Tag: future of ai

The most interesting aspect of Nvidia’s new RTX 50 series is not the GPUs themselves – not even close – and it’s not multi-frame generation either. It’s DLSS 4 Super Resolution upscaling, which has received a substantial update aimed at improving visual quality.

The old CNN AI model has been replaced with a newer, larger Transformer model, which Nvidia claims can achieve a level of quality that wasn’t possible with previous versions of DLSS. So how good is DLSS 4 upscaling? Let’s find out.

Deeply analyzing upscaling technology is a massive undertaking, so today’s focus is on DLSS 4 versus DLSS 3 upscaling at 4K resolution. We have data for 1440p and 1080p upscaling as well, but that’s something we plan to revisit later. The goal here is to determine where DLSS 4 has improved, where it struggles, and what the new acceptable minimum quality level for gaming is.

Previously, we found that for the best experience, you likely wouldn’t want to drop below Quality mode at 4K – maybe Balanced at a pinch – but going all the way down to Performance mode usually resulted in noticeable visual artifacts in motion. Is that still the case with DLSS 4, or are those lower settings now more viable?

To assess this, we will explore image quality across 14 different areas, including texture quality, edge stability, ghosting, disocclusion, foliage, particles, water, and more.

After matching and comparing footage in all the areas where upscaling tends to struggle, we will have a solid understanding of which DLSS 4 modes are visually equivalent to DLSS 3 and how much improvement Nvidia has been able to achieve. We will also briefly examine performance, as the new Transformer model is larger and more demanding to run.

One area of confusion surrounding DLSS 4 is compatibility. DLSS 4 upscaling works on all RTX GPUs, going back to the GeForce 20 series. In other words, it’s not restricted to the latest GeForce graphics cards and you don’t need to buy an RTX 5090 to access DLSS 4 upscaling or ray reconstruction.

The only exclusive DLSS features are single-frame generation, locked to the RTX 40 and 50 series, and multi-frame generation, which is locked to the 50 series. The most useful component of DLSS 4 – the upscaling – is widely available, which is great to see.

All the visual examples in this article and the accompanying videos were captured at 4K using a GeForce RTX 5090. We tested a selection of games, all with settings such as motion blur, film grain, vignette, and chromatic aberration disabled.

For the DLSS 3 examples, each game was upgraded to the final version of DLSS 3 (3.8.10) using the DLSS Swapper utility. For DLSS 4, each game was upgraded to the latest model using Nvidia’s override feature available in the Nvidia App. This allows us to compare the best version of DLSS 3 with the best version of DLSS 4 in every title.

Textures and Blur

The most obvious improvement DLSS 4 has made relative to DLSS 3 is in texture quality. Over the last decade, temporal anti-aliasing has been reducing texture quality and overall clarity in games in an effort to eliminate aliasing.

This has created a signature TAA “blur” that’s especially noticeable in motion, and DLSS 3 – which is essentially a fancy AI-based TAA that also incorporates upscaling – wasn’t immune to this issue. DLSS 4 has made enormous strides in eliminating TAA blur, which raises texture quality relative to DLSS 3 and even native TAA.

For a better representation of image quality comparisons, check out the HUB video below:

Across all the examples we’ve seen, running games with DLSS 4 gives the appearance of using a higher texture quality setting, even though texture settings remain completely unchanged.

DLSS 4 makes games look like they’re running at a higher resolution because the blur is eliminated

Another way to look at it is that DLSS 4 makes games look like they’re running at a higher resolution because the blur is eliminated. At 4K, if DLSS 4 is delivering a true 4K presentation, DLSS 3 almost looks like 1440p in comparison. For those who are highly sensitive to detail, this is a game changer.

In Cyberpunk 2077, both using Quality upscaling, DLSS 4 is not only clearly sharper when stationary, but it also preserves this sharpness when movement starts. When we pause the footage while walking forward, it’s immediately obvious how much better the texture quality is with DLSS 4 and how some pattern artifacts present in DLSS 3 are reduced or eliminated.

