3 minutes ago, strider1 said:

The shimmering scenery is a buzz kill.

They have an internal bet running how many  times they can get you to post the same complaint, someone has a lot of beer running on "at least 500 times" and that guy is holding back the fix for the shimmering scenery.

13 hours ago, SayAgain said:

have different preferences

The funny thing is that you can't tell the difference between those press videos and a brute-force ray-traced version of the same scene. Folks certainly claim they can but it's an emotional and biased lie. There's currently no reason to suspect that NN will go away for these rendering tasks. But don't worry. You'll still be able to turn it off and get your 10 fps.

GPUs are suffering the same scalability problems as CPUs. Maybe on a slight delay but it's not like silicon knows if it's doing an addition or subtraction op to know that it should speed up or slow down or send an electron across the rails to another transistor.

 

12 hours ago, Litjan said:

"thin air"

Indeed. It's called either extrapolation or a poor training data. Either will lead to wild results.

3d gfx is actually the perfect place to NOT see that. The game devs and artists make a complete game/app and run it at 1fps or whatever it takes to do a fully path-traced frame, running through every possible scene and interaction, dumping training data as it goes and voila: you end up with a data set that that won't be prone to "thin air" results. 

 

12 hours ago, Litjan said:

But would it not be better to have real data about the walls and windows? Yes, it would.

With a properly trained data set you won't be able to tell the difference. The thing that you're missing here is that dlss5 still requires that hard geometry info. This particular algorithm isn't creating imagery out of thin air like your favorite GPT image gen search. It's based on the game devs' and artists' designs and imagination. It relies 100% on their real data. The only thing changing here is the lighting and in the case of organic surfaces, subsurface scattering.

3 hours ago, Litjan said:

They have an internal bet running how many  times they can get you to post the same complaint, someone has a lot of beer running on "at least 500 times" and that guy is holding back the fix for the shimmering scenery.

LOL....

400 wins the jackpot!

I envision the dya they'll announce AI-based flight dynamics and weather modelling too 🤣

Just now, jcomm said:

I envision the dya they'll announce AI-based flight dynamics and weather modelling too

and flying, and posting in these forums. then we can finally turn to our non-AI hobbies. and do all the things again we did before Flight Simulator version 1 was ever released. I've lost a few girl friends along the way here and there. priorities. first things first.🤣

1 hour ago, blingthinger said:

This particular algorithm isn't creating imagery out of thin air like your favorite GPT image gen search. It's based on the game devs' and artists' designs and imagination. It relies 100% on their real data. 

So you are imagining the artists for each project will generate enough data to train a specific model for that project ?

11 minutes ago, Matchstick said:

So you are imagining the artists for each project will generate enough data to train a specific model for that project ?

Exactly, if the dev is required to do that extra work then they may as well do the art and bake it vs. hope probability is on there side.

1 hour ago, blingthinger said:

you can't tell the difference between those press videos and a brute-force ray-traced version of the same scene.

I sure can as can many others.  Never done any game/sim at 10 FPS?  30-60 FPS current 40 FPS with MSFS 2024 and 60 FPS with XP12 (both maxed out with add-ons).

But hey, if you think you can react to visual stimulus higher than 75 FPS ... go for it ... just for the record, fastest ever recorded reaction time is 102 ms.  By all means chase JH probability dream as I'm sure he's more than happy to sell you on it.

30 minutes ago, Matchstick said:

So you are imagining the artists for each project will generate enough data to train a specific model for that project ?

Oh heavens no. They'd build the app as they normally would and then run an automated script that moves the camera and characters/objects in thousands (if not 10s or 100s of thousands) of different lighting scenarios and positions, dumping the needed data along the way. Then you train the NN based on that monster data set. In this case the training would focus on generating the resulting lighting distribution for a given frame. Shadows, specular highlights, reflections from off-screen geometry, etc. That latter item was noted by one of the few neutral/positive reviews of dlss5, which is either too costly for full path tracing or simply impossible with pure screen space methods.

You could hypothetically reduce that data set size to some degree with what they're calling PINNs: physics informed neural networks, which might include the energy-conserving refraction, reflection, and light propagation equations in the error-calculation steps during training. That could actually significantly reduce the data set. Maybe still a more experimental method though.

What I'm describing yields essentially a per-frame lighting bake. Except that it's dynamic in its ability to respond to scene changes. And with the amount of data you would generate to train the NN, it's no longer a 'hope'. It's never needing to extrapolate.

14 minutes ago, SayAgain said:

I sure can as can many others

Oh did they also release full path traced samples of those scenes too?

14 minutes ago, SayAgain said:

react to visual stimulus higher than 75 FPS

React? No, I'm talking simply distinguish. The higher res and bigger screens do become more obvious. VR does like a good 90 for the vast majority of the population to not get sick. 30 fps was is for 800x600 CRTs.

1 hour ago, jcomm said:

weather modelling

Do you folks use Deepmind at all? Or is it being looked at?

https://www.noaa.gov/news-release/noaa-deploys-new-generation-of-ai-driven-global-weather-models

Oops, either JH doesn’t know what he’s talking about or the nVidia engineer doesn’t know what he’s talking about.

They are making a blooper here by not addressing the Resident Evil character. Her geometry is different. All the background remains the same. I'm thinking they did rework her underlying 3d. There's another Starfield character in the same situation. But otherwise everyone else is identical. JH isn't wrong, nor is the engineer. The YT'er is wrong. If you factor in the training data which does include underlying 3d, they're both correct.

As for the sideburns? Most hair is shader-based. Not 3d geometry based. It's easy to see where the model would be left to its own measures there. What's more is that these "revealing" YT vids only ever mention the same 2-3 examples. If every frame was equally different, there'd be a story here. "Angry" YT slop otherwise.

1 hour ago, SayAgain said:

either JH doesn’t know what he’s talking about or the nVidia engineer doesn’t know what he’s talking about.

can we ever trust our approach charts and navigation instruments again with DLSS5? when all is fake and nothing but AI slop ....

NVidia has killed MSFS.

scandalous!

1 hour ago, blingthinger said:

Deepmind

Hmmm ... I know not directed at me but NOAA and AI (Google DeepMind's GraphCast model) is going to take you down a path of thread lock or deletion here on AVSIM ... the "heat dome" producing 97F temps (+28F above average for "winter") over Los Angeles these past 3 days was sufficient data to validate C C models (word not to be mention).  Climate and weather prediction AI is NOT the same AI as that used for graphics rendering (thank goodness).