Support NeoGAF

[Romulus]

AI isn't great at generating complex videos or games out of thin air yet. But give AI a specific job such as a filter, and now it has a baseline of the imagery onscreen and it knows how to improve it.




It's still very new and not ready for use just yet but the potential is exciting.

 

[IbizaPocholo]

potential is exciting

Photoreal Ai filter for existing games

This looks sick. It’s still janky and not at all possible in real-time for now, but it gives an interesting glimpse into the future. Letting the filter handle the image while allowing the game-logic to manage simulation consistency (something LLMs still aren’t great at) seems promising.

www.neogaf.com

 

[PauloRoberto]

It is interestingly not different that those Unreal Remakes we can see at You Tube... so tech but so lifeless at the same time.

 

[Mr Moose]

@ 27 seconds. What the fuck. We've gone too far, we need to go back.

 

[Magic Carpet]

How long does it take to generate these videos? I assume this is not realtime but something churning around in a server somewhere for days before it spits out a result.

 

[ResilientBanana]

Do mario 64 now.

 

[Romulus]

How long does it take to generate these videos? I assume this is not realtime but something churning around in a server somewhere for days before it spits out a result.

20 seconds to a few minutes between scenes is what I was reading. It's very new but increasing with blazing speed.

This is from the creator on reddit:

Still exploring with it - I was testing both gen3 and gen3 alpha vid2vid - these are all gen3 as they looked cleaner and more coherent but that being said even the alpha gens looked pretty amazing + they took like 15/20 seconds to generate!! The alpha makes me really wonder about the realtime aspect in the future - they keep improving the speed and I know Nvidia is working on neural filters like this being almost like the next step of DLSS and DLAA - and their RTX Remix tech. With the speed it’s growing I wouldn’t be shocked if we have playable versions of this next year maybe not perfect latency wise but it’ll just be so next level omfg

 

[Bitstream]

20 seconds to generate... a frame? So maybe real time enhancements in 7 years?

 

[Roxkis_ii]

Rip reshade when this becomes real time.

 

[ResilientBanana]

20 seconds to generate... a frame? So maybe real time enhancements in 7 years?

This will be equivalent to waiting for ray-tracing as we also have to wait for AI to maybe not just do 2D rendering but in real-time in a 3D space.

 

[KingHippo1967]

It looks like complete dogshit. There isn't a single legitimate video game artist who will support this.

 

[Romulus]

20 seconds to generate... a frame? So maybe real time enhancements in 7 years?

Steam diffusion is already doing realistic realtime games at 30fps but they're 2d.

The creator and others say we're about a year away from real time 3d. Which probably means 6-10 months.

But with AI most predictions seem to be off the mark. Remember the Will Smith eating spaghetti? People were saying in a few years we'd have video that is photorealistic. It took a few months.

 

[Romulus]

It looks like complete dogshit. There isn't a single legitimate video game artist who will support this.

Their lack of support means absolutely nothing.

 

[Danjin44]

Why would I EVER want this!? It just looks bad.

 

[SaintALia]

Umm, so at 40s in, is that a completely different scene, or did that dude just change jackets colours on the fly? Actually, I'm not sure anyone was even in the car when he even opened it first time....

And at 46, where are those reflections coming from? It couldn't have been from across the street surely. Where is it pulling that reflection data, or does the OG game have the same problem? The earlier scene seemed to have more accurate reflections too.

I'm also interested in how it's doing those clothing physics.

I guess I should be more impressed, it does look pretty I guess. But what's AI doing about that animation though?

 

[Bernoulli]

What would I EVER want this!? It just looks bad.

It's not designed for this but it will be improved quickly

 

[Tiago Rodrigues]

The entertainment industry will be a totally different place in 10 years and clearly we're still pretending it won't happen. Music, movies, games, etc...

 

[thicc_girls_are_teh_best]

AI isn't great at generating complex videos or games out of thin air yet. But give AI a specific job such as a filter, and now it has a baseline of the imagery onscreen and it knows how to improve it.




It's still very new and not ready for use just yet but the potential is exciting.

"Welcome to Burger Shop muthafuckaaaaaaa"

Have to admit though, this does look really good. Lighting reminds me of '90s crime-drama films where there's lots of natural light but also good use of set lights. Even if the filter is a bit too busy visually, the potential's clearly there.

 

[delfrickintree]

Umm, so at 40s in, is that a completely different scene, or did that dude just change jackets colours on the fly? Actually, I'm not sure anyone was even in the car when he even opened it first time....

And at 46, where are those reflections coming from? It couldn't have been from across the street surely. Where is it pulling that reflection data, or does the OG game have the same problem? The earlier scene seemed to have more accurate reflections too.

I'm also interested in how it's doing those clothing physics.

I guess I should be more impressed, it does look pretty I guess. But what's AI doing about that animation though?

Its called jank that most early technology has. Also something Im a fan of, instead of pursuing realism and fidelity. If AI imagery never evolved into indistinguishable from reality, I would be fine with that.

 

[Fafalada]

Umm, so at 40s in, is that a completely different scene, or did that dude just change jackets colours on the fly? Actually, I'm not sure anyone was even in the car when he even opened it first time....

And at 46, where are those reflections coming from? It couldn't have been from across the street surely. Where is it pulling that reflection data, or does the OG game have the same problem? The earlier scene seemed to have more accurate reflections too.

I'm also interested in how it's doing those clothing physics.

It's done the same way Prompt Gen Videos are doing it - AI hallucinations(this just uses source video as 'control net'). It's not like state of the art Sora videos are much better - most of them have completely wrong light-transport if you pay any attention to it - it's just 'close enough' to look believable.
Like all Gen AI really - it's an adversarial attack on human cognition, it's never trying to be actually accurate.

 

[Romulus]

The jank everyone is seeing it just part of it. But always improving. There was a video from two months ago were it was far worse.

 

[cinnamonandgravy]

video's blocked for me, but been watching this stuff lately.

shit's AMAZING
the amount of complexity and fidelity it adds to a scene is just crazy.
it's like jumping ahead 5 generations graphically.

and because it's an AI model interpreting a game, you can take the same game and throw different AI models at it for different results.
just... change the entire visual make-up of a game at a snap. it's wild.

lord leather jensen envisioned AI generating all frames of a game... this is what he meant.
but were years and years away from this being realtime. like ps8 or something.

 

[peek]

Its fine if you squint your eyes a bit. We got a long way to go, baby still learnin to crawl.

 

[llien]

It's still very new and not ready for use

But give AI a specific job such as a filter, and now it has a baseline of the imagery onscreen and it knows how to improve it.

Isn't it a contradiction?

 

[DeepEnigma]

Those framerates tho

 

[Romulus]

Isn't it a contradiction?

It's not ready for use because its not real time yet, but that doesn't mean its not proficient. It just takes time, and time equals latency.