• Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
    Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.

591.74 New DLSS 4.5

I'm still using this set

NVIDIA DLSS 4.5 Dynamic Multi Frame Generation Available from March 31

by btarunr Today, 09:30 Discuss (0 Comments)
NVIDIA at its GDC 2026 GeForce ON community update, announced that its DLSS 4.5 Dynamic Multi Frame Generation feature, first unveiled in January 2026, will be generally available from March 31. The feature lets GeForce RTX 50-series "Blackwell" gaming GPUs dynamically adjust the rate of DLSS Multi Frame Generation (MFG) rate between 0 (no frame-gen), and 6x, depending on a target frame-rate determined by the display's capability. The feature is meant to be enabled through the NVIDIA App, specific to supported games. NVIDIA had earlier announced 6x MFG for RTX 50-series GPUs, going up from 4x MFG that the company launched alongside the GPU family in early 2025.

Multi Frame Generation allows the GPU to draw up to five frames succeeding every conventionally rendered frame completely using AI, effectively multiplying framerates. RTX 50-series which debuted alongside DLSS 4, could originally do up to 4x MFG, but NVIDIA increased this to 6x MFG with its 2025 DLSS 4.5 update that introduced a new 2nd generation Transformer AI model that's more accurate, and can more accurately predict motion vectors allowing NVIDIA to increase MFG factor to six. Frame generation comes with latency costs, and so Dynamic Multi Frame Generation allows the driver to opportunistically lower MFG factor to improve whole-system latencies. At GDC, NVIDIA announced that DLSS 4.5 is coming to at least 20 of this year's most hotly anticipated AAA game titles.“
 

NVIDIA DLSS 4.5 Dynamic Multi Frame Generation Available from March 31

by btarunr Today, 09:30 Discuss (0 Comments)
NVIDIA at its GDC 2026 GeForce ON community update, announced that its DLSS 4.5 Dynamic Multi Frame Generation feature, first unveiled in January 2026, will be generally available from March 31. The feature lets GeForce RTX 50-series "Blackwell" gaming GPUs dynamically adjust the rate of DLSS Multi Frame Generation (MFG) rate between 0 (no frame-gen), and 6x, depending on a target frame-rate determined by the display's capability. The feature is meant to be enabled through the NVIDIA App, specific to supported games. NVIDIA had earlier announced 6x MFG for RTX 50-series GPUs, going up from 4x MFG that the company launched alongside the GPU family in early 2025.

Multi Frame Generation allows the GPU to draw up to five frames succeeding every conventionally rendered frame completely using AI, effectively multiplying framerates. RTX 50-series which debuted alongside DLSS 4, could originally do up to 4x MFG, but NVIDIA increased this to 6x MFG with its 2025 DLSS 4.5 update that introduced a new 2nd generation Transformer AI model that's more accurate, and can more accurately predict motion vectors allowing NVIDIA to increase MFG factor to six. Frame generation comes with latency costs, and so Dynamic Multi Frame Generation allows the driver to opportunistically lower MFG factor to improve whole-system latencies. At GDC, NVIDIA announced that DLSS 4.5 is coming to at least 20 of this year's most hotly anticipated AAA game titles.“
What's sucks is I want to hate it but running 50-60fps with 2x in the new re game feels better then straight 50-60
 
What's sucks is I want to hate it but running 50-60fps with 2x in the new re game feels better then straight 50-60
There is no reason to hate it. The idea that we can just brute-force more FPS and have games with nifty modern graphics is just not realistic. It's not like they could just build faster GPUs if they really wanted to, they are already throwing everything at the wall to make them as fast as possible. So things like framegen are what we need to higher frame rates. If they look good, and feel good, then great. I don't care why a game looks great, I don't care if it is "cheating". Rendering has long been hacks, piled on top of cheats, piled on top of shortcuts to make it work fast. I just want it to look good and feel good.

I'd rather have ray-reconstruction working with DLSS 4.5 than dynamic frame generation.
I imagine they are working on it however who knows how long it takes to train the models. Particularly since I can imagine situations where they train one, but it doesn't actually look better.
 
