Rendering a frame at full resolution is expensive. Rendering a smaller one and reconstructing the rest is cheaper, and that trade is the whole idea behind AI upscaling.
What is actually being reconstructed
A traditional scaler looks at one frame and guesses at the pixels between the ones it has. A learned scaler has more to work with: the previous frames, the motion vectors the engine already computed to move objects around, and a model trained on what the full-resolution image was supposed to look like. That combination is why AI upscaling can recover a thin railing or a line of text that a simple filter turns to mush.
Why motion vectors matter more than the model
The interesting part is not the network, it is the temporal data. Because the engine knows where every object moved between frames, the reconstruction can carry detail forward instead of inventing it. When that data is wrong — a reflection, a particle effect, a shadow with no geometry behind it — you get the shimmering and ghosting that people notice and dislike.
The cost nobody mentions
Upscaling is not free. It occupies the same silicon that could have been drawing the frame, so the saving is real but smaller than the resolution ratio suggests. On older hardware the overhead can eat most of the gain, which is why the feature tends to appear first on the cards that needed it least.
The honest summary: it is a very good trade at high resolutions, a marginal one at 1080p, and it is closer to informed reconstruction than to invention.