Runway Gen-3 AI Videos Are Kinda Blowing My Mind!

Update: Gen-3 is now publicly available. I mentioned a couple weeks... It was only a day! :) Runway released a prompting guide to help you get better outputs here: help.runwayml.com/hc/en-us/articles/30586818553107…

8:21 He was reloading the baguette with cheese.

By the time Sora finally becomes available we’ll probably already have an open source version or free one that’s just as good

While it's really impressive how far we've come with AI video, I'd really like to see how specific we can get with it. Your prompts were kinda generic. What happens if I want a character to wear very specific clothing? Colours, material, fit? Can it do a coat with mother-of-pearl buttons, or do we get any old buttons? Image Gen looks impressive as long as you do generic portraits of people, but breaks if you do anything complicated and out of the ordinary. Try prompting for a handstand or someone dangling from a tree branch heads down. To be useful for serious production work and not being just a toy, you need to be able to have a granular level of control. I know I'm asking for a lot, since we're only at the beginning stages yet, but it'd be nice if you really stress-tested it, so we know where we're standing at the moment. Still, great work, Matt! 🙂

I think it will be better when the image to video option becomes available.

The band playing music beneath the ocean was pretty cool! Even the floating microphone looked kinda real lol!

Low quality clipart video for people who really don't need to have anything particularly good. It'll do because it doesn't really matter because no ones pay much attention to it anyway. This will be great for those adverts at the side of blog posts that everyones trying hard not to accidentally click on so they only focus on the thing they googled in the first place. Marketeers will love it. As long as it moves, it'll do.

next time you see a demo, expect that they generated 100s of videos to make 1 video demo!

That’s cool, but I wish we had more control over the videos created. The videos are starting to look better, but the control is still lacking, which makes these videos not that useful for filmmaking projects. Also, we need ‘Image to Video’ if we want this tool to work with consistent characters, and to create scenes for films. I want to be the director of the film, rather than letting the AI have all of the control over how a scene looks.

comparing this to year-ago generations, it's a big improvement. Next year vid generations should be even better

These funky artefacts heavily remind me of image gen around DALL-E 2, Midjourney 4 and SD 1.5. As we've seen those ironed out and improved in later versions, now we're seeing better hand generation and legible text in the newer versions of those image gens - can we expect similar, along with improvements in temporal consistency, in these video gen models? Pretty excited for the next year, if so. That baguette video blew me away a little bit.

Gen 3 said ‘not available in your country’. LTX was available today and it was like an interface for the usual crappy vid generators. A waste of time and early signup attention. Great vid as always 👌

17:58 the girl on the water is practically perfect, very difficult to see any problems. Impressive!

The color palette it picks is pretty good. Also I like the contrast. I'd say in about another year we'll be looking at some pretty fantastic looking AI videos from all companies. Thank you Matt for putting in the time to show us all!

I would like a video on why the AI videos do what they do and what the blockers are in making them more realistic - I mean can the output be put back through an AI that corrects all the issues?

11:02 Imagine if the man & woman hugging here were ‘consistent characters’. I hope it will become easier to create scenes with ‘multiple consistent characters’.

Just tried this and it's amazing for music videos. When making a real film you're also creating a lot of takes to cherry pick the best one, so I'm okay rerolling until it looks ok

So many of these videos look like they’re beyond CGI, they’re really captivating!

Huh, another AI company showing off their product and preventing the general public from trying it out.

I just started using Runway Gen-3 a week ago and so far I've made a zombie video and now a dancing cats video. I think both turned out great. Yes image to video will make Gen-3 easier to use but if you know enough about cinematography and camera shots, and you're not too picky about the details, you can do pretty well with just prompts alone in Gen-3.