What's sucks is I want to hate it but running 50-60fps with 2x in the new re game feels better then straight 50-60
I've def enjoyed MFG in the few games I've played that support it. Consider also there's many instances where Reflex with MFG still has lower or similar input latency than AMD/Intel cards playing with no MFG as well, or lower/similar latency than just gaming on your same card with Reflex off vs on with MFG. So given your base FR is high enough (around 60) and there's no visible artifacting (which I haven't personally seen yet, only artifacting due to DLSS upscaling, not MFG), I really don't get the hate for FG/MFG.
 
I've def enjoyed MFG in the few games I've played that support it. Consider also there's many instances where Reflex with MFG still has lower or similar input latency than AMD/Intel cards playing with no MFG as well, or lower/similar latency than just gaming on your same card with Reflex off vs on with MFG. So given your base FR is high enough (around 60) and there's no visible artifacting (which I haven't personally seen yet, only artifacting due to DLSS upscaling, not MFG), I really don't get the hate for FG/MFG.
I've had both good and bad experiences with it, sometimes it has issues, but the thing is it isn't mandatory, I'll try it and see. In Hitman 3, it works quite well, let's my use their very performance intensive RT and still get FPS near 200. In Satisfactory, there's something broken and it makes lights flicker, so it isn't usable and I leave it off.

On my monitor I almost always try it if available, since it is a 240Hz monitor and very little can push that kind of FPS. On my TV it depends, that's only 120Hz so there are plenty of games running close enough to that where enabling it wouldn't make a lot of sense, but in other cases it really helps for those games with heavy path tracing that looks great, but hits FPS hard.
 
yeah plus especially with single player games and my GSYNC (FreeSync) monitor I can tolerate down to 48FPS without caring TBH
I find I'm a little less tolerant with OLED. The super fast element response makes low FPS more noticeable to me than on LCD.
 
So having played more with it, Hardware Unboxed is right: Don't sleep on Preset L for other modes. It is heavy, which is why they recommend it for ultra performance, but it does a really good job. I've been messing with Satisfactory and between K and M, I was defaulting to K. While M does clean up ghosting, it makes Lumen light issues worse (they are bad in that game no matter what because Lumen was an afterthought). However I decided to try L in quality mode and... it does pretty good. It does have some issues that K doesn't with shadows on foliage and such, but in general it not only is sharper, and cleans up ghosting, but it also was even able to improve some of the lighting issues. Hits the FPS hard, but worth it in this case.

It is worth trying any time you can spare some FPS to see if it cleans up the image any more. Not going to be worth it in all cases, but it can give some really clean results.
 
So having played more with it, Hardware Unboxed is right: Don't sleep on Preset L for other modes. It is heavy, which is why they recommend it for ultra performance, but it does a really good job. I've been messing with Satisfactory and between K and M, I was defaulting to K. While M does clean up ghosting, it makes Lumen light issues worse (they are bad in that game no matter what because Lumen was an afterthought). However I decided to try L in quality mode and... it does pretty good. It does have some issues that K doesn't with shadows on foliage and such, but in general it not only is sharper, and cleans up ghosting, but it also was even able to improve some of the lighting issues. Hits the FPS hard, but worth it in this case.

It is worth trying any time you can spare some FPS to see if it cleans up the image any more. Not going to be worth it in all cases, but it can give some really clean results.
I’m using L for DLAA for the last week. I’m firmly in the DLAA or nothing camp now. There is a difference from K but I’m at a loss to explain what precisely it is.
 
I’m using L for DLAA for the last week. I’m firmly in the DLAA or nothing camp now. There is a difference from K but I’m at a loss to explain what precisely it is.
Haven't tried it with DLAA but in Satisfactory using quality mode (well actually using the correct scale for it, they label them wrong) I notice a few things (vs M):

  • Less issues with oversharpening. It is sharp as in more detail, but not oversharpened.
  • Less shadow and lighting noise. Those are issues that both the new presets have compared to K, and they aren't gone, but L has less than M.
  • Better handling of RT reflections, though either way it looks better with denoising off.
  • Maybe slightly less ghosting, this one is hard to tell.

For DLAA I imagine it is less noticeable, the higher the input resolution, the less the differences are going to matter though they won't be nothing.
 
