Over the past decade, real-time graphics rendering has faced an increasingly difficult challenge: delivering higher visual fidelity while maintaining playable performance. Technologies like ray tracing and path tracing dramatically improve realism, but they also push GPUs to their limits.
To address this problem, GPU vendors (NVIDIA, AMD, and Intel) have introduced neural rendering and image reconstruction technologies designed to not only improve performance — via temporal upscaling — but also improve visual smoothness — via frame generation/interpolation — and denoising of ray/path-traced effects, via machine learning-powered denoising.
Three major technology suites dominate this space today:
What started as simple spatial upscalers has evolved into full ML-based rendering stacks featuring:
In this roundup, we’ll explore how each technology evolved, what each version introduced, and how they compare in terms of capabilities and hardware support.
Before diving into each vendor’s implementation, it’s worth briefly explaining the key techniques these technologies use.
The earliest technique was used in many early implementations.
The GPU renders the game at a lower resolution (for example, 1440p instead of 4K), then uses spatial upscaling/resampling algorithms — such as nearest neighbor interpolation, Lanczos resampling, or even ML-based neural networks — to upscale the image to the target resolution.
Examples:
Pros:
Cons:
Modern solutions rely heavily on temporal upscaling (also called temporal reconstruction or temporal upsampling), which combines (or “accumulates”) data from multiple, previous lower-than-native resolution frames.
The algorithm analyzes:
This enables the reconstruction of a higher-resolution image with much better quality than spatial methods.
Examples:
Pros:
Cons:
Frame generation inserts interpolated frames (either analytically or with AI/ML) between traditionally rendered ones, dramatically increasing perceived smoothness.
For example:
The technique analyzes motion vectors, depth buffer data, and optical flow in order to estimate what the interpolated frames should look like.
Examples:
Pros:
Cons:
Modern ray-traced games require denoising algorithms to reconstruct the final image from sparse ray samples.
Instead of having multiple “hand-tuned” denoisers, recently emerged ML-based models can now perform this step more efficiently.
Examples:
Pros:
Cons:
NVIDIA introduced Deep Learning Super Sampling (DLSS) in 2018 alongside the RTX 20 Series (Turing) GPUs and their dedicated hardware units called Tensor Cores.
DLSS uses neural networks trained on NVIDIA supercomputers to reconstruct images from lower-resolution renders.
Over time, DLSS evolved from a fancy ML-powered spatial upscaler into a full-on AI-driven neural rendering suite.
What follows is a chronological overview of NVIDIA’s DLSS releases, charting the technology’s evolution from its debut in 2018 to the most recent versions available today.
DLSS 1.0 marked NVIDIA’s first attempt at bringing AI-powered upscaling to real-time gaming. Introduced alongside the GeForce RTX 20 Series in 2018, the technology used neural networks accelerated by Tensor Cores to reconstruct higher-resolution images from lower-resolution renders, improving performance while aiming to preserve image quality. However, the first implementation relied largely on single-frame reconstruction and required per-game training of the DLSS neural network, which limited adoption and often produced highly subpar visual results.
DLSS 1.0 is widely considered a rough first step.
Key features:
Supported GPUs:
Limitations:
DLSS 2.0 marked a major overhaul of NVIDIA’s temporal upscaling technology. It introduced a generalized AI model — based on a Convolutional Neural Network (CNN) — and a new temporal reconstruction approach that combines data from multiple lower-resolution frames to produce sharper and more stable images while significantly boosting performance. DLSS 2 effectively turned the technology into an AI-driven TAAU (Temporal Anti-Aliasing Upscaling) system.
Game adoption increased rapidly after this version.
It’s also worth noting that before DLSS 3 arrived, NVIDIA continued refining its DLSS 2 temporal upscaling model through numerous incremental updates. These improvements were often delivered via updated DLL files, gradually improving image stability and detail retention without a full generational overhaul.
