DLSS vs FSR vs XeSS 2026: Which AI Upscaler Should You Use?

If your GPU can handle 60fps at your target resolution but you want 100fps or better, AI upscaling is the answer — and in 2026, three serious contenders fight for your settings menu: DLSS 4 (NVIDIA), FSR 4 (AMD), and XeSS 2 (Intel). Which one you should enable is mostly decided by your hardware, but quality mode selection, frame generation choices, and game support status all matter too.

This guide explains how each upscaler works, compares image quality at matched settings, and delivers a GPU-by-GPU decision table so you’re not guessing. For context on where upscaling fits in your full graphics config, the game settings explained guide covers everything from resolution scaling to shadow quality. NVIDIA users will also want to pair their upscaler choice with the right NVIDIA Control Panel settings — there are driver-level toggles that interact directly with DLSS behaviour.

What AI Upscaling Actually Does

Standard rendering is simple: your GPU draws every pixel at your output resolution. At 4K that’s 8.3 million pixels per frame; at 120fps, over a billion pixels per second. AI upscaling breaks this down by rendering at a lower internal resolution — typically 50–77% of your target — then running an AI model to reconstruct the full-resolution frame. The model draws on motion vectors (how objects moved since the last frame), depth data, and previous frames to fill in the missing detail.

The maths of why this works: resolution scales as a square. Drop from 4K (3840×2160) to 67% render scale (2572×1447) and you’re rendering just 44% of the pixels — your GPU does roughly half the work on the geometry and shading pass. The AI reconstruction adds overhead, but on modern hardware with dedicated matrix units (NVIDIA Tensor Cores, AMD ML cores, Intel XMX), that overhead is a fraction of the saving. The practical result is 30–80% more fps depending on mode, GPU, and the game engine.

What separates good upscaling from poor is how cleanly the model handles three problem areas: ghosting on fast-moving objects, shimmering on thin geometry like fences and power lines, and temporal stability when panning quickly across complex environments. This is where DLSS 4, FSR 4, and XeSS 2 differ — and where hardware generation matters most.

DLSS 4: NVIDIA’s Transformer Upscaler

DLSS 4 is NVIDIA’s neural rendering suite, exclusive to GeForce RTX graphics cards. The super resolution component uses a transformer AI model — the same architectural shift that drove progress in large language models — replacing the recurrent neural network used in DLSS 3. The practical result: better temporal stability, sharper fine detail on hair and foliage, and noticeably less ghosting on disoccluded surfaces. [1]

DLSS 4 also includes Ray Reconstruction (for ray-traced scenes), DLAA (AI anti-aliasing at native resolution), and two distinct frame generation technologies that depend on which RTX generation you own.

DLSS 4 Quality Modes

Each mode sets how low the internal render resolution goes before the AI reconstructs it to your output resolution. Lower render resolution means more fps, at a quality cost:

Frame Generation vs Multi Frame Generation

Frame Generation (RTX 40 series and later) generates one AI-synthesised frame between each rendered pair, effectively doubling your display frame rate. The trade-off is input latency: the generated frame doesn’t contain new input data, so controls feel slightly less responsive. Always enable NVIDIA Reflex when running Frame Generation — it won’t eliminate the latency cost but reduces it significantly. The hard rule: don’t enable Frame Generation below 60fps base. Artefacts and judder become pronounced at lower frame rates. [4]

Multi Frame Generation (RTX 50 series only) generates up to three additional frames per rendered frame. DLSS 4.5’s Dynamic Multi Frame Generation — available from 31 March 2026 — adjusts the multiplier automatically based on scene complexity, delivering up to a 6x frame count increase on flagship RTX 50-series hardware. Hardware Flip Metering shifts frame pacing to the display engine for smoother delivery, and Reflex 2 meters queued frames dynamically to keep input latency manageable. [5]

FSR 4: AMD’s Open Upscaler for Every GPU

FSR 4 — now called FSR Upscaling as part of AMD’s FSR Redstone suite — is the most versatile upscaler in 2026. It runs on any modern GPU that supports DirectX 12, making it the default choice for hardware that can’t access DLSS or XeSS hardware acceleration paths. The Redstone suite bundles FSR Upscaling with FSR Frame Generation, FSR Ray Regeneration, and the new FSR Radiance Caching. [2]

FSR 4 on RDNA 4 vs Older Hardware

The hardware tier you’re on determines which upscaling path actually runs:

• RDNA 4 (Radeon RX 9000 series): Full ML-accelerated FSR Upscaling 4.1 using dedicated matrix units in the RDNA 4 architecture. Quality at this tier is very close to DLSS 4 at matched presets. Ghosting on disoccluded surfaces and particle handling have improved dramatically over FSR 3. [2]
• RDNA 3 (RX 7000 series) and RDNA 2 (RX 6000 series): AMD’s driver automatically falls back to FSR 3.1.5 analytical upscaling — no ML acceleration. Solid and reliable, but noticeably softer than the ML tier at the same quality preset. Community tool OptiScaler offers an unofficial FSR 4 INT8 path for RDNA 3 that brings partial ML improvements, though it requires manual setup.
• NVIDIA and other non-AMD GPUs: Same analytical fallback. FSR works on any GPU — but on NVIDIA hardware that supports DLSS, always choose DLSS instead.