Haven't tried it with DLAA but in Satisfactory using quality mode (well actually using the correct scale for it, they label them wrong) I notice a few things (vs M):

  • Less issues with oversharpening. It is sharp as in more detail, but not oversharpened.
  • Less shadow and lighting noise. Those are issues that both the new presets have compared to K, and they aren't gone, but L has less than M.
  • Better handling of RT reflections, though either way it looks better with denoising off.
  • Maybe slightly less ghosting, this one is hard to tell.

For DLAA I imagine it is less noticeable, the higher the input resolution, the less the differences are going to matter though they won't be nothing.
NVIDIA DLSS 4.5 Dynamic Multi-Frame Generation and 6x Mode Officially Arrive
 
I can only get it to work in a few games. Has anyone been able to enable dynamic in Battlefield 6? The option os Grayed out for me in Nvidia app for BF6 and RE9 but is available for me in Indiana Jones and Cyberpunk and Oblivion Remake.
 

NVIDIA DLSS 4.5 Leaves Beta, Available Now In War Thunder, Enlisted & Dawn of Defiance

PRESS RELEASE by Nomad76 Today, 09:33 Discuss (1 Comment)
Each week, new games and apps integrating NVIDIA DLSS, NVIDIA Reflex, and advanced ray-traced effects are released or announced, delivering the definitive PC experience for GeForce RTX players. This week, Samson has launched with ray tracing, DLSS Ray Reconstruction, and DLSS Multi Frame Generation. Additionally, Gaijin Entertainment's Enlisted and War Thunder have both introduced support for DLSS Ray Reconstruction when using each game's ray-traced effects, and DLSS 4.5 Super Resolution for those playing with FPS-maximizing rasterized visuals. And Dawn of Defiance recently introduced DLSS support, which you can upgrade to DLSS 4.5 using the NVIDIA app.

And today, our new NVIDIA app update has exited beta, bringing the full suite of DLSS 4.5 updates to every user of our essential application.“
 

Blind Test Shows Gamers Prefer NVIDIA DLSS 4.5 Over AMD FSR 4.1

by AleksandarK Today, 05:50 Discuss (9 Comments)
Back in February, ComputerBase conducted a large blind test comparing in-game screenshots generated using the latest upscaling technologies: AMD's FSR 4.0 and NVIDIA's Deep Learning Super Sampling 4.5. The testing has since been updated, and community votes have been processed, revealing that AMD's updated upscaling technology, FSR 4.1, shows significant improvement over FSR 4.0. However, it still trails behind NVIDIA's DLSS 4.5 in visual quality. In the latest ComputerBase testing, the following games were upscaled using FSR 4.0, FSR 4.1, and DLSS 4.5: Year 117 - Pax Romana, ARC Raiders, Assassin's Creed Shadows, Call of Duty: Black Ops 7, Kingdom Come 2: Deliverance, Resident Evil Requiem, and The Last of Us Part I. In all of these games, the ComputerBase review concluded that DLSS 4.5 was the top performer, a view confirmed by the community in a separate blind test vote.

Similar to the previous test, the ComputerBase team conducted the comparison using videos labeled with three options, without revealing which rendering method was used, to ensure a fully blind test. This resulted in a community verdict with two notable outcomes. First, NVIDIA's DLSS 4.5 remains the leader in image quality, with 6 out of 7 games showing the best results using DLSS 4.5. The only game where AMD's FSR upscaling came out on top was Resident Evil Requiem, where DLSS 4.5 placed second behind FSR 4.1. Overall, DLSS 4.5 is seen as providing sharper visual details and more consistent frame generation compared to AMD's FSR upscaling.“
 


"
NVIDIA DLSS 4.5 Ray Reconstruction Review - Transformer Improved

Review Graphics Cards
NVIDIA DLSS 4.5 Ray Reconstruction boosts image quality with a new AI-powered model that uses 20% more parameters and 35% more compute, all without sacrificing performance. Our testing shows that ray-traced visuals are enhanced, offering crisper, richer textures, especially in path tracing.
Read Review"
 
Very much looking forward to this. From the look of things, it is basically a straight up improvement. The most stunning shot was the one where you could compare native PT to ultra performance PT. Like, you can see the downgrade, it isn't invisible... but fuck me is it WAY less than I would expect for rendering at 1/9th of the resolution.

