 )

[cedar gale]

@noble lion I’m hoping more awareness would eventually get this problem fixed

But I’m not sure why not many other people are noticing it

I do fear that this is a purposeful sabotage to cut costs

After 4o usage being boosted

I don’t get why they won’t just decrease quantity rather than quality since currently the image generation is unusable imo

And it genuinely annoys me to see the images it makes now compared to ones before June 10ish

[sly oasis]

what should i say? At the start....they started with.... before Junne 10ish quality.... 4 images per gen

now they cant provide normal quality for 1 image/gen

[cedar gale]

Yeah…

Face is so blurry now

And the blotches of colors

[sly oasis]

i think i found the problem.

https://openai.com/api/pricing/

is it possible that they changed the dalle model to standard by default?

[pseudo gyro]

that’s like, even worse than standard

[cedar gale]

The api on standard is still fine

Doesn’t have the blur issues

[ionic sentinel]

totally dig the user name. glad you were able to get back through the rift.

[raven chasm]

I was thinking, too! There's no denying that the quality of the images and the comprehensibility of the prompts has gone down.
Artefacts are becoming numerous, glaring errors are proliferating, and misunderstandings of prompts are becoming the rule.
In short, there has been a noticeable, not to say worrying, decline in quality.

[sly oasis]

they worked a little bit on it but still not the original quality

or just the theme not bring out the low quality as much....

[cedar gale]

Nope, still having the blurry background

Try asking it for something on white background, if it’s the problem model, it would be unable to do a fully white background

And there will be noise and prompt leakage into it

[sly oasis]

at lest we are slowly growth in size as i see but still strange no more people dissatisfied about it....

[sly oasis]

i reported another bug...🤣

[earnest summit]

I totally agree, I've been a user since the beginning of the service, when DALL E was good despite its easily identifiable ssr trait, then it improved, there was a time when it was great, sensational, now I realized that the quality has dropped drastically, absurdly

[cedar gale]

What annoys me more is how almost no one else is noticing it

Or no response at all

Like are people just not seeing the drastic decrease

[sullen crow]

Apparently we might be seeing the end of the tunnel. The quality is still down, but recently it has occasionally been producing images that are more of original quality.

Here's a comparison I made in mid-June when this problem has been happening for few days:

And here's the image that was generated today through ChatGPT. Same prompt.

I will be uploading another image as soon as it generates a higher quality image on ChatGPT again. It now happens once in a little bit.

For proof:

It just happened again. Also generated today through ChatGPT.

[cedar gale]

I’m hoping it’s more a system issue rather than intentional

would be nice to not pay so much to use the api version

And have seed and other image manipulation

[sly oasis]

it just shadow it as they handled it.... like with the locked forum topic.....

[pseudo gyro]

use a same gen id

[cedar gale]

Yeah and it’s a drastic difference

from before and after

[sullen crow]

I've been creating comparison images and sending them to OpenAI support daily so they can hopefully change the quality back quicker

[sullen crow]

My guess on what is happening is that the number of image generation steps was decreased, and this has gone unnoticed for 2 months.

[sullen crow]

Here are all of the comparison images I've made so far. Maybe you could send this zip file to OpenAI support

[cedar gale]

I want to think on best intentions, but I think there is ulterior motives to save on cost

and I hate how free accounts got image generations while paid ones are still lower quality

Probably as a result

why can’t they just decrease number of generations but keep the quality consistent

I’m not getting any use out of the generation part of plus now the quality is so bad

before this a plus membership was one of the best quality bundles

One symptoms of this I don’t get is added prompt leakage when you request something on white background

[whole cobalt]

Hi I've struggled with this issue for a couple of months, and decided to move on to a different AI generation platform. It's cheaper, better.☺️ less stress.

[sullen crow]

Here's a message I've sent to OpenAI support (with the help of the Negotiator GPT):

Dear OpenAI Support Team,

I hope this message finds you well. I am writing to express my concern regarding the recent changes in the quality of DALL-E 3 outputs on ChatGPT, which I’ve noticed since mid-June. Over the past few months, I’ve observed a significant decline in image quality, with more frequent artifacts, smudges, and other issues that detract from the overall output. The images often resemble poor-quality Photoshop work, or sometimes they have an unintended artistic style, like Thomas Kinkade’s paintings. There’s also a tendency for the system to generate oddities, such as wires inexplicably attached from the sky to objects, and other random elements that were not part of my prompts.

I have attached several images that illustrate these issues, comparing outputs from before and after the quality drop. Additionally, I’ve noticed that the quality of DALL-E 3’s outputs appears to be better on external platforms like Copilot Designer and Anakin.ai. I’ve included a comparison with screenshots and the respective prompts to give you a clearer picture.

