DLSS 5 looks impressive, but implementation will matter. DLSS 5 can dramatically reshape how games look, but its success will ultimately depend on how carefully developers balance AI generation with their original artistic vision.
When NVIDIA first showed DLSS 5 earlier this year, the reaction was understandably mixed (partly on their own doing by showing really poor examples!). The technology was impressive, sometimes strikingly so, but generative AI changing how a game looks in real time immediately raised a much bigger question: how much of the final image should we really leave to AI?
A lot has happened since then. NVIDIA went into greater detail about DLSS 5 at SIGGRAPH, showing how developers can control the generative model, while recently, a leaked DLL from NBA 2K27 has already escaped into the wild and been hacked into everything from Control and Silent Hill 2 to Hogwarts Legacy. Some of those unofficial implementations look fantastic. Others, not so much.
I recently attended another NVIDIA developer briefing on DLSS 5, where the company went deeper into how developers will actually work with the technology, along with new assets showing the different models, controls and visual comparisons. Most of the underlying technology isn’t new if you watched NVIDIA’s SIGGRAPH presentation, but seeing all of those tools together makes one thing increasingly clear: there isn’t really going to be one definitive “DLSS 5 look”.
How good DLSS 5 ultimately looks is going to depend heavily on what developers do with it. It releases on September 3 with NBA 2K27 being the first game to ship with it.
DLSS has so far primarily been a reconstruction technology. Super Resolution attempts to reconstruct a higher-resolution image from a lower-resolution input, while Ray Reconstruction replaces hand-tuned denoisers to recover cleaner ray-traced lighting.
DLSS 5 crosses into generation. The game engine still renders the world and defines the geometry, objects, characters, camera and basic lighting relationships, but DLSS 5 sits afterwards and generates an additional layer of appearance over the rendered frame. NVIDIA’s model understands what it is looking at, so it can treat skin differently from cloth, foliage, glass, metal or hair.
That allows it to generate visual cues that would otherwise be extremely expensive to calculate conventionally: better subsurface scattering through skin, more convincing light transmission through hair and leaves, improved contact shadows, ambient occlusion, reflections and more complex material responses.
And while all of those things sound fairly minor when written down separately, the combined effect can be quite dramatic.
The foliage comparison NVIDIA provided is probably the clearest example. With DLSS5 disabled, it is already a fantastic-looking 3D scene, but there is a certain flatness to the foliage and lighting that still makes it immediately identifiable as a render. Switching DLSS 5 on changes how light interacts across the leaves, adds more believable separation between overlapping foliage and strengthens the shadowing and material response across the scene.
It is the accumulation of all these small changes that makes the difference. The leaves feel less like individually rendered pieces of geometry and sit more naturally within the environment, while the stone, wood and surrounding vegetation gain another layer of depth. It starts moving away from “very impressive game graphics” and closer towards something that could, at a quick glance, be mistaken for a photograph.
The merchant comparison shows the same idea applied to a character. His skin gains more convincing light response and depth around the eyes, cheeks and facial contours, while his clothing, jewellery and surrounding objects respond more naturally to the lighting. Again, DLSS 5 isn’t replacing the character or changing his geometry. It is changing how the existing character is finally presented.
The ramen chef is another good example, particularly around the skin, ear, facial creases and clothing. The original render looks good, but the additional occlusion, subsurface scattering and material response makes his face feel less like a highly detailed game character and more like actual skin reacting to light.
These are, of course, NVIDIA’s own tech demos. They have been designed, tuned and presented specifically to showcase DLSS 5 at its best, and that needs to be kept in mind. The leaked DLSS experiments have already demonstrated that simply throwing DLSS 5 into a game doesn’t automatically make it look better.
I have seen hacked implementations in games such as Silent Hill 2 and Hogwarts Legacy that look excellent, while others can introduce an overly processed or artificial quality to the image. But those experiments are also unfair to the technology in another way: the developers of those games never designed or tuned their visuals around DLSS 5.
