CEOInterviews.AI
Start App
Get an email when Tim Schiesser gives a new interview
One email per new interview. Unsubscribe in one click.
Tim Schiesser
Co-host, Hardware Unboxed, Hardware Unboxed

DLSS 5: The Good, The Bad and The Ugly (Full Analysis)

📅 Sep 12, 2026 Hardware Unboxed 37 MIN 208034 VIEWS 82 SEGMENTS · 1 SPEAKERS
See MSI GeForce RTX 50 Series: https://www.msi.com/Graphics-Card/GeF... Australians - Get MSI Maestro PC at PCCG: https://www.pccasegear.com/lp/msi-meg... This is our full analysis of DLSS 5 image quality based on the official launch title, NBA 2K27. We explore the good stuff, the bad stuff, and everything in between, including environments, skin quality, artefacts, inconsistencies and more. Support us on Patreon:   / hardwareunboxed   YT Membership:    / @hardwareunboxed   Buy relevant products from Amazon, Newegg and others below: GeForce RTX 5090 - https://geni.us/JyDFBvM GeForce RTX 50...
Tim Schiesser 0:00 ↗
The official implementation of DLSS 5 produces some interesting visual results. There are improvements. There are regressions. There are things that impress me. There are things that didn't. Some things look more natural and realistic. Others don't. Artifacts introduced when using this technology in motion. I'll be exploring all of that and more in today's video. Is DLSS 5 AI slop? Is it the next step in transformative 3D graphics? Well, let's find out.
DLSS 5 introduces what Nvidia calls 3D guided neural rendering. Essentially how this works is an AI model processes the rendered frame of a game and some additional data like motion vectors. The model is trained to identify elements throughout the scene such as character faces, skin, objects, lighting direction, materials, and more. The model then modifies the rendered frame to what it thinks will look more realistic, tweaking some aspects to lighting and material interactions. The model used here is a pixel space diffusion transformer.
There's a couple of interesting aspects to how DLSS 5 works that are potential strengths and weaknesses. DLSS 5 only processes the current rendered frame. It does not work temporally across multiple rendered frames like DLSS super resolution. It does use motion vector data which Nvidia claims is enough to make DLSS 5 temporally consistent. But this design puts DLSS 5 at risk of creating artifacts in motion because it's processing each frame separately without the full context of previous frames.
It is however a deterministic design. What this means is that the DLSS 5 model creates the same output for the same input every time. This avoids one of the biggest problems with generative AI, its probabilistic nature. Creating a new random image from identical inputs. A model that creates random outputs is completely unsuitable for real-time gaming. So, it had to be deterministic. To be clear, though, a deterministic model doesn't mean artifact free. Small changes to the input frame can still lead to slightly different outputs.
Nvidia have also chosen to process DLSS 5 after upscaling, making this an advanced post-processing filter in its design. This is one of the most controversial choices as modders have found performance gains moving DLSS 5 before upscaling. It's not that Nvidia are being idiots and just didn't think of this. Of course, they considered it. But the official implementation runs on the full resolution rendered frame on purpose to provide the best image quality. Putting it before upscaling can exacerbate artifact issues.
Speaking of performance, I've already covered how DLSS 5 performs in depth in a previous video. So, if you want to see the performance hit, frame time costs, and some interesting numbers about expected performance in a variety of situations, go check out that previous video. I'm not going to go over the performance data again today. This is largely a look at the visual impact.
In the official implementation of DLSS 5 that is integrated directly into games, there are no user controls over image quality. Developers have access to intensity sliders, multiple DLSS 5 models and masking tools, but these are purely designed to give developers control over how their game looks. In NBA 2K27, the launch title for DLSS 5, it's a simple on/off toggle listed as DLSS neural rendering in the game. Let's get into the visual analysis after this.
