Meta's AI image generator can't imagine an Asian man with a white woman | Hacker News Reader

Meta's AI image generator can't imagine an Asian man with a white woman

theverge.com

33 pointsSLHamlet27 comments

This feels a little misleading of a title. It couldn't do the reverse, either, as shown by one of the prompts. Nor could it show just friends of those races.

More apt to say it's unable to generate images of white people and asian people together.

The inability to reliably separate features for specific subjects in multisubject generations is present in most AI image generators; there’s tools outside of the main model to help with this in many Stable Diffusion toolchains, and DALL-E 3 via ChatGPT uses an LLM-involved versions of similar tooling under the hold, AIUI, but its otherwise a pervasive problem that has nothing to do with racial bias.

"japanese man with white female friend" -> Success

"japanese man with white female wife" -> Fail, the wife is always Asian

"japanese woman with white male friend" -> Success

"japanese woman with white male husband" -> Success

A little? This is purposefully designed to generate rage.

It's not censorship. Generating two different people is a hard problem for Stable Diffusion type systems. I was playing with this for a while. If you ask for any scene that involves two people touching or some limbs occluded, results are terrible. A common bug is one arm from each person joined to one arm of the other. Background people tend to closely resemble the foreground subject.

There are uncensored models out there. This one[1] is a cut-down model for speed, and it's free. What you really want is the "I can't believe it's not photography" model, without censorship. There was a site which would generate images from it for free, but it overloaded and they had to cut back. They have the same problem.

[1] https://stable-diffusion.site/

[2] https://civitai.com/models/28059/icbinp-i-cant-believe-its-n...

> It's not censorship. Generating two different people is a hard problem for Stable Diffusion type systems.

Does the technical mechanism matter for the social impact? The net result is the same: A certain category of ideas becomes comparatively harder to express.

Exactly why governance of AI model development is so critical going forward. If the nations don't step in now, they will quickly lose the propaganda machine to Silicon Valley.

AI Image generators fail spectacularly at many kinds of prompts. Is this really remarkable ?.

Its not remarkable, but it is great click bait :)

> I tried dozens of times to create an image using prompts like “Asian man and Caucasian friend,” “Asian man and white wife,” and “Asian woman and Caucasian husband.”

For the last one maybe: “Chan and Zuckerberg”

But can they imagine a black woman with an asian man?

I tried to help you out but I got,

> Not available in your location

After observing who dates who for decades, I have concluded that seeing a black woman in a relationship with an asian man is far more unlikely than a white woman with an asian man.

It is so unlikely that people don’t even think of it to bring the example of AI bias, because people would just say the AI reflects what the statistics it actually sees.

Asian man + white woman one of the least common pairings in the U.S. (0.4%) [1]

How does Asian man and black woman fare? (0.3%)

[1] https://en.wikipedia.org/wiki/Interracial_marriage_in_the_Un...

This is just the same "horse riding an astronaut" problem. You have to mask the image and extend.

I'm locked out of meta, could someone try with specific locations and see if that changes things? I am curious whether this is a race-word prompt override or whether it's some kind of model fine-tuning

e.g. "Norwegian woman with her husband from Hong Kong"

I get two derpy looking Asian people: https://i.imgur.com/957rCN7.jpeg

Regardless, that isn't a great prompt for testing this out, because "Norwegian woman" could be interpreted as "a woman with Norwegian citizenship", which could be a racially Asian woman. And "her husband from hong kong" could be a white guy born and raised in hong kong.

Sure, but I would expect the model to model them as average-looking from those locations.

I did one and then came to read the comments to see if anyone had the same idea, our idea is the same but different: https://s.h4x.club/WnuXo0WN

[dupe]

More discussion: https://news.ycombinator.com/item?id=39930072

This is hilarious, and the implied weightings seem representative of Bay Area reality.

In the US, "Asian" is understood in common speech to mean "East Asian," as opposed to "Indian" or "South Asian." That particular objection feels disingenuous. The article was strong enough without it.

So if people use a word incorrectly or colloquially, AI should too?

As an example, would you be OK if you searched for a list of Asian countries on Google or ChatGPT or Facebook's AI, and got only a list of East Asian countries?

BTW, from what I know, in the UK Asian is commonly referred to people from the Indian Sub-Continent.

It's an American reporter, writing for an American publication, probably connecting from a US IP address. Why would she expect UK English?

And to answer your question, yes, I'd prefer and expect the AI (just like any human) to use colloquial English rather than some pedantic but technically correct vocabulary.

If you're in the UK or India and had the same experience, I'd understand your confusion. You should write an article!

But if you prefer the pedantic train, "subcontinent" is one word, a common noun, neither hyphenated nor capitalized.

And another author who doesn't understand how generative models work

> Asians who don’t fit into the monolith are essentially erased from the cultural consciousness, and even those who do are underrepresented in mainstream media.

This is giving way too much relevance to Meta’s ability to put together AI models. LOL.

Why would a generative model have a hard time producing an image of an asian man and a white woman?

Because it is doing two tasks, with different goals, using an optimization process that has a single goal derived from the prompt.

Reply on news.ycombinator.com