Neural network upscaling is based on one simple condition: the algorithm must always have a fresh frame from the game with which to compare its guesses.

Once this connection is severed, the video card begins to honestly improve its own hallucinations until the human face turns into a pulsating fractal of flesh that looks like a terrible dream.

This is exactly the video that went viral on Reddit under the sarcastic caption “DLSS 5 broke out of quarantine.” In the video, the character’s face slowly spreads out, acquires new eyes, mouths and beards, and then finally crumbles into digital dust.

In the comments, people unanimously identified Gandalf, Michael Jackson, or Master Splinter in this mess. We agreed that it looked creepy.

The mechanics of what is happening are quite understandable. Modern upscalers like DLSS don’t just stretch the image, but accumulate information over time: they take the history of previous frames, shift a sample of subpixels, compare movement through motion vectors, and add details that are not physically present in the original resolution.

When this history includes not a new render, but the previous result of the network itself, the error ceases to be extinguished and begins to be raised to a power with each frame.

However, it’s not NVIDIA’s fault here. According to the technical version voiced in the discussion, we are talking about a crooked implementation on the side of third-party tools: OptiScaler replaces upscalers in games, ReShade and RenoDX wedge themselves into post-processing and HDR, and a fork of the DLSS Ray Reconstruction denoizer in this combination sends the processed frame back to the chain of output buffers, where it is not supposed to be.

The result is classic positive feedback, just like when pointing the camera at your own monitor or resaving a JPEG for the thousandth time.

LEAVE A REPLY

Please enter your comment!
Please enter your name here