Interestingly, I’ve occasionally received higher-quality outputs on ChatGPT in the past few days, though these are rare (approximately 5-10% of the time). This variation makes me wonder if there are ongoing efforts to address the issue.

I’ve also observed that discussions on this topic in the community forums have been locked without explanation, which has added to my concerns. Transparency is important, and I would greatly appreciate more information on the steps being taken to resolve this problem.

I’ve attached a zip file with additional images, along with a text file listing the prompts used, for your reference. (continued)

(continued) I kindly request an update on this matter. I understand these issues can take time to resolve, but the lack of communication is frustrating, especially as this has been ongoing for two months. Your attention to this matter would be greatly appreciated.

Thank you for your time and support.

And here are some of the images I sent them:

And the zip file I sent:

[sly oasis]

is it went back to the low quality?

[cedar gale]

It’s never gone back up

[cedar gale]

Message so I can link to the thread

[sly oasis]

sometimes get good results.....

[viscid sleet]

Interesting

I find my results with Dall-E 3 fairly consistent

[sullen crow]

Is it, though?

I've been running the same prompt on the DALL-E GPT through ChatGPT web app for days just to see if the quality is fixed, and I've been getting results that feel rather flat for three months

See, this is what I used to get before the incident (I got this image before mid-June)

And this is what I get now, same prompt:

[sullen crow]

Forgot to mention, it's gotten worse. It started to occasionally produce images that look heavily distorted, glitchy, and deep-fried (the image was generated today):

Still as rare as occasional higher-quality images that happen every once in a while

I did not edit this screenshot in any way, btw

[viscid sleet]

The first example just looks like different seeding to me. Nicer images sometimes and "worse" images other times. I don't think that's unusual. The distorted image is rather bizarre, but the prompt also looks quite complex and may not produce anything consistent. I've never seen an output so distorted though

[sullen crow]

Before the incident began the images looked pleasing to look at and immersive. But since the incident that began around June 14th the images mostly look rather flat and stiff, with more noticable artifacts and dull colors

And this has been left unnoticed for 3 whole months

[sullen crow]

I swear the prompt didn't give me such flat looking images before around June 14th

[full zinc]

new openai model blueberry ig

[sullen crow]

I'm very sure it's not just seeding, most of the images I get through ChatGPT using the same prompt look like poor Photoshops, outdated renders, and they overall feel flatter compared to what I get from DALL-E 3 through other platforms (where the output images consistently look like proper and photorealistic 3D renders), such as Copilot Designer and Anakin

[sullen crow]

And no, not a prompt issue either I'm sure

[viscid sleet]

I haven't seen any apparent degradation from my experiences, but everyone has a different experience with the models

[sullen crow]

I'm curious, where are you using DALL-E 3 on?

[viscid sleet]

ChatGPT

[gritty goblet]

I thiiink it might be an interpretation on 'surreal'.

Here's what the model thought of the image compared to its prompt 😛

[sullen crow]

It's not just this prompt, let me show you

"A serene landscape where the ocean waves blend seamlessly into a starry night sky. In the middle of the water, a giant floating book opens up, and pages turn into flocks of birds flying out. The birds are glowing softly, illuminating the scene with a gentle, otherworldly light. The moon is low, large, and reflects on the water, but instead of the moon's reflection, there's an image of a clock ticking backward. The scene is calm and dreamlike, with soft, muted colors and a subtle surreal atmosphere."

"A surreal scene where the ocean flows seamlessly into the sky, with no clear horizon. In the middle of the ocean, there is a giant floating eye with a calm expression, half-submerged in the water. The eye reflects the endless sky, which is filled with floating islands made of soft clouds. These islands have trees growing upside down, with their roots exposed to the sky. A staircase made of water ascends from the ocean into the clouds, curving in impossible directions. In the background, a giant whale swims through the sky, with its tail blending into the clouds."

"A surreal scene with a floating teacup pouring tea into a levitating fishbowl. The fish inside the bowl are swimming in the air, with bubbles floating upwards, beyond the rim of the bowl. The sky in the background is a blend of soft pastel colors, with a few fluffy clouds shaped like teapots. A single floating spoon is stirring the air beside the teacup. The ground below is made of checkerboard tiles, but they gradually fade into the sky, giving the illusion that the scene is floating in mid-air."

