What's up with photoshop-like generations?

[yellow0]

Hey,

has anybody else noticed that sometimes DALL-E3 outputs images that look like a bunch of crop-outs photoshopped together? If it happens, it’s usually a pair of images, instead of one, and the quality is much lower. Is this some under-the-hood process accidentally revealed on some prompts? Asking it to regenerate fixes it, but it’s still weird.

Example:

Abandoned island, sandy beach, palm trees. An orange overweight cat notices his phone on the beach nearby. Vibrant colors, realistic.

More examples here (add the dot) → imgur com/a/VtLq7j2

[_j]

Yes, they’ve been doing this kind of trial on a new DALL-E “3” for a while now, which I first received in March:

---

It took a prompt like this:

and gave an image like this:

or this:

---

or worse to others:

A variant of the “looks like Photoshop” imager now seems to be used when you specify “natural” style using the API version.

---

Another note: there is seemingly more than one test “DALL-E” in play in ChatGPT besides the one offering two images.

Here is reusing the same gen_id and prompt, normally outputting identical images. In this test today to show white artifacts around glasses, notice the first generation is significantly different, the second with different artifact, before the rest repeat. I’ve also seen 1 different artifact-free of good quality, then 3 repeating with the artifact.

(it also could be that GPT-4o is so poor in GPTs that it can’t emit the same prompted gen_id sequence correctly, while still reporting that it did)