2kliksphilip tests DLSS at 1% internal resolution to show when upscaling becomes usable
A new video from 2kliksphilip focuses on an extreme way to visualize what DLSS is doing, not by swapping presets, but by forcing the 3D scene to render at absurdly low internal resolution and then letting the upscaler rebuild the final image. While the idea on its own sounds crazy, it gives us an interesting point to see the exact moment where DLSS shifts from “this is a blur” to “this is usable” as you raise the render scale.
YouTuber even tested the game (Kingdom Come Deliverance 2) at a 1% render-scale test and DLSS 4K output. At that setting, internal resolution lands around 38×22 pixels, which is roughly 0.01% of the pixels of a 3840×2160 frame. In other words, the game’s world starts as a handful of colored blocks, while UI elements can remain sharp because they are often drawn separately at output resolution.
Needless to say, the image quality is poor, and at that resolution you would have to be determined to play it, as the YouTuber put it.
Source: 2kliksphilip
The video uses the setup to show where DLSS starts to produce recognizable structure again as you move upward from the floor. At 1% scale, the output is still dominated by reconstruction artifacts and missing detail, but it becomes easier to see what DLSS is leaning on: prior frames, motion vectors, and additional per-frame buffers, rather than trying to “invent” a clean image from a single tiny input frame.
Philip says around 20% internal resolution scale is a minimum where games can start to be considered playable, depending on the title and motion. At 4K output, 20% per axis equals 768×432 internal resolution, or 4% of the total pixel count, which is still a large reduction in shading work compared to native. He also gave a try in other games, including Red Dead Redeption at 76p and 92p, which look like this:
Source: 2kliksphilip
While NVIDIA forces its base resolution upon gamers, “DLSS quality” is not a single knob across all content. Different games feed different data into the upscaler, and fine detail, thin geometry, particles, and fast camera pans can expose limits quickly. The same is true for stutter and CPU limits, since reconstruction does not remove frame-time spikes, it only changes how the output frame is produced.
The YouTuber also compares how far DLSS has come since its first implementations. The change over the past few years has been substantial, and early versions were widely criticized for being a blurfest. It still is, but at much lower resolution.
Source:
