01 · THE MECHANISM
You are not stretching an image. You are accumulating one.
Modern upscaling is temporal: the frame you see is assembled from several previous frames, aligned using the engine’s own motion vectors, then resolved into one image at output resolution.
That single fact explains almost every artefact you have ever blamed on an upscaler. Because reconstruction depends on history, anything that breaks history breaks the image:
- Ghosting — history that should have been rejected got kept, usually behind a fast-moving object.
- Shimmer and crawl — thin geometry such as fences, wires and foliage never survives long enough in history to resolve.
- Smearing — effects drawn without motion vectors, like many particles, transparencies and some shadows, have no history to align.
- Disocclusion trails — the area revealed behind a moving object has no accumulated data at all, so it must be invented.
The generation you run matters more than the brand you run. A modern model on a five-year-old GPU beats an old model on a new one, which is why the single most valuable thing in this guide is the section on forcing a newer model into an older game.
02 · THE FIELD
Who can run what, as of September 2026.
Hardware support is the first filter, and it eliminates most of the debate before quality even comes up.
| Technology | Runs on | Approach | Practical position |
|---|---|---|---|
| DLSS 4 / 4.5 (NVIDIA) |
GeForce RTX 20 Series and newer for super resolution; frame generation is limited to newer generations | Transformer-based super resolution and ray reconstruction, running on tensor cores | The quality reference. The transformer model reaches back to the 20 Series, which is unusually generous for a vendor feature. |
| FSR 4 (AMD) |
RDNA 4 cards such as the RX 9070 and 9070 XT; earlier cards fall back to FSR 3.1 | AMD’s first machine-learning upscaler, using the AI units introduced with RDNA 4 | The biggest generational jump of the three. The catch is that the good version needs current hardware. |
| XeSS 2 (Intel) |
Best on Intel Arc via XMX units; a DP4a path runs on most other modern GPUs | Machine-learning reconstruction with a vendor-neutral fallback, plus XeLL latency reduction and XeFG frame generation | The useful middle option. Often the best choice on a non-RDNA-4 AMD card in a game with no good FSR build. |
| TSR / TAAU (engine-provided) |
Any GPU | Engine-side temporal upscaling, notably Unreal’s Temporal Super Resolution | The floor. Better than it gets credit for at high internal resolutions, clearly behind the others at low ones. |
Vendor pages move faster than guides. Model versions, per-generation frame-generation limits and the list of supported cards all change with driver releases. Check the current vendor documentation before you plan a purchase around any row above.
03 · THE ARITHMETIC
Your quality preset is just a render resolution.
Quality, Balanced and Performance are not opinions. They are fixed scaling factors, and knowing the number tells you immediately whether a setting will look acceptable.
| Mode | Scale | Renders at, for 1080p | for 1440p | for 4K |
|---|---|---|---|---|
| Native / DLAA | 1.00x | 1920 × 1080 | 2560 × 1440 | 3840 × 2160 |
| Quality | 0.67x | 1280 × 720 | 1706 × 960 | 2560 × 1440 |
| Balanced | 0.58x | 1114 × 626 | 1485 × 835 | 2227 × 1253 |
| Performance | 0.50x | 960 × 540 | 1280 × 720 | 1920 × 1080 |
| Ultra Performance | 0.33x | 640 × 360 | 853 × 480 | 1280 × 720 |
One rule falls straight out of that table: judge the internal resolution, not the label. Performance mode at 4K renders a full 1080p frame and usually looks excellent. Performance mode at 1080p renders 540p and usually does not. The same word means two very different pictures.
A practical starting point
- 4K display — start at Performance. You are reconstructing from 1080p, and the extra frames are worth more than the difference to Quality.
- 1440p display — start at Quality, drop to Balanced only if you need the frames.
- 1080p display — Quality only, and consider running the anti-aliasing mode instead (DLAA, or native with the upscaler’s AA) if you have headroom.
- Anything below 720p internal — stop and lower a different setting. Shadows and volumetrics usually cost more than the resolution you are trying to save.
04 · GENERATED FRAMES
Frame generation is a smoothness feature, not a performance feature.
Upscaling makes the frames cheaper. Frame generation inserts frames that no input ever touched. The difference decides when each one is a good idea.
