Last night, Nvidia held their keynote ahead of CES 2026, which—as usually—only focused on AI, robotics and large enterprises market rather than consumer PCs and devices. Even so, the company does have something new for gamers. The company is introducing a new generation of their AI-based game rendering—DLSS 4.5. It places even greater emphasis on interpolated frame generation, but improvements go beyond that as well.
DLSS 4.5 introduces two new sets of features. The first is an improved AI model used for the upscaling (or “super resolution”) functionality and the temporal filtering (and thus also anti-aliasing, as temporal filtering in the DLSS, FSR and XeSS upscaling technologies simultaneously takes over the role of temporal anti-aliasing). While DLSS 4 introduced a transformer-type neural network last year (replacing the previous convolutional-type model), DLSS 4.5 deploys its second, improved generation.
New AI model
The second-generation transformer model is trained on a larger dataset, which has been extended with more problematic situations have been incorporated—situations where previous versions produced sub-par results such as artifacts or blurring. The result should again be a somewhat reduced occurrence of ghosting and better temporal image stability than before (less shimmering and flickering). Do not expect these issues to be completely eliminated, however—that is not realistic, as occurence of artifacts and imperfections stem from the very nature of upscaling.
According to Nvidia, the anti-aliasing effect of DLSS 4.5 should also work better—edges and silhouettes in the image should be smoother. This stronger smoothing could, however, have the side effect of making the image slightly softer and less sharp/less detailed (at least judging by some of the sample images shown in the presentation). A proper evaluation of image quality will of course only be possible with more extensive examination, however.
What is important is that this model is supported by all generations of GeForce RTX graphics cards, including the first-generation Turing GPUs. It will, however, be somewhat slower on GeForce RTX 2000 and 3000 series than on newer GeForce RTX 4000 and 5000 cards—in the sense that the processing will require more GPU time, so applying the upscaling will result in a lower FPS increase on these cards.
This is because the AI model uses FP8 values as its data type. These can be processed efficiently only by GPU architectures Ada Lovelace (GeForce RTX 4000) and Blackwell (GeForce RTX 5000), whereas Turing and Ampere cannot use them—likely resorting instead to lower-quality INT8 or slower FP16 instructions.
Better upscaling is already in the drivers
This part of DLSS 4.5 is available immediately according to Nvidia—if you update your GPU driver now, it should already include the required libraries, which the graphics drivers can then inject into games using the DLSS Override function in the drivers.
Update (January 24th, 2026): DLSS 4.5 does not work with Ray Reconstruction
DLSS 4.5 seems to have one limitation that has not been mentioned in the information made available by Nvidia upon launch: At the time there is no way to use it together with Ray Reconstruction. Nvidia apparently doesn’t have an AI model available that would implement this features together. If you have Ray Reconstruction enabled in a game, it will only be able to use DLSS 4.0 and attempts to override it to use DLSS 4.5 won’t work.
Dynamic frame generation and up to 5 generated frames in a row
The second novelty is expanded support for “FPS generation,” meaning the insertion of AI-generated intermediate frames between real game frames. This improves the frame rate shown on the monitor, although the game’s actual frame rate does not change (which has implications for responsiveness). Frame generation first appeared on GeForce RTX 4000 graphics cards, which support inserting one frame between every two real game frames. The result is 50% generated and 50% real frames, and the output frame rate on the monitor is doubled (in reality, it is less than doubled, because the generation processing itself has overhead and the GPU can render slightly fewer real frames due to the performance cost).
With GeForce RTX 5000, Nvidia introduced DLSS 4 last year, which for the first time supported generating two or three frames instead of one. This means that 67% or 75% of frames are generated, and the apparent frame rate increase is correspondingly higher—with three frames, nearly a 4× increase is achieved (again, you get less in practice due to computational overhead).
6× frame generation
DLSS 4.5 builds on this and adds a new mode that increases the frame rate up to “sixfold”—again with an asterisk, as the real speedup is lower due to overhead, and again it applies only to generated frames and the FPS on the monitor, while the game’s actual frame rate remains unchanged.
This 6× mode inserts five additional generated frames between two real game frames. This means that the vast majority of the time you are looking at generated frames (with potential drawbacks due to artifacts and image errors)—only 16.7% of the time do you see a real frame rendered by the game, while 83.3% is AI-generated.
How useful is this? If you need to multiply the frame rate by six just to make a game smooth, then this certainly will not help you, because with final output rates of 60 Hz or 120 Hz your actual game frame rate would be only 10 or 20 FPS, respectively—which is hardly playable and will most likely result in noticeable input lag.
This 6× frame generation therefore probably makes sense only if you want an extremely high refresh rate on the monitor, such as 300–360 Hz (in which case the game’s internal frame rate would be 50 to 60 Hz). Whether padding the output so much with “fake frames” will actually be a better experience, or whether it will suffer from a certain “soap opera” effect, remains an open question.
According to the slides Nvidia shows, there may also be a further increase in latency. An example from the game Black Myth: Wukong shows that 4× generation results in 184 FPS (46 FPS in-game) with a latency of 47 ms, while 6× generation delivers 246 FPS with a latency of 53 ms. This is likely because generating additional frames increases the processing overhead, meaning more performance is consumed by the generation itself. If the output frame rate is 246 FPS, this implies that the game’s internal frame rate has actually dropped to 41 FPS—and because latency increases based on the inverse of the real frame rate (not the apparent display rate after generation), the result is inevitably longer (worse) latency.
Variable generation mode
Alongside this addition comes another, potentially more interesting feature that could be practical even for more realistic, lower-multiplier generation modes. DLSS 4.5 supports a variable mode that automatically changes the number of inserted frames depending on how high the frame rate is without it. This means that when the GPU can easily keep up, it may insert fewer “fake” frames, while in demanding scenes where FPS drops, it will add more.
It will work by setting a target frame rate in the drivers, after which the GPU automatically increases or decreases the number of generated frames to reach it. In practice, this could mean switching between 2×, 3× and 4× modes, for example.
Not until spring
Unlike the new transformer model, DLSS 4.5 with dynamic (variable) frame generation and the option for the static 6× mode of interpolated frame generation will not be available immediately. We do not yet have an exact release date; Nvidia only states that this feature will be made available to users sometime in the spring (i.e., March to June).
And as you may already suspect, this part of DLSS 4.5 will be exclusive to GeForce RTX 5000 graphics cards—it’s not going to be available on any older ones. Nor will 3× and 4× FPS generation be released to older GPUs, this also remains unchanged. Multiple frame generation continues to be an exclusive feature of the latest generation of graphics cards based on the Blackwell architecture. Besides them, multi-frame generation is also offered by Intel on Arc graphics cards via the XeSS 3 technology, but that does not work on GPUs from other brands, so it does not help GeForce RTX 4000 and older cards.
Source: Nvidia
English translation and edit by Jozef Dudáš
⠀
