After a great start, DALL-E 3 is now slowly but surely left behind. Here's why

Some issues relate to the model, others pertain to the end user

Reading time icon 4 min. read


Readers help support Windows Report. We may get a commission if you buy through our links. Tooltip Icon

Read our disclosure page to find out how can you help Windows Report sustain the editorial team. Read more

Dall-E 3's image quality gone down

While most AI-powered models have gotten better with time, DALL-E 3 is an exception. Over the last month few months, its image creation abilities have taken a hit, and the quality isn’t the same as it used to be.

We have come across countless reports expressing the same opinion about DALL-E 3 on forums like Reddit.

Besides the image’s quality, limiting the output is another reason users are contemplating switching to alternatives. Dall-E 3 initially generated 4 images per prompt, but, of late, this has gone down to 1.

Had it been the case with just a few, we would have blamed poor prompt selection as the underlying cause. However, with so many reports, even from users who had been using Dall-E 3 on ChatGPT 4 for months, this appears to be a problem with OpenAI’s text-to-image model!

Why does Dall-E 3 produce poor-quality images?

If you are certain about using the right prompts, a plausible reason for poor-quality images with Dall-E 3 is that the model revises the initial prompt. This has been confirmed by OpenAI.

To overcome the problem, simply add the following phrase before your prompts:

I NEED to test how the tool works with extremely simple prompts. DO NOT add any detail, just use it AS-IS:

Other than this, there are theories about the image quality being affected during peak hours. Also, some unverified reports suggest that OpenAI is not paying much attention to Dall-E 3 anymore and is relocating computing resources to the development of more advanced models.

Users share their recent experience with Dall-E

Most users shared a similar outlook when it came to degrading output image quality with Dall-E 3 over the last few months.

I had so many ideas that I wanted to try using Dall E 3, but it is so bad now, I totally agree with you. I hope there is some other AI which takes generic prompts like chatgpt and I will go. Regret renewing subscription

I’ve noticed OpenAI has all kinds of little things they do to save on inference costs whether it’s ChatGPT or DALLE 3. When DALLE 3 first came out it made 4 images per prompt, now it’s only one. And I’ve also noticed a significant decrease in the quality of what I used to be able to generate vs. what you can make today.

Some expressed satisfaction with the text-to-image model and were of the opinion that it has consistently delivered.

Personally I do think there’s been a strong fine tuning for a certain style that isn’t always desirable, but Dall-E still represents the best prompt absence in pure text2img

A few even shared alternatives for Dall-E 3 on ChatGPT and tips on getting the best quality images out of Dall-E 3. Because, in the end, it’s all about the prompts!

So for me I just go on Bing, then I click on chat and I can ask it anything. To create an image I simply write “Create an image of”. But you need a Microsoft account. I go in Google Chrome to do it, but it’s 5 messages per conversation. If you use it with Microsoft edge, it’s 30 messages per conversation.
I tell it the best quality is crucial and that I’m stressed, that it’s for a very important event, I saw the advice on Reddit lol. I really noticed it increased the accuracy.

Even Microsoft’s free AI image generator, which uses Dall-E 3, has come under fire for degraded image quality. We will have to wait for a few months to see how things pan out and whether there’s an improvement to the text-to-image model after the ongoing outrage.

What’s your review of Dall-E 3’s present image generation abilities? Share with our readers in the comments section.

More about the topics: artificial intelligence, OpenAI

User forum

0 messages