Today's sponsor spot is brought to you by MSI and their GeForce RTX 50 series. Built with Nvidia's Blackwell architecture and enhanced by MSI's premium board design, power delivery, and thermal solutions, RTX 50 series cards deliver major gains in games, AI workloads, streaming, and productivity. Enjoy many of the latest games such as Control Resonant with the full suite of RTX features such as DLSS 4.5 ray reconstruction, path tracing, and dynamic multiframe generation. Ray reconstruction improves image quality in supported ray traced and path traced games offering better lighting accuracy, motion clarity, and stability, while dynamic frame generation helps to better utilize high refresh rate monitors. MSI offers a number of higher quality models from the Shadow range to the Supreme. So, for more information on MSI's entire range, please click the link in the video description.
I want to start here with environment quality. There are certainly some scenes in NBA 2K27 that I feel look subjectively better with DLSS 5 enabled. You might not believe it, but NBA 2K actually has a basketball court in a forest. And it was here that I think DLSS 5 is most transformative. The additional shadowing in the leaves looks better. The surface of the tree trunks and ground look more realistic, and the general contrast in the scene is improved. It looks less like a game. The image does have a bit of shininess to it. The leaves and trunks get brighter and whiter, but overall I do prefer the DLSS 5 on image.
I also think DLSS 5 looks better in this outdoor cityscape with multiple basketball courts. The DLSS 5 image is higher contrast and appears more naturally lit. The surface interactions on metal look superior with DLSS 5 enabled and tree foliage looks less flat. The game does not utilize ray traced lighting except for reflection. So this is mostly a comparison between rasterization and neural rendering. The DLSS 5 off image is certainly not showing us cutting-edge pathed lighting. But nevertheless, I think DLSS 5 here is a big improvement.
Here we're looking at one of the stores in the city area, which is a good example of how DLSS 5 can improve contact shadowing and ambient occlusion. Without DLSS 5, the game is pretty flat in its lighting and shadowing, but with DLSS 5 enabled, it has more lifelike shadows. Of course, this is not a photoreal scene by any stretch, even with DLSS 5 enabled due to other reasons like texture quality and so on, but DLSS 5 is producing a better image to my eye. The polished wood on the roof, for example, is much better with newer rendering enabled.
Most of NBA 2K27 is played in arenas, though, and here I felt DLSS 5 was a good improvement, but not as transformative. In completely still images like this taken in the instant replay mode, you can clearly see what DLSS 5 is doing. An attempt at generating more realistic lighting interactions with various surfaces, a higher contrast overall presentation, more lifelike characters. But in actual gameplay, flicking DLSS 5 on and off in real time using the game shortcut key, I found it much harder to spot the differences.
I still preferred the way DLSS 5 looks, but a lot of the changes are made to the crowd's lighting, which is not something I was focusing on during gameplay. DLSS 5 doesn't really change the lighting of the court itself, and character models are small enough in this view that any changes there aren't very transformative. The biggest change I'd say in these court scenes is DLSS 5 defogging the game to a degree, creating better clarity in the crowd.
Here's another example I took with camera rotation. And I'm flicking DLSS 5 on and off here. But I found it quite difficult to tell which image was which. And then in this example, which is a smaller arena, less of a crowd, different lighting. Again, I feel DLSS 5 is not particularly transformative. If you focus on things like the hoop base, it looks better lit. But overall, in gameplay, it didn't seem like that much of a difference.
And there are lots of these examples where DLSS 5 either isn't that transformative or ends up creating just a different image. Here is a shot of the tattoo parlor in the city area. And I want you to guess which one of these is DLSS 5. Is A or B the presentation from the next generation lighting system? If you guessed B, you were correct. I certainly don't think this is obvious without peeping at specific areas. Not nearly to the degree we saw earlier with the forest scene.
And this continues when viewing this outdoor scene. I mean, there are obviously differences in the way each option is lit, but I personally don't think one option looks considerably more realistic than the other. This scene in this local arena, again, there are differences. I just don't think it's a huge transformation, though.
I tried a wide variety of scenes within the confines of NBA 2K27. Not the best game for testing this sort of feature, but it is what it is. And the changes range from unnoticeable to substantial. It really depends on exactly what you're looking at. There are some types of surfaces and objects that DLSS 5 doesn't improve at all and others where it is clearly much better at recognizing what it's looking at. It's not a universal enable realism button, though. It requires favorable conditions or scenes to really give you a transformative improvement. And to be clear, you get the performance penalty all of the time, regardless of whether DLSS 5 is significantly upgrading the visual experience or not.