What’s super impressive is that DLSS 4 on Performance mode has basically the same texture-preserving properties, meaning that in terms of texture detail, DLSS 4 Performance often looks better than DLSS 3 Quality in motion.

There are plenty more examples of this. In Ratchet & Clank, looking at these barrels, DLSS 4 is clearly less blurry, and this even holds true when comparing DLSS 4 Performance to DLSS 3 Quality. Later, in a cutscene, we can see DLSS 4 once again delivering higher texture quality for Ratchet’s leather helmet. In Horizon Zero Dawn Remastered, we see higher texture quality when comparing DLSS 4 Performance to DLSS 3 Quality for the ground and rocks, with both DLSS 4 Quality and Performance modes delivering a similar texture experience.

In most examples, specifically when talking about blur and texture quality, DLSS 4 is superior to native rendering – even using Performance mode. Of course, there are other issues when comparing Performance to Native that we’ll explore later, but textures in particular are undoubtedly rendered best with DLSS 4.

We even found the experience better using DLSS 4 Performance versus DLSS 3 DLAA. The upgrade is so noticeable that most gamers will finally feel like they’re experiencing the “better than native” image quality that DLSS promised way back at the GeForce 20 series launch. Games look, feel, and – most importantly – play in a way that’s less blurry.

Edge Stability

Another crucial aspect of image quality is edge stability. Not only do we want sharp, clean textures with no blur, but we also want those textures and edges to look the same between frames – with no sizzling or aliasing.

DLSS 4 is generally a significant upgrade over DLSS 3 in this area as well, as seen throughout the Horizon Zero Dawn example below. Looking at the wooden bridge, DLSS 4 maintains better stability for each line within the texture and geometry – even when comparing the typically unstable Performance mode in DLSS 4 to DLSS 3 Quality. In examples like this, even the Performance mode provides superior stability to what the Quality mode offered previously.

For a better representation of image quality comparisons, check out the HUB video below:

Where DLSS 4 sees the biggest benefits is when there is a small amount of movement between each frame. In Black Myth: Wukong’s menu, for example, with DLSS 3 Quality mode, the slight swaying of the character causes some instability with fine details and edges. With DLSS 4, even in Performance mode, the new model is much better at identifying edges, accounting for small movements between frames, and locking those edges down to provide much better stability. In some cases, like this one, DLSS 4 Performance actually comes out ahead of DLSS 3 Quality.

This poor stability in motion is generally why we haven’t recommended using the DLSS Performance mode before, even at 4K. In Ratchet & Clank, for example, even though the motion between frames is consistent and relatively slow, DLSS 3 Performance just can’t maintain stability, leading to ugly artifacts. Not only is DLSS 4 Performance sharper, but it’s also much more stable, which is immediately noticeable in most situations.

That doesn’t mean DLSS 4 Performance mode is always better than DLSS 3 Quality for edge stability. For example, when driving in Cyberpunk 2077, overhead wires and bridges are more stable using the higher render resolution of DLSS 3 Quality.

However, DLSS 4 does have advantages in other areas, like fences on the side of the road and overhead lights. Generally speaking, when comparing DLSS 4 Quality to DLSS 3 Quality, DLSS 4 offers a more stable presentation, and we have yet to come across an example where DLSS 4 edge stability is worse than DLSS 3 when using the same mode.

Ghosting

More recent versions of DLSS 3 are not especially prone to ghosting, and across the games we tested, this was rarely an issue – whether using DLSS 3 or DLSS 4. However, when ghosting does occur, it’s hit or miss whether DLSS 4 will resolve the problem compared to DLSS 3.

The Cyberpunk 2077 is notably better at reducing ghosting from overhead street lights or the tailpipes on the car. This is another case where DLSS 4 Performance can look better than DLSS 3 Quality.

For a better representation of image quality comparisons, check out the HUB video below:

But when switching over to Forza Motorsport, a game that is prone to ghosting with most forms of upscaling – or even native TAA – there’s little change between DLSS 3 and DLSS 4, both using the Quality mode. In fact, if anything, ghosting appears more visible in this example with DLSS 4, though neither option is ideal.