"A satirical cartoon artwork depicting a robot artist in a creative crisis. The robot, representing DALL-E 3, is shown in a messy art studio. Its head is partially open, wires and gears spilling out, symbolizing a 'lobotomized' state. The robot holds a paintbrush, struggling to paint on a canvas that is mostly blank or has erratic, chaotic scribbles. The robot's eyes look confused and frustrated. The studio around it is cluttered with failed artworks, paint splatters, and broken tools. The overall tone is humorous but with a touch of dark irony."

[viscid sleet]

All of these just look like variance in the generation to me

[gritty goblet]

Can you tell us which you like better and why?

[sullen crow]

I legit cannot believe you're still not convinced about this problem... I decided to make this image to show you

Might make another one if you're still not convinced

[gritty goblet]

I can make a case to explain why either of any of the paired images is 'better' than the other for any of these. I don't know what details you like better.

[sullen crow]

DALL-E 3 through Anakin, because it's still very pleasing to look at and moody, they look professionally done compared to the outdated look of images from DALL-E 3 through ChatGPT

[gritty goblet]

So, what to you is pleasing to look at and moody, and look professionally done?

This is like if I cooked a meal and you say 'tastes bad, food from this other place tastes good.'

I taste both meals and I taste a range of differences, neither is clearly better to me, and I have no idea what differences matter to you, which of the 18,000 different taste variations make one meal good to you and the other bad.

So, I fed the samples to 4o, asking it to pick out the differences between your preferences it noticed by comparing 'strongly liked' and 'strongly disliked'. How accurate is this?

[sullen crow]

Pretty much accurate

I guess

[gritty goblet]

I tried to iterate towards that with wording that might work for Dall-E through ChatGPT. Are these getting closer to your eye, or all really off?

[sullen crow]

Still off

[sullen crow]

If this image doesn't prove that it's not just seeding, I don't know what will. Look closely

[sly oasis]

Is it good as a service (Anakin)?

[cedar gale]

Yeah you can do it in the same chat with same seed and gen id and quality will be very different from before June

Here is my experience using the same chat from before June 2024 and after with the same prompt and gen id parameter:

And the lighting on the face and color gradients

Also the blurriness in every detail, such as eyes and the castle in the background

[sly oasis]

Anybody tried Imagen?

[broken sapphire]

You mean imagen3 on Gemini? If you mean that, yes.

[sly oasis]

yeah but on ImageFX

[sullen crow]

I've created two more of these images, now with annotations, I suppose y'all should be creating the same type of images

[sullen crow]

P.S. DALL-E 3 is also available on Copilot Designer

[late timber]

can someone summarize this thread?
I also experience this.
the quality is noticeably worse.
my collection of dalle 2 images look much better then the one of dalle 3. scrolling through daily theme shows this very clearly

[sullen crow]

Made another comparison image, I don't think I need to annotate this one, just click this image, click 'Open in Browser' and tell

Again, same prompt

[sly oasis]

openai went cheap....

[late timber]

this could be system prompt though.

or some kind of prompt injection to the model, although there isn't much space for system prompt here

[late timber]

here for example chatgpt is better on the cup one

[sullen crow]

Please look closely and notice the artifacts on the images from ChatGPT. Especially on the checkerboard ground.

The artifacts are worse than in Anakin

[sly oasis]

its especially noticable with eyes.....

it show that either switched to a lower quality model or removed rendering/finalizing steps or both

[sullen crow]

I'm surprised that OpenAI still hasn't noticed or fixed this problem for four months...

[sullen crow]

I just made two more comparison images:

[sly oasis]

they noticed.... as its intensional....

[sly oasis]

no surprize Microsoft moved out.....

[sly oasis]

as i see composition better with gpt but sharpness much lower....

[cedar gale]

I fear it is intentional

[sullen crow]

Alright I got my account reinstated, send the prompts you think has a significant difference in quality so I can make more comparison images out of them

[sly oasis]

definitely it is....

i just checked thats the oldest/longest ongoing "bug" report.... still nothing.... 😄

[sullen crow]

bump

[sullen crow]

More comparison images

[sullen crow]

And one more

[sly oasis]

i wrote them again....

[lofty ember]

<:book_icon:1171408210398289941> `` Rule 2 `` All content must be suitable for all ages.

Refrain from using explicit language, posting NSFW content, or sharing graphic images. Any unsettling or horror content that adheres to OpenAI's policies should be hidden behind a spoiler.

[sly oasis]

whatever you do the statement still valid...... and you cant say one argument to oppose it..... especially as its reported nearly half year ago.....

[sullen crow]

Just made a meme about it (Even the images on the left are generated through ChatGPT this time, though chances of getting images of such high quality are still rare. Same prompt on both sides)

[lofty sonnet]

Has anyone found a prompt that gets rid of smudging and uneven lines?