And NVIDIA is giving developers considerably more control than a simple On button.
One of the more interesting aspects of DLSS 5 is that developers can choose between different trained models, each producing a slightly different result.
NVIDIA’s comparison shows the exact same ramen chef running through Model A, B and C, all at 80% Structure Intensity. They aren’t radically different, but they aren’t identical either. Model A produces a slightly darker and more contrast-heavy face, Model B looks more balanced to my eyes, while Model C produces a somewhat softer response around the skin and eyes.
I personally prefer Model B from this comparison, but my preference ultimately doesn’t matter. I am not building the game.
Developers can pick whichever model better suits the scene they are creating, and NVIDIA says they aren’t locked into using a single one throughout the entire game. A studio could theoretically select one model for an indoor environment, another for an outdoor scene and another for a cutscene if that produces the look they want.
That’s where DLSS 5 starts becoming less of a graphics toggle and more another tool in the art pipeline.
The model itself is only the first layer of control. NVIDIA also provides separate Structure Intensity and Tone Intensity settings.
Structure Intensity controls how aggressively DLSS 5 applies things such as contract shadows, ambient occlusion, reflections and subsurface scattering. At 25%, the ramen chef remains much closer to the original render. Moving through 50% and eventually 95% gradually increases the additional shading and material response, with the face, eyes, ear and clothing becoming progressively more defined.
There isn’t a “correct” value here. An artist might decide that 95% fits one character perfectly while completely destroying the intended look of another.
Tone Intensity is separated from Structure and controls more of the color and lighting response. If an artist likes the additional material detail DLSS 5 provides but doesn’t want the model touching the color palette or broader lighting response, Tone Intensity can be pulled all the way while retaining the Structure enhancement.
This separation matters because photorealism isn’t automatically an improvement. Games have an intended mood, color palette and visual identity, and sometimes making something more physically realistic could make it look worse. A game designed around exaggerated lighting, stylised characters or deliberately unusual colors doesn’t suddenly become better because an AI model decides what real skin or foliage should look like. The developer’s art direction has to come first.
This becomes even more granular with masking.
DLSS 5 includes something NVIDIA calls Model Automask, where the model semantically understands what is inside the image. In the ramen chef example, it can identify the chef as the base character, including his face, body and hands, and developers can change the intensity on him independently from the rest of the environment.
NVIDIA also exposes developer-created masks from the game engine.
One of NVIDIA’s supplied images demonstrates this particularly well. The pitcher, grapes and bottles are independently highlighted and can each receive completely different Structure and Tone settings. Developers can group objects together too, so all of the foliage, food, clothing or whatever other category makes sense for that particular game can receive its own treatment.
There is effectively no requirement for DLSS 5 to be smeared across the entire frame at one universal intensity.
This is probably the part of DLSS 5 I find most important. NVIDIA is putting a powerful generative tool into the rendering pipeline, but the technology is considerably more useful if the AI remains underneath the artist rather than the other way around.
I don’t particularly want a developer to hit 100% on every slider and call it a day. If DLSS 5 improves the foliage but hurts a character’s face, mask the character out. If a particular model ruins the lighting tone of an environment, use another one. If a game doesn’t benefit from DLSS 5 at all, don’t implement it. AI is a tool. The results depend on how it is used.
That is also why I am far more interested in seeing native implementations than another dozen videos of the leaked DLL being objected into random games.
During NVIDIA’s presentation, NBA 2K27 was used as a real game example, with Visual Concepts tuning DLSS 5 specifically around the game’s goal of recreating an NBA broadcast. Player likeness is particularly important here because these aren’t fictional faces the model has some room to reinterpret; they are recognisable real-world athletes.
The footage NVIDIA showed was extremely impressive. Skin shading, eye reflections, light passing through ears, clothing contact and hair all combine to make some of the player models look remarkably close to real footage.
