NVIDIA Dynamic Multi Frame Gen Ray Reconstruction
NVIDIA’s DLSS has steadily moved beyond its original role as an AI upscaler. With DLSS 4.5, the technology now increasingly sits between the game engine and the final image that reaches the display, deciding not only how pixels should be reconstructed but also how many frames need to be generated along the way. Two features demonstrate that shift particularly well: Dynamic Multi Frame Generation and the latest version of Ray Reconstruction. Dynamic Multi Frame Generation has been out for a while but Ray Reconstruction will be generally available starting today.
We received early access to Ray Reconstruction ahead of their wider availability and tested it in Cyberpunk 2077, a game that remains particularly useful for evaluating ray tracing because Night City is packed with reflective vehicles, glass, wet roads, artificial lighting and complex geometry. To be honest, the DLLs were leaked a few days ahead of today’s announcement so anyone can try out Ray Reconstruction using a DLSS DLL Swapper. For performance testing, our PC used an AMD Ryzen 7 9800X3D, 48 GB of DDR5-6000 memory and an NVIDIA GeForce RTX 5090. Testing was conducted at 4K resolution, with the game configured to place a substantial ray-tracing workload on the GPU.
The two technologies are trying to solve very different problems. Dynamic Multi Frame Generation is essentially about matching rendered output more intelligently to a high-refresh-rate display. Ray Reconstruction is about making the ray-traced image itself look more convincing. And after spending some time with both, Ray Reconstruction may actually be the more visually interesting of the two.
Frame Generation has a fairly straightforward underlying idea. Instead of requiring the GPU to conventionally render every frame displayed on screen, NVIDIA can generate additional intermediate frames using AI. With conventional Multi Frame Generation, the multiplier is selected beforehand. Dynamic Multi Frame Generation changes that model. Rather than operating at one fixed multiplier throughout a game, the system can dynamically move between different Frame Generation multipliers depending on the workload and the target frame rate. DLSS 4.5 supports generation at up to 6X, meaning that at its maximum setting it can generate as many as five additional frames for each conventionally rendered frame. The easiest analogy is an automatic gearbox.
If the conventionally rendered frame rate drops and there is a larger gap between the GPU’s current output and the display’s refresh rate, Dynamic Multi Frame Generation can increase the multiplier. If the GPU starts rendering more frames natively, it can reduce the amount of Frame Generation instead. NVIDIA describes the system as continuously monitoring the difference between GPU performance and the target refresh rate, increasing or reducing generation according to what is required.
A 240Hz monitor does not necessarily require a game to use the maximum possible Frame Generation multiplier at every moment. If the GPU is already producing sufficient frames, generating five additional frames for every traditionally rendered one would be unnecessary. Conversely, a particularly intensive section of a path-traced game could require a larger multiplier to approach the same refresh-rate target. Dynamic Multi Frame Generation therefore attempts to make Frame Generation adaptive rather than simply more aggressive.
That dynamic behaviour was readily visible during our Cyberpunk 2077 testing. With the game running at 4K on the RTX 5090, the Frame Generation multiplier did not remain locked to a single value. We could watch it shift upwards and downwards during the benchmark run, responding as the complexity of the scene changed. The “Dynamic” part of Dynamic Multi Frame Generation is not simply an automated setting chosen when the game starts. The multiplier is being adjusted during gameplay according to the rendering workload.
Across our test run, Cyberpunk 2077 averaged 238 FPS at 4K, with reported latency hovering around 53 to 54ms. For a graphically intensive title running at 4K with ray tracing involved, seeing the displayed frame rate sit close to the 240 FPS territory is impressive from a throughput standpoint. But it is equally important to understand what that number represents. 238 FPS with Multi Frame Generation is not equivalent to a game engine conventionally rendering 238 discrete frames every second. A significant portion of those displayed frames are AI-generated intermediate frames. That also explains why latency does not fall proportionately as the displayed frame rate rises.
If a game were genuinely rendering 238 frames per second conventionally, the time between frames would be a little over 4 ms. Yet our measured system latency remained around 53 to 54ms. Frame Generation increases the visual update rate, but it does not magically make the underlying game simulation, CPU workload and conventionally rendered frame pipeline operate at the same rate. NVIDIA Reflex remains important here because it helps control latency elsewhere in the rendering pipeline while Frame Generation increases displayed frame output.
Essentially, Dynamic Multi Frame Generation is best understood as a smoothness technology rather than a direct substitute for raw rendering performance. On a 240 Hz display, pushing the visible output closer to the refresh ceiling can make camera motion and animation look considerably smoother. But for highly latency-sensitive competitive games, the conventionally rendered base frame rate remains important. For something such as Cyberpunk 2077, though, that trade-off woks for gamers. The game can make extensive use of demanding ray-traced effects while Dynamic Multi Frame Generation uses available headroom to push the visual output towards the capabilities of a high-refresh display.