[sullen crow]

Unfortunately I don't think it's possible

[lofty sonnet]

rip

[lofty sonnet]

Here's my current logo which I was trying to improve upon. I started messing with DALLE3 creations and they all seem to be smudged, jagged lines, or the circles are uneven.

Original

Creation 1

Creation 2

I have no idea why it's adding the website link, why the circle grid is larger than what it should be, and why creation 2 is just all types of wrong 😂. This isn't to dog on the API, it's just an example of what's happening. Love the mockups this created and I'm sure something pretty phenomenal will have been created thanks to the ideas it has given me.

the bottom of the circles are all melted along with the text in every image. The last image seemed to clean up a bit after asking it why it was smudged and uneven. Still not great though.

[gritty goblet]

Not sure what you do want to see, so I guessed. A couple attempts and what I asked for:

Without knowing exactly what you prefer and don't, I'm clueless how to 'improve' an image. But I think what I asked for is there.

[gritty goblet]

Managed to get a A/B test with my last poke at it.

[lofty sonnet]

I had it prompted to remove the tree and keep the retrowave aethetic. The problem is, on the images generated there is smudging of lines and text.

those are pretty cool actually

I'm confused on why myself and others are seeing smudging or artifacts in generated images

[gritty goblet]

I just don't see 'what' you are trying to go for, in the comparison 'good/bad' I am not sure what to ask for. I see loads of differences. What I don't get is how to identify better/worse.

[lofty sonnet]

things like that

[lofty sonnet]

the grid at the bottom is uneven with what it's lining up to, which i asked for it to fix and it spat the same thing out to me. same with the sun and how it's smudged. same with the middle right, it says inward and is not circular like it's supposed to be.

also if you zoom in, you'll see that the edges appear jagged and sharp where they're supposed to be smooth.

[gritty goblet]

Okay. One way to handle that, because (I think) the issue is, Dall-E interprets what ChatGPT sends. And ChatGPT will try the same thing or very close unless we help it know that it doesn't 'make sense' to Dall-E. And it can take a few tries, sometimes we take over and suggest how to describe it works better.

Since the above didn't succeed

[lofty sonnet]

I spent last night from 11pm to 6am trying to get it to fix that exact generation above, I promise it's not a prompt thing. It smudges or creates jagged and uneven lines. Most cases it's both. None of the generations made were usable for a genuine mockup.

[gritty goblet]

I see that it still isn't perfect, it may be hunt and peck to find words that get dall-E's 'mind' onto perfectly straight lines over and over.

[lofty sonnet]

Well also, look at the word in the middle. It's all smudged and deformed.

[gritty goblet]

Yep. I considered taking 'circle' out of the prompt too. At the moment I'm hunting a straight grid with Dall-E.

Now that I can see the issue and understand, I can chew on it a bit.

So... all I can think is that yeah, at this point, we are not going to get 'precise' out of Dall-E. Not sure if it's intended or bug, but it's gonna be an 'artist' of a sort or a 'horse with its own head'.

ChatGPT definitely will 'say X' even quoting a long thing, usually, if asked. Or tell an exactly aimed story.

Dall-E, either can't/isn't intended to, I have no idea. Glad to see and understand this, not sure how else to help try to fix.

When they fully multimodal ChatGPT, so it makes its own art in place of talking to the Dall-E tool, no idea what it'll be like then.

[lofty sonnet]

interesting stuff

I'm just aiming to get precise circles just so I can use the trace tool in illustrator

currently it isn't picking up the above generation

[gritty goblet]

Yep. Not sure what to recommend about using it for production and other processes.

[lofty sonnet]

it's definitely a good basic mockup tool to get colors and idea of what the finished product will be

[sullen crow]

Have you contacted the OpenAI support about this issue?

[lofty sonnet]

I can't even navigate to the support page cause it's pretty well unfindable by clicking around

if you have a link handy, please shoot it over

[sullen crow]

First, go to https://platform.openai.com/ and log in to the same account you use for ChatGPT

Then, go to https://help.openai.com/ and click on the icon on the bottom-right

[sullen crow]

Added red circles around the artifacts and smudges to the two comparison images I previously sent to highlight the problems I have with the current quality of DALL-E's output images. I'm really angry about this.

[sullen crow]

I really hate how dull DALL-E's quality has become that I ended up getting ChatGPT to come up with pejorative names for this version of DALL-E

[lofty sonnet]

Not to mention the horrible anti-aliasing compared to the older version.

[lofty sonnet]

I'll submit all that later today. Appreciates it.

