Just months after the debut of DLSS 4.5, NVIDIA has announced its next major leap in upscaling technology: DLSS 5. Unveiled during the GTC 2026 keynote, this iteration pivots toward full neural rendering, with NVIDIA claiming the technology “infuses pixels with photoreal lighting and materials” to achieve a level of visual fidelity previously reserved for pre-rendered cinema.
Read: Bandwidth Blog & Smile 90.4FM Tech Tuesday: Apple MacBook Neo!
During the keynote, CEO Jensen Huang demonstrated DLSS 5 using titles such as Resident Evil: Requiem, Hogwarts Legacy, and Starfield. The technology showed a marked improvement in fine details, particularly regarding character hair and skin textures.
Technically, DLSS 5 functions by taking a game’s colour and motion vectors as input and processing them through a real-time AI model. This model then “infuses” the scene with lighting and material data that is anchored to the source 3D content, ensuring consistency from frame to frame at resolutions up to 4K.
The power required to drive this neural model is currently substantial. Huang’s demonstration utilized a system equipped with dual RTX 5090 GPUs. While NVIDIA aims to eventually optimize the tech for single-card setups, it is clear that DLSS 5 represents a step toward generative AI-driven graphics. It functions similarly to a generative video model, but one that developers can precisely control via game engines rather than text prompts.
See also
NVIDIA is calling DLSS 5 its most significant breakthrough since the introduction of real-time ray tracing in 2018. By moving away from purely hardware-driven rendering toward AI-produced pixels, NVIDIA hopes to offer “Hollywood-quality” visuals without the massive server farms typically required by film studios.
However, sceptics note that while the on-stage demos were impressive, they were often compared to games with all DLSS features disabled. The true test will be how much of a performance and visual gain DLSS 5 provides over the already capable DLSS 4.5 when it arrives this fall.
