DLSS, FSR and XeSS Explained: How Upscaling Works and Which Mode to Leave It On
Upscaling (DLSS, FSR, XeSS) renders the game at a lower resolution and reconstructs it with AI up to your monitor’s: you gain real frames. Frame generation is a different thing — it invents in-between images and raises latency rather than lowering it. Turn upscaling on almost always; generation only above 60 FPS.
That distinction is most of the story, and it’s exactly the one the boxes work hard to blur. Below is why each one behaves the way it does.
What upscaling actually does
The underlying idea is old: fewer pixels cost less work. If your monitor is 1440p and the game draws internally at 1080p, the GPU computes 56% of the pixels. That’s the saving, and it’s real.
What’s new is the reconstruction. A classic scaler stretched the image and left it blurry. DLSS, FSR 4 and XeSS use a neural network that receives, besides the current frame, the previous ones and the motion vectors from the game engine — meaning it knows where every object is heading. With that it recovers detail that wasn’t in the low-resolution frame but was in the earlier ones.
That’s why it can look better than native on edges, fences and thin cables. It isn’t a sharpening filter: it’s information accumulated over time. And it fails when that information is useless — something appearing out of nowhere, particles, water, shifting foliage.
The modes, and what they really render
This is the number that decides whether you’ll like upscaling. The mode percentages apply to each side of the image, not to the pixel total. NVIDIA documents them as 66% for Quality, 58% for Balanced, 50% for Performance and 33% for Ultra Performance. FSR and XeSS use near-identical scales:
| Mode | From 1080p | From 1440p | From 4K | Actual pixels |
|---|---|---|---|---|
| Quality | 1280×720 | 1707×960 | 2560×1440 | 44% |
| Balanced | 1114×626 | 1485×835 | 2227×1253 | 34% |
| Performance | 960×540 | 1280×720 | 1920×1080 | 25% |
| Ultra Performance | 640×360 | 853×480 | 1280×720 | 11% |
Look at the 1080p row. In Performance mode your game is drawn at 540p, and no neural network fixes that: there’s no detail to recover because it was never computed. Which gives you the rules that matter:
- 4K: any mode down to Performance, which is internally 1080p.
- 1440p: Quality always, Balanced if you need it.
- 1080p: Quality only. Upscaling works better the higher the output resolution.
- Ultra Performance: built for 8K. Ignore it.
If your card can’t hold 1440p even on Quality, the setting isn’t the problem: see which GPU makes sense for 1440p before you keep fighting with menus.
Frame generation: where the number on the box misleads
This is a separate technology and worth not mixing up. Upscaling makes the GPU work less to produce real frames. Frame generation (MFG) takes two real frames and inserts invented images between them. The FPS counter jumps; how it feels to play, less so.
The reason is mechanical: to interpolate between frame A and frame B, the GPU needs B already finished, so it holds it back for an instant before showing it. That instant is latency. And the invented frames don’t read your mouse: not one of them contains a decision of yours.
Tom’s Hardware measured input latency in Cyberpunk 2077 at 4K with path tracing, on an RTX 5070 with DLSS 4.5:
| Frame generation | Input latency |
|---|---|
| Off | 35.0 ms |
| 2X | 46.6 ms |
| 4X | 53.2 ms |
| 6X | essentially the same as 4X |
Off to 4X is 18 ms more, even as the FPS counter multiplies. In their RTX 5080 test at 4K it was 24 ms without generation and 34 ms with 4X. The same outlet puts around 60 ms as the point where image and controller start to feel decoupled in a single-player AAA game.
And here’s the paradox: generation is sold as the rescue for someone stuck at 30 FPS, and that’s where it works worst. With a low base the real frames sit further apart, so the algorithm invents more and guesses wrong more often — TechSpot notes artifacts become obvious in complex scenes, with unpredictable motion or fine detail like foliage. And latency, already bad, gets worse. It works well once you already have 60-80 real FPS and want to fill a 144 or 240 Hz monitor: precisely when you need it least.
Reflex doesn’t fix this, helpful as it is: it trims the render queue, it doesn’t make an invented frame respond to your mouse. Leave it on, but don’t expect it to offset generation.
