Fast frame-consistent video models as "world" models

krea.ai

8 points·dvrp··9 comments

The biggest problem I see in an AI world is tools lacking the precision for craft (e.g the gap in product you get between lovable and a design engineer making something).

Codegen largely solves this problem since you get (code) as the output, but changing video has always been harder. Cool stuff

thank you! i still think there’s a long way to go in both images and videos, but i agree. i think we need more Cursor-es que products in the sense that AI and the craft part go hand in hand, or at least they feel that way to me

Cool! Can you share any details about how you did it? Is the model architecture similar to Genie 3 at all?

it’s an auto-regressive model, but where did you find details about Genie 3 model architecture?

I can't find info about the Genie architecture online -- are there blog posts that detail how these work? I'm not familiar with this space but am curious how we're getting consistency here.

oh nice! i still remember when Krea announced latent consistency models (https://x.com/krea_ai/status/1723067313392320607). we’ve come a long way

thanks! we do have come a long way and i want to think there’s still a long way to go

i also posted our updates on HN back in the day and i remember it got to the top: https://news.ycombinator.com/item?id=38223822

to me its really interesting how Stability started from model and went up the stack, Krea started from the app layer and is now going down. retroactively obvious what was the better way.

hahaha, ironically enough, Emad from Stability just commented on our twitter post!

https://x.com/EMostaque/status/1961083407452000726

Reply on news.ycombinator.com