Nvidia’s Deep Learning Super Sampling technology has consistently redefined the boundaries of real-time graphics rendering, and its latest iteration seeks to push those limits even further by leveraging a more sophisticated AI model. This review explores the evolution of this AI-powered upscaler, focusing on the key features, performance metrics, and image quality improvements introduced in version 4.5. The purpose of this analysis is to provide a thorough understanding of the technology’s current capabilities by comparing it directly to its predecessor, DLSS 4.0, and to assess its potential future impact on the gaming landscape for owners of both new and older hardware.
Understanding DLSS From Upscaling to AI Reconstruction
Nvidia’s Deep Learning Super Sampling (DLSS) represents a paradigm shift from traditional spatial upscaling techniques, which often rely on simple algorithms to stretch a lower-resolution image, frequently resulting in a loss of detail and the introduction of visual artifacts. At its core, DLSS employs a sophisticated neural network, trained on super-high-resolution offline renders of game scenes, to intelligently reconstruct a high-resolution image from a lower-resolution input. This process is accelerated by dedicated Tensor Cores found on Nvidia’s RTX graphics cards, allowing the complex AI calculations to occur in real-time without placing an excessive burden on the primary shading and rendering pipelines.
The technology has matured significantly since its inception. Early versions focused primarily on boosting frame rates, sometimes at the expense of image clarity compared to native rendering. However, with each major revision, the AI model has grown more adept, evolving from a simple upscaler into a powerful image reconstruction suite. It now tackles complex graphical challenges like aliasing, temporal instability, and motion artifacts with remarkable efficacy. The relevance of DLSS in the broader landscape of gaming graphics cannot be overstated; it has made high-resolution, high-frame-rate gaming, particularly with demanding features like ray tracing, an accessible reality for a wider range of hardware. DLSS 4.5 continues this trajectory, aiming to refine the reconstruction process to a point where upscaled images can consistently surpass the quality of native resolution rendering.
Core Technology Breakdown of DLSS 4.5
A More Advanced Second Generation Transformer Model
The principal architectural advancement in DLSS 4.5 is the integration of a second-generation transformer model, a type of neural network architecture that has demonstrated profound success in natural language processing and is now being applied with greater sophistication to image synthesis. This new model is specifically designed to better understand and predict the relationship between frames over time, a concept crucial for maintaining temporal stability. By analyzing motion vectors and historical frame data with greater contextual awareness, the transformer model can more accurately reconstruct details that are temporarily obscured or in rapid motion, which directly addresses long-standing challenges in temporal upscaling. This advanced model is significantly accelerated by the inclusion of FP8 (8-bit floating point) instruction support on the latest GeForce RTX 40 and 50 series GPUs. The use of a lower-precision data format allows for a massive increase in computational throughput on the Tensor Cores, enabling the more complex transformer network to run efficiently without incurring a prohibitive performance penalty. Nvidia claims this combination enhances temporal stability to reduce the shimmering seen on fine details, minimizes the ghosting trails that can appear behind moving objects, and delivers more precise reconstruction of complex geometry like edges and dynamic lighting effects. This represents a fundamental shift in the underlying AI, moving beyond pattern recognition to a more holistic understanding of scene dynamics.
Introduction of New Model Presets M and L
To give developers and users more granular control over the new capabilities, DLSS 4.5 introduces two new model presets: “Preset M” and “Preset L”. These presets are distinct from the “Preset K” associated with the DLSS 4.0 model and are tuned for different use cases. Preset M is positioned as the primary and recommended model for most scenarios, having been specifically optimized to deliver the most significant image quality gains when using the Performance mode. Nvidia’s focus here reflects a desire to make lower internal resolutions more visually appealing, thereby enabling higher frame rates without substantial compromise.
Preset L, in contrast, is an even more specialized model tailored for the Ultra Performance mode. This mode uses the lowest internal rendering resolution and is typically reserved for extremely high-resolution displays like 8K monitors, where achieving playable frame rates is the paramount concern. By training a dedicated model for this specific scenario, Nvidia aims to maximize image coherence and stability even when reconstructing from a very small pool of source pixels. Through the Nvidia App, users retain access to the older Preset K, allowing for direct comparisons and the flexibility to choose the model that best suits their hardware and visual preferences for a given title.
