DLSS 4 and Transformer Frame Generation: A Revolution in Visual Quality

 With the release of the NVIDIA GeForce RTX 50-series graphics cards, built on the Blackwell architecture, a new generation of image generation technologies has entered the market - DLSS 4 with Multi Frame Generator, which supports up to four intermediate frames per rendered frame. This technology is based on a transformer model trained on massive data sets and allows for multiple frame rate increases without visual artifacts. Unlike previous versions of DLSS, it uses a new algorithm called Multi-Frame Generation, minimizing latency and maintaining predictable scene reconstruction even with fast camera movements. The new frame generator is integrated at the driver level and tightly coupled with the graphics pipeline. GPU, which ensures the stability and scalability of the technology in any gaming scenarios.

Blackwell series video cards are already available for sale. A reliable partner for their delivery is the online store Telemart, offering Wide range of NVIDIA GeForce RTX 50-series models, including both gaming and professional solutions.

The video card used for the review was MSI GeForce RTX 5080 16G GAMING TRIO OC, which perfectly coped with all the tasks set.

DLSS: From FPS Accelerator to Core Component of Next-Gen Graphics

DLSS (Deep Learning Super Sampling) is a technology that emerged as a response to the increasing workload of modern graphics systems. Its original goal was do not improve the picture, increase productivity. It was first introduced as a way to display a game on screen at a lower resolution, and then “build in” the missing pixels to the required format - to unload the graphics system and get more frames per second without a radical drop in image quality. But over time, DLSS ceased to be just a temporary solution. It evolved into a full-fledged part of the visual pipeline, and with the advent of DLSS4 a fundamental transition has taken place: now it is not a way to speed up, but way to display correctly.

The first versions of DLSS were strictly utilitarian. They were based on the idea of ​​scaling: GPU renders the image in, say, 1080p, and DLSS upscales it to 4K. Visual distortions were noticeable — blurry textures, flickering lines, motion artifacts. But the key benefit — a sharp increase in performance — justified the compromises. DLSS was used as a switch: if you need FPS, turn on DLSS, if you want clarity, turn it off. This approach existed until rendering technologies stepped beyond the line of photorealism, where regular upscaling was no longer enough.

With the advent of ray tracing, global illumination, light simulation and reflections, it became clear: performance without quality is a disadvantage. GAMES began to use scenes with dozens of light sources, volumetric shadows, refractions and transparent materials. All this is a monstrous load, and previous methods could no longer cope. At this point, DLSS began to change: instead of simply increasing the resolution, it became reconstruct the scene, filling in the missing elements not according to a template, but based on understanding, what exactly does it restore.

This reconstruction was made possible by changing the DLSS architecture to transformer model. Instead of traditional convolutional neural networks that analyze limited areas of an image, DLSS 4 uses a model that can understand the relationships between all parts of a scene in time and space. It takes into account previous and future frames, motion vectors, depth buffers, and geometry — and based on this predicts what the frame should look like. The result is not just an upscale, but a visually accurate, stable, clean picture that does not fall apart when moving.

It was with the RTX 50 generation that a new era began. DLSS 4 was implemented for the first time in this generation. in full, without compromise, as a permanent element of the image output. GAMES, released with DLSS 4 support, are no longer supposed to work without it. It is used as a core component not only for scaling, but also for generation of intermediate frames, restoration of traced lighting (Ray Reconstruction), noise elimination, maintaining visual integrity. From this generation onwards, DLSS is not an option. It is part of the scene display system.

This paradigm shift changed everything. DLSS 4 now does several things at once:

• Increases productivity, as before - allowing the image to be rendered at a lower resolution;

• Generates extra frames between rendered ones., providing stable animation even with unstable FPS;

• Eliminates visual defects: flickering, “trembling” of shadows, unstable reflections, noise in transparent and dark areas;

• Reconstructs ray tracing, eliminating the need for "raw" rays and making lighting photorealistic even at minimal cost.

DLSS 4 is not a gimmick. It is the tool around which modern visual engines are starting to be designed. Developers no longer aim to render “as is.” They create a minimal base for a scene, feed it into the DLSS 4 system, and get an image that superior in quality to native. This became possible only after a paradigm shift - when the system no longer calculates each pixel manually, but predicts it based on the logic of the scene.

Physical materials, reflections, transparency, complex lighting - all these elements were previously either simplified or reduced in resolution for the sake of FPS. Now they are restored by a neural network, without losses. And all this became possible not with the first generation of DLSS, but precisely from the moment when the architecture allowed the transformer to work in real time. That is, from the very architectural point from which RTX 50 began.

