Today, defects in images have appeared that have never been seen before. Other images have small artifacts or white spots, or they show streaks that look like exposure errors in photo development.
If DALL-E has been updated or if any parameters have been changed, something has gone wrong.
Today something went wrong, I got similar to yours today, but not for all prompt.
For example I used following prompt:
An anime-style illustration featuring Nasreddin Hoca, a wise and humorous old man with a long white beard, wearing traditional Turkish attire, including a colorful turban and a long robe. He is standing next to his donkey in a lively village marketplace. Nasreddin Hoca is smiling mischievously, holding his hand up as if explaining something to the crowd around him. His donkey is looking towards him curiously. The marketplace is bustling with villagers in traditional Turkish clothing, with vibrant stalls in the background. The atmosphere is cheerful and warm, with soft natural lighting. The illustration style is influenced by soft pencil strokes with a pastel color palette, giving a soft, friendly, and humorous feel. ar: widescreen.
And I got this image, today.
But now it is OK with same prompt:
Yes exactly!
So i am not the only one. They have recently implemented a bug.
It is not prompt related, it is a generator issue.
I got 2 defect images at the same day, and never before not one. And i made many images.
It is not fixed yet. It repeats again:
A hyper-realistic 3D image of a lion entirely composed of matchsticks. The lion is depicted in a majestic pose with intricate details in its mane, body, and facial expression, all constructed from matchsticks. The mane features varied shades to represent flow and texture, using matchsticks of different colors and orientations. A burning matchstick lies nearby, adding a sense of danger to the lion’s powerful stance. The scene is highly detailed, emphasizing the matchstick construction while maintaining the realistic appearance of a lion. ar: widescreen.
No, Not fixed, i got a other one too, the same artifacts.
It would be nice if a developer could tell us something. There was clearly a change around 2024-09-27.
The probability is around 1:50 to 1:100. I Got 2 in 100, and now 1 in 15.
These artifacts have also appeared, there are images with white dots larger than a pixel and streaks that look like exposure errors, in chemical photo development.
(Did DALL-E consume nightshade? I don’t know exactly how nightshade affects training, I just know it exists.)
We’re trying to find out, but no guarantees we’ll hear anything. They’re likely aware and working on a fix. Thanks for reporting!
Sudden, unusual, and volatile phenomena are usually a sign of important work being done behind the scenes. We may have to wait a few days for a small update in DALLE.
Yes… And it seams to get more worse, yesterday i got 2 of 100, today 5 of 100.
It damaged some nice pictures 
I wish we could use seed numbers, it would make trouble checking and learning so much easier.
I’ve not seen it on the API…yet… and I’ve done probably > 1,000 in the last couple of days…
I know the Choose A or Choose B response now has an “Analyzing…” flash and some other new stuff on ChatGPT proper…
I wonder if the lower quality is just DALLE2 or a beefed up DALLE2 maybe? A test model? It’s kinda funky. But previously if you did a LOT of ChatGPT images in a row, it would spit out 2 lower quality images that appeared to be DALLE2… But it was usually two images, so I don’t know…
One of us will report back if we hear anything official. I’m just speculating here!
I speculating too of course, but it looks like Nightshade, but this would only enter by a new training update. Or it could be a webp encoding issue?
Nightshade seam to be very toxic, it only need few pictures to scramble a the weights of a training.
Nothing has changed in my using pattern or speed. I have seen defect pictures in the past, but nothing like this. It is only the last 2 days, never before.
Now if you create pictures too fast, it gives you a penalty waiting time, but not so bad quality. (something like over 20 to fast.)
(I never used DallE 2 and never API, so i can not compare.)
I got 16 defect images now. And it lock like damaged weights to me.
I checked pictures from some days before, and i could find some with spots like shown above.
In this image is rain.
Now i get this, over and over again.
“Error loading image”
same. 
I can’t get any image.
“Error loading image”
We’ve let them know. Waiting on hearing back, but as Dev Day is upon us, I doubt an answer quickly. I’m sure the DALLE / ChatGPT team are hard at work on the problem, though…
I’m not seeing the problem on the API. It’s a lot more expensive, but you get what you want when you want it…usually!
More than happy to help anyone set-up a super simple script to interact with DALLE API for image generation via a text prompt…
Having no experience with API, I would like to know how to get started in a simple way. I’m hesitant to dive into API because I dislike Python. I code in C# and maybe PHP… I haven’t looked into it at all, so I don’t really know what more it can do.
I also have a few ideas to adjust the mask to my needs.
(Also, I’ve seen people here somehow lose a lot of money with API.)
If you know of a good tutorial, I’d definitely check it out. (But if I see Python, I’m out! 
)
(Could it be that they’re working on a solution for why generation is off? Maybe they’re using Plus users for beta testing, if the API still works 
)
Thanks in advance!
I can throw something super-simple together if you’re interested.
Might be something on GitHub already?
Yeah, testing something on the ChatGPT side I think. Maybe a fill-in for the old DALLE2 that sometimes snuck in if you were generating too quickly. I’m not 100% sure, though. Will definitely pass along any info I can!
But yeah, I’ve not seen it on the API with well over 1,000 generations now in the last few weeks…