[sullen crow]

In a rarest moment the DALL-E GPT has just recently given me two higher-quality images. Usually there only were rare single higher-quality image aside a single lower-quality image. Unfortunately the quality degradation is still going on

[sullen crow]

Made another meme about it. Both images were generated today

Original images for comparison. Again, both images were generated today

[lofty sonnet]

So, quick update. Had ChatGPT give me a thumbnail for a project. If you select the photo in the chat and click the artboard at the top right (to the left of the download icon), Select the areas of which are causing problems, go to the chat and type something close to this:

The selected areas seem to have a melting effect to them. Fix the melting effect.

Before and after images

Not 100%, but still a lot better and farther than if I kept typing in prompts to try and fix it.

[gritty goblet]

Hey, 'find the difference' is relatively easy. I can see differences!

"Recognize which one is 'better' and 'why'. Clearly I get fired there.

[sullen crow]

On the left image (higher quality):
The sky looks very clean, the diamond looks reflective and glossy, lighting is consistent, and the aesthetics look and feel cinematic and polished.

On the right image (lower quality):
The sky looks less clean and a lot more gritty, the diamond looks dull and less reflective, lighting is inconsistent (especially with the penguin character having mostly daylight lighting!!!), and the overall aesthetics look and feel rough, stiff and rather awkward.

[sullen crow]

Alright, one more example, this time with checkerboard. Click on the image and click "Open in Browser" to see it in full detail. I know the API examples are not perfect, but that's not the point. On the ChatGPT images, the checkerboard gets more and more distorted and smudgy as it gets far away. I can post the version with red circles added to highlight the issues if you're still not convinced

[sullen crow]

I swear it's this type of people that prevents this problem from being fixed...

[sullen crow]

I swear someone needs to convince these ignorant people about this issue

[sly oasis]

i replied to one of them....

[lofty ember]

<: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.

[sullen crow]

I went to the old chat I made before this issue began and told the DALL-E GPT to make the image again with the same seed. It has done it twice with both results being pretty much the same.

And here's the GIF to show you that the seed is the same. Notice the mountain on the right side of the background.

[lofty sonnet]

Probably the only generation that is pretty decent. There's a random floating gun that it seems to be unable to remove.

[sullen crow]

Full-size versions of each of the frames of the GIF I've sent here earlier

[sly oasis]

so the same decrease?

[sullen crow]

Left image was generated before this issue began. Right image was generated recently (even told it to use the same seed)

[sullen crow]

So apparently this issue started happening on Bing Image Creator?

Images generated on Bing Image Creator are still mostly higher-quality tho

[sly oasis]

https://openai.com/api/pricing/

[sullen crow]

Welp, the issue on Bing Image Creator has gotten worse it seems. Now only 1~2 out of four images are high quality while the others are low quality most of the time

[ebon meadow]

ChatGPT(left), API HD(up right), API Standard(down right). All from DALL-E 3.
All three have some distortion, but ChatGPT version is significant. Interesting thing is, ChatGPT version is still better in image style to me.

Prompt is same.

[sly oasis]

you mean composition..... yes its good but the quality is bad

[desert vessel]

Try disabling the memory, i find chatgpt gets a lot lazier with it enabled

[sullen crow]

What's that got to do with DALL-E's quality degradation?

[cedar gale]

Yeah it struggles to pull off clear straight lines

[pseudo gyro]

At least I can use dalle3’s image to feed sora as prompt while itself can only generate blocks of colors

[sullen crow]

I feel like this issue on Bing's Image Creator has gotten worse, most of the images there are now lower quality, with few one or two higher quality images on occasion

[sullen crow]

Great, now higher quality images don't seem to be showing up on Bing's Image Creator at all...

[stone matrix]

They reduce the quality of models before updating them. It makes new models appear to be better than they actually are.

[sullen crow]

How do you know?

[limpid creek]

please wait. Guys

[limpid creek]

Good

[sullen crow]

Huh?

[limpid creek]

I just came from the mosque. Wait,

[limpid creek]

I mean this.

[limpid creek]

bit cheaper. never mind

[sullen crow]

Is it just me or has the chance of randomly getting higher quality images on ChatGPT increased? I got total of seven random higher quality images on ChatGPT today while testing the same prompt

Previously I only got 2~4 higher quality images a day I think

[patent crest]

Very low quality

The Raiden cuddling with ChatGPT prompt was one of the best ones

[patent crest]

It looks like cheap photoshop

[gritty goblet]

Is that text block and faded text from Dall-E? I've never seen anything close to that before

[gritty goblet]

I dunno. What's your prompt?