From this point on, DLSS 4 stops being a feature. It becomes necessary. GAMES focused on photorealism no longer work without it. They base their pipeline on upscaling, motion generation, and light restoration. This applies to both realistic action games and magical fantasy, science fiction, and racing simulators. DLSS 4 is not a universal tool for everything, but it is universal in its role. guarantor of visual quality.

Thus, the path of DLSS - from a purely utilitarian FPS accelerator to a key component of visual reconstruction - is complete. And the transition point was the architecture that began a new era of DLSS 4: when performance and quality ceased to be opposites and began to work together. Not because "the hardware became more powerful", but because the model learned to understand the scene. And this is not just a technology - it is a new standard for displaying games.

What is a transformer in DLSS 4 and why is it needed

DLSS 4 is based on the Transformer architecture, which is a fundamental difference from previous generations of upscaling. The Transformer is used not as an auxiliary neural network, but as a central scene analysis system - it replaces classic convolutional approaches that did not have a global understanding of the frame. Instead of processing the image with local convolutions in small fragments, the Transformer takes into account the entire scene at once, including depth, movement, frame history and spatial relationships. This allows for accurate reconstruction of the visual sequence even with complex geometry, sharp movements and dynamic light sources.

The transformer is fed not only with low-resolution frames, but also with geometry, depth, normals, speed, and optical flow data. The system works in a time perspective: for each current frame, several previous ones are used, accumulating context. This gives the model access to full information about what is happening, including the behavior of objects, lighting, refractions, and dynamics of the scene. As a result, the transformer is able to generate not just a scaled image, but a logically complete picture, in which there is no noise, motion artifacts, or unstable zones.

Image reconstruction in DLSS 4 is a series of operations involving the Transformer: first, it interprets the input data, then restores missing details, and finally synthesizes a new visual layer that complements the original rendered scene. Unlike traditional upscaling, where the final frame is formed by direct enlargement, predictive logic is used here: the system restores parts of the image that were not even initially rendered. This is especially important in ray tracing, where a significant part of the visual information relies on calculations of global illumination, reflections, and transparency. The Transformer does not simply build up pixels; it models the behavior of light and materials based on patterns.

This architecture avoids the typical problems of previous DLSS: flickering shadows, trembling lines, visual "noise" at the edges of objects. At high speeds, frames do not fall apart, geometry remains intact, and transparent objects and reflections behave predictably. The transformer provides spatio-temporal stability, creating a visually clean, integral scene that is not inferior to native resolution, and in some cases surpasses it in clarity and consistency.

DLSS 4 cannot be implemented without the transformer model. It is the basis for all its functions: upscaling, frame interpolation, artifact removal, tracing restoration and image stabilization. At the level of game rendering, this means that the transformer becomes an integral part of the pipeline. Modern games no longer create the final frame at the rendering stage - they create basic data, pass it to the DLSS system, and receive the final image after processing by the neural network.

This model is used for the first time in real time with minimal latency due to the architectural features of RTX 50. From this generation, DLSS 4 ceases to be a filter - it becomes a visualization stage, mandatory for the final output of the image. The image quality directly depends on the work of the transformer, and not on the rendering power. The visual integrity of the scene is formed by a neural network that understands what is happening at each point of the frame, how objects relate to each other, and what result is required to match the visual logic of the game.

DOOM: The Dark Ages — DLSS 4 as the basis for visual purity

DOOM: The Dark Ages uses DLSS 4 not as an add-on, but as a necessary support for the entire visual experience. When path tracing is activated, the picture loses stability without upscaling: characteristic grain, dynamic noise and jerky lighting make the game visually heavy. However, after turning on DLSS 4 with the transformer model, everything changes. Image quality reaches a new level - each scene looks like a finished render without digital artifacts.

The game's special feature is the huge number of light sources. From explosions, pools of flame, glowing enemy eyes and lava flows to magical flashes and electrical discharges, the lighting changes dozens of times per second. Without DLSS, this leads to sharp shadow flickering and strong noise in global illumination. With DLSS 4 enabled, these shortcomings disappear completely. A transformer model trained on temporal and spatial patterns stabilizes the behavior of light. Even with multiple intersections of rays, shadows remain clear, the edges are not blurred, and the softness of the lighting is preserved without loss of detail.