Interestingly enough HUB's test found that it actually GAINS performance on the older cards somtimes, which is the opposite of what I'd expect. I thought with the talk of more parameters and such it would be more quantized and thus run as fast on newer cards, but be slower on the older ones that don't have acceleration for the smaller quants. Not the case, they show that on a 5070Ti it is basically always right about the same performance as the old preset. So just free quality upgrade. However on a 3090 or a 2080 sometimes it is the same, but sometimes there is a slight uplift for the new preset.

So it really is just basically straight out better and something you should use, no matter what your hardware.

I'm going to wait until it is out of beta, but I'm excited.
 
It's officially out. Nvidia App just updated which added support for DLSS 4.5 Ray Reconstruction. It will automatically apply Preset F for Ray Reconstruction in games if you set the DLSS Override Model Presets to "Recommended".
 
Very much looking forward to this. From the look of things, it is basically a straight up improvement. The most stunning shot was the one where you could compare native PT to ultra performance PT. Like, you can see the downgrade, it isn't invisible... but fuck me is it WAY less than I would expect for rendering at 1/9th of the resolution.

Interestingly enough HUB's test found that it actually GAINS performance on the older cards somtimes, which is the opposite of what I'd expect. I thought with the talk of more parameters and such it would be more quantized and thus run as fast on newer cards, but be slower on the older ones that don't have acceleration for the smaller quants. Not the case, they show that on a 5070Ti it is basically always right about the same performance as the old preset. So just free quality upgrade. However on a 3090 or a 2080 sometimes it is the same, but sometimes there is a slight uplift for the new preset.

So it really is just basically straight out better and something you should use, no matter what your hardware.

I'm going to wait until it is out of beta, but I'm excited.
Last time I compared ultra performance upscaling was CNN vs. Transformer when DLSS4 came out. It went from eye bleed to tolerable. Makes me think stupid high res screens might actually be workable on high end cards.

Worse quantization doesn't necessarily mean slower. Often it doesn't. Ada added FP8, Blackwell added FP4. But Ampere & Turing can do INT8 & INT4 if my Googling is correct. Compared to FP8, INT8 is less "accurate" but not inherently slower. Could easily be faster. So it's quite possible the image quality isn't quite as good on 20 and 30 series cards compared to 40 and 50 series, but speed is ok. Sacrificing some IQ for speed on older cards seems like a very reasonable trade-off. I wouldn't be surprised if they did that.

Actually, I hope they did do that. Really ought to take it one step further and have different models for different classes of cards. Suppose Player 1 has 5090 + 4k OLED gaming monitor and Player 2 has a cheap 4k TV and a 5050 or 4060. Ideally they shouldn't be running the same upscaling model. The cards may be the same architecture, but the 5090 has lots more tensor cores and a bunch of GB of spare ram for upscaling. They have totally different targets. Player 1 wants the game to look as good as possible on that beast of a setup. Player 2 is just trying to make it work decently on a budget rig.
 
Last time I compared ultra performance upscaling was CNN vs. Transformer when DLSS4 came out. It went from eye bleed to tolerable. Makes me think stupid high res screens might actually be workable on high end cards.
It's also my understanding that it works better on higher rez screens because of higher input resolution. So however good it is a 4k, it is better at 8k. I lack the ability to test that but based on what I understand about these, it makes sense.

I suspect that really it will be the only way forward for ultra high rez gaming. We just aren't going to be able to build cards big enough to brute force high frame rates at high resolutions. So either reduce the resolution, which certainly is an option particularly since many high rez monitors are dual or triple mode and give better FPS, or use more upscaling.

Worse quantization doesn't necessarily mean slower. Often it doesn't. Ada added FP8, Blackwell added FP4. But Ampere & Turing can do INT8 & INT4 if my Googling is correct. Compared to FP8, INT8 is less "accurate" but not inherently slower. Could easily be faster. So it's quite possible the image quality isn't quite as good on 20 and 30 series cards compared to 40 and 50 series, but speed is ok. Sacrificing some IQ for speed on older cards seems like a very reasonable trade-off. I wouldn't be surprised if they did that.
The issue isn't just if they can do them, but can they do them faster and on what part of the card. Like NVFP4 can be made to run on 4090s no issue but at the same speed at FP8, since you use a whole FP8 unit to do it. The gain on the 5090 is that you can effectively split the unit in half and do 2 4-bit calculations at the same time. Not sure how that carries over to integer calculations but ya, just because you can support it doesn't mean it is faster. Likewise I'm not sure if the tensor cores work on integers, or if those have to go to the shaders. That's not a problem, you can totally do inference on shaders, but then it is competing with graphics rendering.

