Nvidia's PiD tech blows up tiny images to 4K in under a second

Nvidia's team just dropped PiD, a new image tech that can take a tiny 512x512 pixel image and blow it up to 2048x2048 pixels in under a second. On a regular RTX 5090 graphics card, it only uses 13 GB of memory.

Normally, when making high-res images, the decoder just tries to rebuild the compressed version, but it's not great at adding fine details. PiD changes that by combining decoding and upscaling into one step. It uses a method called conditional pixel diffusion to actively add texture, structure, and local details as it generates the final image.

PiD is built on PixelDiT and includes a lightweight ControlNet-style adapter that adjusts how much it trusts the compressed data based on how noisy it is. The team also used DMD2 distillation to cut the process down to just 4 steps, plus an early stop trick to keep things fast without losing quality. On a GB200 GPU, it can do the job in as little as 210 milliseconds.

This tech works with both traditional VAE compressed data and newer semantic types like SigLIP and DINOv2, making it pretty versatile. Compared to other upscaling methods, PiD is up to about 6 times faster and delivers better visual quality.