Complex textures – blood puddles, dusty stones, wet metal – are processed especially accurately. DLSS 4 allows you to reproduce their depth and structure even when moving. In older versions of upscaling, such materials turned into noise or a blurry mass when the camera moved. Here, each element retains its shape. Metal chains do not lose their contours, and intricate bas-reliefs on the walls remain legible even with sharp turns.

Movement is a separate zone of stability. Thanks to Multi-Frame Generation and the analysis of consecutive frames by the transformer, when strafing, attacking or quickly finishing off objects do not disintegrate into trails. Even when moving from a dark tunnel to a hall filled with fire, there is a correct adaptation of brightness, without jumps in exposure and overexposure. The shadows under the character's feet do not "tremble", but smoothly change depending on the angle of incidence of light.

The handling of transparent objects and refractions is particularly impressive. Glass, protective shields, energy fields - previously, when traced, they would lose their shape or appear with strong distortions. DLSS 4 treats these areas as independent structures and maintains the accuracy of reflections, even when the camera is moving at high speed. Transparency does not lead to layer conflicts, and even intersecting effects retain their physical properties.

Every frame in DOOM: The Dark Ages with DLSS 4 is not just a “clean” picture. It is a visually stable canvas that does not fall apart in dynamics and does not lose detail under lighting. Even scenes with dozens of enemies, blood particles, smoke, explosions and sharp changes in brightness look holistic. DLSS 4 does not make the game soft - it makes it stable. It does not smooth out sharpness, but restores detail and geometry where classic methods fail.

 

DOOM: The Dark Ages shows a dependence on frame generation, especially with path tracing enabled. At native 2560x1440, the average FPS is only 54, which is clearly not enough for smooth gameplay on Ultra Nightmare. With DLSS Performance and Frame Generator 4X, the frame rate rises to 302 FPS — almost 6 times higher. Even DLSS Quality maintains a stable 252 FPS, and DLAA gives 173. This clearly demonstrates how well DLSS 4 works with full ray tracing load. In this case, the Transformer stably reconstructs up to 4 intermediate frames for each real one, maintaining minimal drops above 224 FPS. DOOM uses the same engine as Indiana Jones, but the load is higher, due to the abundance of effects, particles and intensive lighting calculations. Nevertheless, FG 4X copes without artifacts, maintaining the minimum FPS above 150 even in DLAA.

Indiana Jones and the Great Circle — when DLSS 4 creates atmosphere

visual style Indiana Jones and the Great Circle is based on contrast: ancient temples and dusty ruins are juxtaposed with bright spots of light, candles, torches, spotlights and natural light sources. DLSS 4 acts here as a mechanism for stabilizing this visual environment, giving each frame clarity, structure and depth.

The main challenge is scenes with variable atmospheric density: dust, smoke, steam, glare on ancient stone. Without DLSS, they look loose and noisy. But the DLSS 4 transformer model absorbs not only geometry, but also the behavior of light on the material. Thanks to this, the glow from the flames, the diffuse light from the windows and the sun's rays in the catacombs look like in the film, without jumps and digital noise.

The use of DLSS 4 is especially important here in slow and static scenes - the camera moves smoothly, but the saturation of details is high. Each object in the room - be it a shelf, an artifact or an architectural element - retains its texture even with the slightest movements. Without a transformer, these scenes "pulsate", contours are lost. With DLSS 4, even the texture of sandstone is discernible up close, and the plant ornaments on the columns look hand-carved.

The shadows in the game are complex - they are soft, intersecting, and react to the movement of torches. Thanks to DLSS 4, they remain sharp even in the depths of the scene. Transitions between light and shadow occur without jerks. Tonal gradients are smooth, there are no color jumps or darkening in inappropriate places.

In high-speed scenes (such as chase scenes, falls, jumping over chasms), DLSS 4 shows its best. The edges of objects, including hair, clothing, background elements - everything remains sharp. There are no breaks in animation, no traces of "ghost" frames. This creates a sense of continuity and immersion: everything you see looks complete.

Mirrored and translucent materials in the game are also reproduced with great accuracy. Water, droplets, films on old lenses or pieces of mosaic - all this no longer fragments in motion. Light passing through such materials retains density and shade. DLSS 4 interprets each effect separately, without mixing lighting levels.

In the end, DLSS 4 becomes a tool not for speed, but for precision. Indiana Jones shows how the transformer model can reconstruct a complex visual environment while preserving the concept. Everything looks not just realistic, but artistically pure. The frames do not reveal their interactive nature - they are close to the quality of film animation. DLSS 4 makes this possible not through filtering, but through a full analysis of the scene.

