NVIDIA develops 'PiD,' a new technology that directly converts image generation AI latent representations into high-resolution images

NVIDIA introduced PiD, a diffusion-based decoder that generates high-resolution images directly from latent space.

It aims to replace the usual low-resolution decode plus super-resolution pipeline with a single, higher-quality generation step.

The system reached 2048×2048 output in under a second, with a best reported 210 ms on GB200 hardware.

The approach reframes the decoder as a true high-resolution generation module, potentially improving both speed and image quality.