FSR 4 Quality Modes

FSR 4 uses the same mode names as DLSS at roughly matching render scales:

XeSS 2: Intel’s Cross-GPU Upscaler

Intel’s XeSS 2 (Xe Super Sampling 2) takes the broadest hardware compatibility approach of the three. Like FSR, it runs on any modern GPU — but it splits into two quality tiers based on whether your card has Intel’s dedicated XMX matrix hardware. [3]

On Intel Arc GPUs, XeSS 2 uses XMX (Matrix Extension) cores — Intel’s equivalent of NVIDIA’s Tensor Cores. On Arc hardware, XeSS delivers image quality that reviewers place close to FSR 4 on RDNA 4: clean fine detail, good temporal stability, and solid handling of complex geometry across most game engines.

On non-Intel hardware — NVIDIA GeForce and AMD Radeon — XeSS falls back to DP4a integer math instructions, available on any GPU supporting Shader Model 6.4. This covers NVIDIA GTX 10-series through RTX 40-series and AMD RX 6000/7000 series. The DP4a path produces quality on par with FSR 3 analytical mode on the same hardware — a genuine option, but not competitive with ML-accelerated DLSS or FSR 4 on RDNA 4. [3]

XeSS 2 offers the widest quality mode range of the three upscalers, including Ultra Quality Plus (77% render scale) and Native AA — useful for competitive players who want AI anti-aliasing without sacrificing render resolution. Around 50 games support XeSS 2 as of early 2026, with the XeSS 2 SDK now fully open to developers and XeSS 3 in the pipeline. [3]

Head-to-Head Image Quality: DLSS 4 vs FSR 4 vs Native

Comparing upscalers is hardware-dependent — the same game at the same quality preset will look different depending on your GPU. At 67% render scale on matched hardware:

DLSS 4 on RTX 40/50 leads overall. The transformer model reconstructs fine detail cleanly on hair, wire fences, and fast-moving foliage. Temporal stability under rapid camera movement is best-in-class. Minor ghosting can still appear on very fast particles in some titles, but this is the reference standard at this render scale.

FSR 4 on RDNA 4 has closed the gap significantly. At a single-frame screenshot level, FSR 4 Quality on an RX 9000 series card is hard to distinguish from DLSS 4 Quality on an RTX 40 card in most scenes. In motion, the most visible difference surfaces on disoccluded geometry — where a fast object reveals background that wasn’t visible in the previous frame. DLSS 4 handles this slightly more cleanly, but for most games and most scenes, the difference is subtle. [2]

XeSS 2 on Intel Arc sits close to FSR 4 on RDNA 4 for most content. Wide surfaces and large textures are sharp and temporally stable. Thin geometry — power lines, chain-link fences, dense foliage at distance — can show more shimmering than DLSS 4 on RTX hardware, particularly at Balanced mode or lower.

FSR 3 / XeSS DP4a on older hardware is a clear tier below the ML upscalers. Quality is still better than no upscaling, but at 67% render res you’ll see soft edges and fringing in fast-moving scenes. On this tier, lean toward Performance mode only when you genuinely need the frames — otherwise Quality mode offers the best balance of fps and clarity.

Which Upscaler Should You Use? GPU Breakdown

The decision is mostly made by your hardware. The game settings optimisation guide covers how to pair this choice with GPU driver settings and in-game quality configs for the full setup:

One important edge case: if you have a GTX-era NVIDIA GPU and a game only supports DLSS, you’re locked out entirely — DLSS requires RTX Tensor Cores and won’t run on GTX hardware. FSR is the safety net in that scenario, assuming the game also supports it.

Frame Generation: When to Use It (and When Not To)

Frame generation of any kind — DLSS Frame Gen, Multi Frame Gen, or FSR Frame Generation — produces synthetic frames that your GPU never fully rendered. The AI interpolates between two real frames, which means generated frames don’t contain new input data. Every mouse movement or button press after the last real frame is invisible to the generated frames sitting between them. That’s the input latency cost.