 

Indiana Jones and the Great Circle, despite its apparent simplicity, shows a heavy load in ray tracing scenes. Native performance here is one of the lowest - 42 FPS, which indicates high costs for lighting, reflections and shadows. Connecting FG 4X in combination with DLSS Performance increases the frame rate to 227. This is lower than in DOOM, but higher than in Cyberpunk. In DLSS Balanced mode, the counter shows 211 FPS, and DLSS Quality - 191. Even DLAA in conjunction with frame generation gives 143 FPS - three times more than in the native. This emphasizes the optimization of ray tracing in the new version of the engine, but also its demanding nature. Frame Generator stably scales performance even under increased load, and the minimum FPS in DLSS Performance remains above 200, which proves the effectiveness of FG 4X and the transformer model for heavy story games.

Cyberpunk 2077 - visual reconstruction of the metropolis

Cyberpunk 2077 — a game with an extremely complex visual structure: the metropolis of Night City is full of reflections, neon, weather effects, transparent panels and many moving objects. Without DLSS 4, even at high settings, constant visual artifacts appear: neon flickers, reflections tremble, shadows shift, and textures lose clarity. With the DLSS 4 transformer model enabled, the picture becomes cinematic — clear, stable, structured.

First of all, DLSS 4 stabilizes the light structure of the scene. In Cyberpunk, this is important: there is no single light source — dozens of lights, advertisements, lanterns, cars, and windows interact with each other. Without DLSS, light often conflicts: it creates parasitic glare, causes overexposure, or disrupts shadows. Transformer solves this problem by analyzing past and future frames, aligning the light composition. The lighting stops “jumping,” objects cast precise shadows, and bright areas do not flood the scene.

The second is reflections. In Cyberpunk, there is glass, chrome, water, and shop windows everywhere. Without upscaling, reflections often jitter or disappear in motion. With DLSS 4, their shape is fixed: they match the geometry of the scene, do not blur when the camera is turned, and maintain the correct orientation. This is especially noticeable on wet streets - each light source is displayed stably, without flickering or color distortion.

Third, the density of details. Even in densely populated areas of the city, where there are dozens of objects in the frame at the same time, DLSS 4 preserves microtextures: brick, leather, concrete, digital screens and fonts are not lost. They remain readable, without "blurring". When moving quickly on transport, the architecture does not "collapse", but retains depth and clarity - buildings do not turn into a blurry mass, but maintain geometry to the horizon.

Dynamics is DLSS 4's strong point. Thanks to Multi-Frame Generation, frames in Cyberpunk do not break when changing the angle. When the player rotates the camera, turns, aims or goes on the attack, all objects remain intact. There are no "tails" from lanterns, no halos from light sources, no distortions from glass partitions. Even in the rain, when drops flow down the windshield, the picture looks clear.

Processing transparent objects is another example of the advantages of the transformer. Glass walls, displays, holograms and even films on helmets no longer conflict with light. Previously, layers, noise, and sticking shadows were observed. Now DLSS 4 highlights each type of material separately, correctly processes their physics and optics, allowing you to maintain realism.

Cyberpunk with DLSS 4 is a completely different game visually. Smooth, clean, contrasty, with rich light and clear depth. There are no more “gamey” signs - DLSS 4 makes the scene whole. Objects do not flicker, lighting does not shake, movement does not break the frame. Everything works as intended - with respect for geometry, lighting, composition. This is not an improvement - it is a redefinition of the visual standard.

  

 The video card was tested at a resolution of 2560x1440 with maximum graphics quality settings.

Cyberpunk 2077 demonstrates a sharp jump in performance when using the new frame generation modes. In native resolution, the average frame rate was only 39 FPS, which is not surprising with path tracing enabled and maximum settings. However, switching to Frame Generator 4X + DLSS Performance instantly raises the average FPS to 296, with a minimum of 244 frames - an increase of almost 7.5 times. Even the DLSS Balanced mode provides 262 FPS, and DLSS Quality - 231, showing that even at high scaling quality, the frame rate remains at 230+. Interestingly, even DLAA in combination with FG 4X gives 144 FPS, while without upscaling - only 39. This emphasizes how aggressively the transformer model and multi-frame generation work in DLSS 4. Using FG without DLSS upscaling (in DLAA) increases the frame rate almost 4 times compared to the native one. Playability is only possible due to the full set of 50-series technologies.