A generated frame is interpolated between two rendered frames, which means the renderer must hold a frame back before it can produce one. You gain motion smoothness and you pay latency. Multi-frame generation extends this further — recent DLSS versions push the ratio up to several generated frames per rendered one, which multiplies both the benefit and the cost.
| Base frame rate | Frame generation | Why |
|---|---|---|
| Below 40 fps | Leave it off | The latency penalty lands on an already-slow input path, and interpolation artefacts are most visible when the gap between frames is large. |
| 50 to 70 fps | Good case | This is what the feature is for: a high-refresh display fed by a frame rate that cannot reach it natively. |
| Above 100 fps | Optional | Useful for a 240 Hz panel, pointless below that. Diminishing visual return, same latency cost. |
| Competitive shooters | Off | You want the lowest latency, not the smoothest picture. Always. |
Always pair frame generation with the vendor’s latency reduction — Reflex on NVIDIA, Anti-Lag on AMD, XeLL on Intel. Without it you are adding queue depth on purpose. The input-lag guide covers the whole chain and the frame-cap arithmetic that goes with it.
05 · THE DECISION
Pick in this order, and stop when you reach a yes.
Almost every “which is better” argument dissolves once you sort by what the game actually offers and what the card actually accelerates.
- Does the game offer the upscaler your GPU accelerates in hardware? Take it. An NVIDIA card takes DLSS, an RDNA 4 card takes FSR 4, an Arc card takes XeSS 2. This resolves most cases immediately.
- Is the offered version old? Check which model the game ships. A 2022 game with an early DLSS or an FSR 2 build is a candidate for a newer DLL or a driver-level override before you judge its quality.
- Non-RDNA-4 AMD card, and FSR looks poor? Try XeSS 2 if the game has it. The vendor-neutral path frequently beats an older FSR build on the same hardware.
- Plenty of headroom already? Use the anti-aliasing mode rather than the upscaling mode. DLAA and its equivalents render at native resolution and use the same reconstruction to clean the image — the best-looking option in most games.
- Nothing available looks right? Lower shadows, volumetrics or ray-tracing steps first. They usually buy more frames than another notch of upscaling, and they do not damage motion clarity.
Forcing a newer model into an older game
This is the highest-value trick in the whole category and it is free. NVIDIA’s app exposes a DLSS override that swaps a game’s bundled model for the current one; DLSS Swapper does the same by replacing the DLL per game. Because the transformer super-resolution model runs on every RTX card since the 20 Series, an old game and an old GPU both benefit. Expect less shimmer on thin geometry and noticeably less ghosting behind moving objects.
Do this before buying anything. Overriding the model in the five games you actually play costs an evening and changes more than most upgrades in this price range.
06 · MISTAKES
Six settings people get wrong, in rough order of frequency.
Upscaler sharpening, an in-game slider and a driver-level filter all at once produces the crunchy over-processed look people then blame on the upscaler. Pick one, keep it low.
That is a 540p frame. Use Quality, or lower a different setting. No reconstruction model recovers detail that was never sampled.
It smooths what is already acceptable. Below roughly 40 fps it makes the game feel worse while the counter goes up.
Two games with the same setting name can ship models years apart. Check and override before concluding an upscaler looks bad.
Upscaling reduces the render resolution but not texture memory. If you are stuttering from a full frame buffer, lower texture quality — upscaling will not save you.
Every artefact that matters is temporal. Judge in motion, ideally while turning the camera past a fence or a stand of trees.
07 · QUICK ANSWERS
Upscaling on PC, briefly.
On image quality, DLSS with the current transformer model is generally ahead, and it runs on every GeForce RTX card since the 20 Series. FSR 4 closed much of the gap but requires RDNA 4 hardware such as the RX 9070 series; on older AMD cards you are running FSR 3.1, which is a different product. In practice the answer is decided by which one your GPU accelerates.
Look at the internal resolution rather than the label. At 4K, Performance mode renders a full 1080p frame and usually looks excellent. At 1080p, Performance renders 540p and usually does not, so stay on Quality. A useful floor is to avoid any mode that renders below about 720p internally.
Yes. A generated frame is interpolated between two rendered frames, so the renderer holds a frame back to produce it. Always enable the vendor latency reduction alongside it, and skip frame generation entirely below roughly 40 fps or in competitive shooters.
No. DLSS runs on NVIDIA tensor cores. On AMD use FSR, on Intel use XeSS, and if a game offers only DLSS, a translation layer such as OptiScaler can map the game DLSS calls onto FSR or XeSS instead.
Use the DLSS override in the NVIDIA app, or swap the DLL per game with a tool such as DLSS Swapper. Because the current super-resolution model runs on every RTX card since the 20 Series, this usually improves shimmer and ghosting in older titles at no performance cost.