DLSS

Rasterization and aliasing

When a frame is displayed on a screen, it appears as a two-dimensional image composed of a grid of pixels that simulates three-dimensional depth. However, in applications such as video games, this image is not created directly in two dimensions. It is generated from a mathematical description of a three-dimensional space. In a video game each frame begins as data describing a three-dimensional scene containing objects, surfaces, lighting, and a virtual camera. The GPU converts this information into the two-dimensional grid of pixels shown on the screen, in a process called rasterization. The image is said to be rendered when the final image appears on the target display device.

Because pixels form a fixed rectangular grid, they cannot reproduce curved or diagonal edges perfectly. These edges can therefore appear jagged or stair-stepped, an effect called aliasing.

Anti-aliasing: MSAA, supersampling, and DLSS

Traditional anti-aliasing techniques reduce jagged edges by taking additional samples (measurements at different locations within a pixel) when determining how pixels should appear. Multisample anti-aliasing (MSAA), for example, samples several points within a pixel to produce smoother edges.

Another approach is supersampling anti-aliasing (SSAA). In this technique the GPU renders a game at a resolution higher than the one being displayed and then shrinks the image to the target resolution. The extra information produces smoother edges and finer detail, but rendering all those additional pixels requires considerably more processing power.

Explore Britannica Premium!

The trusted destination for professionals, college students, and lifelong learners.

SUBSCRIBE

DLSS takes a different approach. It renders the image at a lower resolution than the final image and uses an AI model to reconstruct a higher-resolution image from the lower-resolution rendering. Instead of rendering an image at a higher resolution and then downscaling it to the target resolution, DLSS begins with a lower-resolution rendering. For example, DLSS might render a game internally at 1,920 × 1,080, about 2 million pixels, while producing a 3,840 × 2,160, or 4K, image containing about 8 million pixels. Its AI model uses information from the rendered image and patterns learned during training to estimate the additional detail needed for the higher-resolution output. Because the GPU initially renders far fewer pixels, the game can run faster.

DLSS

Released in February 2019, DLSS approximated the visual quality of supersampling while maintaining high performance. NVIDIA trained the model on high-resolution reference images provided for each individual game. The trained network would then process the game in real time, inferring missing detail frame by frame to produce a higher-resolution output. The technology rendered only a fraction of the pixels required for the target resolution but used deep learning to draw on the patterns of shading, texture, color, and light in its training data to infer how to scale each lower-resolution frame to a higher resolution.

The biggest limitation of this early release was that it relied on game-specific training data and required game developers to provide high-quality visuals in advance. It also did not effectively manage information from previous frames, which led to inconsistent image quality with noticeable blurring and loss of detail.

DLSS 2

DLSS 2, released in March 2020, significantly improved both image quality and performance by introducing a more advanced, generalized approach to AI upscaling. It was able to produce images comparable to native resolution (an image rendered directly at the output resolution) while rendering frames with only a fraction of the pixels required for the target resolution, using information from previous frames to accumulate detail across multiple frames and improve stability.

Unlike the first version, DLSS 2 no longer required per-game training. NVIDIA trained a single, generalized neural network with a wide range of high-quality reference images, allowing it to work across different games without needing developer-specific input.

DLSS 2 also improved efficiency, making better use of Tensor Cores to deliver higher frame rates, broader compatibility across RTX GPUs, and support for more resolutions. It introduced multiple user-selectable modes—“Quality,” “Balanced,” and “Performance”—to render the game internally at different resolutions, allowing users to balance visual quality and performance.

DLSS 3

DLSS 3, released in October 2022, expanded the technology beyond upscaling by introducing frame generation, allowing it to create an entirely new frame between two traditionally rendered frames for smoother motion and higher apparent frame rates. The system analyzes how objects and individual pixels are moving between the surrounding frames and uses that information to construct the intermediate image.

To maintain responsiveness, DLSS 3 was paired with NVIDIA Reflex, which reduces system latency by synchronizing CPU and GPU workloads, ensuring that the higher frame rates do not come at the cost of input lag.

DLSS 3.5

DLSS 3.5, released in September 2023, introduced a feature called ray reconstruction, which improves the quality of ray-traced lighting. Instead of relying on traditional techniques to clean up noisy ray-traced images, DLSS 3.5 uses a trained neural network to reconstruct more accurate lighting, reflections, and shadows from the limited information produced by ray tracing while continuing to benefit from the performance gains introduced in DLSS 3.

DLSS 4

DLSS 4, released in January 2025, introduced a major upgrade to the underlying AI models, marking the most significant improvement to the technology since DLSS 2. While retaining basic upscaling and features such as frame generation and ray reconstruction, this version enhanced how these systems operate by adopting a new class of neural networks known as transformers.

Transformers allow DLSS to better analyze patterns across frames, improving temporal stability (stability between frames) and reducing common visual glitches. This results in sharper detail, particularly in motion, where earlier approaches struggled to maintain consistency. DLSS 4 also improved the efficiency of its frame generation model, increasing performance while reducing memory usage on supported GPUs.

DLSS 4.5

DLSS 4.5, released in January 2026, further improved both image reconstruction and frame generation, building on the transformer-based models introduced in DLSS 4. Upscaling quality was significantly enhanced through a second-generation transformer AI model, trained using substantially greater computational resources, resulting in sharper detail, reduced visual glitches, and improved stability between frames.

This release also expanded frame generation capabilities, enabling the creation of multiple intermediate frames—up to six times the number of traditionally rendered frames—to further boost performance. At the same time a new Dynamic Frame Generation system was introduced to apply frame generation selectively, using AI-generated frames only where beneficial and relying on fully rendered frames where visual accuracy is critical.

DLSS 5

In March 2026 NVIDIA’s CEO Jensen Huang announced DLSS 5, and it was released in September that year. The technology built on the transformer-based models and frame generation techniques introduced in earlier versions. Notably, DLSS 5 incorporated generative AI for the first time, with a more direct role in producing final image output rather than only reconstructing or interpolating frames.

DLSS 5 drew criticism from some gamers, who argued that, rather than appearing more photorealistic, the technology produced visuals resembling AI-generated filters, particularly in displaying character faces. These reactions reflected broader concerns that generative AI might interfere with the artistic intent of developers. The criticism also emerged amid wider skepticism toward AI in gaming, alongside ongoing concerns about rising costs of hardware components such as RAM and graphics cards.