Why modern games render below your monitor’s resolution
Native 4K is brutally expensive. A game that renders every pixel at 3840 by 2160 has to shade roughly four times as many pixels as it does at 1080p, and that cost climbs again if you push a 120 Hz or 144 Hz refresh rate. Developers have spent years squeezing more performance out of each frame, but there is a hard physical ceiling: at some point you cannot render more pixels without either lowering settings or buying faster hardware.
Upscaling and frame generation exist to sidestep that ceiling. The core trick is to render fewer pixels and then reconstruct or invent the rest, so a game looks close to native resolution while running significantly faster. Three competing implementations dominate the PC market today: NVIDIA DLSS, AMD FSR, and Intel XeSS. They share the same broad goal but differ sharply in how they work, which GPUs they run on, and how they handle the trade-offs involved.
This article explains what each technology does, how upscaling differs from frame generation, and what you should check before toggling one of them on. Every technical claim below is drawn from the three vendors’ own documentation, linked inline and collected at the end.
The shared idea: render low, reconstruct high
Upscaling starts by rendering the game at a lower internal resolution, then uses past frames, motion information, and in many cases a trained neural network to rebuild the image at your display resolution. The reason this works better than a naive resize is temporal data: because a moving scene changes only slightly from frame to frame, an upscaler can combine samples across several frames to recover detail that a single low-resolution frame cannot contain on its own.
This differs from lowering your resolution. Lowering resolution to 50 percent and letting the monitor stretch the result produces a soft, aliased image. An upscaler that tracks motion can recover edges, text, and fine geometry far more convincingly, which is why a quality-mode upscale can look close to native while still saving a large amount of rendering work.
Frame generation is a separate idea layered on top. Instead of reconstructing a frame at higher resolution, a frame generator inserts entirely new, AI-generated frames between the frames the GPU renders, raising the displayed frame rate without raising the rendering cost. This is why the two features are often bundled but must be understood independently: upscaling improves performance by reducing work, while frame generation improves smoothness by adding frames, and it carries its own distinct trade-off in the form of added latency.
NVIDIA DLSS
NVIDIA describes DLSS as “a revolutionary suite of neural rendering technologies that uses AI to boost FPS, reduce latency, and improve image quality.” The name covers more than one feature. Super Resolution is the upscaler, reconstructing a higher-resolution image from a lower-resolution render using a neural network that runs on the Tensor Cores inside GeForce RTX GPUs. DLAA applies the same reconstruction technique at native resolution to improve image quality rather than performance, and Ray Reconstruction uses the network to clean up the noisy output of real-time ray tracing.
On top of upscaling, NVIDIA offers Frame Generation, which inserts an AI-generated frame between rendered frames, and Multi Frame Generation, which can generate “up to five frames per rendered frame” on GeForce RTX 50 Series GPUs and fifth-generation Tensor Cores. NVIDIA’s DLSS 4.5 release introduced Dynamic Multi Frame Generation alongside a second-generation transformer model for Super Resolution and Ray Reconstruction, and DLSS 5 added what NVIDIA calls 3D-Guided Neural Rendering, which “infuses game scenes with lifelike lighting and materials” and is tuned by developers per title. DLSS is the most hardware-bound of the three: its neural features run on NVIDIA Tensor Cores, so they are available only on GeForce RTX GPUs.
See NVIDIA’s DLSS technology page for the full feature breakdown.
AMD FSR
AMD’s FidelityFX Super Resolution, or FSR, took a different path. Earlier versions such as FSR 3.1 used a hand-tuned spatial and temporal algorithm rather than a trained neural network, which let them run across a much wider range of hardware, including GPUs from other vendors. FSR 4 changes that: it is machine-learning based, and AMD positions it as delivering “a significant image quality improvement over AMD FSR 3.1 upscaling,” with an ML algorithm designed to improve temporal stability, preserve detail, and reduce ghosting.
FSR 4 keeps a fixed set of quality modes that control how aggressively the source image is scaled. Native AA renders at 100 percent resolution for improved quality, Quality mode scales by 1.5x per dimension (67 percent screen resolution), Balanced by 1.7x (59 percent), Performance by 2.0x (50 percent), and an optional Ultra Performance mode by 3.0x (33 percent). Each step trades image quality for a larger performance gain. AMD also notes that FSR 4 no longer requires the reactive and transparency masks that earlier versions asked developers to supply, which simplifies integration.
FSR 4 can be combined with AMD’s existing frame generation and with AMD Radeon Anti-Lag 2 to reduce latency, which AMD describes as delivering “ultra-smooth, ultra-responsive gaming at incredible frame rates on AMD Radeon RX 9000 Series graphics cards.” The technology is distributed through AMD’s open GPUOpen program as prebuilt signed DLLs, and it includes an Unreal Engine 5 plugin spanning versions 5.1 through 5.6. See AMD’s FSR 4 page on GPUOpen for the mode table and integration details.
