On July 20, 2026, NVIDIA demonstrated developer-facing controls for DLSS 5 during its SIGGRAPH 2026 keynote, showing how generative AI is being integrated into real-time rendering. Developers can choose from three types of models on a per-scene basis, and adjust everything from the overall magnitude of on-screen change down to the intensity applied to specific characters or objects. At its initial unveiling in March, NVIDIA had only said that intensity adjustments and masks would be provided to address concerns that post-generation footage might alter the original art direction. This is the first time the granularity of that control has been shown.
DLSS 5 plays a different role from conventional super-resolution processing, which reconstructs images from lower-resolution input. Instead, it is a post-processing stage that takes frames already rendered by the game engine and regenerates the appearance of lighting and materials using a generative model. What NVIDIA presented at SIGGRAPH was a design that preserves the freedom of generative AI while still tying character identity, object shape, and lighting intent back to decisions made on the game side.
Determining the Degree of Change with Three Models and Two-Stage Masking
DLSS 5's demo screen showed three model options side by side: Model A, Model B, and Model C. Each model is trained with different parameters, so even when given the same character and the same intensity setting, the resulting look of skin, fabric, and lighting differs. Developers are not required to lock the entire game to a single model—they can use different models for regular gameplay scenes, other scenes, and cutscenes. NVIDIA has not disclosed the training data, parameter counts, or speed differences for each model. At this point, developers can only choose based on the visual output.
Across the whole screen, there are two sliders: "Structure Intensity" and "Tone Intensity." Structure adjusts high-frequency detail, affecting the strength of effects such as ambient occlusion, subsurface scattering on skin, and reflections. Tone governs low-frequency components and drives larger changes in lighting and color. During the presentation, Structure was shown at 25%, a mid-range value, and 95%, comparing output ranging from close to the original rendering to a result where the model's effect was strongly applied.
There are two further ways to isolate what the effect is applied to. On the model side, an automatic mask recognizes semantic content within the frame and varies the applied intensity individually for elements such as characters. On the engine side, developers can designate specific objects—such as a water pitcher, grapes, or a bottle—and independently adjust the Structure and Tone values for each. This makes it possible for game developers themselves to decide choices such as preserving facial features while intensifying the material texture of clothing, or adding only translucent light effects to plants.
This kind of control is necessary because there is no one-to-one correspondence between input and a single "correct" output image. In super-resolution, the goal is to approach a high-quality reference image, but for photorealistic materials and lighting, multiple plausible results can exist. NVIDIA's Edward Liu explained DLSS 5's philosophy as follows: "You can change the image, but you must not change the story." Which look to choose, and where and how strongly to apply it, is a line drawn by the creators—not by the model.
Internal Information for Generating Frame by Frame
Many generative video models process multiple frames together, looking both forward and backward in time. But in games, a response to player input must appear in the very next frame, so the system cannot wait for future footage. DLSS 5 uses motion vectors already available from the game engine, feeding the model information about which direction objects have moved between adjacent frames. According to NVIDIA, the process is causal—it does not look ahead to future frames—and operates by receiving one frame and returning one frame.
The inputs used to lock in artistic intent have also expanded. At the March announcement, each frame's color and motion vectors were presented as the main inputs, but at SIGGRAPH, NVIDIA explained that internal renderer buffers such as albedo, surface normals, and partial lighting information are also used. Albedo refers to an object's inherent color with lighting effects separated out, while surface normals represent the direction a surface faces. Rather than having the model infer everything from on-screen color alone, shape and lighting cues are mapped to pixel positions and passed along, which helps suppress artifacts such as scars disappearing or objects morphing into something else.
NVIDIA states that using these inputs, the system can preserve character identity and pose while generating effects such as light scattering beneath the skin, light passing through hair or leaves, contact shadows, and material reflections. The company also claims this approach reduces temporal artifacts such as shimmer, drift, and pattern "swimming." However, these are all claims based on NVIDIA's own presentation demos, and whether the same stability holds across different games or during rapid camera movement will need to be verified by third parties.
Not the Next Step After DLSS 4.5, But "Generation" After Rendering
The DLSS name has, until now, been strongly associated with super-resolution and frame generation techniques aimed at boosting performance. This time, NVIDIA divided AI usage in real-time rendering into three categories. Super Resolution and Ray Reconstruction fall under "reconstruction"—recovering a reference image that could theoretically be obtained given sufficient compute. Neural Radiance Cache and Neural Materials fall under approximating expensive calculations with small networks.
The third category, which DLSS 5 adds, is "generation." Here, a model takes the world and composition already produced by a conventional renderer as input, and generates the final appearance. The pathway of improving the original rendering with more rays and higher-resolution textures still remains, and its output becomes the starting point for DLSS 5. Rather than replacing the existing rendering pipeline, this design adds a generation stage after it.
The execution model is not simply a general-purpose video generation model dropped in as-is. NVIDIA explains that it distilled a large foundation model trained on how the real world looks, compressing it into a "one-step pixel-space diffusion transformer" whose role is narrowed specifically to enhancing the appearance of real-time rendering. By narrowing the use case, NVIDIA aimed for a size and speed that could run alongside the game itself. Because it uses the same Streamline framework as existing DLSS and NVIDIA Reflex, the integration path is also shared with prior technologies.
This distinction is essential to understanding the relationship with DLSS 4.5. DLSS 4.5 covers the second-generation transformer for super-resolution and Ray Reconstruction, along with dynamic Multi Frame Generation. DLSS 5, on the other hand, generates photorealistic lighting and material representation on top of already-rendered frames. Even though the version numbers continue in sequence, the processes being replaced are not the same.
What's Inside the 16-Millisecond Budget, and What Remains Unknown Before Fall
A single 4K frame contains roughly 8 million pixels, and to exceed 60fps, an entire frame must be completed in under 16 milliseconds. Moreover, most of that time budget is already consumed by the game engine. NVIDIA has indicated its intent to fit the distilled model into whatever time remains, but has not disclosed the processing time of DLSS 5 alone or its impact on frame rate. The 16-millisecond figure should not be mistaken for the inference time available to DLSS 5.
Supported GPUs, minimum VRAM requirements, and the load imposed by each model also remain unknown. While NVIDIA states that DLSS 5 runs in real time up to 4K and will ship in fall 2026, it has not specified which GPU generations, or at what quality settings, this will actually be achievable. Nor has it been disclosed whether choosing between Model A/B/C only changes appearance, or also involves trade-offs with speed and memory consumption.
The technology itself also has boundaries. Edward Liu did not claim that NVIDIA has solved real-time photorealism, and acknowledged that current models improve animation less than they improve appearance. As the scope of generation expands, so too does the number of decisions creators must make. The per-object control shown this time is a starting point, not a finished form.
DLSS 5 is scheduled to launch in fall 2026 with support for titles including Assassin's Creed Shadows, Hogwarts Legacy, Resident Evil Requiem, and Starfield. Only then will it be possible to verify, in actual games, the temporal consistency and input latency that static comparison images cannot reveal—along with the visual choices creators have made. What will ultimately determine adoption is not the number of models, but the results of the shipped product, backed by supported GPUs and measured frame times.