NVIDIA PiD in ComfyUI: 4× upscaler setup

This tutorial explains a ComfyUI workflow for NVIDIA PiD, a pixel diffusion upscaler built on Flux1. The presenter describes a four-step process that regenerates detail at four times the input resolution. The workflow normalizes the image to a 1024-pixel long edge before the upscale, a step the presenter says helps prevent broken or green-tinted results across different aspect ratios.

Florence-2 generates a caption from the input image for automatic prompting. The workflow uses an LCM scheduler, saves the full upscale, and applies Lanczos downscaling to produce a supersampled 1024-pixel version. A built-in comparison slider lets viewers inspect the original against the result. Claims about sharper, faithful detail come from the presenter; the supplied transcript does not establish a comparison with Topaz.

Setup requires a Gemma 2B text encoder, a VAE and the PiD Flux1 four-step diffusion model in their respective ComfyUI model folders. The tutorial also covers five custom node packs and loading the workflow JSON from CivitAI. The metadata recommends about 16 GB VRAM for the full 4096-pixel pass and a recent ComfyUI build with PiD support. It also identifies the model license as NVIDIA NSCLv1, restricted to non-commercial use.