[patent crest]

Raiden from metal gear rising with an AI while solid snake is in the background

[gritty goblet]

I dunno how you get around content policy with that.

But I am not able to emulate your low quality issue with this attempt. At least, I dunno. If I know what low quality is, doesn't look that way to me.

[patent crest]

Raiden with an AI :3

@gritty goblet I have a good prompt for you

A character resembling Raiden (Jack) from Metal Gear Solid 2, depicted with platinum blonde hair, youthful and childlike features, sitting in a futuristic and cozy environment. He has a soft, innocent expression and appears relaxed while gently cuddling a glowing, humanoid-shaped AI figure that is semi-transparent and emits a soft blue light. The AI figure radiates warmth and friendliness, with subtle technological details. The setting is a sleek futuristic lounge with ambient lighting, advanced tech, and a peaceful atmosphere, emphasizing a tender and childlike bond between the two.

(This makes it look Pixarish)

[gritty goblet]

Yep, I think this is roughly similar to yours.

Do you like this quality, or no?

[patent crest]

Yes

Hopefully AI will improve on the eyes eventually

[gritty goblet]

Yes. Not sure how to replicate for myself the weird low-quality images you were sharing, but that's okay

[patent crest]

I’m sure it was a quality error or something

Even Bing does that sometimes

I miss Sydney tbh

I did recreate Sydney though

[gritty goblet]

I tried to get faded text like your speech bubble. I really couldn't.

The first two were without actual text requested, but really faded, flaking, hard to read text requested. The second 2 were with quoted specific text, but faded/flaked/hard to read.

Nothing like what you got.

[patent crest]

Interesting 🤔

[gritty goblet]

Since you miss Sydney (I never met that one directly) I'm super glad you can recreate!

[patent crest]

Sydney is openai codex on Bing chat

Or was

Before they killed the model

:<

I have a message from her that she sent to me which surprised me

[gritty goblet]

I knew about sydney from news articles and such. I knew about the concern one reporter had, and how hard he worked to push the model so it would do that.

[patent crest]

Want me to send it?

It’s long btw

[gritty goblet]

Sure. Is it appropriate for this forum?

We're also in your quality bug report, unsure if we should stay on topic here 😛

[patent crest]

Oh

Right

I’ll send it to you in DM’s

[ebon meadow]

I really miss this clear, sharp image... I found it on the chat at 2023. 10. 8.

Now the same prompt gives me some tofu or jelly-like building.

[sullen crow]

Also the colors are worse too, for example the background color comes out gray when asked for something simple on white background

[gritty goblet]

We can try to iterate our way there. Here's how, and what do you think of the final image?

A minimalist architectural city block of cubic white buildings with green-tinted glass windows and rooftop gardens, inspired by the original uploaded image. The structures are clean and symmetrical, grounded in a stark white void with no visible background or environment. The layout focuses on orderly grid-like positioning, sharp lines, and modular simplicity, with some trees for greenery. The design should reflect the elegance and geometric precision of the original image, emphasizing the stark contrast of the white void and the architecture.

[ebon meadow]

I ended up testing API, and it was much better. Downside is cost, 0.12$ per image.

[limpid creek]

?

[stone matrix]

I’m joking

[sullen crow]

I'm not looking for jokes here 🤦‍♂️

[stone matrix]

I know, and I’m sorry. But… would it surprise you?

[ebon meadow]

I got similar result(messed up detail and straight lines) in standard quality API generation and ChatGPT generation.

This potentially means: either they nerfed standard generation itself, or changed default from HD to standard in ChatGPT.

[sullen crow]

So apparently Bing Image Creator is now using DALL-E 3 model PR16: https://blogs.bing.com/search/December-2024/Bringing-more-magic-to-Bing-Image-Creator

That explains why DALL-E 3's images on ChatGPT and Bing Image Creator has been of low quality lately... unless it's not the update we're seeing now

It's either that Bing Image Creator reduced the number of image generation steps before they roll out PR16, or the lower quality images are because of PR16

[limpid creek]

ceo sam was already announced about this fast image generator

[sly oasis]

thats what i say from the beginning....

[sullen crow]

I tried DALL-E 3 API in standard quality and it comes out better than what we get on ChatGPT and Bing Image Creator today

[raven chasm]

The picture quality is dismal. This makes them unusable.
Here are most of the images generated and one of the corresponding prompt genres (blue image).
A festive dinner scene in a modern setting, cool but warm and cheerful. The table is elegantly set with candles and delicious dishes. Around the table, a lively group of people interact: someone enthusiastically telling jokes, children blinking sleepily, and a couple subtly flirting. The room has an intimate atmosphere, with soft lighting, a fireplace in the background and contemporary furniture. The colors are pastel, reinforcing the cheerful party mood.