Key innovations:
Supported GPUs:
DLSS 3 introduced the first major expansion beyond temporal upscaling: Frame Generation. Alongside DLSS Super Resolution (which was previously called DLSS 2.0), the new technology uses motion vectors and NVIDIA’s Optical Flow Accelerator to generate entirely new frames (via interpolation) between traditionally rendered ones, significantly increasing perceived smoothness, at the cost of extra latency and some visual artifacts. This feature debuted with the GeForce RTX 40 Series GPUs and marked a shift from simple image reconstruction towards AI-assisted frame interpolation in game rendering.
New features:
Supported GPUs:
DLSS 3.5 shifted NVIDIA’s focus toward improving ray-tracing quality with a new feature called Ray Reconstruction. Instead of relying on multiple hand-tuned denoisers, DLSS 3.5 introduced a single AI model trained on large datasets to denoise ray-traced lighting and effects more accurately. This approach allows games to produce cleaner reflections, global illumination, and shadows while maintaining relatively high performance.
Benefits:
Supported GPUs:
DLSS 4 represents NVIDIA’s next major leap in AI-assisted rendering, building upon previous iterations with a new transformer-based neural model for both Super Resolution and Ray Reconstruction, not to mention the introduction of Multi Frame Generation (MFG). Unveiled alongside the GeForce RTX 50 Series, the MFG technology can generate multiple AI-interpolated frames (up to 3) for every traditionally rendered frame, dramatically boosting visual smoothness at the cost of even more visual artifacts and latency vs the previous DLSS Frame Generation technology.
DLSS 4 also upgrades Super Resolution and Ray Reconstruction with new transformer-based AI models. Replacing the previous Convolutional Neural Network (CNN) architecture, these models better capture complex image relationships, improving temporal stability, preserving fine surface detail, and reducing artifacts such as ghosting in motion and noisy ray-traced lighting. However, this did come at the cost of slightly lower performance vs the previous CNN-based models.
DLSS 4 also introduces an improved frame generation model that is faster and more efficient than previous implementations. The new approach reduces VRAM usage while increasing performance, and no longer relies on the dedicated Optical Flow Accelerator used in DLSS 3, instead generating the optical flow data directly through neural networks running on an RTX GPU’s Tensor Cores.
New features:
Supported GPUs:
DLSS 4.5 builds on NVIDIA’s latest neural rendering stack with further improvements to both image reconstruction and frame generation. The update introduces a second-generation transformer model for Super Resolution, improving temporal stability, edge detail, and motion clarity across hundreds of supported games. It also expands Multi Frame Generation with dynamic scaling and higher multipliers —up to 6× frame generation on RTX 50-series GPUs — allowing the system to automatically adjust the number of AI-generated frames to maintain smooth performance.
It’s also worth noting that DLSS 4.5 isn’t a complete overhaul of NVIDIA’s neural rendering stack. While Super Resolution moves to a second-generation transformer model, Ray Reconstruction has not yet received the same upgrade, meaning the visual improvements are limited to the upscaling component of DLSS. Moreover, the new DLSS 4.5 Super Resolution presets cannot currently be used alongside Ray Reconstruction, as enabling the latter forces the game to fall back to the older combined upscaling and denoising model.
New features:
Supported GPUs:
AMD FidelityFX Super Resolution (FSR) is AMD’s answer to NVIDIA’s DLSS, initially designed to boost gaming performance by rendering frames at lower resolutions and reconstructing them to higher output resolutions. First introduced in 2021, FSR takes a different approach from NVIDIA’s solution by prioritizing broad compatibility and an open ecosystem, allowing the technology to run on a wide range of hardware, including AMD, NVIDIA, and even Intel GPUs. Over time, the technology has evolved from a simple spatial upscaler into a more complete ML-based neural rendering stack that now includes temporal upscaling, frame generation, and even AI-assisted denoising of ray/path-traced effects in its latest iteration.
The following constitutes a chronological overview of AMD’s FSR releases, from its debut with FSR 1 in 2021 to FSR Redstone in 2025.
