Deep Learning Super Sampling
Trademark
UnclaimedNVIDIA’s suite of real-time AI image reconstruction, upscaling, anti-aliasing, ray-reconstruction, and frame-generation technologies.
Last updated August 26, 2026
Overview
Deep Learning Super Sampling, commonly abbreviated as DLSS, is a family of real-time image-processing technologies developed by NVIDIA for video games and other graphics applications. Its central purpose is to improve the relationship between visual quality and rendering performance. Instead of rendering every part of a scene at the final display resolution, a supported application can render much of its graphics pipeline at a lower internal resolution and provide DLSS with information such as motion vectors, depth data, exposure data, and previous frames. An AI model then reconstructs an output intended to approach, and in some situations visually exceed, the detail and stability of conventionally rendered images at the target resolution. The saved GPU resources can be used to increase frame rates, enable more demanding effects, or support higher output resolutions. NVIDIA introduced DLSS as a major feature of its GeForce RTX 20-series graphics processors in 2018. The first implementation required game-specific neural-network training and received mixed reviews because some games showed softness, ghosting, or other reconstruction artifacts. NVIDIA substantially revised the technology with DLSS 2 in 2020. The newer approach used temporal information from multiple frames and a generalized model that did not require a separate basic training process for every game. DLSS 2 became available across a much broader range of titles and was integrated into engines including Unreal Engine and Unity. NVIDIA also introduced DLAA, which uses the AI anti-aliasing component at native resolution without upscaling. DLSS 3, associated with GeForce RTX 40-series hardware, added Frame Generation. This feature creates an additional intermediate frame between conventionally rendered frames using motion and optical-flow information. DLSS 3.5 expanded the suite with Ray Reconstruction, an AI-based approach intended to replace or supplement several conventional denoisers in ray-traced and path-traced scenes. DLSS 4, introduced alongside the GeForce RTX 50 series, adopted a vision-transformer-based model for upscaling and introduced Multi Frame Generation on supported RTX 50-series GPUs. The technology continued to distinguish between features supported by different hardware generations: basic DLSS functions span supported RTX generations, while frame-generation capabilities are more restricted. DLSS is primarily a software technology, but its performance and feature set depend on NVIDIA hardware. Most implementations use Tensor Cores and related accelerators in RTX graphics processors. A game must also implement the relevant DLSS software interface; installing a newer library alone cannot add support to a title that was not designed to use DLSS. As a result, DLSS operates both as a consumer-facing GeForce feature and as a developer technology adopted in individual games, engines, and graphics workflows. NVIDIA has promoted it as an important complement to hardware-accelerated ray tracing, since reconstruction and frame generation can offset some of the performance cost of advanced lighting. The technology’s market position rests on a combination of image quality, performance gains, developer adoption, and hardware differentiation. Its principal limitations include dependency on compatible games and GPUs, possible temporal artifacts, latency considerations when generated frames are used, and variation in quality between implementations and presets. Despite those limitations, DLSS has become one of NVIDIA’s most visible gaming technologies and a recurring feature in the company’s RTX product launches.
History
NVIDIA marketed DLSS as a defining capability of the GeForce RTX 20-series launch in September 2018. The original system used convolutional auto-encoder networks and was trained separately for each supported game. It combined a low-resolution frame with motion information and attempted to reconstruct a higher-resolution result. Early support included titles such as Battlefield V and Metro Exodus, but reviews were uneven. Observers reported softness and visual artifacts, and the game-specific training requirement limited the technology’s initial reach. Control later shipped with a related image-processing approach that approximated some DLSS behavior without relying on Tensor Cores in the same way. In April 2020, NVIDIA released DLSS 2 through a graphics-driver update. This version changed the technical approach substantially. Rather than relying primarily on a single frame and a game-specific model, DLSS 2 used temporal information from earlier frames, sub-pixel jitter, motion vectors, depth buffers, and exposure information. Its neural network was designed to handle recurring temporal reconstruction problems across different games, reducing the need for a separately trained model for every title. The technology could therefore be integrated more broadly into games and engines. Control and Wolfenstein: Youngblood were among early examples, followed by adoption in commercial game engines and many subsequent releases. DLSS 2 also clarified the distinction between upscaling and anti-aliasing. Temporal reconstruction can recover information from samples accumulated across frames, while traditional single-frame upscaling attempts to infer detail from the current image alone. NVIDIA’s DLAA mode applies the AI-based anti-aliasing process at native resolution, allowing users to seek improved edge quality without reducing the internal rendering resolution. DLSS 2 generally reduced processing overhead relative to the original implementation, although image quality continued to depend on motion, game assets, resolution, and implementation choices. DLSS 3, introduced with the Ada Lovelace generation and GeForce RTX 40-series products, added Frame Generation. The system uses two conventionally rendered frames and motion information to synthesize an intermediate frame. This can increase displayed frame rates, although generated frames do not replace the need for conventionally rendered frames and can introduce latency or visual errors in difficult scenes. The new feature used a newer optical-flow accelerator and was not made available in the same form on older RTX generations. NVIDIA announced DLSS 3.5 in 2023. Its Ray Reconstruction component targeted ray-traced and path-traced rendering by using a trained AI model in place of multiple conventional denoising processes. Early supported examples included Cyberpunk 2077 with its Phantom Liberty content, Portal with RTX, and Alan Wake 2. Unlike Frame Generation, Ray Reconstruction was positioned for availability across RTX generations, subject to software support. DLSS 4 arrived with the GeForce RTX 50 series. Its upscaling model moved from a convolutional neural-network design toward a vision-transformer architecture intended to improve temporal stability, reduce ghosting, and preserve detail during motion. Multi Frame Generation expanded the frame-generation concept by producing multiple interpolated frames for each traditionally rendered frame on supported RTX 50-series hardware. NVIDIA subsequently described DLSS 4.5 as adding Dynamic Multi Frame Generation, a newer transformer model, and broader controls for balancing image quality and computational cost. The brand has continued to evolve as a software layer linking NVIDIA GPU architecture, game-engine integration, and AI-based rendering.