Ray Reconstruction tackles an entirely different part of the graphics pipeline. Ray tracing works by tracing rays through a scene to determine how light interacts with surfaces. The problem is that tracing enough rays per pixel to produce a completely clean image in real time would be extraordinarily expensive. Games therefore work with a limited number of ray samples. That produces noisy information, which then needs to be reconstructed into a usable image.
Traditionally, ray-traced games employ multiple denoisers designed for specific effects such as reflections, shadows and indirect lighting. Those denoisers try to infer what the clean result should look like from incomplete ray-traced data. Ray Reconstruction replaces much of that conventional denoising process with an AI model that reconstructs a higher-quality ray-traced image.
The DLSS 4.5 version uses NVIDIA’s second-generation transformer architecture. NVIDIA says the new model uses a larger denoiser, deeper spatial awareness and finer control over temporal accumulation, with the aim of improving lighting accuracy and producing sharper ray-traced imagery. That sounds rather abstract until it is placed next to the same scene without Ray Reconstruction. With out tests, we looked at three scenes in Cyberpunk 2077, concentrating particularly on reflections.
Our first comparison involved a parked car with a signboard reflected across its bonnet. At a glance, both images looked broadly similar. Looking closer revealed several differences. With Ray Reconstruction enabled, the reflection across the bonnet was better defined, particularly around fine detail. The lettering from the reflected signboard was noticeably sharper and easier to distinguish.
More interestingly, there was graffiti sprayed across concrete handrails further into the background. Without Ray Reconstruction, that graffiti was barely discernible. There was enough information to suggest something was present on the concrete, but much of the detail was lost. Turning Ray Reconstruction on made the graffiti substantially clearer. This is where the improvement becomes more meaningful than simply increasing sharpness. The feature was recovering small scene details that were otherwise being smoothed away during reconstruction. That also helped the reflected surface retain more of the visual information present in the environment instead of producing a broadly correct but comparatively indistinct reflection.
The second scene concentrated on a building façade containing a large number of windows. This proved to be an even better showcase because glass creates a difficult mixture of reflections, transparency, interior illumination and geometry. Towards the centre and upper-right portion of the image, environmental reflections in the windows were visibly crisper with Ray Reconstruction enabled. Without it, the reflected environment was still present, but details tended to blend together more readily. The difference became even more obvious towards the left and upper-left portions of the scene.
With Ray Reconstruction switched on, tube lights inside the building could be seen as well-defined individual light sources. Their shape and placement were substantially easier to distinguish. Without Ray Reconstruction, several of these elements became blurred or simply appeared duller, reducing the sense that there was a properly illuminated interior behind the glass. This is an important distinction because ray-traced image quality is not purely about producing sharper reflections. Lighting needs to remain spatially coherent. A light source seen through or reflected by another surface should retain enough shape and intensity for the viewer to understand what is producing that illumination. In this scene, Ray Reconstruction did a noticeably better job of retaining that structure.
The third comparison was the most revealing. We examined reflections across a puddle on the road. Without Ray Reconstruction, the reflected image appeared to spread much more broadly across the pavement. There was a considerable amount of reflection bleed, with bright reflected information appearing even across areas where there did not seem to be enough standing water to justify such a strong reflection. The result was visually striking but not especially natural. Enabling Ray Reconstruction changed the character of the surface.
Where the puddle was deeper and had a more substantial layer of water, reflections remained strong and almost mirror-like. Crucially, some of the underlying pavement texture could still be seen through the reflected image. Moving towards the edges of the puddle produced a more gradual transition. As the water became thinner and started breaking up across the road surface, the intensity of the reflection dropped correspondingly. Instead of a large reflected image seemingly painted across the pavement, the strength of the reflection appeared much more closely related to the physical characteristics of the surface. That made the result look considerably more realistic.
And this is probably the strongest demonstration we saw of what Ray Reconstruction can contribute. If it simply sharpened the reflection everywhere, the technology could effectively be described as a better denoiser. What we observed instead was a more plausible reconstruction of how different parts of the reflective surface should behave. The thicker portion of the puddle acted more like a mirror. The thin wet edges produced comparatively muted reflections. Dry or almost-dry portions of the pavement did not appear to carry the same reflected image. The improvement was therefore not only one of detail, but of visual coherence.
Testing both features together also demonstrates how broad DLSS has become. Dynamic Multi Frame Generation is principally concerned with motion and throughput. It analyses how much additional frame generation is useful at a particular moment and alters its multiplier to help push displayed output towards the monitor’s refresh rate. Ray Reconstruction is less obvious on an FPS counter, but arguably more significant to the final appearance of a ray-traced game. Across all three of our comparison scenes, its effects were consistent even though the type of improvement changed.
The goal of better ray tracing should not simply be to make every reflective object shinier or every reflection sharper. It should be to make light behave more convincingly according to the material and geometry being represented. Personally, my gripe with the way things are is that you need to use multiple apps to tweak the settings to make use of these features in a game. Hopefully, with time, the configuration should become a little easier to handle. As it stands now, it’s more like tweaking the knobs of a nuclear reactor.