Nvidia's upcoming DLSS 5 upscaling technology, intended for its RTX 50-series graphics cards, has been inadvertently showcased through fan experimentation following the discovery of its rendering model files in an early access build of NBA 2K27. Initial reports suggest that while DLSS 5 demonstrates improvements over previous iterations' tendency to excessively alter image detail, it continues to exhibit an "uncanny valley" effect in certain applications.
Earlier this year, DLSS 5 faced criticism after Nvidia released promotional footage for titles like Resident Evil Requiem. Nvidia asserted that the technology solely enhanced lighting, but players accused it of altering character designs to conform to conventional attractiveness standards. Nvidia's defense, citing full backing from game studios, proved counterproductive when several studios known for opposing AI implementation denied being consulted. Sentiment further soured when Nvidia's CEO dismissed critics as "completely wrong" for rejecting DLSS 5.
The recent leak has brought DLSS 5 back into the spotlight, this time without Nvidia's curated demonstrations. Players have been testing the technology on games such as The Last of Us Part 2, Final Fantasy 7 Rebirth, The Elder Scrolls 5: Skyrim, and Grand Theft Auto: San Andreas.
While DLSS 5 can enhance certain environmental details, such as sharpening greenery in The Last of Us Part 2, its performance with human characters remains inconsistent. Reports indicate that faces and skin tones can appear unnaturally smooth or oily, with one comparison to Skyrim suggesting characters were rendered in a style reminiscent of beauty influencers. Similarly, a modified CJ from Grand Theft Auto: San Andreas appears significantly altered, resembling a different individual.
Commenters on user-uploaded videos suggest that the AI's training data, potentially derived from heavily edited HDR photos rather than natural imagery, contributes to the artificial look, resulting in high contrast and unnatural lighting. The current implementations are acknowledged by users to be unfinished and not representative of the final product. However, the technology's reputation for producing "yassified" visual outputs may prove difficult to overcome upon its official release.