- 2025
DLSS 4 introduces Multi Frame Generation
[1]The fourth generation adopted a transformer-based model and enabled multiple generated frames on RTX 50-series hardware.
- 2023
DLSS 3.5 adds Ray Reconstruction
[1]An AI denoising and reconstruction model was introduced for ray-traced and path-traced rendering.
- 2022
DLSS 3 adds Frame Generation
[1]NVIDIA added optical-flow-assisted generation of intermediate frames for RTX 40-series graphics processors.
- 2020
DLSS 2 released
[1]A generalized temporal reconstruction approach replaced the original game-specific model architecture and broadened developer adoption.
- 2018
DLSS introduced with GeForce RTX 20 series
[1]NVIDIA launched the first DLSS implementation as an RTX-era AI image reconstruction feature requiring game-specific training.
Products and positioning
Premium AI-assisted graphics technology for improving real-time image quality and performance on NVIDIA RTX hardware.
AI upscaling · 2018
DLSS 1
The original DLSS implementation used convolutional auto-encoder networks to enhance and upscale images. It relied heavily on game-specific training data and single-frame information, which could produce visible softness, hallucinated detail, and other artifacts in difficult scenes. Its principal value was reducing the internal rendering workload while targeting a higher output resolution.
Temporal reconstruction · 2020
DLSS 2
DLSS 2 uses information accumulated across frames, including motion vectors, depth, exposure, and sub-pixel samples, to perform temporal anti-aliasing upsampling. Its generalized model made integration practical across many games and engines. It also provides the foundation for DLAA, which applies the AI anti-aliasing process without upscaling.
Frame generation · 2022
DLSS 3
DLSS 3 combines DLSS 2-style reconstruction with Frame Generation. The frame-generation component synthesizes an intermediate image between rendered frames using motion and optical-flow information. It is associated with RTX 40-series hardware and can improve displayed frame rates, while its benefits and latency characteristics depend on the game and rendering workload.
Ray reconstruction · 2023
DLSS 3.5
DLSS 3.5 added Ray Reconstruction, an AI model intended to replace or improve several conventional denoising stages in ray-traced and path-traced scenes. The feature was designed to produce sharper, more stable, and less noisy lighting results and was supported on RTX hardware in compatible games.
Transformer reconstruction and frame generation · 2025
DLSS 4
DLSS 4 introduced a vision-transformer-based upscaling model and Multi Frame Generation. The transformer approach was designed to improve stability and reduce ghosting, while Multi Frame Generation produces more than one interpolated frame per conventionally rendered frame on supported RTX 50-series GPUs.
AI anti-aliasing · 2020
DLAA
Deep Learning Anti-Aliasing uses the AI anti-aliasing and reconstruction technology associated with DLSS while keeping the game at native rendering resolution. It is intended for users who prioritize edge quality and temporal stability over the performance gain created by upscaling.
Brand decisions
2022
Add Frame Generation in DLSS 3Product launchBackground
Higher-fidelity ray tracing and increasingly demanding game workloads created demand for additional frame-rate techniques.
Action taken
NVIDIA added optical-flow-assisted generation of intermediate frames and limited the feature to compatible RTX 40-series hardware.
Aftermath
DLSS became both an image-reconstruction system and a frame-rate-generation platform, with new discussions about latency and generated-frame artifacts.
2020
Replace the original approach with DLSS 2Generation changeBackground
The first generation’s game-specific training requirement and inconsistent image quality constrained its usefulness.
Action taken
NVIDIA released DLSS 2 with generalized temporal reconstruction using information from multiple frames and additional rendering buffers.
Aftermath
DLSS support expanded to more games and engines, establishing the version as the practical foundation for later DLSS features.
2018
Position DLSS as an RTX launch technologyProduct launchBackground
NVIDIA introduced its first real-time deep-learning upscaling system alongside the GeForce RTX 20-series architecture.
Action taken
The company marketed DLSS as a way to reduce rendering resolution while targeting a higher-quality output image.
Aftermath
Initial adoption was limited by game-specific training and mixed image-quality results, creating pressure for a more general architecture.
Recent events
- 2025
DLSS 4 debuts with RTX 50 series[1]
Product launch · Product generation
DLSS 4 brought a transformer-based upscaling model and Multi Frame Generation for supported GeForce RTX 50-series hardware.
- 2023
DLSS 3.5 introduces Ray Reconstruction[1]
Product launch · Product generation
Ray Reconstruction used an AI model to improve denoising and image stability in ray-traced and path-traced scenes.
- 2022
DLSS 3 adds Frame Generation[1]
Product launch · Product generation
The third generation added AI-generated intermediate frames and tied the new feature to the optical-flow hardware in GeForce RTX 40-series GPUs.
- 2020
NVIDIA releases DLSS 2[1]
Product launch · Product generation
DLSS 2 introduced a substantially revised temporal reconstruction approach that generalized across games and expanded adoption beyond the initial launch titles.
- 2018
NVIDIA launches DLSS with GeForce RTX 20 series[1]
Product launch · Product generation
DLSS was presented as a core RTX feature, using neural networks to reconstruct higher-resolution game images from lower-resolution rendering.
Sources2 sources
Cite this profile: Cite the canonical profile. /brand-wiki/deep-learning-super-sampling · Editorial policy · How profiles are compiled