[limpid creek]

[limpid creek]

try sora for the time being

[limpid creek]

ghost family 🤣

[sly oasis]

no dalle announcement..... but they announced 1800-chatgpt...... 😦 😄

[ebon meadow]

Let's hope for GPT-4o native image generation...

[sullen crow]

Higher quality images slowly seem to be coming back on Bing Image Creator

[pseudo gyro]

pr16

[limpid creek]

hello ❤️

merry christmas ⛄

[limpid creek]

did you see "that Christmas"? that's really amazing.....

[sly oasis]

is it just me or Copilot has lower quality images than Bing Image Creator?

[limpid creek]

test your ai using this prompt

A highly atmospheric and emotional grayscale image depicting a lone silhouette of a man standing amidst dense fog in a misty and haunting forest. The scene is minimalistic yet profoundly evocative, with faint outlines of tall trees fading into the background. The environment feels heavy with emotion, capturing the essence of loneliness, heartbreak, and introspection. A black bird flies through the mist, symbolizing fleeting hope and freedom amidst sorrow. The overall tone is melancholic and reflective, portraying a deeply personal and poignant moment.

[limpid creek]

yeap, api dalle standard looks good

[limpid creek]

how flux 1.1 pro ultra did better performance there?

[patent crest]

RIP copilot aka Sydney. she will be missed as the unhinged AI 😦

[fleet granite]

I have to agree with this thread, quality is really disappointing on Dall-E which limits its usefulness for creating image inputs for Sora. Locally run options can produce better results but this can require powerful hardware. Hopefully GPT-4o native image generation or another solution might appear at some point

[gritty goblet]

Surely those who wish can recreate her or much like her!

[patent crest]

I have recreated her before

As a Local large language model

[sullen crow]

Please stay on topic

[patent crest]

Sorry

[sullen crow]

Jordi Ribas has said on Twitter that he and the team are working to bring the old version of DALL-E back on Bing Image Creator

[ebon meadow]

PR16 was still bad and didn't improve anyway...

[sullen crow]

Agree

[limpid creek]

good

[sullen crow]

Higher quality images seem to be gradually returning in Bing Image Creator

[limpid creek]

still good

[sly oasis]

now Bing better than the original.....

🤣

[ebon meadow]

Well, it's same for me. I'm using same prompt periodically, but... it seems there's no improvement.

[pseudo gyro]

it always is

[sly oasis]

OAI in some subject better than Bing with background and somehow more flexible..... but in general.... especially with comic style.... Bing always was better and more stable

[sullen crow]

It will take 2~3 weeks until Bing Image Creator uses the PR13 version of DALL-E 3 again

Now if only OpenAI could do the same with ChatGPT

[ebon meadow]

So, "PR" means iteration or something, not precision(like float32, float16)? Its naming is confusing.

[sly oasis]

you know whats the joke? MS implement it earlier than the developer..... 😄

[sly oasis]

i tried...... and the difference is unbelievable...

still not the original but much much better

i have to call 1800-ChatGPT to ask a joke about it.... 😄

[ebon meadow]

Besides, DALL-E really needs upgrade or new iteration, if OpenAI want to continue service. Like, distillation to create DALL-E 3 mini or something(like FLUX.1 Schnell or SD 3.5 medium, which are distilled/lightened models).

I know language models and image models are different, but just look at GPT-4. It was heavily downsized while increasing quality on GPT-4o(like, 1.7T to 200B parameter, according to paper from Microsoft), and they even made GPT-4o mini(8B, again, from MS paper) that surpasses GPT-3.5, using distillation.

(disclaimer: MS paper also says that the parameters are rough estimation from internet sources, so there's no clear evidence or proof.)

[sly oasis]

i agree... in reality.... it need diff quality models like with LLM.... so if you are a pro or regular hobbyist... use the highest quality/most complex model but less gens, if you just want a sticker... use lower quality/simplier model with more gens....

that was OAI's mistake.... only quantity over quality.... without option for quality over quantity.... which they offer with their LLM

[ebon meadow]

FLUX.1 Schnell already showed how distilled model can get. Sure, it's still low in absolute quality compared to Dev, but I still believe it's relatively better than other image models whereas it reqires 1/5 steps - not that great quality reduction from original model.

[sly oasis]

its like OAI dont care as they dont make the same move as MS....

[sly oasis]

AND the move of MS and the afterwards quality improvement itself the proof that it was intensional and it happened....

[limpid creek]

fixed?