Stellar Blade - Combat Movement Without Visual Degradation

Stellar Blade combines highly detailed character models with sharp movements and combat collisions. This combination usually causes graphical problems: skin blurs, metal parts shimmer, backgrounds collapse when the camera is tilted. DLSS 4 eliminates these effects completely, making even the fastest scenes clean, crisp, and cinematic.

The game's special feature is the huge contrast between bright lighting and shadow. On the heroine's skin, for example, even in dynamics, pores, shine, and relief are visible. The metal elements of the suit reflect light correctly: when rotating, dodging, rolling, the reflections do not twitch or blur. The transformer maintains the correct geometry of all moving elements, preventing layering or loss of contrast.

During a fight, there are dozens of objects in the frame at the same time: particles, blows, weapon marks, energy effects. All of this is displayed without noise, even with a sharp change of scene. Thanks to DLSS 4, every spark retains its shape, every blow is drawn to the end. The effects interact with each other without conflict.

The background is another critical point. In classic upscalers, background elements often "collapse" when turning: they blur, disappear, and are replaced by noise. In Stellar Blade, the transformer ensures that the geometry is fixed: the background does not fall through, does not shake, and does not lose color saturation.

Visual integrity is achieved by analyzing not just one, but several frames. It is Multi-Frame Generation and temporal context allow DLSS 4 to keep the picture moving even in the most crowded scenes. The game turns into a cinematic - but a controlled one. No glitches, no loss of sharpness, no degradation when moving.

 

Stellar Blade is the only one of the five without ray tracing, and therefore demonstrates the highest results. In the native resolution of 2560x1440, the average FPS is already 175, and with DLSS 4 and FG 4X in Performance mode - 566. This means almost 3.2 times higher productivity, despite the lack of ray tracing. Even in DLSS Quality, the result reaches 527 FPS, and DLAA produces 447. The minimum values ​​are also impressive: from 402 (DLAA) to 517 (Performance). This speaks to the high efficiency of multi-frame generation based on the Stellar Blade engine, especially in scenes with many effects and animations. Although the increase here is not so critical for playability, it emphasizes how much the FG 4X is able to scale even with high base FPS values. Most importantly, the lack of ray tracing allows the engine to use the frame buffer almost perfectly without artifacts and frequency drops.

Dune: Awakening - scale controlled by DLSS 4

The wide open spaces of the Arrakis desert, sand storms, and complex particle interactions all create the perfect environment to showcase the benefits of DLSS 4. Depth stability, air transparency, and lighting accuracy are especially important here. Without DLSS 4, scenes often lose clarity: textures blend, dust causes noise, and lighting becomes overexposed. With the Transformer, everything looks stable.

Sand is a key element. As the player moves across the dunes, each grain of sand retains its physical behavior. There is no blurring at the edges, no loss of detail - the structure is preserved even in flight. When a storm arises, the dust does not turn into noise - it forms volume. Visibility decreases, but without artifacts.

The shadows on the sand are smooth and continuous. Even under a moving ship or a walking giant worm, DLSS 4 maintains an accurate light and shadow boundary. This is important for the sense of space: the player does not get lost in the visual chaos. Lighting from the sun, glare from stones, reflections on transport - all this is displayed stably. There are no gaps in gradients, no sticking of light spots. Transparent particles, such as dust and fog, no longer conflict with the distant background: the transformer calculates scenes in layers and restores them accurately.

DLSS 4 transforms Dune: Awakening into a fully-fledged interactive visual universe. There is no loss of information, even in the most intense scenes. Scale is no longer the enemy of quality - it is its ally. Every frame looks like a complete composition: stable, clear, clean.

  

Dune: Awakening, being an MMO sandbox, shows a less pronounced load compared to ray tracing games, but also benefits from DLSS 4. In native 1440p, the average frame rate is 77 FPS, which is already enough for comfortable play. However, when you turn on Frame Generator 4X + DLSS Performance the counter goes up to 342 FPS, with a minimum value of 307. Balanced and Quality provide 323 and 302 respectively, and even DLAA - 229. It is significant that even DLSS Quality gives a performance increase of almost 4 times compared to the native. This is explained by the fact that the Dune engine does not use heavy tracing, but scales very effectively through FG 4X. It is especially clear how DLSS 4 minimizes lags in scenes with many NPCs and animations. In this case, the load goes towards reconstruction and multi-frame buffering, where 5080 gives the maximum.