We also saw that in Ratchet & Clank, DLSS 4 can introduce ghosting where there previously was none, when comparing the same quality settings. When Ratchet moves across the red carpet, there’s a subtle ghost trail left behind in the DLSS 4 image that isn’t present with DLSS 3, which is a disappointing regression in worst-case scenarios. But like we said, after examining 11 games in detail for this article, ghosting generally isn’t going to be a major issue or concern when enabling DLSS 4.

Disocclusion

Disocclusion is up next, and we’re primarily going to focus on the artefact you get around your character in a third person game when moving. Essentially, disocclusion occurs when something on screen moves to reveal the background behind it, and in the frame where this disocclusion occurs there usually isn’t any temporal data to draw from, so it’s prone to artefacts. Usually this is seen as a sizzling or lower resolution halo around your character in motion where the immediate area lacks a bit of definition that you normally see elsewhere on screen.

For a better representation of image quality comparisons, check out the HUB video below:

While DLSS 4 is generally an improvement in texture detail and stability, the technology struggles to deliver an improvement in disocclusion, often creating more artefacts than DLSS 3 when comparing Quality mode vs Quality mode. This was perhaps the most noticeable in The Last of Us Part I when Joel’s head moves to reveal the water and grass behind, there’s a bit more sizzling in the DLSS 4 image versus DLSS 3 as disocclusion occurs. I suspect this is because DLSS 4 is trying to preserve more detail than DLSS 3, but has the same single frame of lower resolution data to work with, whereas with DLSS 3 the disocclusion artefacts are hidden somewhat by the overall blurrier image.

I also found disocclusion to be slightly worse in Horizon Zero Dawn and Black Myth Wukong in areas where motion was relatively fast and backgrounds were detailed. However the impact was less pronounced in Dragon Age The Veilguard and Star Wars Jedi Survivor, where really there was little difference between the DLSS 4 and DLSS 3 Quality modes. In the best cases, DLSS 4 Balanced mode delivers a similar experience to DLSS 3 Quality mode in this area, but more common I would say the reverse is true, where a lower DLSS 3 mode shows less disocclusion artefacts than a higher DLSS 4 mode.

The good news is that a lot of this is nitpicking, and disocclusion artifacts, in general, are hard to spot at 4K. To really see the changes, we basically had to step through the footage frame by frame or view zoomed-in shots in slow motion. In most of these examples, it’s difficult to notice a difference in real-time gameplay, even if there is a small variation when looking closely.

The Last of Us was the only game where we found the downgrade noticeable in real time. In most other cases, whenever artifacts appear, they are confined to a small area and disappear within one or two frames.

Hair

Hair is one of the most difficult elements to upscale due to its dense, fine detail. Unfortunately, we didn’t see much improvement when comparing DLSS 3 and DLSS 4 in this area.

In games with high-quality hair rendering, like Dragon Age: The Veilguard, there is little difference in the level of detail between the two, and aliasing remains an issue with any level of upscaling – including Quality mode – relative to native rendering with TAA or DLAA. We also noticed no difference in Black Myth: Wukong, where upscaling still results in a reduction in quality compared to native rendering.

The challenges with hair rendering also apply to fur (as tested in Ratchet & Clank). While DLSS 4 is able to extract a higher level of detail from basic textured elements, fur rendering remains largely unchanged.

It’s not a situation where DLSS 4 Balanced is able to match DLSS 3 Quality – it’s very much a like-for-like comparison. Occasionally, we spotted examples, like in The Last of Us, where hair quality appeared slightly sharper, but in terms of aliasing, there is little improvement with the new Transformer model.

Particles

Particle reconstruction is another area that can be difficult for upscalers to handle, as particles are often small, fast-moving, and inconsistent. One area where DLSS 4 improves upon DLSS 3 is particle ghosting.

In the Starfield example (video below), if you look closely, you’ll see streaky trails following particles in the DLSS 3 image, which are eliminated using DLSS 4 at the same Quality mode. Even when dropping DLSS 4 to Performance mode, ghosting is cleaned up significantly, leading to a more stable presentation.