FSR 1.0 marked AMD’s first foray into real-time upscaling technologies for gaming. Released in 2021, the technology relies on a Lanczos resampling-based spatial upscaling algorithm that reconstructs higher-resolution images from a single lower-resolution frame, followed by a sharpening pass to enhance detail. Designed as an open and GPU-agnostic solution, FSR 1.0 could run on a wide range of GPUs from AMD, NVIDIA, and even some integrated graphics, helping it achieve rapid adoption across many PC games.
Features:
Pros:
Cons:
FSR 2 represented a major evolution of AMD’s upscaling technology, moving beyond the purely spatial approach of FSR 1 to a full temporal upsampling pipeline. Introduced in 2022, the new algorithm combines information from previous frames, motion vectors, and other rendering data — such as depth/color buffers and camera jitter/camera jitter removal — to reconstruct higher-resolution images with improved detail and anti-aliasing, delivering significantly better image quality than FSR 1 while maintaining broad GPU compatibility without requiring dedicated AI hardware.
Much like NVIDIA’s DLSS 2, AMD continued refining FSR 2 through several incremental updates —gradually improving temporal stability, ghosting/disocclusion behavior, and overall image quality before eventually transitioning to FSR 3.
Features:
Supported GPUs:
FSR 3 marked the next major step in AMD’s upscaling technology, building on the temporal upscaling pipeline introduced in FSR 2. The new version further improved image fidelity and stability while introducing FSR Frame Generation, AMD’s answer to NVIDIA’s DLSS Frame Generation. This feature uses analytical frame interpolation techniques — derived from AMD’s Fluid Motion Frames technology — to generate additional frames between rendered ones, dramatically boosting perceived visual smoothness (at the cost of latency and visual fidelity) in supported games.
Features:
Supported GPUs:
FSR 3.1 introduced several important refinements to AMD’s temporal upscaling/frame generation ecosystem, focusing on image quality improvements and greater flexibility for developers. One of the most notable changes was the decoupling of FSR Frame Generation from the upscaling component, allowing the frame generation technology to work alongside other temporal upscalers such as NVIDIA DLSS Super Resolution, Intel XeSS upscaling, or even a game’s own native temporal anti-aliasing upscaling (TAAU) solution. The update also improved temporal stability and reduced ghosting, delivering higher-quality upscaling compared to earlier FSR versions.
Features:
Supported GPUs:
FSR 4 marked AMD’s first major shift toward machine learning-powered upscaling, moving beyond the purely analytical approach used by previous FSR versions. The new technology introduced an ML-based Super Resolution model designed to improve reconstruction quality, temporal stability, and overall image fidelity compared to FSR 3.x. Shortly after its debut, AMD expanded the concept into a broader neural rendering stack called FSR Redstone, effectively rebranding FSR 4 as part of a larger suite of AI-driven neural rendering technologies rather than a standalone temporal upscaling solution.
Features:
Supported GPUs:
Intel Xe Super Sampling (XeSS) is Intel’s AI-driven suite of neural rendering technologies that are designed to improve gaming performance while maintaining high image quality. First introduced alongside Intel’s Arc A-Series (Alchemist) GPUs, XeSS upscaling renders games at a lower internal resolution and then reconstructs the final image using machine-learning models to approximate a higher-resolution output. Much like NVIDIA’s DLSS and AMD’s FSR, the goal is to deliver higher frame rates with minimal visual loss, but XeSS distinguishes itself by offering two execution paths: an AI-accelerated mode for Intel Arc GPUs using dedicated XMX units (equivalent to the Tensor Cores in NVIDIA GeForce RTX GPUs), and a fallback “DP4a” mode that allows the technology to run on GPUs from other vendors.
The following sections present a timeline of the various Intel XeSS technology releases, tracing its evolution from the original XeSS 1.0 launch in 2022 to the latest XeSS 3 version.