What you actually have, by card
| Upscaling | Frame generation | |
|---|---|---|
| NVIDIA | DLSS 4.5 with the 2nd-gen transformer model, on all RTX cards (20, 30, 40 and 50) and 400+ games and apps | MFG up to 6X: RTX 50 only |
| AMD | FSR 4.1 on RX 9000 (RDNA 4) and, since June 2026, on RX 7000 (RDNA 3). RX 6000 promised for 2027 | RDNA 4 only (RX 9000) |
| Intel | XeSS 2 on Arc with XMX cores; earlier versions in 250+ games | XeSS 3 with MFG 2X-4X on Arc and Core Ultra. XeSS FG also runs on GeForce GTX 10 and Radeon RX 5000 and newer |
Three things in that table that aren’t obvious:
- The new DLSS model doesn’t switch itself on. Plenty of games still load the old one; you change it in the NVIDIA app, under DLSS Override. If you own an RTX card and have never touched it, you’re missing the biggest quality jump in two years, and it’s free.
- AMD split upscaling from generation. With an RX 7000 you’ve had the good upscaler since June 2026, but AI frame generation stays exclusive to RDNA 4.
- XeSS is the wildcard. Its frame generation runs on older NVIDIA and AMD cards left out of DLSS and FSR 4. On a GTX 1660 or an RX 5700, it’s your only route.
On which looks better: DLSS with the transformer model is still ahead on fine detail and motion stability. FSR 4 closed a lot of the gap — Digital Foundry placed it above DLSS’s old CNN model — but the difference exists if you compare still and zoomed in. While playing, almost nobody would spot it. It isn’t a buying criterion on its own: the rest of the card matters more, as we break down in NVIDIA vs AMD.
DLSS 5 isn’t “DLSS 4 but better”
The number confuses people. DLSS 5, which arrived on 3 September 2026, is not a new upscaler: it’s 3D-guided neural rendering, a technique that improves lighting and materials. It doesn’t give you frames, it costs them. And it’s on RTX 50 and, for now, in one game (NBA 2K27).
If you own an RTX 40, you are not missing the new upscaler — the DLSS 4.5 transformer is yours. What you’re missing is a lighting feature almost nobody uses yet.
Decision table
| Your situation | What to turn on |
|---|---|
| Gaming at 4K and short on FPS | Upscaling on Performance. Best quality-per-frame trade there is |
| Gaming at 1440p | Upscaling on Quality, always. Visually near-free |
| Competitive shooter | Upscaling yes, generation no. Latency rules |
| Single-player and already at 60+ FPS | Generation yes, to fill the monitor |
| Sitting at 30-40 FPS | Generation no. Drop settings or resolution; this is where it behaves worst |
| You don’t know your base FPS | Measure first: how to monitor FPS and temperatures free |
When it isn’t worth it
Three cases where the honest answer is no:
- Going from 30 to 60 FPS with frame generation. The counter will say 60; the controller will say 30, with lag.
- As a reason to buy. Picking a card because “it has 6X MFG” is paying for a multiplier that only works if the card already performs. Look at raw performance and VRAM; if the maths don’t add up, check the best value graphics cards.
- If the bottleneck is the CPU. Upscaling relieves the GPU. If your frames are capped by the processor, lowering internal resolution changes nothing: the diagnosis is in what is a bottleneck.
And something that saves money: upscaling cuts the pixels computed, but not VRAM in the same proportion. It eases the render buffers; textures load the same, and textures are what fill the memory. If your card runs out, this won’t rescue it — that’s what how much VRAM you need is for.
Frequently asked questions
Which is better, DLSS or FSR? DLSS with the transformer model looks slightly better on fine detail and motion stability. FSR 4 is very close and the difference is hard to see while playing. In practice you don’t choose: you use whichever your card supports.
Does frame generation add real FPS? It adds displayed frames, not performance. The inserted ones don’t respond to your input, so the game looks smoother but responds as before, with a bit more latency.
Does turning on DLSS hurt image quality? On Quality at 1440p or 4K, almost never noticeably, and on edges and fine detail it improves. On Performance at 1080p it shows, because the game renders at 540p.
Can I use DLSS on a GTX 1060? No: DLSS needs the tensor cores in RTX cards. Your alternative is XeSS, which runs on GeForce GTX 10 and newer, or FSR in its non-AI version.
Should I enable upscaling even with FPS to spare? You don’t need to for performance, but at 4K it usually pays off: lower power draw and temperatures for a minimal visual cost. And check your monitor is set up properly.
Sources consulted
- NVIDIA Developer — internal render resolutions of the DLSS modes (accessed October 2026)
- Tom’s Hardware — hands-on with DLSS 4.5 Dynamic Multi Frame Generation and 5X/6X latency (2026)
- Tom’s Hardware — official FSR 4.1 support for Radeon RX 7000 (RDNA 3) (June 2026)
- TechSpot — a review of NVIDIA’s DLSS 4 Multi Frame Generation
- PC Gamer — XeSS 2 with Frame Generation and XeLL
- NVIDIA — DLSS 5 and 3D-guided neural rendering (September 2026)