NVIDIA is addressing the biggest concerns regarding DLSS 5 while presenting the tools it offers to developers to retain "artistic integrity" at SIGGRAPH.
NVIDIA Says DLSS 5 Is Not "AI Replacing Graphics" but rather "AI Expanding Graphics", As It Promises More Control To Preserve Artists' Vision While Delivering Enhanced Visuals
At GTC 2026, NVIDIA unveiled DLSS 5, its next major update to the DLSS technology suite. The first DLSS off vs DLSS on comparisons showed a day-and-night difference in visuals, massively uplifting the quality of existing assets through Neural Rendering, but as soon as the technology was unveiled, there was some major backlash surrounding the tech.
The main concern was that NVIDIA was using AI to change the artistic vision. The company was quick to clarify that this wasn't the case and devs/artists would have full control over the changes made with DLSS 5. Some even claimed that DLSS 5 replaces the original models with AI-generated assets, and that was also not the case. However, the initial reaction wasn't what NVIDIA expected, and they have been tight-lipped ever since.
At Computex, NVIDIA announced DLSS 4.5 Ray Reconstruction, but made no comments regarding DLSS 5. We talked to several representatives about what happens to DLSS 5 moving forward, and they only told us to "stay tuned". Well, it looks like NVIDIA is once again ready to talk about DLSS 5, and at SIGGRAPH, they are debunking some of the rumors surrounding the GenAI claims made around DLSS 5, and further diving into the tech, explaining how they overcome major challenges while delivering world-class tools to developers to maintain the artistic standards originally set by them.
NVIDIA Conceptualized DLSS 5 All The Way Back In 2018 With Turing RTX
Since DLSS 1, NVIDIA's architectures have evolved immensely. With new Tensor Core and Neural technologies, DLSS has also evolved tenfold. NVIDIA's Director of Applied Deep Learning Research, Edward Liu, says that what we saw at GTC was just a short glimpse of DLSS 5, and at SIGGRAPH, they further expand on the technology. He says that the journey for DLSS 5 began back with Turing, the first RTX GPUs with DLSS 1 support. It was always projected that graphics and AI would have to advance together to achieve photo-realism one day.
DLSS 5 brings NVIDIA closer to the goal. DLSS 4 and DLSS 4.5 with Path Tracing have brought games, movies, and content close to reality, but NVIDIA says they still aren't at the point where graphics look indistinguishable from reality. Traditional graphics are still hand-curated by artists who author the scene, simulate the light, simulate light bounces through physics, & then frame them with a camera. This is modern-day rendering.
But Generative Models were quick to understand photorealism, and once those models are fed with enough of the real world as a reference, they would learn almost everything about what real-life looks like, everything, every bit from lighting to physics to simulations. While that is great, these models are slow to generate and giant in complexity and size. So NVIDIA decided to work out a way so that both traditional rendering and GenAI could be combined. That's what DLSS 5 is meant to achieve.
DLSS 5 Reborn At SIGGRAPH: A Refined Approach to Maintain Artistic Style While Enhancing Visuals Through AI
We say DLSS 5 has been reborn at SIGGRAPH because after its initial reaction, NVIDIA had to make some major changes to the technology. Now the concept remains the same: to harness Neural Rendering capabilities that modern RTX platforms (GPUs, PCs) have to offer, & use them to deliver a leap in graphics fidelity.
Now DLSS 5 sure looks great, but three core challenges needed to be overcome. These are:
- Challenge 1 - Preserving Artistic Intention
- Challenge 2 - Frame-by-Frame Streaming
- Challenge 3 - Real-Time 4K
Achieving realism is only half the problem; the other half is staying true to the original vision and content. With GenAI, this is hard, sure you can ask for something once, and it will generate a very real and nice asset, but the second you ask it to repeat the process, it will come with a different asset. Let's take a portrait as an example: GenAI is probabilistic and won't stick with one model each time; every time, it will make a different portrait that sways away from the original content.
Imagine you are playing a game, you turn on DLSS, and it makes your character's face look great, but the moment you turn it off and on again, or the model regenerates the character's face, it'll look different. That's not a wise choice to do in real-time rendering, and that gets in the way of the original artist's vision. This is why NVIDIA had to come up with a precondition for the trained model that is more constrained and rigorous.
