NVIDIA Announces DLSS 5 with 3D-Guided Neural Rendering, Updates to ACE and RTX Kit

NVIDIA announced DLSS 5 on September 22, 2026, on its developer blog, introducing 3D-Guided Neural Rendering and fine-grained controls for game developers. The positioning of this generation of technology is not to replace engine rendering, but to treat the frame rendered by the game engine—including artist-authored geometry, textures, and lighting buffers—as an unshakable foundation. NVIDIA's phrasing is: the engine frame determines what must be preserved, and developers determine what may be changed.

This means neural rendering is placed at the end of the graphics pipeline, adding lighting and material detail as a final stage while preserving scene structure, character identity, and artistic intent. For developers, this is both a new layer of expressiveness and a clear constraint line: the model does not rewrite what the engine has already determined.

DLSS 5 uses a strict one-frame-in, one-frame-out model, combined with game engine motion vectors, to produce deterministic, temporally stable output. NVIDIA explicitly distinguishes it from offline generative models: the latter process blocks of multiple frames, whereas DLSS 5 does not.

This trade-off directly serves real-time 3D rendering: results need to remain consistent as the player moves. This compact, dedicated model runs locally on a single GeForce RTX 50 Series GPU, with support for up to 4K. Local operation and the single-card constraint are the hardware-level premises of this approach.

The controls NVIDIA gives developers include: model selection, Structure Intensity and Tone Intensity adjustment, semantic AI masks, and engine-level masks, and the ability to blend different models across scenes, gameplay, or cutscenes. Structure Intensity is used to adjust high-frequency detail, while Tone Intensity is used to adjust broader lighting and color response.

Semantic AI masks can apply or suppress effects on recognized scene elements, while engine-level masks can isolate props or asset groups, such as glassware, water droplets, and vegetation. The existence of these controls shows that NVIDIA is handing part of the decision-making over the final output back to developers, rather than letting the model unilaterally determine the image.

NVIDIA says DLSS 5 can noticeably improve traditional rasterized imagery, but that if richer source data is provided as input (such as ray-traced or path-traced lighting), the results will be significantly more accurate. This is a qualitative vendor description, with no quantitative data provided.

Inferring from engineering logic, this means DLSS 5's effectiveness is not independent of the input pipeline: the richer the lighting buffers the engine provides, the more basis the model has for the detail it can add. But the specific magnitude of "noticeably improve" and "significantly more accurate" currently rests only on NVIDIA's own account, lacking independent third-party evaluation.

DLSS 5 is already available in NBA 2K27, developed by Visual Concepts and published by 2K, for all GeForce RTX 50 Series desktop and laptop GPUs. GeForce NOW Ultimate members can also experience it when streaming from NVIDIA-operated cloud gaming devices equipped with GeForce RTX 5080.

In NBA 2K27, DLSS 5 enhances skin subsurface scattering, transmission of light through hair and ears, and contact shadows while preserving scanned facial geometry. Visual Concepts uses overall tone and style controls as well as per-pixel enhancement control masks to fine-tune character detail while respecting player likenesses.

These effect descriptions come from NVIDIA and Visual Concepts' collaboration statement and are vendor self-reports. DLSS 5's availability on non-RTX 50 Series hardware or other platforms is not addressed in the sources.

NVIDIA ACE adds Nemotron Speech 3.5 Streaming, a 600-million-parameter automatic speech recognition model that uses a streaming architecture to minimize latency while maintaining accuracy. It also adds Qwen3 TTS, a 600-million-parameter text-to-speech model that can generate high-quality audio and supports custom fine-tuning.

Updates to the NVIDIA In-Game Inferencing SDK include: RTX Spark support (developer preview), integration of Gemma4 in the GPT plugin, a new Stable Diffusion plugin and sample code, and incorporation of the latest llama.cpp updates to improve inference performance. RTX Spark is described as NVIDIA's new AI and graphics platform for thin-and-light laptops and ultra-efficient desktops.

The NVIGI SDK executes through in-process C++, providing a high-performance, streamlined path for deploying AI models locally alongside game graphics. RTX Spark support is still in developer preview, and its scope of availability and timeline are not stated in the sources; the latency and accuracy performance of the new ACE models in real games also lacks independent verification.

RTX Mega Geometry 2.0 adds streaming of continuous level-of-detail (LOD) clusters to support high-density meshes, demonstrated with a textured glTF version of Zorah, and is coming soon to Gears of War: E-Day.

"Coming soon" means no specific date is given. This update, like DLSS 5, belongs to the RTX Kit's rendering workflow updates, but the two solve different problems: the former targets streaming of high-density geometry, while the latter targets neural rendering enhancement of the final image.