The other big change you will have seen as it's been a big focus of a lot of DLSS 5 marketing is how neural rendering alters characters and skin. How much of an improvement DLSS 5 is depends on how you are viewing character models. You might have noticed that a lot of Nvidia's advertising focuses on relatively close-in shots of people where the face takes up a significant portion of the screen or the scene is framed so that the person dominates the field of view. That's because this is the best case scenario for DLSS 5 and how characters are rendered in other scenarios can be quite different.
Most of the time when you've got these nice still close-up shots of character models, DLSS 5 delivers an improvement to visuals. Skin rendering appears more realistic. Characters go from having more plasticky skin to more natural skin. There's better subsurface lighting interactions. Detailing and shadowing across the face is better, and hair is noticeably improved with better lighting and more depth. I think if you look at any of these still close-up comparisons in NBA with the models often representing real people, you'll think the DLSS 5 image is better. I certainly prefer the DLSS 5 output most of the time.
Of course, you can argue about whether these sorts of visuals have been seen in other games without DLSS 5, but if you're just taking the input into DLSS 5 and what it produces out at the end of it, well, I think it does look better with the way the developers have implemented DLSS 5 in this game. There's no AI slop look here. Nowhere near the levels of atrocious overprocessing we saw from the initial embarrassing DLSS 5 reveal.
But the problem with these examples is that in most games, you don't view characters like this outside of cutscenes in the photo mode. It's much more likely in real gameplay that you'll be viewing characters or NPCs from a distance. And in third-person games, you'll see the back of your own character most of the time. This weakens one of the more transformative aspects of DLSS 5. It does a lot of work to improve skin and face rendering, but that only takes up a small portion of the screen in most games. And DLSS 5 simply isn't as good at improving character rendering when the character is smaller on screen.
I tended to find that DLSS 5 makes more exaggerated and unpleasant changes to characters when they are smaller and further from the camera, whereas closer in, the improvements are more measured, considered, and pleasing. There are some especially weird differences when you look at a character in fixed lighting and start zooming in or out.
Take this referee for example. When the character's face is close to the camera, DLSS 5 does a decent job of enhancing the lighting, making the skin more realistic and better shaded. The tone of the skin doesn't change substantially before and after applying the effect. But when you zoom out, making the model smaller on screen, DLSS 5 doesn't apply the same changes despite lighting being exactly the same in this scene. It makes this character darker both in terms of shadowing applied and in terms of apparent skin tone. It also accentuates any of the wrinkles and blemishes on the skin which has the effect of making the character appear older at a further distance.
In case you were thinking this is just your eyes tricking you, here is a comparison of DLSS 5 on and off with the same character in the same position close and far from the camera. When closer to the camera, DLSS 5 is clearly better able to maintain the complexion of the person. And when further away, the skin tone is darkened. When even further away, the skin tone change is even more significant. Despite this change not occurring when DLSS 5 is disabled.
Here's another example. When close up, DLSS 5, I think, looks better. This is what Nvidia usually shows in screenshots. But when further away, the same character in the same position in the same lighting becomes darker in skin tone, shinier, like some sort of oil has been applied to the skin, and slightly older with more accentuated skin wrinkles. Not only that, but the playing kit becomes darker even though in the close-up shot it has the same color both with and without DLSS 5 enabled.
And you'll see the exact same issue looking at the reverse angle where again the playing kit becomes a darker shade of blue and his skin becomes shinier. Across all of these examples with DLSS 5 disabled, the character looks basically identical whether you are viewing up close or from a distance outside of obvious resolution and draw distance differences.
On characters that already have darker skin, this effect is less pronounced. I'm guessing because less darkening can happen. But even in this example, DLSS 5 makes the skin look shinier at a distance, which isn't apparent when viewing close up. Same issues on this female character. At a distance, DLSS 5 produces darker, shinier, and older skin relative to close-up shots of the same person.