And they found an even better solution: to use the frame the game just rendered. The frame isn't an input, prompt, reference, or proxy scene; it's a renderer's own output, which is universally available across all engines. So NVIDIA is simply taking that rendered frame and making it more realistic.
DLSS 5 Model Is Trained To Preserve The Original Asset, Not Change It
But even that had its own challenges, which brings us to Challenge #1: "Preserving Artistic Intent". NVIDIA acknowledged that they had a hard time with this approach early in DLSS 5's development, as each time they gave the model a render frame, the model changed it.
To overcome this, a hard line was drawn to train the model in a way that the artistic vision was respected. NVIDIA used the renderer's own internal buffers, surface normals, lighting information, and other aspects to make sure that the model stays consistent with the original intent.
With the model now bound to a specific set of rules, it found ways to freely enrich "what sells realism". It started doing better subsurface scattering, finer material responses, transmissions of light across materials, hairs, foliage, better contact shadows, and better environment lighting. So the asset would never be changed by the model, but the aspects around it will improve in a way that the scene is a better rendition of what it was before that abides by the original artist's vision.
Think of how Shader Models evolved graphics, or how tessellation and ray-tracing make the visuals richer and more detailed. DLSS 5 is doing the same thing; it is applying AI to graphics rather than replacing graphics with AI.
DLSS 5 Leverages Motion Vectors To Tackle Temporal Streaming
The second challenge is "Frame-by-Frame Streaming," and what it means is how the model will interpret the frames in real-time. The difference between the GenAI and DLSS 5 model is that GenAI works in chunks, and the model reasons across multiple frames and generates a batch of frames at once.
But DLSS 5 is doing that in real-time, and it cannot simply apply the same GenAI technique here since the model has to respond immediately to user input in real time and render frame by frame.
The solution comes from within the game engine itself. DLSS 5 leverages the engine's very own capabilities, such as motion vectors, so that the model does not have to infer across pixels and frames to figure out what the motion is going to be.
The DLSS 5 pipeline starts at the Render, which the DLSS 5 model takes in and renders out, and repeats. This smooth and consistent process allows the scene to remain temporally stable and coherent. NVIDIA says that DLSS 5 faces no shimmering, drifting, or swimming drawbacks, which are very much a case in GenAI models.
DLSS 5 Is 4K, High Frame Rate, Low Latency Capable & Runs on A Single GPU
The third and final challenge for DLSS 5 was "Real Time 4K". DLSS models have improved vastly, and with features such as Multi Frame Generation, now delivering up to 5 generated frames for each rendered frame, the tech has made 4K 240Hz+ gaming accessible with the best visual tech around, such as Path Tracing.
NVIDIA aims to achieve the same with DLSS 5, and they are targeting 4K 60+ FPS. The DLSS 5 model is unlike GenAI models, which are large and slow; DLSS 5 is a real-time renderer, and as such, it needed to be small and more efficient. NVIDIA distilled the DLSS 5 model by taking other models that knew about world appearance and compressing them into a much smaller "one-step pixel space diffusion transform" model that is made to do just one thing, and that is to make real-time rendering look more real.
With the new compressed model, DLSS 5 turns out to be ready for 4K, High-Frame Rate, low-latency gaming. And there's one more thing.
Remember how NVIDIA showed its initial DLSS 5 demos running on a dual RTX 5090 setup? Well, they did promise DLSS 5 to run on a single GPU, and that remains the case. The model will run on a single GPU, and even more important, it will be very efficient in terms of VRAM. Now the final performance remains to be seen as NVIDIA isn't disclosing any figures or VRAM usage yet, but it will all come due time as the company continues to polish DLSS 5 for gamers and creators.
DLSS 5 Toolset For Artists Shown For The First Time
So the challenges have been outlined with their solutions but that just overcomes one layer of control, the second layer is just as important and for that, NVIDIA's Content Artist, Gav, made it to the stage to showcase the toolset NVIDIA will provide to game developers and artists to ensure that they have more than one outcome for mapping between a render frame and a more realistic version of it.
