Hacker News Reader

Comment by cptroot | Hacker News Reader

For all that this is lauded as a "prototyping tool", it's frustrating to see Genie2 discarding entire portions of the concept art demo. The original images drawn by Max Cant have these beautiful alien creatures. Large ones floating, and small ones being herded(?). Genie2 just ignores these beautiful details entirely:

> That large alien? That's a tree. > That other large alien? It's a bush. > That herd of small creatures? Fugghedaboutit > The lightning storm? I can do one lightning pole. > Those towering baobob/acacia hybrids? Actually only two stories tall.

It feels so insulting to the concept artist to show those two videos off.

View on HN

Yes, and it should be treated as a front-and-center limitation. Generative text models can kinda ape creativity, because the amount of creativity in the training data is so huge. They still are interpolating across text and cannot generalize well, but the interpolation works to most of us because the data is so varied. It's quite easy to write text so if you have a thought you think is original, odds are someone on the internet wrote about it at some point, which makes the model seem quite capable of originality!

But these video game models I think are a lot less capable, because there just aren't that many video games out there, they aren't all that different from one another, and they're all just finite state machines. WASD, desert, jungle, ruins, city. Hell half of them share the very same game engine!

How many massive, cohesive, open world games are there? Red Dead and GTA5... Gee, I wonder why so many of their examples look like that?

That's an odd thing to complain about. Focusing on such a minor issue feels overly critical at this stage, like anything less than a pixel perfect 3D world representation of the source image is unacceptable. Insulting? Come on... Max Cant works at DeepMind so I'm sure he's fine.

> That's an odd thing to complain about. Focusing on such a minor issue feels overly critical at this stage

Welcome to HackerNews.

Yeah, those two demos fell flat for me. The model performing badly on inputs far outside the training data is fine, but those two videos belong in the outtakes section or maybe a limitations section, not next to text lauding the "out-of-distribution generalization capabilities". The videos show the opposite of what's claimed.

Reply on news.ycombinator.com