I also saw inconsistencies like this. Pay close attention to this character's skin on his face and hands. With DLSS 5 disabled, the skin tone is relatively consistent between the face and hands. They're supposed to be attached to the same body, but there's slight lighting differences due to the scene's shadowing. When I enable DLSS 5 at this distance, the face is sort of similar in skin tone, but blemishes are exaggerated quite significantly, which makes the character look older.
But now look at the hands. They are a completely different skin tone to the face, having shifted redder since DLSS 5 has been applied. If we flick back and forth, you can clearly see the skin tone change in the face versus the hands.
All of these issues with inconsistent skin tones and lighting occur because DLSS 5 isn't simulating the lighting in the scene, and it doesn't actually know the full context of the scene and the objects in it. It's merely detecting things in certain areas and applying localized changes. In this scene, the model doesn't appear to know that the hands and face are connected to the same person. So with slight lighting differences on the areas in the rendered frame, it interprets what the output should be differently for each body part.
And in the examples where skin tones change when zoomed in or out, that's because the model doesn't actually know that the character is the same and is being lit in the same way in both scenarios. It reanalyzes the scene at each frame, and the model changes its decision about what lighting tweaks to make in the zoomed out versus zoomed in frames. This doesn't happen when lighting is simulated, which creates more consistent results at different camera positions.
And this applies to objects as well. While DLSS 5 is broadly consistent, so you won't notice massive changes as you pan the camera or walk around in the scene, objects definitely change depending on how close they are to the camera. Take this Gatorade tub for example. With DLSS 5 disabled, it has basically the same shade of orange when viewed up close versus further away. But with DLSS 5 enabled, the orange color darkens noticeably at a distance, almost becoming red. The tub's plastic finish also becomes shinier in my opinion, similar to what happens to skin when zoomed out.
In this example, we're in one of the stalls, and I want to focus here on the green hat. With DLSS 5 enabled, I think the material quality is improved. The hat looks more realistic and is better lit when close up. But then if I walk out of the store and view the same hat in the same lighting from a distance, the DLSS 5 image has become darker and the surface of the hat no longer looks like the fabric I saw previously. It's shinier. Whereas with DLSS 5 disabled, it basically looks like the same hat, same color, same material, but further away. Of course, the zoomed in quality of the hat is not as good, but the consistency is superior, and in my opinion, it looks better without DLSS 5 enabled from a distance.
And this isn't nitpicking certain scenes or items. It's pretty common throughout the game. You might have noticed the tracksuit on the left being darkened when DLSS 5 is enabled. And yes, it's darkened more at this distance than in this other closer angle. The ball in the game becomes shinier at further distances when the texturing on the surface is no longer resolved. Items can become slightly more shiny or slightly less shiny depending on the camera position. And this happens despite what I'd describe as a relatively toned down approach to DLSS 5 compared to the initial showcases.
Of course, you may simply not notice these inconsistencies that I'm referring to here. They can be hidden in game motion. You're not always going to be thinking consciously about whether an item looked exactly the same color at different angles. And to be fair, a lot of the time DLSS 5 is actually very consistent in what it produces, especially when on-screen items are larger or stay in roughly the same position. You can leave areas and come back. DLSS 5 will look pretty much the same for a large portion of the image.
But to me, inconsistencies are one of those subconscious immersion-breaking things that suddenly remind me what I'm looking at is actually not realistic or photoreal, but a game with graphical flaws. And when these flaws occur on faces, something that humans are very tuned into, it becomes more immersion-breaking. I can very much see this being an issue in games where you're walking around seeing NPCs that look a certain way only for their skin tone to change when you enter a cutscene or closer up conversation.
DLSS 5 also just doesn't know how to handle some items. This plastic mannequin, for example, it looks pretty normal when DLSS 5 is disabled, but after enabling neural rendering, for some reason, it adds muscle tone and inaccurate shading to the plastic item. The model seems to think this is a human with gray skin, so it looks wrong.
I also saw this example with a player character statue which appear in the city world to denote the current leaders for certain in-game competitions. These are supposed to look exactly like the characters real people have designed. But what is going on with this one? When DLSS 5 is enabled, it looks all sorts of wrong, very bizarre results.
