Every Upscaler Explained: DLSS, FSR, and XeSS Head to Head

A few years ago, "upscaling" was a dirty word among PC gamers. It meant your console was faking a resolution it couldn't actually push. Today it is the single most important performance lever most of us have, and turning it off usually means leaving a third of your frame rate on the table for no visible gain. The trouble is that there are three competing systems with confusingly similar names, and the marketing does everything it can to blur the differences. Here is how DLSS, FSR, and XeSS actually work, and how I decide which one to switch on.

What an upscaler is actually doing#

Every one of these technologies solves the same basic problem. Rendering pixels is expensive, and the cost scales roughly with how many of them you draw. So instead of rendering your full 3840x2160 panel natively, the GPU renders a smaller internal image, say 2560x1440, and then reconstructs the missing detail up to your display's real resolution.

The naive way to do that is a simple stretch, which looks soft and blurry. Modern upscalers are far cleverer. They are temporal reconstructors: they pull detail from previous frames, using motion vectors from the game engine to figure out where each object moved between frames. Over several frames they accumulate enough sub-pixel samples to rebuild an image that can genuinely rival native resolution, and sometimes exceed it because the accumulation acts as a form of anti-aliasing.

That temporal nature is why all three share the same weaknesses. Fast motion, particles, transparent surfaces like foliage and chain-link fences, and anything the motion vectors don't describe well are where reconstruction gets confused. When you hear people complain about "ghosting," "shimmer," or "smearing," they are describing the temporal accumulation making a wrong guess.

DLSS: the Nvidia-only benchmark#

DLSS, Nvidia's Deep Learning Super Sampling, is the one everyone benchmarks against, and honestly it earns that position. The reconstruction is driven by a neural network that Nvidia trains on its own hardware, and the model runs on the dedicated Tensor cores built into RTX cards. That hardware dependency is the catch: DLSS only works on GeForce RTX GPUs. If you own a GTX card, an AMD card, or an Intel Arc card, DLSS is simply not an option.

Within its walled garden, though, it is the strongest of the three at equivalent settings. The image is cleaner in motion, holds fine detail better, and its handling of thin geometry is the best I have used. With the newer transformer-based model that shipped in the DLSS 4 era, the stability improvement over the old convolutional model is genuinely noticeable, especially the reduction in shimmer on fences and vegetation.

A few things worth separating out, because Nvidia bundles them all under the DLSS brand:

• Super Resolution is the core upscaler, the part that competes with FSR and XeSS.
• Frame Generation inserts entirely AI-generated frames between real ones. It boosts your on-screen FPS number but adds latency and does nothing for responsiveness. Treat it as smoothness, not performance.
• Ray Reconstruction replaces the denoiser in ray-traced games and can meaningfully clean up noisy reflections and lighting.

I keep Super Resolution and Frame Generation mentally in separate boxes. Frame Gen feels great in a slow, pretty single-player game and terrible in a twitch shooter.

FSR: runs on almost anything#

AMD's FidelityFX Super Resolution took the opposite philosophy. It is open, it is vendor-agnostic, and for most of its life it ran on nearly any modern GPU regardless of brand, including older Nvidia cards and even some integrated graphics. If you are on hardware that DLSS locks out, FSR has usually been your lifeline, and that alone makes it enormously valuable.

The spatial-versus-temporal split#

FSR's history matters here because the versions are wildly different in quality:

1. FSR 1 was purely spatial. It upscaled a single finished frame with no temporal data, which made it cheap and universal but visibly soft. Skip it if anything newer is offered.
2. FSR 2 added temporal reconstruction and closed most of the gap, becoming the genuinely usable version that shipped in hundreds of games.
3. FSR 3 kept improving reconstruction and added AMD's own frame generation, called Fluid Motion Frames.
4. FSR 4 moved to a machine-learning model, and this is the important shift. It looks dramatically better than FSR 3, but the ML path leans on newer AMD RDNA 4 hardware, so it partly gives up the "runs on everything" promise that defined the brand.

In practice, the older temporal versions of FSR still shimmer more than DLSS, particularly on fine detail and in motion. FSR 4 narrows that gap a lot. The honest summary is that FSR's ceiling has risen sharply, but its best results now want current AMD silicon.

XeSS: Intel's clever middle path#

Intel's XeSS is the one most people forget exists, and that is a shame because its design is smart. XeSS ships in two flavors from the same code. On Intel Arc GPUs it runs on the dedicated XMX AI cores and delivers its best image quality. On every other GPU it falls back to a more compatible instruction path (DP4a) that runs slower and looks a touch worse, but still works.

That means XeSS gives you something unique: a neural-network upscaler with real quality that you can run on an Nvidia or AMD card. When a game offers XeSS but not DLSS and only an old version of FSR, XeSS is frequently the better pick even on non-Intel hardware. On Arc cards specifically, the XMX version is the reason Intel's GPUs punch above their raw horsepower.

Its weaknesses mirror the others: some ghosting on fast transparent effects, and the fallback path costs more performance than the native one. But as a "works well almost everywhere" option, it has quietly become my default recommendation for mixed or older systems.

Reading the quality presets#

Whichever one you use, they all expose the same ladder of presets, and the names describe how aggressively they drop the internal render resolution. At a 4K output target the rough internal resolutions look like this:

• Quality renders around 1440p internally. This is the sweet spot and where I start every time.
• Balanced renders a bit lower and trades a little clarity for more frames.
• Performance renders around 1080p internally, a big FPS jump with visible softening.
• Ultra Performance renders around 720p internally. It is a last resort for pushing 4K on a weak card, and it shows.

The single most important thing to understand is that these ratios are fixed to your output resolution. "Quality" at 4K reconstructs from 1440p and looks superb. "Quality" at 1080p reconstructs from only 720p, and there simply isn't enough source information, so it looks noticeably rougher. Upscalers are at their most convincing at 4K and get progressively harder to hide the lower your monitor's native resolution goes.

How I actually choose#

After enough hours of squinting at fences and waterfalls, my decision tree is boringly simple.

• On an RTX card: use DLSS Super Resolution, almost always on Quality. There is little reason to reach for anything else on Nvidia hardware. Add Ray Reconstruction in ray-traced titles, and treat Frame Generation as an optional smoothness topping for slower games only.
• On a current AMD card: prefer FSR 4 where the game supports it. If a title only has an older FSR version and also offers XeSS, try XeSS and compare.
• On an Intel Arc card: use XeSS. The XMX path is exactly what the card was built to feed.
• On older or mixed hardware DLSS won't run on: XeSS first, then whatever FSR version the game provides.

Two habits are worth building regardless of brand. First, compare at Quality before dropping lower, because the jump from native to Quality is usually invisible while the jump from Quality to Performance is where artifacts start to show. Second, if a game shipped with an old, blurry version of an upscaler, check whether you can swap in a newer DLL. Because these systems ship as replaceable library files, updating one is often a matter of dropping a newer file into the game folder, and it can meaningfully improve an old title.

The bottom line#

There is no universal winner, only the best tool for the card in your case. DLSS remains the quality leader if you already own RTX hardware, XeSS is the surprisingly capable option that runs almost anywhere and shines on Arc, and FSR is the broad-compatibility workhorse whose newest ML version has finally closed most of the quality gap. What unites all three is that turning one on at Quality is nearly free performance, and refusing to use them on principle just means a slower, uglier game for no reward. Pick the one your GPU runs best, leave it on Quality, and spend the reclaimed frames on something you'll actually notice.