Product1 distinct publisher3 min readPublished
The leaked neural renderer costs roughly 35 frames a second in Control and repaints character art while it runs, and Nvidia has already conceded the model does not read the rendering pipeline it decorates.
The Product Desk · Product desk

build
China's accelerator swap makes Cambricon supply, not export policy, your ship-date risk1 distinct publisher
product
Nebius funds $4.5bn of AI capacity on terms that pay lenders mostly in stock2 distinct publishers
build
NVIDIA put a number on agent skills: 300+ verified, two harnesses, baselines under 50/1001 distinct publisher
invest
Legal tech's two leaders are buying the field, and nobody has disclosed a price1 distinct publisher
Compiled by The Product DeskSomething wrong?How this is made
A filter that cannot read material assignments has to infer material from pixels. Nvidia has said the model applying these visuals does not take in much of the rendering pipeline, including the physically based rendering data that tells an engine which surface is wood and which is metal, and that DLSS 5 reinterprets a scene with AI [6]. Guessing is how you get a protagonist with new laugh lines and hair that drifts from red toward burgundy [10]. Gizmodo's reading of Control is the one that will land with anyone who has sat through an art review: the game's palette is built on office grays with sparks of orange and red, and Remedy colored Jesse's hair to carry that [11]. Players said much the same about the first renditions after the March reveal, complaining the model ignored artistic intent [17][1]. The mod has since been dropped into Cyberpunk 2077, Red Dead Redemption 2 and The Last of Us Part II [12], and one Redditor's Wuchang: Fallen Feathers build gave its protagonist more of a supermodel look [18].
The performance figures are worth converting, because percentages flatter the faster card. Control goes from above 70 fps to about 35 on an RTX 5070 Ti [4], and the recorded footage shows above 90 fps dropping to the low 50s on an RTX 5080 [5], a cut of 50 percent and roughly 42 percent [15]. In frame time, which is what a studio actually budgets, that is about 14.3 ms per frame becoming 28.6 ms on the 5070 Ti and 11.1 ms becoming 19.2 ms on the 5080, so neural rendering is charging somewhere between 8 and 14 ms on top of everything else the frame already pays for [16]. That gap eats directly into any 60 fps target a studio has set.
The pitch is a rendering advance, but this build post-processes character models against a preset, with an on/off switch, an intensity slider and a style slider [8], seven styles, and characters only so far [9]. The style slider is the part a producer should look at harder than the frame time. Frame cost can be negotiated down with internal resolution. A slider that hands players seven readings of a character the art team specified once turns direction into a default [9].
Nvidia told SIGGRAPH 2026 it had toned down the harshest aspects and got DLSS 5 running on lower-end RTX 50-series cards [7], and this build is not the shipping product [13]. Both statements can be true at once, and a studio still has no finished build to plan around.
The test that travels to your own vendor layer is simple: whether it reads your inputs or repaints your outputs, and whether control sits with your team or the end user. DLSS 4.5 sits on the first side of both axes, and its standing as an upscaler is measured in frames it gives back [14]. Anything landing in the repaint-outputs quadrant with the slider in the player's hands belongs in the settings menu beside film grain, costed at its worst measured frame time rather than its demo.
Ranked by verification strength, evidence, and original report placement.
Nvidia first revealed its DLSS 5 "neural rendering" feature in March.
Modders on the RenoDX Discord server retrofitted what appears to be a newer leaked version of DLSS 5 into several modern games, including Control.
The DLSS 5 leak apparently stems from the NBA 2K27 beta, which contained some DLSS 5 files inside it.
Modders on the RenoDX Discord say Control ran above 70 fps on a system with an RTX 5070 Ti, and around 35 fps with neural rendering enabled.
Video of the mod shows Control running above 90 fps normally on an RTX 5080 and dropping to the low 50s with DLSS 5 enabled.
Nvidia admitted the AI model applying DLSS 5 visuals does not take into account much of the rendering pipeline, such as the physically based rendering that designates material types like wood and metal to a game engine; DLSS 5 takes a scene and reinterprets it with AI.
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 28, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
One desk, one Discord
Everything traceable in this story passes through Gizmodo, and the sharpest numbers pass through unnamed Discord participants before that. The visual claims are at least inspectable — you can watch the Control footage and the side-by-side clips — but the performance claims arrive with no resolution, no settings, no build identifier and no second run. Nvidia's pipeline admission and its SIGGRAPH 2026 remarks are the only parts anchored to the company on the record.
Spreading fast, shipping nowhere
Real usage exists and moved quickly — a DLL scraped out of an NBA 2K27 beta was running in five or six major games inside about a day — but every instance of it is an enthusiast mod of an unreleased component. Nothing is available to a normal player, no engine or studio has adopted it, and the only shipping product in the picture is DLSS 4.5's new Ray Reconstruction model, which is a different thing entirely.
The verdict outruns the test
"Makes games look and run worse" is a judgment on a stolen pre-release binary, benchmarked by strangers, with the story's own admission two paragraphs later that this is probably not what ships — and with Nvidia having already said it had softened the effect and broadened GPU support. The direction of travel looks right, and the pipeline concession is a real structural problem rather than a tuning one. It is the certainty and the neat halving that are ahead of the measurement.
Nobody in this story is on the record
The people supplying the numbers are anonymous modders whose standing rises with the clip's reach; the artifact itself is contraband, so no one involved can be corrected or contradicted. On the other side, Nvidia's only voice here is managed messaging — a concession framed as progress and a SIGGRAPH claim of having toned things down. And an AI feature that ages a beloved character is a story that writes itself before the benchmarks are checked.
Confident about the flaw, not the figures
We would stand behind two things: the model reinterprets a finished frame without material data, and running it costs real milliseconds. Almost everything downstream — how much it costs, on which settings, in which build, and whether shipping DLSS 5 will look like this — rests on a single outlet relaying an unattributed test of leaked code, which is not enough to be firm about.