Aside from rendering inconsistencies, the other main issue with DLSS 5 is noise and artifacts. As DLSS 5 only processes the current rendered frame, it's prone to noise in motion. Like a lot of artifacts, whether you notice this depends on how sensitive you are to noise, but generally I've found DLSS 5 to add noise to the image that wasn't there previously. The main cause of noise in the DLSS 5 image is a subtle difference in generated lighting between frames.
In this scene, focus on the pedestals in the background behind the court. In motion with DLSS 5 disabled, there's very little noise on this surface. When I enable DLSS 5 and perform the same tracking shot, you'll notice this flickering and boiling in the lighting on that surface. This is a clear stability issue introduced by DLSS 5 in its attempts to more accurately represent surface interactions in that area.
Like the previous issues we saw with DLSS 5 inconsistencies, this is generally cleaned up by simply viewing the object from a closer distance. If I perform the same tracking shot, but with the in-game camera closer to these pedestals, you'll notice pretty much no visible noise. The AI model appears better able to identify the surface when it's more dominant in the frame compared to the shot from further away.
These boiling artifacts are seen on most DLSS 5 faces from a distance. In this example, the slight shifts in the face in motion, the small amount of animation triggers instabilities on the surface of the skin. When I disable DLSS 5, obviously the quality of the face is reduced in some ways and it becomes flatter and more waxy, but the boiling artifacts are removed and the image becomes more stable.
In this example of a speaker, I'm enabling and disabling DLSS 5 while panning the camera. And you can see the boiling artifacts and instabilities appear when DLSS 5 is switched on, then disappearing when DLSS 5 is switched off. All of these examples are at 4K. This orange shelf is slightly boiling in motion. It's almost like when ray tracing isn't done at a sufficient sample count and there's too much denoising going on. With DLSS 5 off, you get a very smooth, very consistent surface. And while DLSS 5 is attempting to improve that surface here, it can only do so with noise in motion.
Here's another example where as we jog towards this door, we see flickering and pulsing in the door frame, but only with DLSS 5 enabled. It's otherwise stable with the feature disabled. Here's a slow panning shot of the crowd with DLSS 5 switched on and off. You can clearly see additional boiling and pulsing when DLSS 5 is enabled on shirts and skin. You'll also see this on your character's clothes at times, even standing still, because a slight swaying animation can trigger instabilities.
Not even close-up shots of characters are saved from these issues. And this was one of the better examples of DLSS 5 improving image quality when there is no motion. Focus here on the ears and lips as we move around the head. You'll clearly see boiling on the surface of the skin that simply isn't there when DLSS 5 is disabled. And again here where we see lighting inconsistencies on the hand as we move the camera angle.
I have no shortage of these examples because these issues kept cropping up in the majority of scenes that I looked at. The flicker on this tree trunk, for example, this was a pretty common one despite DLSS 5 improving the look of trees when stationary. These artifacts are visible in actual matches within the main arenas as well around the basket and the crowd.
Anything that's smaller, things that have a flat and smooth surface or typically darker areas in the scene or darker surfaces are more prone to DLSS 5 noise. Sometimes I saw quite bad lighting flicker where the changes in lighting between frames were quite noticeable on certain surfaces and repeating the same moves or motion without DLSS 5 would eliminate this flicker.
The other thing that DLSS 5 really struggles with is disocclusion. If the AI model doesn't have a consistent uninterrupted picture of the surface or material it's dealing with, this causes issues. For example, focus on this section of concrete which is seen through a wire mesh. Without DLSS 5, we get a relatively consistent image. That concrete remains the same color in motion. When DLSS 5 is enabled, you'll see the tone of that concrete change. It gets brighter when I move the camera and it settles to a darker state when I stop moving.
Here in this store, the lighting inside this display area is much more consistent and stable when DLSS 5 is disabled. With neural rendering enabled, not only do we see flicker, but also disocclusion problems where the amount of lighting changes as the character passes in front of this area. I saw a similar issue in this test where the tone of the hoodie changes depending on how much of the hoodie I've revealed. With DLSS 5 disabled, the tone is completely consistent in motion and doesn't change depending on how much is shown.
