leSLIe
Fully [H]
- Joined
- Oct 18, 2004
- Messages
- 17,566
Nice!
Follow along with the video below to see how to install our site as a web app on your home screen.
Note: This feature may not be available in some browsers.
The official DLSS 5 will come with less tunables. Will come with more or less stock settings. You can clearly see how much brighter and sharper the left image is. Some would say too sharp.
View attachment 822614
If NV is going to try to foist this on us the least they could do is let up pick what sort of adjustments we want. DLSS5 needs a retro Tomb Raider cheat code modeI find it funny that it makes characters look older across the board. Ah well, hopefully they'll get it right soon enough.
I did a few comparisons in Stellar Blade with DLSS 5 NR.
https://imgico.com/c/qNGS0dW
https://imgico.com/c/nUwGbvw
https://imgico.com/c/wPw5CyX
https://imgico.com/c/40N4UJM
That's actually been one of the things I've been most interested in based on what we've seen so far is it looks like it handles AO and self shadowing much better. In the first shot the two things most noticeable to me, even if not the most different are her hair and her hand. The hair looks MUCH more correct, like the strands are properly shadowing neighbors, and her hand has way more depth, and a much better appearance of being in front of her body. That's the part that has me more interested/excited than faces. I think faces get shown off because the changes can be super dramatic there, which can be a bad thing depending on how it goes. However the improved shadows, that looks great. Not sure about "half the FPS" great but it's a noticeable improvement.DLSS 5 here looks like it alters ambient occlusion for the most part and environment shadow intensity depending on the light source angle. The tint or hue also changes a bit.
Well I'd say most people have been using significantly stronger settings to show a massive difference, I have been using settings subtle settings ~1.02 strength it scales from 0.0-2.0. 2.0 is a dramatic change, but also causes more errors with the final image which leads to the "AI slop" effect.DLSS 5 here looks like it alters ambient occlusion for the most part and environment shadow intensity depending on the light source angle. The tint or hue also changes a bit.
Well I'd say most people have been using significantly stronger settings to show a massive difference, I have been using settings subtle settings ~1.02 strength it scales from 0.0-2.0. 2.0 is a dramatic change, but also causes more errors with the final image which leads to the "AI slop" effect.
Here are some images with slightly stronger strength ~1.05, the selfie one is at 1.15.
https://imgico.com/c/A3SFgaD
https://imgico.com/c/7RqsYda
https://imgico.com/c/MsYc043
Edit: Here is one image at 2.0 Strength (not recommended)
https://imgico.com/c/OeTlRfX
I think the sweet spot is somewhere between 1.0-1.2. Personally I prefer to drop the tone change down to around 0.2 so you don't lose the games original feel.
Some people have a fear/hate of everything A.I. and will try and push false narratives and misrepresent things because their cogntive powers are impaired by bias.Yeah holy crap 2.0 strength looks insanely bad, borderline nightmare fuel. Unlike some people who are just shitposting memes you are providing some useful insight into DLSS5.
Some people have a fear/hate of everything A.I. and will try and push false narratives and misrepresent things because their cogntive powers are impaired by bias.
Eg. posting DLSS5 videos where settings have been twaked beyond reason or dumb emes trying to sell that as a fact.
But if anything the last 48 hours on the internet shows that these people are a minoiry that can be ingored due to them being a crying bunch of dishonest Dunning-Kruger's.
Well some of it is budgets need to come down. When they are sky high it creates issues where even a game that seems like a success isn't because it doesn't make enough money. Like a game that sells 2 million copies and gets good reviews should be a success right? If a couple million people paid full price and enjoyed something that is doing well. However if that is the case entirely depends on how much was spent. Let's say the game sells for $60, and the studio sees $40 of that on ever sale on average (how much they actually see depends on what channel something is sold in). That's $80 million in revenue, of course that is pre-tax and yes big companies DO pay taxes, even if not as much as they should. If the game was developed and marketed on a modest budget, say $10 million, that's a great success. Even after taxes that's plenty of money made. However if the game cost $100 million that's a huge failure, even with a long tail of discount sales it might never make enough to make up the initial spend.I'm just gonna cross my fingers and hope DLSS5 lets Indy studios kinda sorta almost keep up with AAA on graphics so the big budget studios will have to actually make a good game. Games could be way better if the budget stayed the same, the budget for graphics got cut in half, and the other half got spent on gameplay, storyline, etc. Probably wishful thinking because they'll try to just halve the graphics budget and pocket the rest, but maybe a few indy blockbusters will reshape PC gaming.
Optomistic at best but one can hope still. I won't hold my breath for it though considering the way things have been for this long.I'm just gonna cross my fingers and hope DLSS5 lets Indy studios kinda sorta almost keep up with AAA on graphics so the big budget studios will have to actually make a good game. Games could be way better if the budget stayed the same, the budget for graphics got cut in half, and the other half got spent on gameplay, storyline, etc. Probably wishful thinking because they'll try to just halve the graphics budget and pocket the rest, but maybe a few indy blockbusters will reshape PC gaming.
