The new Kernel Predicting Network places around 90% of its parameters at 540p or below and does little work at 4K

The ELI5(ish) is that they now do as much image enhancement work in low resolution (before upscaling) as possible so the AI has far fewer pixels to process.

The upgraded model also stops predicting the final color directly. Instead, the KPN generates blend weights while conventional filtering handles the actual blend. This avoids storing and blending HDR color in 8-bit form and allows the network to spend less capacity on those operations.

At the same time they aren’t using the expensive model for handling colour, because doing it the “old fashioned” way is more efficient and produces better results.