Performance Analysis and the Question of a Penalty
Benchmarks on Modern Architecture The RTX 50 and 40 Series
For owners of the latest Nvidia hardware, the performance impact of engaging the more sophisticated DLSS 4.5 model is, as Nvidia suggests, a minor tradeoff. Testing on a GeForce RTX 5070 Ti reveals that the performance delta between DLSS 4.0 (Preset K) and DLSS 4.5 (Preset M) is consistently small across various titles and settings. In Cyberpunk 2077 at a 4K resolution with the Ultra Ray Tracing preset, the difference is negligible; DLSS 4.5 is approximately 5% slower in Quality mode and only 3% slower in Performance mode. This slight reduction is not enough to fundamentally alter the user experience, as the frame rate uplift over native resolution remains substantial.
This trend holds in other demanding titles. In MafiThe Old Country, the performance cost is slightly more pronounced, with the Quality mode running 9% slower and Performance mode 5% slower at 4K. This equates to roughly half a tier of performance, meaning DLSS 4.5 Balanced mode delivers a frame rate situated between DLSS 4.0’s Quality and Balanced modes. A similar pattern emerges in Horizon Zero Dawn Remastered, a title with a higher native frame rate where the overhead of upscaling is more apparent. Here, the performance hit can approach a full tier, with DLSS 4.5 Quality running 13% slower than its DLSS 4.0 counterpart. Despite this, the cost is manageable and often justifiable given the corresponding improvements in image quality.
Benchmarks on Previous Generations The RTX 30 and 20 Series
The narrative changes dramatically when examining performance on older architectures like the RTX 30 and 20 series, which lack the native FP8 acceleration that the new transformer model is designed to leverage. On a GeForce RTX 3090, the “heavier performance impact” becomes immediately apparent. In Cyberpunk 2077 at 4K, activating DLSS 4.5 in Quality mode results in a significant 19% drop in frame rate compared to DLSS 4.0. The Balanced and Performance modes are not spared, each seeing a 15% reduction. While still faster than native rendering in this GPU-bound scenario, the performance cost is substantial enough to make a user question the benefit.
This performance penalty is even more severe in titles where the GPU is less strained at native resolution. In Horizon Zero Dawn at 4K on the RTX 3090, the results are stark. DLSS 4.5 Quality mode runs a staggering 26% slower than DLSS 4.0 Quality, to the point where it delivers a lower frame rate than native TAA rendering. This effectively negates the primary purpose of using an upscaler. The impact represents more than two full tiers of performance loss, with DLSS 4.5 Performance mode running slower than DLSS 4.0 Quality mode. These findings confirm that for owners of pre-RTX 40 series GPUs, the advanced model introduces a computational burden that their hardware is not equipped to handle efficiently, making the older and faster DLSS 4.0 model the more practical choice.
In Depth Image Quality Comparison DLSS 4.5 vs DLSS 4.0
Vast Improvements in Disocclusion and Ghosting
Two of the most pronounced and welcome visual upgrades in DLSS 4.5 are the drastic reductions in disocclusion artifacts and ghosting. Disocclusion artifacts, which often manifest as a shimmering or sizzling effect around a moving character as they reveal the background behind them, were a noticeable issue in DLSS 4.0. In titles like The Last of Us Part I, this effect could be quite distracting. With DLSS 4.5, these artifacts are substantially suppressed. While not entirely eliminated, the remaining sizzling is far less perceptible, leading to a much cleaner and more stable image during camera and character motion. This improvement is so significant that DLSS 4.5 in Performance mode frequently presents a cleaner image in terms of disocclusion than DLSS 4.0 in Quality mode.
Similarly, the trailing images or “ghosting” that could plague fast-moving objects against contrasting backgrounds have been largely mitigated. In Ratchet & Clank: Rift Apart, a visible motion trail left behind the protagonist on certain surfaces when using DLSS 4.0 is almost completely absent with DLSS 4.5. The new transformer model’s enhanced ability to interpret motion seems to allow it to more accurately differentiate between an object’s new position and its previous one, preventing remnants of past frames from lingering. This leads to clearer object definition during fast-paced action, contributing to a sharper and more coherent presentation overall.
