Hacker News Reader

Comment by Closi | Hacker News Reader

The same wasn't true for images - it has been easier (so far) to directly compute the images than it has been to train them to use photoshop or illustrator.

Similarly, neural networks have already been shown to be able to directly simulate physics simulations (smoke, particles) with suprising accuracy.

I agree that logically your statement makes sense, however I'm not sure it matches what has been successful in the space so far. Tools may make sense for these NN's to use, however they probably won't look anything like the tools we use today.

View on HN

>> it has been easier (so far) to directly compute the images than it has been to train them to use photoshop or illustrator.

possible because training models are fed by an abundance of existing human-generated content, which is ever increasing in quality, dirt-cheap, and easily accessible; which doesn't make AI Generated Content disruptive in this respect (my view).

not sure what you mean by "easier", since computers are better at following rules and people better at reasoning.

>> I agree that logically your statement makes sense ...

Isn't that the best case for AI generated content, logical-sense? or are you suggesting more (e.g 'source of meaning')?

> Isn't that the best case for AI generated content, logical-sense? or are you suggesting more (e.g 'source of meaning')?

I just mean that the approach sounds logical, although the best approach isn’t always the one that appears the most logical. This appears particularly true in AI - for instance diffusion models feel more like an incredibly surprising discovery rather than an approach that’s immediately obvious by applying logic.

> Not sure what you mean by "easier", since computers are better at following rules and people better at reasoning.

By easier I just mean we have had better results so far. Maybe easier wasn't the right word. On your second point - computers will be better at most reasoning tasks shortly.

Well, all the problems obviously caused by the lack of an underlining physical model (like the hands thing) disagree. Your comment just says that up to now it has been easier to ignore those problems than to fix them.

But video has much stricter physical constraints than image, so it's not clear we can ignore the problem at all.

Not sure if that had been ineffective or other approaches than current one didn't align with user goals. The premise of diffusion-based image generators was that the user only has to verbally describe an image, requiring absolutely no skill, while the trend since their inception seems moving away from that fast.

The future is stacking models in feedback loops. One model could detect bad anatomy and tell the other models to rerender the frame. AutoGTP is an example of this

I've definitely been seeing things trending in this direction.

Reply on news.ycombinator.com