Nvidia used SIGGRAPH to show a newer version of DLSS 5, its next upscaling and image-processing system for games, with a feature that could matter more than the usual beauty-shot demo: developers can pick between three AI models with different levels of visual detail and performance cost.
According to Nvidia’s presentation, a game does not have to run one DLSS 5 model for every frame, object, or scene. Developers can apply different models to different scenes, and can tune how much the system alters individual parts of an image, such as characters versus the surrounding world. Nvidia also said the system can change models in real time without adding latency.
That is Nvidia’s answer to the complaint that followed DLSS 5’s earlier debut this year. Some press and PC gaming enthusiasts argued that Nvidia was pushing beyond upscaling into AI-driven image alteration. Nvidia chief executive Jensen Huang defended the system at the time, saying DLSS 5 would preserve artistic intent and arguing that gamers misunderstood what the company was building.
The SIGGRAPH version tries to draw a cleaner line between upscaling and post-processing. Nvidia described the conventional upscaling stage as the part that establishes core image structure, including lighting and geometry, from the original rendered frame. The more controversial AI enhancement happens after that upscaled frame exists.
That distinction helps explain the architecture, but it does not settle the control question. Tom’s Hardware reported that it remains unclear whether developers, or players, will be able to disable the AI “beautification” stage while keeping the rest of DLSS 5 active.
How Nvidia says DLSS 5 is being constrained
Nvidia identified three main engineering problems during the SIGGRAPH presentation. The first was artistic control, meaning the system must avoid turning a game into an AI remix of itself. That is the obvious sore spot after the early backlash.
The second was latency. Nvidia said DLSS 5 is designed to process one frame at a time. That differs from many generative AI approaches, which evaluate multiple frames together and can introduce delay. For games, especially anything involving fast input, delayed prettiness is still delay.
The third was hardware efficiency. Nvidia’s slides said DLSS 5 can run on a single GPU and described it as “VRAM efficient.” That is a meaningful claim because an earlier demonstration reportedly used two RTX 5090 cards. Nvidia did not say whether the final version will require the top-end Blackwell gaming GPU to run well.
Nvidia also said the system uses a smaller model trained from a larger diffusion model, with the goal of real-time 4K performance. That is the usual AI compression bargain: train or derive from something heavy, then ship something smaller enough to run during play rather than in a lab demo.
DLSS 5 arrives in a market where upscalers are no longer optional garnish. Nvidia DLSS, AMD FSR, and Intel XeSS now often serve as anti-aliasing systems and performance tools, especially in games with ray tracing. Nvidia’s pitch is that DLSS 5 can add more visual reconstruction and enhancement without wrecking responsiveness.
There is still no official release date. Tom’s Hardware reported that DLSS 5 is expected during the third quarter of 2026, while Nvidia continues adjusting models and parameters based on community feedback.
This story draws on original reporting from Tom's Hardware.
