The official release of DLSS 5 runs on games that never received formal support and delivers up to 90% better performance than the leaked version that circulated beforehand, according to reports of early testing. That gap is unusually wide for two builds of the same upscaling technology, and it says something about how much work remains hidden between an internal Nvidia build and a shipped one.

By Tom Whitfield

What is actually being measured?

The comparison pits two versions of Nvidia's deep learning super sampling stack against each other on titles outside the official support list. The leaked build, which appeared online before Nvidia sanctioned its release, was an unfinished artifact. The official build is the code Nvidia intends the public to run. A 90% performance difference between the two is the headline figure, and it comes from running the official release on unsupported games.

The mechanics matter here. DLSS reconstructs frames at a lower internal resolution and uses a trained neural network to output an image close to native quality. Because the model does the heavy lifting, the same technique can be forced onto games that never shipped with a DLSS toggle. Results vary, but the testing behind this report shows the official model is far better at that job than the leaked one.

The plain takeaway: the finished software is not a small polish of the leak, it is a materially faster product.

Why do unsupported games matter commercially?

Nvidia maintains a whitelist of titles that get official DLSS integration, and that list has always been a competitive weapon against AMD's FSR and Intel's XeSS. A stack that performs well even when forced onto unlisted games changes the arithmetic for players with large back catalogues. The value of an RTX card no longer stops at the edge of the supported library.

This also reframes the leak itself. Leaked builds circulate because enthusiasts want early access, but they carry unfinished models and unoptimized inference paths. Anyone who benchmarked the leak and dismissed the technology was measuring scaffolding, not the building.

In short, the official release quietly widens the gap between Nvidia's platform and the software its rivals can offer today.

What should buyers take from the 90% figure?

One number deserves caution. A figure as large as 90% almost certainly reflects best-case scenarios, specific titles, specific resolutions and specific GPU tiers. It is not a promise of a 90% uplift on every unsupported game. It is an upper bound reported from testing, and lower results on other titles are to be expected.

The pattern, however, is consistent and directionally clear. The official DLSS 5 model handles games outside its support list far better than the leaked code did, and it does so without developers lifting a finger on those titles.

The sensible read is that the official build turns forced upscaling from an experiment into a practical option, while the leak was never a fair preview.

What comes next?

The interesting question is whether Nvidia formalizes any of this. If the official stack keeps performing this well on unlisted games, pressure will grow for a sanctioned compatibility mode rather than community-forced installs. Until then, the 90% figure stands as a measured result, and a reminder that leaked software rarely represents the finished product.