DALL-E 3 reverting to DALL-E 2 in certain prompts

Not sure if anyone else had this issue. But in 1 out of 10 prompts (ish) the prompt will generate 2 images that look like ass generated what look like DALL-E 2 or even 1. This is annoying because it wasts a prompt on subpar results. Considering the prompt cap is already very stingy it doesnt make things better.

Is this a bug?

I got one of those also yesterday. Two results even though one was requested by ChatGPT. (I haven’t asked for pictures since.)

The two images did have me doing the same wondering because of their visual distinctiveness from what is expected: trying out something new on us? Trying out real-looking people instead of airbrushed game renderings, perhaps? Watermark with missing fingers and toes?

It’s definitely not DALL-E 2, at least not without a whole bunch of tweaks or rewriting. Here’s the results of me sending the same prompt as what ChatGPT sent internally, but to the dall-e-2 API:

I would just give lots of “conversation good or bad” or “thumbs down” feedback from the options available to you.

1 Like

Single Sora frames I wonder :thinking:

1 Like

I see what you mean, maybe I should have added an image to show what I get. Here is an example:

So every x prompts I get a result like this with much lower quality images

Could be AB testing I suppose.

(extra words for minimum post)

More likely load related? High-load give a few DALLE2 rather than nothing? Transparency would be good, though. Maybe a message, hey, our servers are stressed at the moment, but here’s your idea in the older DALLE2 model before we’re ready to give you our latest and greatest…

Could’ve just been a hiccup or glitch or even A/B testing of something like you said.

I wonder if it’s based on high usage on a personal $20/month account?

They really need to open Labs back up maybe? Or they likely want it all in the same ecosystem which is smart… We shall see. Interesting times!


I got another of these right off the bat with two images.

I’m going to call it DALL-E-Turbo, because it also is beneath expectations.