Hacker News new | ask | show | jobs
by Datenstrom 1357 days ago
The

> A golden retriever eating ice cream on a beautiful tropical beach at sunset, high resolution

example is terrifying.

4 comments

In case anyone else had problems finding it - it's at the bottom of the page:

https://makeavideo.studio/assets/a_golden_retriever_eating_i... (webp)

That grasp though.

These things still feel a bit like e.g. Google/GCP services to me: Super appealing at first glance, quite close to what you want, but somehow never quite there. Maybe they'll asymptotically get there, eventually? Perhaps that statistical model can't really make it to the level we want it to?

I’ve found that replacing the bad parts with new ones, like Dalle Outpainting, can remove the worst parts of the image, like the hands here… doesn’t make it perfect, but certainly removes the worst offenders that instantly bring attention to themselves.
It may be that it's the deep learning tech which will never quite get there. GPT-3 has similar shortcomings in its mimicry. We're 95% there, I guess, but may never quite reach 100%.
Nah, the current issues are just because we're trying to do everything in one step. Because we've built tools that have so much of a stimulus-response approach, few efforts have been made toward interfaces that ask for clarification ('when you say X, do you mean XYZ or XXX?').

Image-to-image and tuning already addresses many of these issues; just as inpainting works really well, it won't be long before we have select-and-repair, where you add an additional prompt like 'improve this part - the ice cream is fine, just work on the dog's muzzle.'

The mistakes the AI makes are too numerous and hard-to-define for this to work I think. They could perhaps be addressed by having two different models trained differently, each fixing the errors of the other. When humans draw a realistic artwork, it's not 'single-pass'; they have to iterate on the details to get it right.
> the deep learning tech which will never quite get there

Never say never, we've come a long way since GPT-2! All this was unthinkable back then

It's certainly possible. I find it somewhat unlikely though, despite the fast-paced progress.
I get the same feeling as well. This approach may well be eternal demo-ware, and you'll actually need AGI (or manual direction by a real human) to get to 100%.
On the home page of HN at the moment is something like GPT but much better. It's at character.ai
It is, but to make up for it we got a man drifting a horse.[1]

[1] https://makeavideo.studio/assets/A_knight_riding_on_a_horse_...

Just needs to show the part where he ends the drift and gets a speed boost
Yeah, put some blue sparks on those hooves and you've got a meme
The hands throw me off. The same with the cat holding the remote... never thought that hands on animals would be able to trigger my uncanny valley response, but here we are
Eugh, thanks for warning me.
They forgot ‘trending on onlyfans’

But actually, this technology is super exciting. Imagine a future where movies and games are choose your own adventure.

We'll have procedural generation that will be hard to distinguish from human-made content. Goodbye repetitive Skyrim filler caves!
I've never played Skyrim - was that the immersive 3d version of a maze of twisty little passages?

Besides just textural content, it's intriguing to consider the possibilities of full-3d roguelikes.

if people weren’t so repressed, this could also be used to severely reduce exploitation in the porn industry. what’s the point in making and selling exploitative porn when it can be auto-generated at will?