For a better representation of image quality comparisons, check out the HUB video below:

This holds true in other games like Ratchet & Clank, which was a bit less prone to ghosting, but still, any time a particle ghosted in the DLSS 3 image, it was clean in the DLSS 4 image.

A particularly stressful test is airborne spore particles in The Last of Us, and while DLSS 4 doesn’t completely eliminate ghosting here, it noticeably reduces it, resulting in a better-looking image.

However, there are still notable differences in particle quality outside of ghosting that limit the advantages of DLSS 4. Upscaled particle quality tends to degrade at lower render resolutions, like in Performance mode, so in most cases, DLSS 4 Performance compared to DLSS 3 Quality gives an edge to DLSS 3 in particle resolution at the expense of ghosting.

At equivalent quality modes, we wouldn’t say there’s a huge difference in particle resolution and edge quality, though this does depend on the game. Generally speaking, DLSS 4 Balanced mode is on par with DLSS 3 Quality mode for particle quality, with the added benefit of less ghosting in the DLSS 4 image.

Transparency

Transparent items are typically pain points for upscalers, and with DLSS 4, we didn’t see significant improvements in this area. For the most part, the quality of upscaled transparencies is heavily linked to the render resolution, so these elements appear more detailed and less pixelated when using a higher DLSS mode. In most cases, this means DLSS 4 Quality mode is equivalent to DLSS 3 Quality mode.

For a better representation of image quality comparisons, check out the HUB video below:

There were some instances where DLSS 4 Balanced was able to match DLSS 3 Quality in transparencies. But in other areas, like the holographic map in Cyberpunk, it’s more of a parity situation.

We also tested DLSS 4 Balanced vs. DLSS 3 Quality in titles like Dragon Age: The Veilguard, but generally found that DLSS 3 Quality mode delivered smoother, less aliased transparencies with better reconstruction. The output quality here seems more closely linked to render resolution than the upscaling model, whereas the reverse is true for stability and texture detail.

The exception to this is when standing still, where DLSS 4 generally has the edge, producing a more detailed image – similar to the texture quality advantages we’ve seen. However, this advantage usually disappears in motion.

Fine Detail

Fine detail reconstruction is yet another challenge, as pixel-level or near-pixel-level detail can be lost, aliased, or appear sizzled in motion when the render resolution is too low for the upscaler to handle. This issue is most noticeable on wires and other fine-line details.

Like with edge stability, this is an area where DLSS 4 has an advantage over DLSS 3. One thing that became immediately clear during testing is that DLSS 4 is less likely to introduce weird patterns in finely detailed grates or meshes, as seen in Cyberpunk 2077. This holds true even when comparing DLSS 4 Performance to DLSS 3 Quality, where DLSS 4 shows no major artifacts of this kind.

For a better representation of image quality comparisons, check out the HUB video below:

Wire detail, such as overhead power lines, is only marginally improved. Based on testing games like Cyberpunk 2077 and Indiana Jones and the Great Circle, we’d say that in this area, DLSS 4 Balanced is on par with DLSS 3 Quality.

Alternatively, when using the same mode (Quality vs. Quality), DLSS 4 provides a small improvement in reconstruction. However, these modes are still prone to sizzling and other artifacts along the edges of wires in motion, and aliasing remains a problem. The best image quality is still achieved with native rendering, like DLSS 3 DLAA, which outperforms DLSS 4 Quality mode.

For other types of fine details, DLSS 4 Balanced gets close to matching DLSS 3 Quality, though there are some cases where it falls short of the reconstruction power of a higher render resolution. Typically, these configurations trade blows, and for those who prefer Quality upscaling, there is an improvement in fine details – though it’s not as significant as the gains in texture detail and general stability.

Trees

Foliage is a key part of visual presentation in most games, so we tested tree rendering quality in different scenarios. Generally, results fall into two categories.