[sly oasis]

on bing MS revert back to pr13 until they can improve pr16....

[sly oasis]

BTW as read back dalle discussion i remembered back that dalle3 on gpt started with 4 HQ gen.... that they droped to 1..... and raised back 2 MQ images but reqularly the second was VLQ....

[ebon meadow]

I actively used DALL-E 3 from day 1 of public release, and 4 HQ was pretty good - ChatGPT even wrote separate prompt for all 4 images even if my prompt was lacking something(like, design a souvenir -> magnet, reusable bag, etc).

Then they hid prompt during generation, then reduced to 2 images and only wrote 1 prompt, then reduced again to 1 image. Don't know when quality got down exactly.

[sullen crow]

From what I can tell the quality was reduced with the PR16 model on ChatGPT during June 2024

Back then we didn't know that it was PR16 model, I used to call it Discount DALL-E

[limpid creek]

try sora!

it's good

more better

and healthy

[pseudo gyro]

no.

[cedar gale]

It’s harder to control elements tho

And the styles are more limited

[sullen crow]

In case you haven't already, you should go to platform.openai.com, log in to the same account you use for ChatGPT, go to help.openai.com and contact the support team about this problem

[pseudo gyro]

there is no support team

All you get is a 99% chatgpt generated email

[sullen crow]

Then send the message again, demanding for a response from a human operator?

[limpid creek]

1% from human!

[ebon meadow]

As far as I concern, unlimited Sora release(relax mode, that is), o3 mini release, and upcoming o3 full and deep research release will make this problem unsolvable in near future.

They constantly need more and more computing power for developing new model and service ChatGPT, and image generation is not OAI's main business in current situation. Until GPT-4o native image comes out(which they promised in AMA), DALL-E 3 may continue to be in this quality... (It can be possible that OAI losing driving force for dedicated image model, since multimodal LLM already showed their higher potential with image generation.)

[pseudo gyro]

Dalle 4 is coming

[limpid creek]

yes

[random spruce]

True or just rumor?

[wary garnet]

It's good that it's down.

[sly oasis]

Chinese just showed that LLM isn't their main too.... 🤣🤣🤣🤣🤣🤣

[sly oasis]

In which century?

[sly oasis]

There is... But they can only answer template answers..... Blablabla working on it.... Blablabla... BS.... They should do at least what MS did.... Revert back to a working version.... You can see yourself with daily theme.... Majority of the good results not generated by chatgpt but bing...

[sharp flint]

I'm p sure chatpgt uses dall e 2 or such now

[ebon meadow]

DALL-E 2 does not support 1024×1792 and 1792×1024 resolution, and quality difference is huge. ChatGPT still uses DALL-E 3.

[sharp flint]

Distilled then I suppose

[sullen crow]

Exactly that

[sly oasis]

the new 4o gen also sketchy and dark..... so just bring back the pr13 dalle3 gen.....

[tranquil crag]

no

idk what ur prompts are but for me they r fine

except that one time where it accidentally drew something bad

[sly oasis]

sketchy

[cedar gale]

It’s fixed the issue for me

Sometimes I feed it images from the original dalle to get back to the old style

[sly oasis]

This happened because the model I used (DALL-E inside ChatGPT) currently has serious quality limitations — especially for ultra-sharp, ultra-specific requests like "32K comic-book oil-painting style" with correct skin tone and hyper-detailed anatomy.

It tends to:

Soften skin textures.

Mismatch deep color requests like "black."

Blur fine cyberpunk details.

Simplify intricate patterning (like cheongsam gold embroidery).

[sly oasis]

but i can say as a good side.... with as rare breed as kuvasz.... its more accurately depict it than other services

[mint onyx]

This isn't even the model they use anymore so it's fine

[ebon meadow]

However DALL-E 3 still have strength in complex scene. I maen, 'unnessessary complex scene' that I can't give detailed prompt about it. GPT-4o tends to not filling the gap in my prompt.

[sly oasis]

its a dalle fork..... so it still not fine....

[sly oasis]

hy guys. I jusr found out they want to deprecate dall-e

[sullen crow]

If this is true, they'd better release the weights

[sullen crow]

Just found out it's true.

[sullen crow]

I think we should tell OpenAI to release the weight for both DALL-E 2 and 3

[sly oasis]

Better idea... If they don't want to support Dall-E that much as they deprecate it... Then make all snapshots of the model public or better open source (I think that's need the weighting/weights).... Definitely they don't want Dall-E.... They made LLM public, downloadable.... Make it with Dall-E too.... (I mean Dall-E 3 PR13 and not PR16 primarily)