Merserk

Merserk / ComfyUI-PiD

Public

ComfyUI custom node for NVIDIA PiD pixel diffusion decoding and upscale workflows

comfyui comfyui-custom-node comfyui-node diffusion-models dinov2
43
3
85% credibility
Found May 30, 2026 at 43 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

ComfyUI-PiD is a free add-on for ComfyUI that lets you use NVIDIA's PiD (Pixel Diffusion Decoder) technology to dramatically upscale AI-generated images. Instead of using standard image decoding, you connect your small images to PiD nodes, and it transforms them into stunning high-resolution pictures—potentially 4x or even 8x larger with much better quality. The project includes special features for people with less powerful graphics cards, allowing them to achieve the same results by processing in stages. Everything downloads automatically when you first use it, and it works with popular AI image models like Flux, Stable Diffusion 3, and others.

How It Works

1
🎨 You create an image in ComfyUI

You start with a small AI-generated image, like a 512x512 picture you made with Flux or Stable Diffusion.

2
🔌 You connect your image to PiD

Instead of using a normal decoder, you connect your image to the PiD nodes and add a text description of what you made.

3
✨ PiD transforms your image

NVIDIA's PiD technology takes your small image and transforms it into a stunning high-resolution masterpiece, up to 4x or 8x larger.

4
Choose your memory option
🚀
Quick mode (powerful GPU)

Let PiD run everything at once for fastest results.

🧠
Memory saver mode (limited VRAM)

PiD breaks the work into stages, releasing memory between steps so even 16GB cards can handle 4K images.

5
📸 You get your high-resolution image

Your picture comes out looking crisp and detailed, ready to print or share at massive sizes.

🏆 Professional-quality upscaling achieved

You've successfully transformed a small AI image into a gallery-worthy, high-resolution photograph using cutting-edge NVIDIA technology.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 43 to 43 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is ComfyUI-PiD?

ComfyUI-PiD is a Python plugin that brings NVIDIA's Pixel Diffusion decoder into ComfyUI workflows. Unlike standard VAE decoders that convert latents to images in one step, PiD takes a different approach: it needs the latent, a text caption, a sigma value, and optionally a baseline image to produce the final output. This node set lets you plug that capability directly into your existing ComfyUI pipelines. You get nodes for direct decoding, a text prompt handler that splits output for CLIP and PiD, and a KSampler variant that captures intermediate latents with matching sigma values for better results. The plugin supports multiple backbones including Flux, Flux2, Stable Diffusion 3, DINOv2 RAE, and SigLIP Scale-RAE.

Why is it gaining traction?

The hook here is upscale quality and VRAM flexibility. PiD can take a 512x512 or 1024x1024 base latent and produce 2048x2048 or 4096x4096 outputs with better detail than standard upscaling. The staged workflow is the real differentiator: it splits the decode into Prepare, Sample, and Finalize steps, with the heavy sampling running in a separate Python process. This means CUDA memory gets released between steps, making 4K outputs feasible on 16GB cards that would otherwise OOM. There's also sequential block offload for even tighter VRAM budgets, trading speed for memory. Auto-download handles fetching the PiD source code and model checkpoints on first run, so setup is minimal.

Who should use this?

ComfyUI users running Flux, SD3, or Z-Image workflows who want high-resolution outputs without buying a workstation GPU. If you've hit OOM errors trying to decode large images inside ComfyUI, the staged subprocess approach solves that. Artists focused on quality upscaling rather than speed will benefit most, since sequential block offload and aggressive cleanup add latency. Researchers experimenting with different backbones and sigma values will appreciate the capture sampler and configurable parameters.

Verdict

This fills a real gap for high-resolution ComfyUI work, but the 43 stars and experimental tag mean you're an early adopter. The 0.8500000238418579% credibility score reflects a small but active project with solid documentation. Try it if VRAM constraints have been blocking your upscale workflows, but budget time for troubleshooting on first setup.

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.