🚀 Standalone PiD Upscaler (No Prompt Required + Auto-Aspect Ratio & Color Match)
This is a universal, production-ready ComfyUI workflow for the revolutionary NVIDIA PiD (Pixel Diffusion Decoder) . This workflow functions as a true standalone image-to-image upscaler .
You can drop ANY image (Midjourney, photo, generate image, old render, or web-export) and it will look like a high-end cinematic remaster with incredibly sharp, non-blurry skin pores, teeth, and textures.
---
🌟 Key Features
True Standalone: Just load any image file and upscale.
Zero Prompts Needed: You don’t need to type or guess prompts to describe the image. The workflow works entirely out-of-the-box using the image data itself.
Integrated Color Match: Includes a dedicated color matching correction step. This completely eliminates "color drift" (unwanted tint shifts, gamma corruption, or washing out of colors) that often happens during heavy diffusion upscaling, keeping your final image 100% true to the original palette.
Auto-Aspect Ratio & Resolution Fitting: You don't need to manually calculate resolutions or crop your images. The workflow automatically analyzes the input image's dimensions and dynamically adjusts tiles to fit the PiD model requirements, preventing any stretching, squishing, or artifacting.
Smart Tiling: Configured with efficient tile overlap to completely eliminate visible seams during the multi-stage upscale process.
Visual Comparison Included: Features the excellent rgthree-comfy Image Comparer node so you can inspect your results instantly right inside ComfyUI.
---
⚠️ IMPORTANT: Tile Size Configuration (512 vs 1024 models)
By default, this workflow is pre-configured for 1024 to 4096 PiD models (with Input Image Tile - Width/Height set to 1024 and overlap to 32 ).
If you decide to use 512 to 2048 PiD models , you MUST manually change the tile settings in the PID Upscaler node:
Set Input Image Tile - Width to 512
Set Input Image Tile - Height to 512
Note: You can leave the tile overlap at 32 —it works perfectly for both 1024 and 512 models and will provide even smoother blending on 512 scales!
---
⏳ Performance Note: Tiling & Render Time
Please note that because this workflow automatically segments the image into precise tiles to handle any resolution and aspect ratio flawlessly without distortion, the upscale time will be increased . The model processes tiles sequentially to ensure seamless, ultra-high-resolution results, so it requires a bit of patience for larger files!
---
💡 Model Recommendations (Tested & Verified)
Based on extensive testing, here is how you should choose your PiD model depending on your art style:
For Realism / Cinematic Photos: Use PiD SDXL . It excels at hallucinating hyper-realistic micro-details like skin pores, fabric textures, eyes, and cinematic bokeh.
For Anime / Illustrative Art: Use PiD Qwen . It handles flat colors, clean line art, and stylized shading much better without forcing unwanted realistic textures onto 2D characters.
---
📦 Required Models & Download Links
To make this workflow work, you need to download three essential files and place them in your ComfyUI directory:
1. PiD Model (Choose SDXL or Qwen based on style)
File Example: pid_sdxl_1024_to_4096_4steps_bf16.safetensors (or Qwen / 512 alternatives)
Path: ComfyUI/models/diffusion_models/
Link: [Download PiD Models]
2. Text Encoder (Clip)
File Example: gemma_2_2b_it_elm_fp8_scaled.safetensors
Path: ComfyUI/models/clip/
Link: [Download Gemma 2 PID Text Encoder]
3. VAE Target
File Example: sdxl_vae.safetensors (or any compatible with the base PiD model of your choice)
Path: ComfyUI/models/vae/
Link: [Download SDXL VAE]
---
🧩 Custom Nodes Needed
Make sure you have ComfyUI Manager installed. The workflow utilizes:
rgthree-comfy (for advanced routing, seed management, and the Side-by-Side Image Comparer).
If you see any red missing nodes upon loading, simply click "Fetch Updates" and "Install Missing Custom Nodes" inside the ComfyUI Manager and restart.
