DLSS 5, NVIDIA Talks About It Again After Months: How Developer Controls and the Generative Model Work
On the occasion of SIGGRAPH, NVIDIA offered the first practical demonstration of the integrated control tools in DLSS 5, following the announcement a few months ago and ahead of the debut scheduled for next autumn. This technology, which had raised several concerns among gamers and developers for being considered too impactful on the original vision of artists—potentially generating excessively artificial or distorted graphics—now presents itself in its operational form, highlighting the degree of customization granted to developers.
Unlike previous iterations focused on image reconstruction and frame generation, DLSS 5 represents a new category of neural rendering, termed generative type, intended to complement—and not replace—traditional rasterization, ray tracing, and path tracing pipelines. The operation of DLSS 5 involves the inclusion of a generative stage that intervenes immediately after the production of a conventionally rendered frame. Geometry, camera, lighting settings, and overall composition continue to be defined deterministically by the game engine. Subsequently, the DLSS 5 model processes this frame to refine materials, reflections, shading, and surface details.
NVIDIA DLSS 5: Three Fundamental Engineering Challenges
To respond to criticisms and ensure maximum visual fidelity, NVIDIA has reiterated its intent to preserve the artistic vision of titles. Unconstrained generative models tend to alter critical elements such as the facial features of characters or the details of costumes. To anchor the visual output to the original scene, DLSS 5 not only analyzes the final image but also draws directly from the internal buffers of the engine, including albedo (which defines a surface's base color before lighting calculations), surface normals (vectors found on surfaces that help understand their nature), and lighting data.
DLSS 5 Off
DLSS 5 On
Another element that NVIDIA has focused on is causal temporal stability. While traditional video generation models analyze blocks of sequential frames (both previous and subsequent), video games require immediate responsiveness to user commands. DLSS 5 operates in a causal mode, processing a single frame at a time without accessing future frames. The removal of dynamic artifacts such as shimmering, drifting, and swimming relies on the analysis of motion vectors provided by the game. Finally, NVIDIA has concentrated on computing efficiency and memory usage. Complex diffusion models have been distilled into a one-step pixel-space diffusion transformer, optimized solely for the refinement of already rendered frames. NVIDIA's stated goal is to ensure execution on a single GPU with efficient VRAM allocation.
It is worth noting that the demo shown at the announcement was "run on two GPUs," one NVIDIA GeForce RTX 5090 dedicated to rendering the game and a second NVIDIA GeForce RTX 5090 dedicated to executing the DLSS 5 model.
Three Profiles for Different Output
As is already the case with current technologies within the DLSS orbit, DLSS 5 also includes different models trained with diverse parameters. During the demonstration, three profiles (called Model A, Model B, and Model C) were showcased, each capable of delivering a different visual interpretation of the same rendered scene. Developers will not be restricted to using a single model throughout the entire title: it will be possible to switch models based on the environment, context, or cutscenes. Developers can intervene on structural and tonal intensity. The first variable affects high-frequency details, such as ambient occlusion, micro-reflections, and subsurface scattering. The second controls macroscopic parameters such as overall lighting, color response, and the global look of the scene.
Furthermore, the neural network integrated into DLSS 5 can autonomously identify specific elements, such as a character's silhouette, selectively applying or excluding the effect. In addition to automatic masking, developers can define custom masks within the engine for individual objects or groups of assets (such as bottles, containers, or scene elements), assigning each independent settings for structure and tone. This allows, for example, to completely disable DLSS 5 on faces or certain objects and intensify its effects on environmental surfaces.
DLSS 5 Off
DLSS 5 On
In summary, aside from the controversies from a segment of the community—stemming from a highly impactful and less detailed announcement—NVIDIA aims to innovate game graphics without stifling creativity or limiting developers' artistic vision. At least, that seems to be the intent, but only the official debut will clarify the technology and its final impact.