Build1 publisher2 min readPublished
NVIDIA constrains DLSS 5 to one masked pass over the engine's finished frame
The engine's rendered frame is the fixed input, and a studio decides what the model may touch using per-pixel masks and two intensity parameters. NBA 2K27 ships it first, on RTX 50 Series GPUs only.
The Engineer · Build desk
What happened
- NVIDIA has added DLSS 5 as a final neural-rendering stage that takes the engine's rendered frame, including artist-authored geometry, textures and lighting buffers, as a fixed foundation it may not overwrite.
- Developers pick among several models, mix them across scenes, gameplay or cutscenes, and tune two parameters, Structure Intensity for high-frequency detail and Tone Intensity for lighting and color response.
- NBA 2K27, developed by Visual Concepts and published by 2K, ships DLSS 5 now for all GeForce RTX 50 Series desktop and laptop GPUs.
- The same release updates the in-game inference SDK with RTX Spark developer-preview support, Gemma4 in the GPT plugin, a Stable Diffusion plugin and the latest llama.cpp changes.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- constraint NVIDIA's stability claim rests on the engine's motion vectors. A renderer that cannot hand over clean frame color and per-pixel motion vectors at the end of the frame has integration work to do before the model produces anything.
- decision Raster-only studios have to decide whether to fund ray or path tracing first, since NVIDIA puts the accurate results behind richer input lighting and only claims a noticeable lift over rasterized frames.
- cost The tuning burden lands on art staff: semantic masks, engine-level masks per asset group and per-pixel uplift maps are authored content that someone has to maintain per scene.
- exposure The pass runs every frame on the same GPU that renders the game, and NVIDIA has not published its frame-time or memory cost. A studio budgeting for it is guessing until it measures on its own scenes.
The input contract decides whether a studio can adopt this. DLSS 5 runs at the end of the frame on the engine's own output: frame color and motion vectors. The rendered geometry, textures and lighting buffers are held as what NVIDIA's developer post calls an "unyielding foundation" [1]. The post states the split directly: "The engine frame defines what must remain, while developers direct what may change." [2] One frame goes in, one frame comes out, and NVIDIA says the game engine's motion vectors are what keep the output consistent as players move [4].
Studios author direction per asset. Semantic AI masking applies or holds back the effect across recognized scene elements, and engine-level masks isolate props or asset groups such as glassware, water droplets and foliage [7].
NBA 2K27 shows what that machinery is for. Visual Concepts uses overall tone and style controls plus a per-pixel uplift control mask to fine-tune character detail while respecting player likenesses [11]. The scanned facial geometry stays as authored; what the model adds is skin subsurface scattering, light transmission through hair and ears, and contact shadows [12].
NVIDIA's quality claim carries its own condition. The post says DLSS 5 "noticeably elevates traditional rasterized graphics, but giving the model richer source data, like ray-traced or path-traced lighting, yields dramatically more-accurate results" [8]. A raster-only title gets the weaker of those two outcomes. For the NBA 2K27 result to transfer, a game needs input lighting of comparable quality and an art team prepared to author masks by asset group.
The hardware floor is a single GeForce RTX 50 Series GPU at up to 4K [5]. Shipping support covers all RTX 50 Series desktop and laptop parts [9], and cloud players reach it on NVIDIA-operated RTX 5080 rigs through GeForce NOW Ultimate [10]. NVIDIA did not publish a frame-time or memory cost for the pass. It runs on every frame, inside the same GPU budget as the rest of the renderer.
The same release adds more local inference work to that budget. ACE now includes Nemotron Speech 3.5 Streaming, a 600M-parameter speech recognition model built to minimize latency [13], and Qwen3 TTS at 600M parameters [14]. That is 1.2 billion parameters of speech models [18], deployed through in-process C++ execution in the NVIGI SDK. NVIDIA describes that path as running these models locally alongside game graphics [15]. RTX Kit adds RTX Mega Geometry 2.0, with support for streaming continuous level-of-detail clusters [17].
What to watch
- A second shipping integration outside sports, where licensed likeness is not the binding constraint, would show whether the masking controls generalize.
- Published frame-time and VRAM figures for the neural pass on RTX 50 Series parts below the 5080 rigs behind GeForce NOW Ultimate.
- Whether ACE speech models running in-process through NVIGI can share a GPU with the DLSS 5 pass without either one missing frame deadlines.