Intel XeSS
Intel’s Xe Super Sampling, now at version 3, is a set of “real-time AI-based technologies that drastically boost your frame rate at the highest visual quality while keeping your game responsive.” It is split into three parts. XeSS Super Resolution (XeSS-SR) is the upscaler and, notably, runs on all GPUs with Shader Model 6.4 DP4a support, which covers a wide range of hardware beyond Intel’s own. XeSS Frame Generation (XeSS-FG) inserts generated frames and runs on discrete and integrated Intel Arc GPUs as well as non-Intel GPUs with SM 6.4 support. Xe Low Latency (XeLL) reduces input lag and can be enabled independently of Super Resolution, while Frame Generation requires XeLL to be integrated. Intel also publishes plugins for Unreal Engine and Unity to make integration easier. See the Intel XeSS SDK repository for the full developer documentation.
A concrete example: what Quality mode renders
The scale factors matter because they determine how much work the GPU saves. Using FSR 4’s published table as a reference, a 4K display running Quality mode renders internally at 2560 by 1440, or 67 percent of native, before upscaling to 3840 by 2160. Performance mode drops the internal resolution to 1920 by 1080, or half of native, and the optional Ultra Performance mode renders at 1280 by 720, one third of native. The gap between the internal and output resolution is the performance headroom the feature buys, and the visual quality of the reconstruction is what separates a good upscaler from a simple stretch. The same logic applies to the other vendors even though they do not always publish their scale factors as plainly.
Frame generation and the latency problem
Frame generation creates a subtle but important trade-off. An inserted frame is displayed before the player’s input for that moment exists, so a generated frame does not make the game more responsive, and in practice it can add a small amount of latency. This is why all three vendors pair frame generation with a latency-reduction feature: NVIDIA Reflex, AMD Anti-Lag 2, and Intel XeLL. The intended use case is a game that already runs at a comfortable base frame rate, where the extra generated frames mainly add smoothness rather than responsiveness.
The practical consequence is that frame generation is a poor substitute for a low base frame rate. If a game runs at 30 frames per second before frame generation, the added frames will make motion look smoother but will not fix the underlying responsiveness problem, and the extra latency can even make fast-paced play feel worse. Treat upscaling as the tool for recovering performance, and frame generation as the tool for adding smoothness once performance is already acceptable.
How the three compare
The most consequential difference is hardware compatibility. DLSS’s neural features require NVIDIA Tensor Cores, so they are locked to GeForce RTX GPUs. FSR has historically been the most portable option, running on a broad range of GPUs including those from other vendors, though FSR 4’s machine-learning upscaler is aimed at AMD’s latest Radeon hardware. XeSS splits the difference: its Super Resolution runs broadly through the DP4a path on SM 6.4-capable GPUs, while Intel Arc GPUs get the faster dedicated-acceleration path.
The quality-mode naming is also inconsistent enough to matter. NVIDIA, AMD, and Intel each label their presets differently, and “Quality” does not mean the same internal resolution everywhere. FSR 4 publishes its scale factors explicitly, which is useful: Quality scales by 1.5x per dimension, Performance by 2.0x. When comparing performance claims between games or reviews, the mode name alone is not enough; the underlying render resolution is what determines how much work is being saved.
Beyond raw compatibility, the choice usually comes down to which vendor’s stack is best integrated into the games you play. DLSS is the most deeply embedded in high-end PC titles with ray tracing, FSR is the most open and console-relevant, and XeSS offers a middle path that works broadly while giving Intel hardware the best results. None of the three is universally the right answer, and a given game may support one, two, or all three.
What to check before you enable it
First, confirm which of the three technologies your specific game supports, because support is per title and per feature. A game can ship with FSR upscaling but not FSR frame generation, or with DLSS Super Resolution but not Multi Frame Generation, so the toggle in the settings menu is the authoritative source. Second, match the feature to your GPU: DLSS neural features need an RTX GPU, XeSS frame generation needs SM 6.4 or Arc hardware, and FSR 4’s upscaler is aimed at current Radeon cards. Third, check your base frame rate before enabling frame generation; if it is low, prefer a more aggressive upscaling mode or lower settings first.
Finally, be aware of the known visual and behavioral limits. Upscaling can still introduce artifacts such as ghosting on fast-moving objects or flickering on thin, bright detail, which is exactly the class of problem the vendors describe their newer models as reducing. Frame generation adds latency even when a low-latency mode is active, so it is not a good fit for every competitive setting. These are documented, expected trade-offs rather than bugs, and the right configuration is a balance between image quality, responsiveness, and frame rate rather than a single best setting.