When trees are relatively still, DLSS 4 is a big improvement over DLSS 3 – to the point where in some games, DLSS 4 Performance looks as good as, or better than, DLSS 3 Quality. One example of this is The Last of Us Part I, which benefits from DLSS 4’s increase in sharpness and detail.

For a better representation of image quality comparisons, check out the HUB video below:

Similar results were observed in Horizon Zero Dawn, where DLSS 4 Balanced is a strong match for DLSS 3 Quality. The DLSS 4 image is generally sharper and more stable in motion, though at a lower base render resolution, it can struggle with the finest details. Performance mode isn’t always a perfect match, but in a title like Indiana Jones, the foliage benefits more from the stability improvements of DLSS 4. Even in Performance mode, trees can look less pixelated than in DLSS 3 Quality mode in motion.

The second category involves faster motion, such as trees blowing in the wind or dense fine-detail branches. Here, DLSS 4 vs. DLSS 3 quality results align more with what we saw in fine detail reconstruction.

In Star Wars Outlaws, for example, DLSS 4 Balanced clearly renders at a lower resolution than DLSS 4 Quality when viewing trees swaying in the wind. This is a particularly difficult test for upscaling, and we only saw quality parity at the same mode (Quality vs. Quality). DLSS 4 is slightly more stable and less prone to sizzling, but there were times when DLSS 3 handled dense foliage better.

In Black Myth: Wukong, looking at fine tree branches, we observed similar behavior to Star Wars Outlaws, where upscaling struggles to resolve these details properly. DLSS 3 DLAA still has an edge over DLSS 4 Quality upscaling.

Typically, the best match was at the same setting, such as Quality vs. Quality. The one exception is Performance mode, which in both Black Myth: Wukong and Outlaws was a notable improvement in DLSS 4 – probably getting closer to the level of DLSS 3 Balanced.

Stability in Performance mode has improved, and for this type of foliage, it’s now in a more usable state. Reconstruction of these details is much cleaner, even if it can’t quite match the output of DLSS 4 Quality. Surprisingly, it gets quite close when viewed side by side.

Grass

We were more impressed with how DLSS 4 handles grass. With DLSS 3, grass often had a grainy look in motion because there was too much variation between frames for the upscaler to handle. This is improved with DLSS 4, as seen in Star Wars Outlaws.

Even when comparing DLSS 4 Performance to DLSS 3 Quality, the DLSS 4 image is less grainy and pixelated when grass moves in the wind. And switching from DLSS 4 Performance to Quality results in even better grass resolution and detail, with less aliasing and improved stability over DLSS 3.

For a better representation of image quality comparisons, check out the HUB video below:

This was also evident in Indiana Jones, where DLSS 4 Performance mode provided more stable and less grainy grass than DLSS 3 Quality in some situations. However, in this extremely demanding test, some detail is lost in Performance mode due to the lower render resolution, so in the worst cases, DLSS 4 Quality is still needed to match the level of detail.

Results depend on the game, though. In Dragon Age: The Veilguard, which has very fine grass detail, we didn’t see much improvement from DLSS 3 to DLSS 4, and in some areas, there were even regressions. Generally, using DLSS 4 Performance results in an image similar to DLSS 3 Performance, so a lower DLSS 4 mode won’t match higher-quality DLSS 3 settings. In Horizon Zero Dawn, DLSS 4 Balanced matched DLSS 3 Quality, with DLSS 4 providing additional sharpness and reducing blur, as previously discussed.

Across all the games we tested, DLSS 4 Balanced or Performance modes handled grass upscaling well – something that wasn’t always the case with DLSS 3. Previously, DLSS 3 Performance mode was often too grainy and unstable in games with dense foliage.

Fences

DLSS 4 does a much better job of reconstructing fences and grates than DLSS 3, giving lower modes an advantage. One of the most noticeable improvements is how DLSS 4 handles repeating patterns without producing moiré artifacts.

In The Last of Us, for example, DLSS 4 Performance has an advantage over DLSS 3 Quality, despite the much lower render resolution of the Performance mode. We saw a similar benefit in Starfield when looking at meshes, where DLSS 4 was much less likely to produce an ugly moiré artifact in motion.