XeSS 1.0 marked Intel’s first entry into AI-assisted upscaling for real-time gaming. Introduced alongside the Intel Arc GPU architecture, the technology renders games at a lower internal resolution and reconstructs the final image using a machine-learning model to deliver higher performance with minimal loss in visual quality. A key differentiator of XeSS is its dual execution path, using dedicated XMX AI units on Intel Arc GPUs while also providing a fallback DP4a implementation that allows the technology to run on GPUs from other vendors.
It’s also worth noting that XeSS 1.0 continued to receive several incremental updates before the arrival of XeSS 2. Over time, Intel released improved versions such as XeSS 1.1, 1.2, and 1.3, gradually refining image quality, stability, and quality presets through SDK and DLL file updates without fundamentally changing the underlying upscaling model.
Features:
Supported GPUs:
XeSS 2 expanded Intel’s upscaling technology into a broader AI rendering stack, adding frame generation and latency reduction technologies alongside the existing Super Resolution upscaler. The update introduced XeSS Frame Generation (XeSS-FG), which uses AI-based frame interpolation to generate additional frames for visually smoother gameplay (at the cost of latency and visuals), as well as Xe Low Latency (XeLL) to reduce latency and improve responsiveness. Together with the original XeSS Super Resolution, these features form a more complete performance-enhancement pipeline similar to the feature sets offered by DLSS and FSR.
Intel also released XeSS 2.1 which later expanded the technology’s compatibility, enabling XeSS Frame Generation and Xe Low Latency to run on GPUs from other vendors such as AMD and NVIDIA.
Features:
Supported GPUs:
XeSS 3 represents Intel’s latest evolution of its AI-driven upscaling stack, introducing Multi Frame Generation (XeSS-MFG) to significantly boost visual smoothness. Building on XeSS 2’s frame-generation pipeline, the new XeSS version can insert multiple AI-interpolated frames (up to 3) between traditionally rendered ones.
Features:
Supported GPUs:
XeSS-MFG is supported on Intel Arc A-Series (Alchemist), B-Series (Battlemage), and Xe-based integrated GPUs in modern Intel Core Ultra processors.
The table below provides a quick side-by-side overview of the technologies discussed above, summarizing their core features, capabilities, and hardware support across each vendor’s ecosystem:
| Technology | DLSS | FSR | XeSS |
|---|---|---|---|
| Developer | NVIDIA | AMD | Intel |
| AI Hardware Required | Yes | No (pre-FSR 4/Redstone) Yes (post-FSR 4/Redstone) | No (DP4a path) Yes (XMX path) |
| Upscaling Type | ML-based spatial (DLSS 1) ML-based temporal (DLSS 2+) | Spatial (FSR 1) Analytical temporal (FSR 2-3) ML-based temporal (FSR 4/Redstone) | ML-based temporal |
| Frame Generation | Yes | Yes | Yes |
| ML-Based Denoising | Yes | Yes (starting with FSR 4/Redstone) | No |
| GPU Support | NVIDIA GeForce RTX GPUs | Most GPUs | Most GPUs |
| Open-Source? | No | Mostly | Partially |
VIA: wccftech.com
Η Sony Pictures δήλωσε ότι το πρώτο trailer για το "Spider-Man: Brand New Day" είναι…
Σύμφωνα με τη δέκατη έκθεση State of Marketing της Salesforce, το 75% των επαγγελματιών μάρκετινγκ…
Λάβετε το καθημερινό ενημερωτικό δελτίο Popular Science💡 Καινοτομίες, ανακαλύψεις και συμβουλές DIY που στέλνονται έξι…
Η Google θα κυκλοφορήσει νέες ενημερώσεις Υπηρεσιών Play και Play Store τον Μάρτιο του 2026,…
Αρχές από τις Ηνωμένες Πολιτείες, τη Γερμανία και τον Καναδά κατέστρεψαν την υποδομή Command and…
Η AbbVie Ελλάδας, η παγκόσμια βιοφαρμακευτική εταιρεία, κατακτά την 1η θέση στη λίστα Best Workplaces™…