You can’t actually believe DLSS is only manipulating light, right? The example comparisons clearly show more than light tuning.
I just wish they would have called it something other than DLSS. There’s no super sampling going on here.
NVIDIA's new real-time neural rendering with DLSS 5 - fxguide .
“Equally important is that the system remains controllable. Developers are not presented with a binary choice. Instead, DLSS 5 provides a set of parameters that allow the effect to be adjusted spatially and artistically. Different materials or regions of a frame can be treated independently, and colour characteristics can be tuned to match a desired look. This keeps the system aligned with established workflows, rather than replacing them.
While NVIDIA positioned DLSS 5 primarily for characters in games, its broader relevance is clear. It works on the whole scene. In virtual production, the possibility of achieving more final-pixel quality lighting in real time could reduce the reliance on heavier offline re-rendering. For interactive digital humans or installations, training environments, or live performances, – the ability to deliver more convincing material and lighting response at interactive frame rates is significant.”
I am pretty sure that dlss5 will be trained per game and therefore the choice of the character replacement and style will be fine tuned per game. I don’t think this is a complete generic ai stuff applied just in time. That’s the only way to make it runnable. Probably for the non-tuned game something generic will be applied and here the result will be less spectacular. Anyway, for games where the character face is not such important like msfs, I think we will get much more realistic images than the models currently provided out of the box by Asobo. They should do the same to fix photogrammetry building to be honest…. Here it will be a great improvement!
Reshade can do that. 
How about real looking clouds???
One can dream…
PACO572
Hardware – Software - Hardware - Software ie those of us of a certain age have been done this road a few times.
As hardware is getting more expensive due to scarcity, if software can improve our flight sim experience via machine learning AI, then I’m all eyes are ears to see what the next gen brings.
The mock-up by SimSoaringClub above is a great example of what proper neural rendering can do. Imagine the MSFS engine providing inputs of decent quality, that are not too heavy to stream and process locally, and then neural shaders outputting an extremely convincing, if not almost perfectly accurate, rendering of the world and atmosphere at a much lower overall cost.
As Seb explained in the past, an accurate ray-marching technique for the clouds would be prohibitively expensive for the GPU, and scaling such techniques will quickly run into diminishing returns as GPUs are terrible at ray-tracing by their own nature. We are slowly moving into ML approximation of lighting (neural radiance caching) and complex materials (neural texture compression), so only neural rendering with a tight integration to the game engine can provide the leap we are looking for.
But DLSS 5 is not it. Despite using inputs like motion vectors from the game engine, what it is doing is modifying the engine’s output image, just like Instagram beauty/artistic filters. The results on human faces in particular is horrendous, and I’m baffled that Nvidia approved of such footage for showcase. Look forward to neural shaders integrated via DirectX Cooperative Vectors and Linear Algebra, not this vendor-locked, post-processing, black-boxed garbage.
Software running where? The software they are describing works because it has massive data that it can work from, the results of its “research” so what good is that code execution if it’s happening at some data center somewhere? The input and the output still have to get to your device to complete the rendering which is streaming. We’ve seen with 2024 just how restrictive streaming can be if there are any blips in the data path. I get worse flight results here in Europe if I fly during hours the US is also online and all that’s streaming is basic data and graphical points not the hairs on someone’s head or the twinkle in their eyes. Our PC’s or Xboxes don’t have the power to run neural networks locally at the same time as our rendering engines are producing what we have now unless we spend massive amounts of money essentially building a data center in our living room.
That is gospel truth.
nah, I don’t want this in the sim. Disregard that the demo had to be run on 2x5090’s, it just makes everything look fake, removes detail, and hallucinates in other detail that wasn’t intended by the developers of the game. Main thing I noticed in the showcase is that the entire atmosphere of many of the scenes was changed.
NvidAI DLSuperSlop
Once models have been trained, the data is no longer necessary. If you’re using DLSS upscaling or frame generation, you’re already running a neural network that has been trained on an exorbitant amount data on Nvidia’s supercomputers, in parallel with 3D rendering on your GPU.
Thanks for this clarification. This is how I understood their YT on DLSS 5. The beauty of this solution from a game developer perspective is that NVIDIA does the heavy lifting and then tools become available to the devs and users as each game is made DLSS 5 ready. I take this as evidence that NVIDIA (and probably AMD for their own new solutions) get that us end users cannot keep trading up GPUs to get better game graphics and performance, and that AI ML solutions running on whatever GPUs gamers actually own (meaning most of us not just power users) must be developed. Some will take this as slop and others will use whatever comes along. Wait and see, all matters of taste in the end?
Precisely. It would be extremely difficult to make ray-tracing performant without fundamentally changing how GPUs work, but since they’re good at GEMM, it makes a lot more sense to push towards ML approximations of ray-tracing instead.
DLSS 5 can be ignored though. Besides the matter of taste, being locked to a single vendor means it cannot be used to decide the final look of a game. It feels like a rushed announcement to counter Project Helix to me.
Putting aside matters of gamer taste, we are all now marginalized by the market demands for data center growth and development. The gaming solutions that follow are all likely to be rushed and partial, whatever compromises that can be cooked up on the sidelines of AGI press and other follies. Where we live, electric bills have gone up almost $100 a month year over year due to nothing more than competing with data centers for service. This needs a regulatory solution but I bring it up because it is not just us as gamers who have been marginalized by the big 7 AI enterprises. OTOH anything that works a little better in the sim on the same GPU I already have will be welcome LOL.
Galileo was a skeptic. You’re in good company.
It’s pretty bleak for PC gaming. My COVID build is trucking along and can still run everything just fine, but it’s 5 years old already. I lost a stick of RAM due to failure two days after prices skyrocketed a few months ago. It’s only a matter of time… At least I still have 32GB of RAM installed and a spare stick from the pair I took out. If my 3080 throws in the towel, I’m not sure what I’ll do.
I hope things stabilize at some point. It’s getting harder to enjoy hobbies in general because of cost and accessibility.
This time I think I’ll skip dlss5 and wait for dlss6, I always do this with new technologies
A future Sim Update ought to bring in support for this!