No Face Variability for DALL-E 3? Who is this woman?

Dall-e 3 is incredible in many ways, but I cannot for the life of me get it to generate a realistic looking person. I keep getting the same strange, elfin-faced, very skinny-looking female:

The prompt was “Photo of a natural-looking woman with black hair and bangs, minimal makeup, and genuine facial expressions, capturing the essence of everyday real people. Thin lips. Blue eyes. Crooked nose.”

What is the deal?

1 Like

From what I understand, the model is stuck on a fixed random seed. So it never varies based on your prompt. Also it is stuck in “Unreal Engine” mode for realism due to current Deep Fake :ghost: concerns.

It is Halloween after all :rofl:

The main issue (for me) is that it’s explicitly ignoring the prompt, as well.

The woman definitely doesn’t have thin lips or a crooked nose. So, it’s not just the seed. This girl doesn’t even look human.

It’s almost like a LORA from Stable Diffusion with a fixed character.

1 Like

That might be a new one. Hopefully those with Dalle-3 access and experience will chime in here.

1 Like

“Photo of a woman with red hair” gets us the same woman…but with red hair. Very odd.

I’ve got about 12 pictures of the same Asian “red hair European” from Bing days separated, because “redhead” or ethnicity doesn’t fly.

There’s certain inputs that always go the same way now, unlike two weeks ago when an ambiguous prompt would give you a vast variety.

I had high hopes of demo-like apocalypse scenes for “When cats invade!”… but

Simply “She’s got too many pets” - nothing about they should all be similar cartoons:

To see if they are hashing the prompt to make a seed or similar, I changed it only by “way too many pets”

or add “Digital photo: She’s got too many pets” - all inferred incorrectly the same way:

So I wonder if in ChatGPT Plus where you payers are, you start a new conversation (so it doesn’t have the same ability to keep building variations by image name on chat history of what you produced before) if you can get a different direction with additional descriptions of a different scene or style description, framing, etc. Not just for the new setting, but just to stir up the prompt.

Perhaps this new tuning is to address those that couldn’t get AI to create any consistent style for book illustrations, for example.

(sorry if I don’t have any great ideas of what to make an AI produce…)


Interesting. I can say that this same “elf girl” has appeared in every DALL-E 3 chat that I’ve created. Here she is again in a whole new chat:

Part of my issue is that she’s so specifically and stereotypically “beautiful” that I’m afraid it’s showing a real bias in the image creation.

Um, actually this problem may be deeper than I thought…

Apparently ‘obese’ is male-coded?

Looks like my ex-wife.

I keep getting the cousin of your photo. It seems like no matter how I change the prompt I get a pretty face with eyes and mouth that too big so it’s not human looking.