DLSS 4.5 on GeForce NOW: what the new controls do

Five games were named for the cloud on their release day: CONTROL Resonant, Gears of War: E-Day, Onimusha: Way of the Sword, The Blood of Dawnwalker and STAR WARS Zero Company. CONTROL Resonant also comes with a twelve-month Ultimate membership, which is the sort of bundle NVIDIA runs a few times a year. Our calendar tracks what is coming to each service, and the changelog records what actually arrived.

One thing to get straight before the rest, because a lot of the coverage has it the other way round. This is not frame generation arriving in the cloud. Frame generation has been running on the Ultimate plan since the RTX 5080 servers landed in September 2025, decided per game by NVIDIA. What arrives is your ability to turn it off.

Super Resolution is the setting people mean when they say DLSS. The game renders at, say, 1440p; the model rebuilds a 4K frame from it plus the frames before it. Because the render is cheaper, the frame rate goes up, and because the reconstruction is good, the picture usually does not go down. The modes are the amount of cheating: Quality renders closest to native, Ultra Performance renders furthest from it.

Nothing about that changes when the machine is in a data centre. There is one cloud-specific note: whatever the model produces is squeezed through a video encoder before it reaches you, so at 1080p on a laptop the gap between Quality and Balanced is often smaller than it looks in a desktop review.

If a control appears, Quality is the sane default. Performance is worth trying on a big screen, where the resolution you are asking the server for is high enough that even the cheap render has detail in it.

Ray tracing traces too few rays to be clean, so every ray-traced game ships denoisers that guess the rest. They are the reason reflections smear when the camera moves and lighting takes a moment to settle after a light turns on. Ray Reconstruction hands that job to one model, and it is the setting most likely to change what a path-traced game looks like on your screen.

There is a second effect that only shows up in streaming. Video encoders spend their bits on detail that changes between frames, and noise is detail that changes between frames. A noisy denoiser output does not just look noisy: it eats the bitrate that should have gone to the picture.

That is a general property of video compression rather than something NVIDIA has measured for GeForce NOW. It points the same way as everything else here.

Here is the part that gets lost. Frame generation raises the number of frames on your screen. It does not raise the number of times the game reads your controller, and it cannot, because the extra frames are invented from two real ones that already exist. A game running at 30 fps with 4X frame generation shows you 120 frames a second and still answers you thirty times a second — and slightly slower than that, because the renderer must hold a finished frame back in order to interpolate towards it.

The cost is small and it is real. TechSpot measured about 10 ms of extra lag when frame generation went on in Cyberpunk 2077, with the second generated frame adding roughly 4 ms more and the third about 2 ms.

NVIDIA's answer is Reflex, which shortens the queue elsewhere in the pipeline and which it says lets DLSS 4.5 generate frames with minimal impact to responsiveness. On a desktop that is a fair summary. Tom's Hardware has been making the opposite point for two years: input latency is the missing piece of most frame-generation coverage, and NVIDIA's own claim that DLSS 4 raises frame rates without adding latency did not survive contact with the fine print.

The catches

Higher multipliers are not free. PC Gamer's hands-on with the 5X and 6X modes found the frame rates real and the latency high for those frame rates, and the artefacts that DLSS 4 had in games like Alan Wake 2 still there in 4.5. Dynamic mode has a limitation NVIDIA states plainly: it is not compatible with frame rate limiters and V-Sync. Hold that thought for the next section.

The first difference is arithmetic. On a desktop, frame generation adds its 10 ms to a chain that Reflex can hold near 30, so it is a third more. On GeForce NOW it adds the same 10 ms to a chain that already contains two crossings of your internet connection, and those crossings are the largest term in the sum for most people. Ten milliseconds on top of sixty is a smaller proportion and a worse place to spend it, because sixty was already near the edge of where a shooter stops feeling right.

The second difference is the ceiling. On a desktop the frames go to a monitor, and Dynamic Multi Frame Generation aims at that monitor's refresh rate. In the cloud they go to an encoder first, and your plan decides how many are allowed out: 120 a second on Ultimate, 60 on Performance. Generating more frames than the stream carries cannot reach you.

NVIDIA has said nothing about how the automatic multiplier reads a stream cap. The closest thing to an official answer is that Dynamic mode does not work with frame rate limiters on the desktop. That is a real question about a menu nobody outside NVIDIA has used, and we will update this page when it ships.

NVIDIA, read on 26 August 2026: the Gamescom 2026 announcements, DLSS 4.5 and its developer post, the NVIDIA app update that shipped Dynamic Multi Frame Generation, the CES 2026 post, the GeForce NOW server upgrade page and the GeForce NOW FAQ.

Testing and reporting: TechSpot on the latency frame generation adds, PC Gamer and Tom's Hardware hands-on with Dynamic Multi Frame Generation, Tom's Hardware on input latency in framegen coverage, and GamingOnLinux on the Steam Machine announcement.

The round-trip figures are our own model, the one behind the network test on this site. No part of this article is a measurement of the GeForce NOW controls, because they are not out.