Photoreal now costs you latency.

DLSS 5 went out on 3 September, on every RTX 50 card plus the laptop parts and GeForce Now, with NBA 2K27 as the first title. The headline feature is what Nvidia calls 3D-guided neural rendering. It takes the frame the engine has already drawn, plus the engine data behind it, colour, motion vectors, lighting, surface information, and re-renders the lighting and materials with a model. Developers get semantic masking so they can point it at the foliage and the glass and leave the faces alone. The demo number is an RTX 5090 at up to 370 frames per second, 4K, Ultra preset.

The number that actually matters is the other one. Nvidia has said plainly that turning neural rendering on costs 50 to 60 per cent of your frame rate. The 370 figure is what comes out the far side after super resolution and multi-frame generation have put the frames back. So the pipeline is: render, throw away half your budget to make the lighting better, then synthesise the missing frames to make the counter look right again.

A frame counter is not a response time

Generated frames are not free and they are not the same as rendered ones. They are interpolated from frames the engine has already produced, which means the picture is smoother but the moment your input reaches the screen has not moved forward, and in most implementations it has moved back. You get a high number on the overlay and a slightly later reaction under your hands.

For a basketball game on a sofa, that is a fair trade and most people will take it. The visual gain is real, the model has got roughly five times faster in six months, and it will keep getting faster. This is not a note about the tech being bad. It is a note about which budget you are spending.

Where we sit on it

We build things that people walk up to. A wall that reacts when you step into a zone, a table that responds to a hand, a screen that follows a queue. In that setting the felt quality of the piece is almost entirely the delay between a person doing something and the surface acknowledging it. Under about 100 milliseconds it reads as the thing responding to you. Past that it reads as a video that happens to change sometimes, and people stop playing with it. Nobody has ever walked away from an install saying the subsurface scattering was disappointing.

So our default on a live piece is the opposite of the gaming default. Spend the frame budget on the loop, not the lighting. Keep the render honest, keep the input path short, and let the art direction do the work that photorealism was going to do. Most of the installations we are proudest of are stylised, and that was a decision, not a compromise.

When it is worth switching on

• 01Non-interactive content on a screen. Attract loops, ambient walls, anything pre-rendered or running without a person in the loop. There is no latency to protect, so take the picture.
• 02Product visualisation where the interaction is a slow orbit or a material swap. A hundred milliseconds on a rotate is invisible.
• 03Never on the responsive layer. If a sensor, a camera or a touch is driving it, leave the neural pass off and put the headroom into the tracking and the render loop.

There is a hardware consequence too. If a build needs an RTX 50 part to hit its look, that is now a fixed line in the budget and a specific machine sitting in a plinth for the life of the install, being maintained by somebody who did not commission it. We would rather spec a piece that looks correct on hardware we can replace in an afternoon.

The useful question for anyone building interactive work this autumn is not whether the new pass looks better, because it does. It is whether the thing you are building is watched or used. Watched, spend it on the image. Used, spend it on the response, every time.