DLSS

DLSS, short for Deep Learning Super Sampling, is an NVIDIA technology that renders a game at a lower internal resolution and reconstructs a sharper, higher-resolution image with a neural network running on the GPU’s Tensor Cores. Among graphics settings in gaming, it is one of the most effective ways to raise frame rates without a large loss in visual quality, especially when demanding features such as ray tracing are enabled.

Over successive versions, the name DLSS has grown into a family of features: super resolution (upscaling), DLAA (anti-aliasing at native resolution), frame generation and ray reconstruction. A game’s DLSS menu can therefore control several different things, which is a frequent source of confusion.

How DLSS reconstructs the image

DLSS super resolution is a temporal upscaler. Each frame is rendered at a reduced resolution with a tiny sub-pixel camera jitter, so that consecutive frames sample slightly different points of the scene. The game also passes motion vectors and depth information to DLSS, which tell it where every pixel moved since the previous frame.

The neural network combines the current low-resolution frame with accumulated detail from previous frames, decides which history is still valid and which must be rejected, and produces the final output image. Because it gathers samples over time, the result can contain more detail than any single low-resolution frame, and in some scenes it rivals or exceeds native rendering with conventional anti-aliasing.

The first version from 2019 needed a network trained per game and was widely criticized for blur. DLSS 2 in 2020 introduced the generalized temporal approach that defines the technology today, and DLSS 4 in 2025 replaced the earlier convolutional network with a transformer model that improves stability and fine detail at the cost of more computation.

DLSS quality modes and internal resolution

Games usually expose DLSS as a set of quality modes, each tied to a fixed scale factor per axis. The lower the internal resolution, the higher the frame rate and the more the network has to reconstruct. The table shows the standard factors and what they mean for a 4K output of 3840 by 2160 pixels.

Quality mode renders only about 44% of the output pixels, and Performance mode only 25%, which explains the large frame rate gains in GPU-limited games. The higher the output resolution, the more information the network receives, so Performance mode at 4K looks far better than at 1080p, where the internal image is only 960 by 540 pixels.

Components of the DLSS family

Each DLSS component solves a different problem and has its own hardware requirements. Super resolution and DLAA run on every GeForce RTX graphics card, while frame generation features depend on hardware added in later generations.

Frame generation raises the displayed frame rate, but the generated frames do not respond to player input. Input latency therefore follows the rendered frame rate, and frame generation is normally paired with NVIDIA Reflex to keep latency under control.

DLSS compared with FSR and XeSS

Two competing upscalers follow the same basic principle. AMD’s FSR and Intel’s XeSS also render at a lower resolution and use temporal data, but they differ in how the reconstruction is computed and in which hardware they run on.

In side-by-side comparisons, DLSS has generally been regarded as the image quality reference, particularly for fine geometry and stability in motion, while FSR’s older versions had the advantage of running on almost any GPU. The gap varies by game and by implementation quality.

Choosing DLSS settings and avoiding pitfalls

The right mode depends on the output resolution, the GPU’s headroom and the target frame rate. A good rule is to start with Quality mode and move down only if the frame rate is still too low.

• At 1080p, prefer DLAA or Quality, because lower modes reconstruct from very few pixels.
• At 1440p, Quality or Balanced usually gives the best compromise.
• At 4K, Performance mode is often visually acceptable and doubles headroom.
• Enable frame generation only when the base frame rate is already about 60 fps or higher, to limit latency and artifacts.

Typical artifacts include ghosting behind fast-moving objects, shimmering on thin details such as fences, and smearing on particles that lack proper motion vectors. These are often caused by the game’s integration rather than by DLSS itself. DLSS also cannot help in CPU-limited situations, because lowering the render resolution does not reduce the processor’s workload.

A non-technical view

DLSS lets a graphics card draw a smaller picture and then enlarge it intelligently. Instead of simply stretching the image, it remembers details from the previous moments of the game and fills in what is missing.

The result is that the game runs noticeably faster while still looking sharp. Think of an artist who sketches quickly at a small size and then uses notes from earlier drafts to paint a detailed large version. You get speed with most of the quality kept.

A concrete example

A player with a mid-range RTX graphics card connects it to a 4K television and starts a recent action game with ray-traced reflections. At native 4K with ray tracing, the game runs at about 30 frames per second and feels sluggish.

The player switches DLSS super resolution to Performance mode, so the game renders internally at 1080p. The frame rate rises to around 55 to 60 fps. Enabling ray reconstruction cleans up noisy reflections, and turning on frame generation with Reflex lifts the displayed rate to roughly 100 fps on the television’s 120 Hz mode.

From a normal viewing distance, the image remains close to native 4K in sharpness, and only small trails behind fast particles reveal the reconstruction.