More developer input:
https://x.com/OverJumpRally/status/2094176137266253837
Yup, that "artistic intent" turned out to be a real "foot-in-mouth" syndrome.The KCD2 developer said the same thing. From his testing, the DLSS 5 output was closer to his artistic vision.
I'm not really against it, i think people on the internet just have bad taste. Also, some folks at nvidia when they announced it as well. If its subtle like in this developer's example, its fine. I'm still not convinced I'd ever turn it on but it looks good in the environment. It's a mixed bag for me so far, a lot of the mod examples look terrible. But yeah, if the underlying game looks good, and it's a bit of extra polish as a toggle, go nuts i guess.The KCD2 developer said the same thing. From his testing, the DLSS 5 output was closer to his artistic vision.
Yup, that "artistic intent" turned out to be a real "foot-in-mouth" syndrome.
So on the one hand we have the artist saying "This is what I tried to achieve"
On the other hand we have the "A.I slop" forums users.
I go with the developers, not the muppets.
Unfortunately, I think that's likely the cost, at least at 4k. We'll see but the DLL for it is BIG implying it is a pretty hefty neural matrix that the card has to run. Since it runs on the final image, it is running at output rez (after upscaling), not at render rez. So it hits hard. I'm not saying it is impossible they'll figure out ways to speed it up, or maybe integrate it with the upscaling but it may just be a pretty costly tech.The more concerning thing to me is the ~50% performance loss there seems to be in a lot of examples. Hopefully that's just because it's based on a leaked DLL and not optimized for each game.
They did say in the original videos that it had to run on 2x 5090s, so....yeah.Unfortunately, I think that's likely the cost, at least at 4k. We'll see but the DLL for it is BIG implying it is a pretty hefty neural matrix that the card has to run. Since it runs on the final image, it is running at output rez (after upscaling), not at render rez. So it hits hard. I'm not saying it is impossible they'll figure out ways to speed it up, or maybe integrate it with the upscaling but it may just be a pretty costly tech.
I doubt it will reduce performance by 50% at launch. There is upscaling built into the dll. The problem is the way the current "mod" works is by injecting it through DLSS, so the model is redrawing a full-res frame. In the official release I expect it to redraw the frame at the base resolution then upscale. So the performance would be drastically different.They did say in the original videos that it had to run on 2x 5090s, so....yeah.
I mean, if that's the case then it's sort of DOA, at least for this GPU generation. We already struggle to get high framerates with top-end cards in a lot of AAA games today, removing 50% performance is not something I'm going to do in like 99% of cases.
This, plus devs will choose what it applies to in the scene unlike the full scene redo we're seeing with this shoehorned nonfinal dll. It won't need to do as much work once it's integrated.I doubt it will reduce performance by 50% at launch. There is upscaling built into the dll. The problem is the way the current "mod" works is by injecting it through DLSS, so the model is redrawing a full-res frame. In the official release I expect it to redraw the frame at the base resolution then upscale. So the performance would be drastically different.
Even currently it isn't 50% performance hit on a 5090, its around 35%-40%. I'd assume it will be a 15%-20% performance hit when it releases.
I think the shoehorned dll is quite characther heavy and leaving everything else a bit alone, leaving the environment a bit alone.This, plus devs will choose what it applies to in the scene unlike the full scene redo we're seeing with this shoehorned nonfinal dll. It won't need to do as much work once it's integrated.
I think the shoehorned dll is quite characther heavy and leaving everything else a bit alone, leaving the environment a bit alone.
Chiplet design maybe, instead of 2 distinct gpu.I think we are going to get to the point you need two gps one does the rendering the other the AI stuff.
I run the Witcher 3 at 4K DLSS Quality with EVERYTHING maxed (even 8xAA for HairWorks) at 75-80 FPS on a RTX 4090, so I think you are mistaken...I think we are going to get to the point you need two gps one does the rendering the other the AI stuff.
Wouldn't shock me if we had cards just designed for AI workloads that run in parallel. Just big boy NPUs i guess?I run the Witcher 3 at 4K DLSS Quality with EVERYTHING maxed (even 8xAA for HairWorks) at 75-80 FPS on a RTX 4090, so I think you are mistaken...
For video games, chance that they will put the NPU on the discrete GPU themselve (either like tensor core now or a chiplet), it is some amount of data that would need to go back and forth hundreds of time per second and that how they will do it for console SOC anyway.Wouldn't shock me if we had cards just designed for AI workloads that run in parallel. Just big boy NPUs i guess?
The very limited current version of DLSS5 over it could push that fps under 40...I run the Witcher 3 at 4K DLSS Quality with EVERYTHING maxed (even 8xAA for HairWorks) at 75-80 FPS on a RTX 4090, so I think you are mistaken...