More importantly, Visual Concepts is able to tune player characters separately and use per-pixel control masks to keep their likeness intact. That’s much closer to how I want to judge DLSS 5 than an unsupported mod where somebody has dropped a leaked runtime into a game whose developers never accounted for it.
If native implementations consistently look like this, there is plenty to be excited about. But implementation is going to matter, and studios that lazily use DLSS 5 deserve exactly the same criticism we already give developers for poor optimization, shader compilation stutters, and using upscaling as a clutch for “performance”.
The other issue is performance, because none of this AI-generated lighting comes for free.
NVIDIA has made substantial progress here. The company says DLSS 5 initially required two RTX 5090 GPUs when it was first demonstrated, before eventually being reduced to a single 5090, and the current implementation is now more than five times faster than where it started.
Its NBA 2K27 performance figures look ridiculous at first glance. At 4K with Ultra settings and RT, NVIDIA showed an RTX 5090 reaching 370fps, while the RTX 5080 reached 232fps. At 1440p Ultra, the RTX 5090 was shown at 594fps, the 5080 at 413fps, the 5070 Ti at 352fps, and the RTX 5070 at 261fps. At 1080p High, even the RTX 5060 was shown at 258fps, rising to 324fps on the 5060 Ti and 285fps on the 5070.
However, every single one of the benchmarks were running DLSS MFG at 6X. The 4K results also use DLSS Super Resolution Performance Mode, while the 1440p and 1080p results use Quality Mode.
The early unofficial results from the leaked DLL also show how expensive the neural rendering pass itself can currently be. One widely circulated Control test on RTX 5070 Ti at 4K dropped from 71fps to around 35fps after the leaked DLSS 5 Neural Rendering implementation was enabled, effectively cutting performance in half. Other community tests have shown similarly large drops.
Those are unsupported hacked implementations and absolutely should not be treated as representative of a properly integrated shipping game. But they do demonstrate that DLSS 5 isn’t a free visual upgrade.
And NVIDIA’s own decision to show every NBA 2K27 performance result with 6x MFG makes it fairly obvious that MFG is going to play a significant role in clawing that performance back.
That also creates an accessibility problem. NVIDIA’s polished performance demos are based entirely around RTX 50-series cards and 6x MFG, which immediately leaves a huge chunk of the PC market unable to reproduce the same performance strategy. RTX 20-and 30-series users don’t have FG at all, while RTX 40-series hardware doesn’t have MFG.
Modders have already forced the leaked technology onto some older RTX cards, but performance there ranges from rough to downright unusable. So while DLSS 5 may technically represent NVIDIA’s next step after PT, it isn’t going to be equally practical for everyone, at least in its current form.
I also don’t particularly want all of these developer controls exposed to players.
I don’t need Model A, Model B, Structure Intensity, Tone Intensity character masking and another twelve sliders sitting inside a graphics menu. Those are artistic decisions and should remain with the people creating the game. If the developers believe Model B at 70% Structure Intensity gives a character the correct appearance, that’s their call.
What I do want is DLSS 5 On and Off.
Even if the developers tune everything perfectly, the final image is still being altered by a generative model, and some players are simply going to prefer the conventional render. That choice should remain with them rather than DLSS 5 eventually becoming inseparable from Super Resolution or being baked permanently into a game’s rendering pipeline.
From what I have seen so far, though, I am more excited about DLSS 5 than worried about it. AI itself isn’t inherently good or bad; it is another tool, and lighting is one of the most computationally expensive problems developers continue to deal with. If neural rendering can help artists achieve the image they wanted faster and more efficiently, while preserving their creative intent, I don’t have a problem with how the pixels got there.
The important bit is keeping that order intact. Developers should decide what their world is supposed to look like, DLSS 5 should help them get there, and the player should still be able to decide whether they want the AI-generated version on their screen.
If that balance holds, DLSS 5 could be an extraordinary graphics technology. Its success will ultimately depend on whether developers use it with the same care and restraint they bring to every other part of a game's visual design.
