Hacker News Reader

Comment by sdwr | Hacker News Reader

I think you would see a quality difference between 'a monkey mascot' at $40 and $200. The $200 designer is shining a light on his personal brand.

The mascot is friendly, vaguely memorable, well-proportioned, soft, and not attention-seeking. Its expression tells a story, adds humanity, and creates unresolved tension.

The AI ones are sharp and confident and eye-catching, zero subtlety, completely missing the point. I'm willing to bet a $40 designer would drop the ball in a different, equally bad way (probably make it too corporate, or miss the precise "cute but low-effort" spot the original designer hit).

View on HN

It's funny because as I read the article, one thought was "the slop results look like they came from very old models"

Use a model released this week and the results are (to my untrained eye) no longer distinguishable from a human artist: https://imgur.com/a/tgEsXq8

And it wasn't some pro image prompting magic. Even compared to 12 months ago, the text encoders have very good. I was able to use wording that came to me naturally and get an aesthetically pleasing result in seconds.

You mentioned a model released this week. Is it Lumina new model by any chance?

Imagen 3 002

The reason AI slop is slop isn't because models are not advanced enough. The slop factor is inherent in the way AI works and cannot be fixed by brute force compute power.

AI is likelihood optimization under the hood. It draws the most statistically likely image. The human brain and eye is very good at picking out "average" pictures. Turns out our in-built AI detection capabilities are very, very good. (This might actually kill the AI industry very quickly. I imagine in a couple years AI-generated pictures or texts will be the lamest thing ever, and AI companies will lose a shitload of money in this arms race.)

a) I linked to an entire album of pictures that 99% of the population would not identify as AI, including the population that knows what AI slop is.

b) This is a gross oversimplification to the point of being unhelpful and pretty much wrong. You don't seem very familar with how these models work.

There is no inherent reason why sampling from the latent space of a model limits us to average of any concept.

Not to mention the models are learning what average means and what exceptional means and can increasingly produce both at will. As they get larger the degrees of seperation between those "sub-concepts" of each concept grow larger and larger.

The reality is humans are good at convincing themselves they're good at things. The false positive and false negative rates are already going up for AI art, and it's only going to accelerate from here on out.

> that 99% of the population would not identify as AI

Absolutely not true about the "99%". Give it a couple years and not being able to spot AI slop will permanently mark you as an "okay boomer" tech illiterate slob.

AI == lame, square, uncool.

We already see this with the AI assistants they keep putting into OS updates. People hate them, and not for "privacy" reasons. They just don't want to be lame.

Again, can you point out to even a single giveaway for that album?

I'm honestly convinced you plain didn't realize the entire Imgur album was AI generated and that's why you haven't been able to address it.

AI is lame, kids hate that stuff. Positioning yourself as some sort of "AI expert connoisseur" will only make you look lamer.

You had 3 days and that was the best you could do?

This comment can't be further from reality. If you show this (https://imgur.com/a/tgEsXq8) to people, their "very, very good built-in AI detection" won't bat an eye.

> This might actually kill the AI industry very quickly

Yes it will, but not in the way you implied. In a few years AI assisted image creation will be so common that no one will bother mentioning "AI" anymore, effectively "kill" the AI hype. Just like you can't sell built-in webcam as a feature of a laptop: every laptop has it.

Reply on news.ycombinator.com