NVIDIA defends its DLSS 5 technology, but reportedly developed it in utmost secrecy

Thibaut had dissected it shortly after his official announcement: the next – fifth – version of DLSS by NVIDIA will focus on using generative AI to enhance the realism of scenes, in line with the company’s neural rendering technologies. While it is not yet publicly testable (the demonstration requiring two RTX 5090s, optimization not being a priority at this stage of development), the camcorder has already announced an expected availability for autumn, demonstrating internal validation of feasibility.

However, if the changes in this DLSS 5 are very noticeable – unlike the beginnings of Ray Tracing, where the public struggled to understand what the technology brought compared to traditional rendering – many criticisms have arisen regarding the order of things. While the gain in photorealism is undeniable, the final rendering can vary from impeccable to completely off-topic depending on the proposed demonstration and, we must emphasize, the player’s personal sensitivity; especially since the “AI” grain is often clearly felt. It is safe to say that in the midst of an overdose of deepfakes and other brainrots based on genAI on the web, the prospect of using them in video games has shaken more than a few.

According to NVIDIA CEO Jensen Huang, the critics of the technology are completely wrong. During a question-and-answer session, the man – not surprisingly – defended that DLSS 5 was “finely tunable” by developers, and allowed to “merge the controllability of geometry and textures” through generative AI. This is sure to stir up controversy! The first statement seems to refer to fine tuning, a method of training networks to better adapt to slightly different tasks from their initial purpose; with the downside of requiring a target dataset and computational power. If this holds true, the variability of the effects of DLSS 5 will then be very large depending on the working methods of development studios.

As for the fusion of geometry and textures, we are awaiting more details regarding the rendering stage in which DLSS 5 sits before passing judgment. According to the latest communications, the network would not operate like a simple post-processing of the image (as the previous versions did), but directly at the geometry level. This could allow for transformations like “cartoon style” or “glass style”, according to NVIDIA. This could pave the way for quality modding! While some will argue that the classic rendering may disappear in favor of AI methods, breaking retro-compatibility and performance on previous-generation graphics cards; this was also the case with Ray Tracing. However, after four generations of GPUs, the impact is now more than noticeable on visual quality.

Given the technical revolution that this DLSS 5 brings, we suspected that silence was maintained throughout its development. In fact, it appears that the publishers and development studios themselves (at least, at CAPCOM and Ubisoft) did not participate in the refinement of the renderings and discovered the result at the same time as the general public: only an agreement regarding the integration of DLSS 5 in their title was signed. This explains the changes in artistic direction induced by the technology in the first demonstrations and the AI level dials placed in the red zone, particularly visible on Resident Evil Requiem. While such a method may seem surprising at first, the green team’s strategy is more than sensible: to limit leaks as much as possible (in 2026, good luck), it is simply best to reduce the number of people involved. As for the AI touch, Bethesda also seeks to reassure: the rendering shown in Starfield is only a first draft, the team of artists will have complete control over the effect and will adjust its settings case by case before the playable version, to best match the atmosphere of the title. Check back in a few months to judge the final result! (Source: Insider Gaming and Tom’s Hardware)