Earlier this week, Nvidia unveiled DLSS 5: an “AI-powered breakthrough” in visible upscaling tech that takes a “sport’s coloration and movement vectors as enter for every body, then infuses the scene with photoreal lighting and supplies.” The web instantly reacted very poorly to its announcement, decrying it as an AI-gen slop filter.
Nvidia CEO Jensen Huang rejected that framing at a dwell occasion later within the week, saying everyone seems to be “fully improper” and DLSS 5 isn’t really “post-processing on the body stage.” That may counsel a finer diploma of nuance and management than the alleged “slop filter” that’s modifying the ultimate 2D picture based mostly on broad web coaching information.
However new particulars from Nvidia’s personal “GeForce Evangelist” advertising specialist Jacob Freeman seem to contradict Huang’s framing of the controversial know-how. PC gaming {hardware} YouTuber Daniel Owens requested Freeman if DLSS 5 is “successfully taking a single 2D body as an enter (with movement vectors) to create the output body?” The Nvidia rep’s response was: “Sure, DLSS5 takes a 2D body plus movement vectors as an enter.” They continued, “DLSS 5 is educated finish to finish to grasp complicated scene semantics comparable to characters, hair, cloth, and translucent pores and skin, together with environmental lighting situations like front-lit, back-lit or overcast – all by analyzing a single body.”
The much less technically inclined amongst chances are you’ll be asking what the gotcha is right here. The problem is that this immediately contradicts Jensen Huang’s earlier assertion on March 17. “It’s not post-processing, it’s not post-processing on the body stage, it’s generative management on the geometry stage,” Huang mentioned to Tom’s Hardware throughout a Q&A. “All of that’s within the management — direct management — of the sport developer. That is very completely different than generative AI; it’s content-control generative AI. That’s why we name it neural rendering.”
Principally, the NVIDIA worker is saying it’s a generative AI filter that makes use of a single picture as a reference, and Huang is saying it’s not utilizing a single body because the reference; it’s utilizing each side of the information, together with the 3D geometry.
Briefly, as Owens places it, DLSS 5 is simply utilizing a screenshot and slapping a filter over prime. For this reason folks on-line, already in backlash mode over the unique demo, at the moment are crying foul and accusing Huang of lying about DLSS 5’s capabilities in his most up-to-date assertion. It wouldn’t be the primary time he was accused of deceptive customers.
It feels like DLSS 5 isn’t really pulling any further data past that. That additionally sort of explains why a few of the lighting results within the first demonstration appear like rubbish, as a result of DLSS 5 is simply utilizing a picture of the lighting and nothing else to generate one thing new. DLSS 5 isn’t some new-fangled geometry stage rendering tech; it’s simply AI slop 2.0, as a result of it’s not doing something a bog-standard generative AI filter doesn’t already do.