Superior Reconstruction of Foliage Fences and Fine Detail
DLSS 4.5 demonstrates a remarkable leap forward in its ability to reconstruct complex and fine geometric detail, an area where previous upscalers often struggled. This is most evident in the rendering of foliage, such as trees and grass. In games with dense forests like MafiThe Old Country, the telltale shimmering and fizzing of tree canopies during motion is significantly reduced. The internal structure of tree branches appears more stable and coherent, resolving more detail without introducing aliasing. This improvement in temporal stability makes swaying trees and wind-swept fields look far more natural and less distracting, creating a more immersive visual experience.
This newfound stability extends to other challenging geometric patterns, like mesh fences. In stress tests, such as those found in Spider-Man 2, where overlapping fences could become a blurry, aliased mess with DLSS 4.0, the new model in DLSS 4.5 provides a far clearer and more stable reconstruction. The individual wires of the fence are better defined, and the clarity of objects seen through the fence is vastly improved. In many such scenarios, the image quality of DLSS 4.5’s Performance mode can easily be mistaken for, or even exceed, that of DLSS 4.0’s Quality mode, showcasing the powerful reconstruction capabilities of the underlying transformer model.
Nuanced Changes to Textures Stability and Hair
While some areas see dramatic improvements, the changes to general texture quality and fine-detail stability are more nuanced. DLSS 4.5 generally produces a perceptibly sharper image across the board, which can enhance the appearance of surface details and textures, especially when a scene is in motion. This extra sharpness often contributes to a more detailed look, but it is not a universal improvement. In some specific instances, such as on certain metallic floor gratings in Cyberpunk 2077, this increased sharpness can introduce moiré patterns that were not present with the slightly softer reconstruction of DLSS 4.0.
Regarding overall stability on fine lines and edges, DLSS 4.5 can be seen as more of a sidegrade. While it often holds an edge, there are scenarios where the older model produces a more stable image with less flicker on distant, high-frequency details. Hair rendering, however, sees a subtle but consistently positive improvement. Benefiting from the overall increase in sharpness and better temporal stability, hair often appears less grainy and more solid with DLSS 4.5. The individual strands seem better resolved, giving characters’ hair a more realistic and higher-resolution appearance, a benefit that is most apparent at lower internal resolutions like those used in Performance mode.
A New Approach to Lighting Particles and Transparencies
Nvidia has altered its rendering pipeline with DLSS 4.5, with the new model now performing its reconstruction calculations in linear light space before the final tone mapping is applied. This technical change is intended to preserve the full color and luminance range of the scene, and the subjective result is that bright highlights often appear more vibrant and impactful. Light sources like neon signs, specular reflections, and weapon effects in games like Cyberpunk 2077 tend to pop more, creating a visually richer image. However, this approach comes with a tangible trade-off: an increased propensity for highlight flicker. The older method of processing in logarithmic space naturally dampened this kind of instability, and while the new transformer model is powerful, it does not completely mitigate the issue, resulting in more noticeable flicker on bright elements during motion.
The impact on particle effects and transparencies is marginal. Particle quality sees a slight benefit from the reduction in ghosting, but the reconstructed detail and resolution of the particles themselves are largely indistinguishable from DLSS 4.0. Similarly, transparent elements like fire, smoke, and holographic displays, which were already a strong point for DLSS 4.0, continue to look excellent with the new model but do not show any significant leap in quality. These areas were clearly not the primary focus of the new training data, with efforts concentrated on the more problematic aspects of temporal reconstruction.
Practical Application and Known Limitations
Putting DLSS 4.5 into practice is a straightforward process for users. The new model presets are accessed through the Nvidia App, which allows for game-specific profiles. In the Graphics tab, a user can select a title and, under the Driver Settings, find the “DLSS Override – Model Presets” option to set Super Resolution to Preset M. This can also be applied as a global setting for all supported games. It is important to note, however, that the default “Recommended” setting will only apply the new DLSS 4.5 model to the Performance and Ultra Performance modes, reverting to the older Preset K for Balanced and Quality modes. Users must manually select Preset M to ensure the new model is active across all settings.