At high base frame rates this latency is imperceptible. At low base frame rates it becomes noticeable. Minimum base fps before enabling frame generation, by game type:

• Single-player games (RPGs, open world, narrative): 45–50fps base. Below this, artefacts on moving edges become visible even in cutscenes.
• Fast-action combat (FromSouls, hack-and-slash, action RPG): 60fps base. At lower fps, generated frames create artefacts on fast weapon swings that break visual clarity exactly when you need it most.
• Competitive multiplayer (shooters, MOBAs): 80–100fps base minimum. The latency cost at lower base fps hurts reaction time more than the fps gain helps your performance.

The underlying rule: frame generation multiplies what you already have — it doesn’t rescue poor performance. If you’re struggling to hit 30fps, frame gen won’t save the experience. If you’re comfortably sitting at 60fps and want 100fps for a smoother feel, frame gen is the right tool. Think of it as the final fps multiplier after upscaling has already done the heavy lifting.

Game Support: How to Check What Your Game Uses

Not every game supports every upscaler, and many support more than one. DLSS 4 is available in over 250 games. FSR (across all versions) covers the widest range — hundreds of titles from FSR 1 through FSR 4, with 28 new FSR 4 titles added in a single AMD update in early 2026. XeSS 2 is the most limited at around 50 games as of early 2026, though the SDK is fully open and game support is growing steadily.

How to check what a game supports in 30 seconds:

1. Open the game’s Graphics or Display Settings menu
2. Look for an “Upscaling,” “Super Resolution,” or “Image Scaling” dropdown — DLSS, FSR, and XeSS will appear here if supported
3. Older games may list these under “Anti-Aliasing” rather than a dedicated upscaling category
4. If nothing appears, check whether the game supports FSR via community tool OptiScaler, which can inject FSR 4 or DLSS 3 Super Resolution into unsupported titles

Where a game supports multiple upscalers, the hardware rule applies: RTX — pick DLSS; RDNA 4 — pick FSR 4; Intel Arc — pick XeSS 2. For hardware older than RTX 30 / RDNA 3 / Arc, FSR is the most consistent choice across titles because its analytical fallback runs on everything and produces reliable results even without ML acceleration. [1] [2]

Recommended Settings by GPU Tier

Upscaler, quality mode, and frame generation recommendation combined into a single reference table:

FAQ

Is DLSS or FSR better for the RTX 4060?

DLSS 4, without question. The RTX 4060 has NVIDIA Tensor Cores that run DLSS’s transformer model natively — FSR on NVIDIA hardware uses the analytical fallback path, which produces noticeably softer results at matched quality presets. In any game that supports both, enable DLSS. For titles that only support FSR, FSR Performance mode is a reasonable fallback on the 4060 at 1440p.

Does FSR work on NVIDIA GPUs?

Yes — FSR runs on any modern GPU including all NVIDIA GeForce cards. On NVIDIA hardware, FSR uses the analytical (non-ML) upscaling path rather than AMD’s RDNA 4 acceleration. In practice, FSR on an RTX 4060 produces quality comparable to FSR 3 analytical mode — better than no upscaling, but in any game that supports DLSS, DLSS will look noticeably better. Use FSR on NVIDIA only when DLSS isn’t available in that specific title. [2]

Which upscaler has the least ghosting in 2026?

DLSS 4 on RTX hardware leads. The transformer model specifically addresses ghosting on disoccluded surfaces — where fast-moving objects reveal previously hidden background pixels. FSR 4 on RDNA 4 is a close second; AMD made reducing ghosting a headline improvement in the RDNA 4 ML path, and the results are clearly visible in motion testing. XeSS 2 on Intel Arc handles ghosting well in most content but can show more artefacts in high-velocity scenes. On older hardware using analytical paths (FSR 3, XeSS DP4a), ghosting is more pronounced — the main practical reason to stay at a higher quality mode with more render resolution rather than pushing Ultra Performance for maximum fps.

Sources

1. NVIDIA. DLSS 4 — AI Neural Rendering Technology. NVIDIA Corporation.
2. AMD. FidelityFX Super Resolution 4 (FSR 4) — Quality Modes and GPU Support. AMD GPUOpen.
3. Intel. XeSS Gaming Technology — Xe Super Sampling Explained. Intel Gaming Access.
4. NVIDIA. DLSS 4 Introduces Multi Frame Generation and AI Enhancements. GeForce News, NVIDIA.
5. Tom’s Hardware. NVIDIA Introduces DLSS 4.5 and Multi Frame Generation 6X at CES 2026. Tom’s Hardware.