When looking at this blue car through a wire fence, we see slight haloing inside the diamond segments that you don't see when DLSS 5 is disabled. And this example is probably the worst for disocclusion problems, as DLSS 5 produces a bright halo-like effect around the character as it runs over grass. When I first saw this, I thought it might have been a rendering error with the game's engine, but it's clearly not there when DLSS 5 is disabled.
And interestingly, this specific issue seems tied to the color of your character. I messed around with different outfits and skin tones, and it was clear that some combinations produced less of a halo effect.
Maybe I'm just overly sensitive to noise, but once I discovered the problems with boiling, flicker, pulsing, inconsistent lighting, subtle rendering changes from frame to frame, they became hard to unsee. Games are played in motion. And while DLSS 5 can look amazing in the right conditions when there is no motion at all, like in a photo mode or screenshot tool, the artifacts present in motion detract from its transformative effect.
DLSS 5 is also unable to improve many aspects to the overall presentation, which limits its ability to properly create photoreal visuals, a claim I've often seen in association with this technology. DLSS 5 really only adjusts the lighting on various surfaces throughout the game, attempting to generate better material properties and light interactions. It cannot improve a low-quality texture into a realistic texture. It cannot fix a noisy or low-resolution reflection to make it look higher quality. It won't solve a flickering or low-resolution shadow. It can't make low poly geometry look realistic. All of these things are passed through the DLSS 5 system untouched, so to speak.
And in a game like NBA 2K27, you often run into these problems, especially low-quality textures and low geometry detail, even on max settings. Nvidia claims that the output from DLSS 5 improves when the input is better, which is why they recommend path tracing as an input into neural rendering. It's quite clear from this game that this isn't going to be a technology that can be used to cheat visual quality, at least in its current form.
The last thing I was curious about is whether the way DLSS 5 post-processes the frame is something that can be replicated with a traditional filter. So, I fired up Nvidia's filter tool they include with the Nvidia app overlay and got to work trying to create my own filter-based version of DLSS 5.
I'm going to show you three examples here, and I'd like you to guess which one of these images is DLSS 5 off, which is DLSS 5 on, and which is DLSS 5 off, but with game filters applied. So, here we have the examples. We have A first up, then a look at B, and lastly, a look at C. Let's run through those again. A first up, then B, and then C. If you guessed that C was the DLSS 5 image, you would be correct.
Let's show another set of examples. Here is option A, here is option B, and here is option C. Once again, option A, option B, and option C. In this situation, option A was DLSS 5 enabled. Actually picking out the DLSS 5 image in some of these scenes is not particularly easy, which might say something about how transformative DLSS 5 is.
However, in my opinion, after examining the impact of filters closely, I don't think you can fully replicate the look of DLSS 5 with a filter. DLSS 5 does selective adjustments to different surfaces that aren't possible without a filter that can also mask and apply changes to different objects in real time. It also applies some contact shadowing and ambient occlusion that cannot be replicated with simple tone mapping or contrast adjustments.
With that said, increasing the clarity filter and increasing the contrast do partially replicate the overall look of the DLSS 5 output, which tends to be a bit more contrasty, not as flat, better clarity than the standard image. It'll be interesting to go through and dial in more advanced filters to see what the overall delta is between DLSS 5 post-processing and something lower in processing cost that is trying to replicate the same look, but that's for another day.
And this brings me to my final thoughts on DLSS 5 as seen in the official NBA 2K27 implementation. Bit of a mixed bag this one. DLSS 5 does some things really well and in the right circumstances it can provide transformative visual improvements. But it's only able to do this with some pretty significant downsides, worse stability in motion, lighting inconsistencies, and the huge performance cost that makes it practically unusable on all but the highest end GPUs of today.
I've spent a lot of this video going through some of the issues that can crop up using DLSS 5, which may give the impression that I think this technology is awful, but that's not the case. When DLSS 5 works well, I actually think it looks really good. That first presentation that I saw back in March had all the AI slop alarm bells going off, but the launch version as implemented into a real game by the developers doesn't have that same look.
