What does DLSS or FSR upscaling actually do?

They render the game at a lower internal resolution and then reconstruct a higher-resolution image from it, so you get most of the visual quality at a fraction of the rendering cost.

Why that works better than it sounds. Naive upscaling — stretching a smaller image — looks soft and blurry. Modern upscalers use temporal reconstruction: they combine information from multiple previous frames, using motion vectors supplied by the game engine to know where every pixel moved between frames. Because each frame samples the scene slightly differently, several frames together contain genuinely more detail than any one of them. The upscaler is assembling, not inventing.

The main implementations:

DLSS (Nvidia) uses dedicated machine-learning hardware on RTX cards and is generally regarded as the highest quality, but only runs on Nvidia GPUs.

FSR (AMD) has historically been an open, hardware-agnostic approach that runs on almost anything including competitors' cards and consoles, with later versions adding machine learning.

XeSS (Intel) sits between the two, running best on Intel hardware but functional elsewhere.

Quality presets — typically Quality, Balanced, Performance, Ultra Performance — set how low the internal resolution goes. Higher output resolutions tolerate more aggressive upscaling, which is why Performance mode looks acceptable at 4K and poor at 1080p.

Known artefacts: ghosting or trailing behind fast-moving objects, shimmering on fine detail like fences and foliage, and softening of small text or HUD elements. These have reduced substantially across generations but are not gone.

Frame generation is a separate feature, often bundled with these. It inserts synthesised frames between rendered ones, which raises the frame counter and smoothness but does not reduce input latency — the inserted frames contain no new input. Worth knowing before treating the number as equivalent.