In addition to this, DLSS 4 is typically better at reconstructing super fine mesh detail – down to the pixel level – and it improves the visibility of items behind fences or grates, which can appear blurred or obscured with DLSS 3.

Fences are generally more stable and less prone to sizzling as well, especially when comparing DLSS 4 Balanced to DLSS 3 Quality. The Performance mode also sees a significant improvement, with much better clarity and stability in reconstructing these elements, even if DLSS 4 Performance can’t always match DLSS 3 Quality.

Cloth

Cloth quality benefits from DLSS 4 upscaling, as DLSS 4 is better at preserving texture quality in motion – especially on character clothing, which is often moving constantly. As we saw in the mesh and fence analysis, DLSS 4 also does a better job of reducing moiré patterns when upscaling.

One of the biggest issues with DLSS 3 Performance mode was the constant moiré patterns on detailed cloth textures, but that is largely eliminated with DLSS 4 Performance mode, giving it a big quality advantage.

For a better representation of image quality comparisons, check out the HUB video below:

In general, to achieve DLSS 3 Quality-like cloth detail, you can use DLSS 4 Performance, though sometimes Balanced is required depending on the level of detail and motion. In third-person RPGs where characters wear capes, DLSS 4 is simply the better way to play, as cape elements are frequently visible on screen.

Water

We didn’t see much difference in water quality when comparing DLSS 3 and DLSS 4 – generally, the same quality modes in both deliver a similar visual experience. In Jedi: Survivor, for example, this was definitely the case, and we even noticed a regression in visuals when comparing DLSS 3 Performance to DLSS 4 Performance, with DLSS 4 surprisingly producing a less stable image.

In Horizon Zero Dawn, we saw little difference in water quality. The Performance mode has been slightly upgraded with DLSS 4 in this scenario, but overall, it’s a wash.

Rain

Lastly, we have rain. Upscaling tends to struggle here in two ways: rendering the rain particles without aliasing and preserving background detail as rain occludes it.

When it comes to the rain particles themselves, there’s little difference between DLSS 3 and DLSS 4. For example, in Ratchet & Clank, we were only able to match raindrop detail when using the same mode – so Quality vs. Quality. When dropping DLSS 4 down to Performance, there was a slightly noticeable loss of detail in the rain compared to DLSS 3 Quality.

As for the stability and quality of the image behind the rain, in Horizon Zero Dawn, we noticed that DLSS 4 Balanced was a close match for DLSS 3 Quality. Technically, we were looking at snow here – but snow is just a type of rain, right? Either way, we observed about one quality step of improvement provided by DLSS 4.

Performance Benchmarks

Let’s now take a look at performance. While the focus of this article is on visual quality, based on what we saw when testing Ray Reconstruction, we expect the new DLSS 4 Transformer model to perform slightly worse on older GeForce 30 series cards. However, we’ll need to investigate that further in a future article.

In Starfield at 4K using max settings, enabling the DLSS 4 upscaling override cost about one tier of performance. That is to say, the performance we previously achieved using DLSS 3 Quality mode is now only available with DLSS 4 Balanced mode. When comparing Quality modes directly, there was an 8% drop in frame rate. However, DLSS 4 Quality still provided a 24% performance improvement relative to native TAA rendering and a 29% improvement relative to DLSS 3 native DLAA.

In Dragon Age: The Veilguard, we saw similar results while testing the Ultra preset at 4K. Performance dropped by 7% using DLSS 4 Quality mode versus DLSS 3, meaning a full tier of performance impact across the modes. This also means that FSR Quality upscaling, which was previously slightly slower than DLSS Quality, is now slightly faster – though with a noticeably different level of image quality.

In Ratchet & Clank: Rift Apart, the upgrade from DLSS 3 to DLSS 4 results in about half a tier of performance loss. DLSS 4 Balanced now performs between DLSS 3 Balanced and Quality modes, with a 7% frame rate hit when comparing Quality modes. Upscaling is more effective in this title than in the previous two, with DLSS 4 Quality achieving a 46% higher frame rate than native TAA rendering, and DLSS 4 Performance running 79% faster.