Turning down DLSS 5 has maybe made it less transformative than those original demos, and less transformative visuals makes it harder to justify the performance cost, but avoiding AI slop is absolutely the right move. Had NBA 2K with DLSS 5 looked like that Hogwarts Legacy or Resident Evil demo, there is absolutely no way I'd use it.
What Nvidia has actually been able to produce is a technology that in perfect circumstances makes the game look better in my opinion. Of course, there are some scenes where DLSS 5 does very little and some scenes where DLSS 5 changes visuals without being an obvious improvement, but on the whole, I think DLSS 5 makes NBA 2K27 look better. Strip away the artifacts and performance cost, and I'd prefer to play with DLSS 5 enabled.
Where opinions will differ is on how transformative DLSS 5's visuals are at the end of the day and whether that justifies the massive performance cost. In some scenes, I'd say it absolutely makes a big difference. And with some level of performance cost, I'd be willing to enable this feature. However, on the whole, examining DLSS 5 across a variety of different scenes, I'd say that DLSS 5 right now is both not quite transformative enough in NBA 2K and too performance taxing to actually enable in game.
And that's in scenarios where artifacts are not problematic. I'm just talking here about the general graphical improvement. The stability and consistency issues with DLSS 5 further hurt its usability. The way DLSS 5 is able to enhance surface lighting comes at a clear artifact cost. You'll spot boiling, flickering, and pulsing effects in motion. The neural rendering system doesn't handle disocclusion very well, creating weird halos and lighting shifts in some situations. And the way materials and surfaces are enhanced changes depending on how close an object is to the camera. The further away the object, the darker and shinier it tends to appear and the more prone it is to artifact issues.
Without an acceptable level of stability in motion, the initial release of DLSS 5 isn't a great fit for actual gameplay. It's great for screenshots and photo modes where you can pause the game and avoid artifacts. This is why DLSS 5 tends to look reasonable in screenshots you might have seen on forums or social media. It's also great for cutscenes where character models are more likely to be closer to the camera, which avoids most skin tone shifts. In actual gameplay though, once you start noticing the issues with DLSS 5 in motion, it's hard not to see them everywhere.
Nvidia really needs to get this technology into a position where the image quality that you see in still images and screenshots is what you also get in motion. That would make me much more likely to enable it.
This actually reminds me a lot of the ray tracing era, which we're currently still in. Ray tracing can substantially enhance visuals if it's implemented well, but often these implementations have been very noisy. The boiling we see on DLSS 5 enhanced surfaces is quite similar to boiling from denoising in some ray traced games. This is one of the aspects I hate the most about modern game visuals. And right as it seems like we're finally making some progress to remove these artifacts with technologies like DLSS 4.5 ray reconstruction, the industry takes another step that introduces those same artifacts back into game visuals.
When you combine everything that we're seeing here, DLSS 5 feels like a tech demo that's not ready for everyday use. Like with ray tracing, there's some promising aspects to this technology. How it works is quite impressive. I can definitely see a path to this being usable at some point. But when that will happen is very much up in the air. Is neural rendering going to be relevant to the majority of PC gamers anytime soon? Probably not.
Nvidia claims they are continuing to optimize the DLSS 5 model to improve performance and expect further gains by the end of this year. That's nice. It would be good to see some progress there. But performance is not the only issue with DLSS 5. If the model ran well, I'd still be annoyed with the artifact and inconsistency issues which appear related to how DLSS 5 is designed to work. It only processes the current rendered frame and some motion vectors.
Can Nvidia both significantly improve the performance of the neural rendering model while also solving the stability issues and while further improving DLSS 5's visual output? I'm not sure. What seems more likely to me is that neural rendering will actually require more powerful GPU hardware, which is basically the same thing we saw with ray tracing.
Yes, there were optimizations compared to the initial release in Battlefield V, but ray tracing only really showed its true potential on much more powerful hardware than the RTX 20 series. And even today, 8 years later, mainstream hardware struggles to run the technology. My guess is that DLSS 5 neural rendering models will need to become more complex to resolve the visual issues and improve the output, which will outweigh any optimization gains and be generally just harder to run in future iterations.