Either way, whatever they did worked really well here. If I had an older card I'd be pleased as hell to see that this not only supported it, but also didn't tank performance.

Actually, I hope they did do that. Really ought to take it one step further and have different models for different classes of cards. Suppose Player 1 has 5090 + 4k OLED gaming monitor and Player 2 has a cheap 4k TV and a 5050 or 4060. Ideally they shouldn't be running the same upscaling model. The cards may be the same architecture, but the 5090 has lots more tensor cores and a bunch of GB of spare ram for upscaling. They have totally different targets. Player 1 wants the game to look as good as possible on that beast of a setup. Player 2 is just trying to make it work decently on a budget rig.
What I'd like to see is something like a "DLSS HQ Mode" or the like that uses bigger models that take more RAM and run slower, but is elective. I mean we kinda already have that with non-RR 4.5 it is heavier than 4.0 and you can choose to use it or not. However I'd be interested in having multiple models trained to do the same thing, but just some with more parameters than the others for use on bigger cards.

That said, it might not be worth nVidia's time to do that. They clearly spend a fuck ton of time and compute resources to train these things, perhaps it wouldn't be worth it to train a HQ and regular model as the HQ one wouldn't be better enough to really justify it, and instead it would be better to spend more time and resources improving the normal one.

The other issue with bigger models is that they use more time, which limits max FPS which is another thing you might want with a bigger card. For something that is doing path tracing, it isn't a huge deal as FPS is low already. However for other things it can be an issue. Like if I made a heavy model that takes 10ms of processing time, that means absolute max FPS for a game using it would be 100, and realistically less. No matter how simple the graphics or how fast the game's rendering, if it takes the upscaler 10ms then you can never render faster than 100fps. Realistically since games take a fair bit of time themselves to render it could cap FPS at something much lower.
 
It's officially out. Nvidia App just updated which added support for DLSS 4.5 Ray Reconstruction. It will automatically apply Preset F for Ray Reconstruction in games if you set the DLSS Override Model Presets to "Recommended".
Looks like it is still only out on beta channels to me. My app isn't interested in updating and it's version (11.0.8.299) is the same as the one you download from nVidia.
 
I didn't need beta channels or anything. I have older drivers (596 something) and simply setting preset to "recommended" (using nvidia inspector, not nvapp though) enabled preset F in games. I used the overlay to check the preset, but nvinspector allows shows that the latest DLSS dll was correctly downloaded.

I only have Indiana Jones that I play with RR atm, and it looks great for sure, the sharpness of DLSS 4.5 is welcome to my eyes.
 
Last edited:
I didn't need beta channels or anything. I have older drivers (596 something) and simply setting preset to "recommended" (using nvidia inspector, not nvapp though) enabled preset F in games. I used the overlay to check the preset, but nvinspector allows shows that the latest DLSS dll was correctly downloaded.

I only have Indiana Jones that I play with RR atm, and it looks great for sure, the sharpness of DLSS 4.5 is welcome to my eyes.
Well can't say why mine isn't updated, but it isn't. I cannot change the RR setting I only have use app setting or recommended.
 
I didn't need beta channels or anything. I have older drivers (596 something) and simply setting preset to "recommended" (using nvidia inspector, not nvapp though) enabled preset F in games. I used the overlay to check the preset, but nvinspector allows shows that the latest DLSS dll was correctly downloaded.

I only have Indiana Jones that I play with RR atm, and it looks great for sure, the sharpness of DLSS 4.5 is welcome to my eyes.
I had to enable the beta, Nvidia post on it also says you must enable the beta/early access. Then you can manually apply the override model.

Tried it in cyberpunk, recently started playing it last month and was overall pretty disappointed with it, even with the pretty path traced graphics on -- BUT, the new model definitely cleans up multiple artifacts I'd notice it before, it is very clean now and not near as distracting. Although, it is such a shame they never bothered to model V properly so you still see floating bodies and V-shaped holes in reflections -.- Oh well
 
Back
Top