Horizon Zero Dawn Remastered is another title where DLSS 4 causes one tier of performance loss compared to DLSS 3. Now, DLSS 4 Balanced delivers frame rates similar to DLSS 3 Quality, with about a 7% performance hit when comparing Quality modes. In the area we tested, DLSS 4 Performance provided a 39% FPS improvement over native TAA rendering.

In Black Myth: Wukong, tested using the Very High preset without path tracing, DLSS 4 caused about half a tier of performance loss. DLSS 4 Balanced sits between DLSS 3 Quality and Balanced, with a 6% frame rate drop when moving from DLSS 3 to DLSS 4. However, the Performance mode is able to nearly double the frame rate compared to native rendering.

Lastly, we have Cyberpunk 2077. Here, we observed the smallest performance impact – just a 5% drop in frame rate when comparing DLSS 4 and DLSS 3 Quality modes. Each tier remains quite comparable in this game, likely due to the lower overall frame rates from running at 4K with the Ultra ray tracing preset.

Performance Summary: DLSS 4 vs. DLSS 3

Here is the geometric mean average across all six tested titles. Typically, DLSS 4 results in about half a tier of performance loss. That is to say DLSS 4 Balanced now sits between DLSS 3 Quality and Balanced for FPS improvement, and DLSS 4 Performance now sits between DLSS 3 Balanced and Performance.

On average, the performance impact was 7% when comparing the same mode across versions. However, switching from DLSS 3 Quality mode to DLSS 4 Performance mode provided a 14% performance boost on average.

What We Learned

Overall, DLSS 4 Super Resolution upscaling is an impressive improvement for 4K gaming. Nvidia has been able to deliver noticeably higher image quality at each DLSS tier in the majority of scenarios we tested, making lower modes – like DLSS Performance – genuinely viable at this resolution without distracting or ugly artifacts.

We were blown away by how Nvidia has managed to fix TAA blur with DLSS 4, resulting in a clearer, sharper presentation with higher-quality textures. TAA has always caused a signature loss of clarity in motion, but even with DLSS 4 in Performance mode, this issue is almost entirely eliminated.

The outcome is that games with DLSS 4 enabled exhibit sharpness typically associated with running at a higher resolution, along with texture quality and detail comparable to high-quality texture packs. As an analogy, if DLSS 4 represents 4K gaming with ultra textures, DLSS 3 is more like 1440p gaming with high textures.

Anyone using DLSS 4 for the first time will immediately notice its superior clarity. While for some, this may be a more subconscious improvement, nearly everyone will prefer the DLSS 4 experience.

We were also impressed with how DLSS 4 cleans up some of the common pain points of upscaling, particularly at lower modes. DLSS 4 is much more stable in motion, showing a significant improvement over Performance modes, and is less prone to annoying moiré artifacts. Grass is typically more stable and less grainy in motion, fine detail reconstruction is improved (particularly for fences and grates), and particle ghosting is less likely to occur.

Not every aspect has improved, though. Hair and water rendering remain largely unchanged compared to DLSS 3, and there is a regression in some forms of disocclusion, though this is difficult to notice in real-world gaming scenarios. Additionally, in scenes that exhibited heavy ghosting with DLSS 3, ghosting may be slightly worse with DLSS 4, though in general, most scenes show either no ghosting or an improvement.

Assessing all aspects of visual quality, we believe that for 4K gaming, DLSS 4 provides a one-to-two-tier improvement. This means that the experience previously achieved with DLSS 3 Quality mode is now possible using DLSS 4 Performance, or in more demanding scenarios, DLSS 4 Balanced.

Previously, we could not recommend the DLSS Performance mode for 4K gaming due to its instability, but now we can. Across many hours of gaming, we were consistently satisfied with the image quality in this mode. Naturally, image quality improves further when using DLSS Balanced, DLSS Quality, or even native DLAA modes – especially for fine detail reconstruction – but DLSS 4 Performance is more than good enough for everyday gaming without distracting artifacts.

