DLSS 5's real product is an approval decision: someone at a studio has to sign off that a rendered face is still the character the art team drew, and put their name on that call. Nvidia's model, which it calls 3D-guided Neural Rendering, reads a game's motion vector data and identifies character faces, clothing, objects and settings [14]. A team picks one of three AI models [15], then uses semantic masking that recognises scene objects automatically to lift the environment while holding characters back, with engine-level masking to isolate props such as glassware or foliage [16]. Nvidia's position is that this preserves artistic intent, and that shipping the feature is the developer's call [27]. Fair enough. It is also a new approval queue, and the reason the queue exists is that the March demo turned Resident Evil Requiem's Grace Ashcroft into a meme [24].
Then the frame rates, which want dividing. PCMag notes that every benchmark Nvidia published ran Multi Frame Gen 6X, five generated frames for each one the GPU actually renders [5]. So the "well over 200fps" at 1080p on an RTX 5060 [6] rests on roughly 33 rendered frames a second [8], and the 370fps at 4K Ultra on a 5090 [7] on roughly 62 [9]. The 6X path has already drawn complaints about blur and odd artifacts [25], which is something a QA lead has to approve at the rendered cadence rather than the displayed one. The Verge measured performance cut in half running an unofficial mod in Shadow of the Tomb Raider [26].
The hardware story is firmer than the software story. Nvidia says it optimised DLSS 5 to run five times more efficiently, which is what lets it run on 50-series cards at all [21]. Its mid-range claim, a single RTX 5060 playing NBA 2K27 at 1080p with ray tracing and DLSS 5 at ultra, carries two conditions: one of the fastest gaming CPUs available, and 6X generation switched on [30]. Some of the cheaper 5060s are still going for over $400 [29].
Studios tell themselves that players will switch this on and see the game the artists intended, only cleaner. Players, going by the meme cycle that greeted the March reveal [24], push the slider past anything an art director would sign and post the screenshot.
If you are costing this, two axes are enough. First, the share of your players on 50-series hardware. Second, whether the look you want requires the model to touch characters at all. Low share, props only: ship it behind the toggle and spend nothing on approvals. High share, props only: that is the case Nvidia is genuinely selling, and sports and photoreal catalogues live there. Low share, characters touched: the art pass costs more than the audience it reaches. High share, characters touched: you need one named owner for faces before the build goes out, and a review pass budgeted in hours, not in frames. The number worth reading afterwards is retention of the setting, meaning how many players still have it enabled in a third session, rather than how many flipped it on once.
