A player opens the graphics menu, sees a setting with their GPU vendor's name on it, and flips it on. Halve 60fps and you get 30, which moves frame time from about 16.7ms to about 33.3ms [1]. Then they look at the protagonist's face, which is exactly what happened in March, when Resident Evil Requiem's Grace Ashcroft turned up filtered and the verdict was AI slop [1].
The reason the number goes the wrong way is in the plumbing. The Verge describes DLSS 5 as a generative filter that reads and modifies the stream of 2D frames after the engine has finished producing them, imagining light, shadow, colour and detail rather than changing shapes or inventing objects [3]. That is additive GPU work, and The Verge says it can easily cut framerate in half [4]. What the work buys is grounding: in a Yakuza Kiwami capture, contact shadows under each character make them read as standing on the floor [6], and pasty characters stop looking like they float above surfaces [7]. What it costs when tuned badly is a character who appears to have aged 20 years in a split second [5].
So the interesting control surface is the mask list. A studio can apply the filter to a whole scene or restrict it to a character's face and skin, foliage, or the furniture, with an automatic Photoshop-style cutout and a Developer Masking checkbox for hand-cut masks [11]. Nvidia's Gabriele Leone told journalists there is no limit to how many masks or groups a team can implement [12]. Nvidia's Edward Liu described the goal of GeForce as photorealism in real-time rendering [13], which is a fair description of where the defaults point and a poor description of a game drawn to look like animation. The Verge notes that realistic art styles survive the filter better, while even realistic games look wrong with the sliders set too aggressively [14], and the deranged modded faces circulating since the build leaked came from pointing those sliders at art that never wanted them [15].
One claim makes any of this reviewable. Nvidia says the filter is deterministic, so a game should look the same every time, unlike image generators that vary run to run [16]. Approving a screenshot only means something if the same frame yields the same pixels; without that, every art sign-off is provisional.
Here is what teams tell themselves: players want a photorealism toggle and will settle on their own preference. Here is what players do: they judge the lead character's face in the first minutes and file the result against the studio [1].
The two axes worth drawing are how photoreal the art already is, and whether the frame budget can absorb the filter at all. Photoreal with headroom is the NBA 2K27 cell, where detailed skin and a jersey that casts shadow on the body flatter the subject [17]. Photoreal without headroom means shipping it off by default and letting players spend their own frames. Stylised with headroom means masking faces and skin out and keeping only the grounding. Stylised without headroom means leaving it alone and accepting that mods will do it anyway [15]. The release checklist line that resolves it: the frame-time budget in milliseconds the filter may spend, the model and slider values per mask group, and the name of the person who signs off on faces. If nobody wants that third field, the setting ships off.