Despite its impressive capabilities, DLSS 4.5 arrives with a significant limitation: it is currently incompatible with Nvidia’s Ray Reconstruction technology. Ray Reconstruction is another AI-driven feature designed to denoise and improve the quality of ray-traced lighting, and when it is enabled, it utilizes a combined AI model that also handles the Super Resolution upscaling. This combined model is, for now, still based on the older DLSS 4.0 architecture. Consequently, players face a choice in games with heavy ray tracing implementations. They must either disable Ray Reconstruction to benefit from the superior motion clarity and foliage rendering of DLSS 4.5, or they must use Ray Reconstruction for cleaner lighting at the cost of falling back to DLSS 4.0’s upscaling model.
The Trade Offs Where DLSS 4.5 Falls Short
While DLSS 4.5 marks a clear step forward, it is not without its own set of challenges and occasional regressions. The most frequently observed issue is the introduction of an overly sharp aesthetic in certain games. While the increased sharpness generally enhances detail, it can sometimes exaggerate high-frequency information, leading to a grainy or aliased appearance on certain textures, particularly grass and foliage, that was not present with the slightly softer DLSS 4.0. This can give parts of the image a processed look that some users may find less natural.
Another specific regression is the appearance of moiré patterns on fine, repeating textures like metal grates or mesh. This artifact, which creates a shimmering, wavy pattern, seems to be a side effect of the new model’s aggressive approach to detail reconstruction. In these specific instances, the older DLSS 4.0 model provides a more stable and artifact-free image. The most significant challenge, however, remains the substantial performance cost on older hardware. The heavy frame rate penalty on RTX 20 and 30 series GPUs makes DLSS 4.5 an impractical choice for a large segment of the user base, potentially limiting its widespread adoption until those users upgrade to newer, FP8-capable architectures.
The Future Trajectory of AI Upscaling
The advancements brought forth by DLSS 4.5 provide a clear indication of the future trajectory for AI-powered rendering technologies. By dramatically improving the visual fidelity of lower-resolution modes, Nvidia is making presets like 4K Performance more viable and appealing than ever before. This trend effectively pushes the boundaries of what is considered a playable frame rate at high resolutions, allowing gamers to prioritize fluidity without a severe compromise in image quality. The underlying message is that with a sufficiently intelligent AI model, the internal rendering resolution becomes less critical to the final perceived quality, shifting the focus from raw pixel counts to the sophistication of the reconstruction algorithm.
This progress also sets a new, higher bar for competing technologies. With AMD’s FidelityFX Super Resolution (FSR) on the cusp of its next major update, the stage is set for a renewed competition in the upscaling space. The industry will be watching closely to see how FSR 4 and other solutions compare to the stability and reconstruction prowess of DLSS 4.5’s transformer model. Looking ahead, it is likely that future DLSS models will continue to leverage more complex AI architectures, potentially integrating AI-driven processes even deeper into the rendering pipeline to address everything from lighting and shadows to character animation, further blurring the line between rendered and reconstructed graphics.
Conclusion A Meaningful Step Forward With Caveats
The analysis confirmed that DLSS 4.5 stands as a significant and commendable evolution of Nvidia’s upscaling technology. For owners of GeForce RTX 40 and 50 series GPUs, the new model is a recommended upgrade. Its profound improvements in handling disocclusion, ghosting, and complex foliage result in a tangibly cleaner and more stable image in motion. These benefits, particularly in third-person games or titles with dense natural environments, often outweigh the minor performance tradeoff and occasional regressions like highlight flicker or moiré patterns.
However, the technology’s value proposition is highly dependent on the user’s hardware. For those on older RTX 20 and 30 series graphics cards, the heavy performance penalty associated with the lack of native FP8 support makes DLSS 4.5 a much less attractive option. In many cases, the significant drop in frame rates can negate the purpose of upscaling entirely, making the well-established DLSS 4.0 a more balanced and practical choice. Ultimately, DLSS 4.5 is a net positive that successfully makes lower-resolution upscaling modes more visually compelling than ever, but its benefits are, for now, best realized on Nvidia’s most recent architectures.