• Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.
  • The Politics forum has been nuked. Please do not bring political discussion to the rest of the site, or you will be removed. Thanks.

After DLSS, what else can we expect from AI/Machine Learning for games in the future?

Elysion

Member
Jan 11, 2020
264
432
275
We already have DLSS, allowing high quality upscaling without the need for super high internal resolutions. And as another recent thread showed, AI generated voices are already in the works. The next thing is probably AI upscaling for textures, meaning a game could ship with low res textures, which would then be drastically improved by AI algorithms while playing the game. I think one of Microsoft’s studios have even said that they’re already experimenting with the idea. Maybe some kind of audio ‚upscaling‘ is possible too? Imagine if low fidelity midi tracks could be ‚upscaled‘ by an AI algorithm to MP3 quality? And what about tesselation? Maybe AI can be used to dynamically increase the polygon count of game assets?

If resolution, textures, assets and audio files can all be generated or improved by AI algorithms, then this could not only reduce the need for super powerful hardware, but could also have the potential to drastically reduce game sizes, possibly by an order of magnitude or more. Imagine if future games could fit on a single DVD again! Who knows, maybe game sizes can be reduced to such a degree that we won‘t even need game installs anymore (at least for physical media). From what I understand, the main reason we need installs is because Blu-Ray drives simply don‘t stream enough data compared to HDDs (and especially SSDs). But if games sizes are back to just a few gigabytes again, then it might be possible to stream the necessary data directly from the disc again, as was the case for all disc-based consoles prior to PS4 and XBox One. And this would be even more true for cartridges! The read speed of the current Switch cartridges for example aren‘t very good, but if game sizes were reduced by an order of magnitude or more, we might see the return of N64 loading times even without a SSD.

I‘m not an expert by any means, but it seems to me Machine Learning might be the perfect antidote to ballooning game sizes, which in turn could reduce the need for large storage, ram and bandwidth. Or am I totally off about this? What else can we expect from ML in the future when it comes to gaming?
 
  • Thoughtful
Reactions: mrMUR_96

Elog

Member
Aug 28, 2016
728
2,964
470
I think two key areas are animations and physics. With enough ML generated data it should be possible to make that really good with less development effort ( cost) or computational requirements. Not sure it can surpass motion capture or exact physics calculations but both areas are very heavy on either development cost or in the case of physics calculation i.e. silicon.
 

Sean Mirrsen

Member
Sep 30, 2020
139
233
250
Togliatti, Russian Federation
Seed-based content reconstruction, maybe. Compress several gigabytes of model mesh, voxel, and material data into a few megabytes of hash, and have AI reconstruct the whole thing to any detail level specified. Would need hilarious amounts of RAM though, I guess.

Then you could do things like predictive physics, have a deep learning system work over examples of rigid and soft-body interactions, and run physics based on what would appear realistic as run through the deep learning algorithm rather than trying to simulate the actual interactions to any degree of actual realism.
 
Jan 7, 2018
1,992
2,269
590
Seed-based content reconstruction, maybe. Compress several gigabytes of model mesh, voxel, and material data into a few megabytes of hash, and have AI reconstruct the whole thing to any detail level specified. Would need hilarious amounts of RAM though, I guess.

Then you could do things like predictive physics, have a deep learning system work over examples of rigid and soft-body interactions, and run physics based on what would appear realistic as run through the deep learning algorithm rather than trying to simulate the actual interactions to any degree of actual realism.
The bottom one sounds interesting. Believable physics is the biggest issue with games right now. [apart from Rockstar and Bethesda open world games]
 
Last edited:
  • Like
Reactions: mrabott

Elysion

Member
Jan 11, 2020
264
432
275
Maybe framerate? Not sure how feasible it is.

Yeah, that‘s something I‘ve been wondering about too. Some kind of ML-based frame interpolation algorithm that predicts how the next frame(s) should look like based on player input, and injects those generated frames in-between the actual frames. I know that many TVs have some kind of interpolation mode to make movement look smoother, but that‘s not suited to gaming since it introduces a huge amount of input lag.
 

TheRobot

Neo Member
Dec 3, 2019
46
17
120
Yeah, that‘s something I‘ve been wondering about too. Some kind of ML-based frame interpolation algorithm that predicts how the next frame(s) should look like based on player input, and injects those generated frames in-between the actual frames. I know that many TVs have some kind of interpolation mode to make movement look smoother, but that‘s not suited to gaming since it introduces a huge amount of input lag.
There are some TVs with a really low input lag with interpolation like the Q90T. (21.6ms)
 

Topher

Gold Member
Sep 25, 2020
1,053
3,060
490
Atlanta, Georgia, USA
large GIF
terminator GIF
 

Kholinar

Member
Nov 9, 2020
242
624
300
Asset creation, world generation, voices, etc. Basically anything that reduces developer workload. Bringing those Rockstar level open worlds from 10 years to possibly 5 years.
 
  • Like
Reactions: martino

Excess

Member
Dec 8, 2020
199
265
245
NPC routines & behaviour is the biggest one for me.
This.

Imagine an RPG where character responses and actions aren't pre-determined, but rather by such complex algorithms as to solicit real-time responses to whatever you say.
 

Sentenza

Member
Dec 3, 2011
14,885
2,352
970
Pretty much anything physics or AI related can be expected to rely on ML/neural networks at some point.

Here's a glimpse into the future:


The Youtube Channel Two Minutes paper is basically entirely dedicated to this topic.

Selected examples:



An AI-driven liquid simulation that takes one tenth of the computational power required by conventional models for comparable results.



An AI-driven system that learned how to animate complex models (humanoids and non) starting by their physical parameters without any direct input.



More physics simulation at a fraction of the computational power compared to traditional systems.



AI-driven simulation of virtual hands.

I could go on for dozens of videos, but I think you may get the point.

What's more impressive is not just the current quality of these systems but the ridiculous rate of exponential improvement they are experiencing every few months, too.
There will be soon a time where the idea to leave some of these tasks to human input will sound laughably unreliable and inefficient.
 
Last edited:
Jul 24, 2006
2,225
32
1,280
Machine learning is the future of videogame.

Play AI Dungeon and dream about a future with npcs capable of chit chatting indefinitely about random shit
 
  • Like
Reactions: CamHostage

Rikkori

Member
May 9, 2020
2,305
4,263
425
This.

Imagine an RPG where character responses and actions aren't pre-determined, but rather by such complex algorithms as to solicit real-time responses to whatever you say.
Yup, think google assistant.


 

hlm666

Member
Feb 25, 2021
92
84
175
Some interesting AI stuff from nvidia gtc on physics done with ai


and some other cool things that have a couple of examples that may help game development some day.
 
  • Like
Reactions: CamHostage