With the current pace of GPU hardware development, it's going to be quite a while before neural rendering truly ushers in a new generation of PC visuals. That cannot happen until the most popular mainstream models can run a better version of the technology. And we seem miles off that occurring.
The good news is that none of this is really a problem right now because it's fine for DLSS 5 to just be a tech demo. It's great that owners of higher-end RTX 50 series GPUs can play around with this feature in photo modes and of course all the mods that have exploded in the last few weeks. Like with ray tracing and path tracing, you can get a glimpse of how PC visuals are evolving over time. This is how things should be. New tech comes to market. It's optional. You can experiment with it. We wait patiently until it's fully usable and life goes on.
What would be really bad is seeing Nvidia weaponize DLSS 5 like they have done with many technologies in the past, chiefly ray tracing, where they tried to make gamers pay up to access a feature that absolutely was not ready for prime time. That was our number one biggest issue with ray tracing. The RTX 20 series offered a poor upgrade in the vast majority of circumstances because you were being sold on a future of ray tracing that never really arrived for those cards.
These are technologies that take a decade or more to fully establish themselves. Using them to squeeze cash out of gamers is bad. Just let this technology cook over time. Don't try and make gamers pay more to access a feature that isn't ready.
The other thing I do wonder about DLSS 5 is what other things could be done if we accept such a huge performance hit. Could we have the FPS and solve pop-in, improve draw distances, significantly improve geometry and world detail, do more ray tracing? We don't usually see these frontiers pushed because developers don't want to have FPS. But what could games do if that performance hit was on the table? Is neural rendering the best use of that processing time?
Unfortunately, I can't answer those questions because we have DLSS 5, but we don't have performance destroying options chasing other aspects of visuals. I hope this has all made it clear why I'm not hugely excited about DLSS 5. I love practical features that most gamers can enjoy. I love high value features that I'd choose to use while gaming. So, things like upscaling are what I get excited about.
DLSS 5 is a very interesting technology. How it works is impressive. Exploring its strengths and weaknesses has kept me busy and entertained, but ultimately it's not something I'd personally enable in my gaming sessions right now, and I think it will be a slow burn until I actually want to turn it on during gameplay. We'll see though. Maybe Nvidia can surprise me with a crazy pace of development. Hopefully, the software doesn't outpace the hardware.
So, anyway, that's it for this look at DLSS 5, the image quality, the trade-off between, I guess, the visual improvement that you're seeing and the performance. If you do want to see more of those performance numbers, we have that separate video that we covered looking at all the RTX 50 series GPUs in NBA 2K27 and the performance hits when enabling DLSS 5. So, go check that out if you're interested.
If you do want to support Hardware Unboxed and all of the independent testing that we do, it takes a lot of time and effort to put these sorts of videos together. I've spent quite a long time on this one. It's pretty much since the launch of NBA 2K27, I've been non-stop looking at this feature. So yeah, if you want to support us, Patreon is the best way to do that. You can sign up in the description below. If you sign up, you get access to all the usual benefits, monthly live streams, BTS content, the Discord chat, great place to chat about these things, a free version of our podcast if you're a podcast listener.
So anyway, that's it for this one. Thanks for watching and I'll catch you in the next one.

Cite this transcript

APA, MLA, BibTeX
APA

Schiesser, T. (2026, September 12). DLSS 5: The Good, The Bad and The Ugly (Full Analysis) [Interview transcript]. Hardware Unboxed. CEOInterviews.AI. https://ceointerviews.ai/interview/3228994/

MLA

Tim Schiesser. "DLSS 5: The Good, The Bad and The Ugly (Full Analysis)." Hardware Unboxed, 12 Sep. 2026. Transcript, CEOInterviews.AI, https://ceointerviews.ai/interview/3228994/.

BibTeX
@misc{schiesser2026_3228994,
  author       = {Tim Schiesser},
  title        = {DLSS 5: The Good, The Bad and The Ugly (Full Analysis)},
  howpublished = {Interview transcript, Hardware Unboxed. CEOInterviews.AI},
  year         = {2026},
  month        = {sep},
  url          = {https://ceointerviews.ai/interview/3228994/},
  note         = {Speaker-attributed transcript with timestamps}
}