DLSS 4 is a heavier algorithm than DLSS 3, so there is a performance cost associated with using it. On a GeForce RTX 5080, this resulted in about a 7% FPS loss at the same quality mode, or roughly half a tier of impact.

The overall improvement from DLSS 4 equates to approximately one and a half tiers of quality enhancement at the cost of half a tier of performance, which averages out to about a full-tier improvement. Essentially, the new version of DLSS provides the visual fidelity of the Quality mode while delivering the performance uplift of the Balanced mode.

Essentially, the new version of DLSS provides the visual fidelity of the Quality mode while delivering the performance uplift of the Balanced mode.

Another way to think of this is that Nvidia have basically delivered a 15% gain via a software update, at least for more recent GPUs. That may not sound like a whopping improvement at first, but it’s pretty massive just from software.

In fact, it’s a larger gain than Nvidia achieved in hardware when moving from the RTX 4080 Super to the RTX 5080. It also surpasses the performance gain seen from the Windows 24H2 update compared to Windows 23H2 for Ryzen CPUs, which we previously described as a major upgrade. Delivering a driver/software update that allows gamers to either enjoy better visual quality in their favorite games or gain a performance boost while maintaining a similar visual level is excellent work from Nvidia.

It’s surprising that Nvidia focused so much on multi-frame generation instead of highlighting the improvements in DLSS 4 upscaling. Sure, multi-frame generation is designed to sell the new GeForce 50 series GPUs, but DLSS 4 upscaling is far more impressive.

Instead of trying to position an RTX 5070 as an RTX 4090, Nvidia could have emphasized how they have practically solved TAA blur with DLSS 4 – a benefit available to all RTX GPU owners. We bet gamers would be more likely to keep buying Nvidia GPUs if they knew they were being supported and taken care of with meaningful, broadly useful software updates over time.

Finally, where does this leave AMD and FSR upscaling tech? We don’t want to dive too deep into this considering FSR 4 is only a few weeks away. We plan to conduct a thorough analysis once it’s released, however as things stands today, FSR 3.1 (and especially FSR 2.2) is not competitive with DLSS 4. In many cases, DLSS 4 Performance mode upscaling looks significantly better than FSR 3.1 Quality mode, while also being 15 – 20% faster.

In many cases, DLSS 4 Performance mode upscaling looks significantly better than FSR 3.1 Quality mode, while also being 15 – 20% faster.

FSR 4 will need to be a massive leap over FSR 3.1. Simply matching DLSS 3 will not be enough – it would only maintain the generational gap that has existed between DLSS and FSR for years. Beyond that, AMD is in serious trouble if it cannot get FSR 4 into a large number of games quickly.

Nvidia laid the groundwork for driver-level DLSS upgrades years ago by using a DLL that can be intercepted and upgraded on the fly, either through the Nvidia App or third-party tools. This allows the vast majority of DLSS-supported games to be easily upgraded to DLSS 4. AMD, however, only started using DLLs with FSR 3.1. This means that in many games, Nvidia users can upgrade to DLSS 4, while AMD users may be stuck with FSR 2.2 upscaling. In those cases, Nvidia will deliver a decisively better visual experience, and how AMD responds to this challenge will be critical.

We’ve already seen FSR 4 in person at CES, and based on that experience, it appears to be a solid upgrade over FSR 3.1. Will it be enough to match DLSS 4? That remains to be seen and will require detailed analysis. What we do know for sure is that, as of right now, DLSS 4 upscaling is the best way to experience gaming. It is a significant improvement over DLSS 3, and we are eager to see what it can achieve at lower and more mainstream resolutions like 1440p.

Shopping Shortcuts:

• Nvidia GeForce RTX 5070 Ti on Amazon
• Nvidia GeForce RTX 5080 on Amazon
• Nvidia GeForce RTX 5090 on Amazon
• Nvidia GeForce RTX 4070 Super on Amazon