Gav states that DLSS 5 offers control over everything that is on the scene: characters, elements, props, and the environment. He first goes on to show an example of a character that has DLSS 5 applied. Once enabled, DLSS 5 retains the original assets, models, and geometry, but adds things such as subsurface scattering, better contrast, ambient occlusion, and contact shadows, all of which uplift the game visuals. The character's face gains better reflections, and DLSS 5 does so by respecting the render frame.
DLSS 5 - Multiple Models To Choose From
DLSS 5 offers granular controls to developers, and the first set of controls is the models which come in a variety to select from. There are models A, B, and C, and each has a different output as they're trained on different parameters.
When the developer picks up a certain model, they won't be locked in to that model. The models can be mixed and matched to whatever a developer wants.
There are two global controls available in DLSS 5: Structure intensity as well as Tone intensity. What structure intensity does is modulate the high-frequency detail, such as ambient occlusion, subsurface scattering, and better reflections, of the frame, and tone intensity modulates the low-frequency details, such as lighting, of the frame. The sliders can be moved to change the intensity of each feature. For example, the tone slider changes the colors because it changes the subsurface scattering and lighting.
Structure Intensity (Side by Side)
Tone Intensity (Side by Side)
DLSS 5 also offers engine-side and model-side controls. The model automasking understands the base character, and with the slider, you can raise or lower the intensity of DLSS 5 for this base character alone. So when DLSS 5 is turned off and on, it only applies to the base character, as the model is able to understand the semantics of the scene. At 25%, the structure intensity remains close to the render frame with a midpoint, while at 95%, developers can achieve the maximum visual uplift that DLSS 5 can provide.
The other control is for the environment itself or specific items in the scene. For these, the game engine provides the control, and here, NVIDIA masks the objects in the scene such as the pitcher, the grapes, and the bottles. Developers will be free to mask as they see fit. This gives flexibility to devs to enhance or not enhance multiple items in a scene. This time, instead of the character, the mentioned objects are being uplifted with better lighting and effects such as translucency, metal reflections, etc.
The last example that NVIDIA shows is that of the foliage, essentially uplifting the environment instead of characters and objects. The reference render presented does look nice and uses ray tracing, but NVIDIA calls it a bit "plasticky". With DLSS 5, this plastic nature is overcome with life-like and photorealistic visuals, and all DLSS 5 is doing is enhancing the subsurface scattering & lighting without touching anything else.
Once again, developers will have the freedom to select the intensity of DLSS 5, but doing the same scene with traditional rendering approaches will be very costly on the hardware, while DLSS 5 uses neural rendering for a faster and more refined approach to enhanced game visuals.
Deeper Collab With Game Devs & More Controls Coming
On a closing note, NVIDIA states that this is just a glimpse of the level of controls DLSS 5 will have on offer for developers. It will be fully controllable, and they are working with various partners to integrate their feedback in DLSS 5 before its launch while building even more controls.
And since DLSS 5 will be tightly integrated with the game engine, that means NVIDIA will have even deeper collaborations with game developers and publishers, further expanding the DLSS feature suite beyond an already enormous list of games and apps. NVIDIA DLSS 5 launches this fall, and more details are to follow.
NVIDIA also says that they aren't claiming to have solved real-time photorealism with DLSS 5, infact, true photorealism is still a long way to go as the animations still need a lot of work to be uplifted in the same degree as appearance so there's a lot to come as far as AI approaches in the realm of digital media are concerned, but NVIDIA's Deep Learning and Graphics teams are showing what they do best, to solve problems and bring future-looking technologies to its audiences way ahead of the curve.
About the author: A Software Engineer by training and a PC enthusiast by passion, Hassan Mujtaba serves as Wccftech's Senior Editor for hardware section. With years of experience in the industry, he specializes in deep-dive technical analysis of next-generation CPU and GPU architectures, motherboards, and cooling solutions. His work involves not only breaking news on upcoming technologies but also extensive hands-on reviews and benchmarking.
Follow Wccftech on Google to get more of our news coverage in your feeds.
