Read Nvidia’s own description of DLSS 5 and the limit is written into it. The model takes a game’s colour and motion vectors each frame, and it infuses the scene with photoreal lighting and materials that are anchored to source 3D content and kept consistent frame to frame. Nvidia’s GeForce write-up of the NBA 2K27 implementation is blunter still in its own URL, where the thing is called 3D-guided neural rendering.
Hold that phrase for a second, because Valery Kryzhanovsky has spent the last few weeks pointing it at games that have no 3D in them at all.
Pong, 1972
The test set runs wide — tens of titles, from 2D platformers and RPGs up through The Witcher 3 and Marvel’s Spider-Man. The end everybody is sharing is the DOS end: Duke Nukem, the 1991 Apogee platformer, 320x200 in sixteen EGA colours. And Pong.
Getting there takes a stack. The DOS games run in DOSBox, ReShade carries the injection, and RenoDX supplies the interface for the parameters. Which injection point you pick matters — render stage, after upscaling, or final output — and that it matters this much is worth noticing on its own. A pass that genuinely understood the scene would be less sensitive to where in the chain you bolted it.
What comes out is stranger than broken. Rectangular level geometry gets embossed, so flat tiles read as raised surfaces. The pass finds the player sprite, the enemies and the bullets, and largely declines to scribble inside them. Turn the strength up and the whole thing slides into an acid trip, which is the expected failure and the least interesting one.
Nothing to anchor to
The library doing this is the one that fell out of an NBA 2K27 build in August — a generative pass that runs after the frame is finished, rather than a renderer feature anybody shipped on purpose. In a modern engine it at least gets the inputs it was trained on. In DOSBox it gets none of them. No motion vectors. No depth buffer. No source 3D content, which is the exact thing Nvidia’s page says the output is anchored to.
So the model has a flat image and a prior about what images of the world look like, and it uses the second to invent the first. The embossed bricks are the clearest evidence. Nothing in Duke Nukem’s tile art encodes that a brick protrudes — the artist drew a face, not a volume. The model raised it because bricks are usually raised. It was right, and it was right by coincidence of what it had been shown, not by reading anything in front of it.
That is a different claim from “it looks bad”. Some of it looks genuinely good. I have not run the stack myself; this is read off published footage and two write-ups, and anyone telling you how the crawl behaves in motion from a compressed YouTube clip is guessing a little too.
The temporal weakness follows from the input list rather than from any bug. Kryzhanovsky’s conclusion is that the technology is promising and undercooked, and that the absent temporal element is its key drawback. Of course it is. Frame-to-frame consistency in DLSS 5 is bought with motion vectors, and DOSBox has never had any to hand over. Each frame gets its own opinion about where the light is.
The cost that stops mattering
There is one accidental joke in all this. The reason neural rendering is a hard sell in a current game is the bill: it takes something like half the frame rate in NBA 2K27, which is why people started handing the pass to a second GPU. Duke Nukem is 64,000 pixels of sixteen colours. There is nothing to lose. The most expensive post-process Nvidia has ever shipped is free the moment you aim it at 1991.
Which makes this the cheapest honest test bench we have for what the model knows. Run it on Cyberpunk and you cannot separate the model’s contribution from the good lighting that was already there. Run it on sixteen colours and every single thing in the output that was not in the input came from the weights.