PMRF Ultra Fast Upscaler | Low VRAM ComfyUI

1. What is the ComfyUI PMRF Workflow?

The ComfyUI PMRF workflow integrates the revolutionary Posterior-Mean Rectified Flow algorithm into the ComfyUI environment for photo-realistic face restoration. Based on cutting-edge research from Technion—Israel Institute of Technology (ICLR 2025), ComfyUI PMRF addresses the fundamental challenge in image restoration: achieving minimal distortion while maintaining perfect perceptual quality. Unlike traditional methods that rely on posterior sampling or GAN-based approaches, ComfyUI PMRF approximates the mathematically optimal estimator that minimizes Mean Squared Error (MSE) under a perfect perceptual quality constraint.

2. Benefits of ComfyUI PMRF:

⚡ UNPRECEDENTED SPEED PERFORMANCE ⚡

This ComfyUI PMRF workflow is ULTRA-FAST - delivering results in mere seconds! At 1.29 seconds for 2x upscaling, ComfyUI PMRF is faster than ANY upscale workflow currently available on RunComfy platform. While other methods take minutes, ComfyUI PMRF completes face restoration in the time it takes to blink!

For other upscale workflows, please see the bottom of this page

• Ultra Fast Processing: ComfyUI PMRF achieves 2x face upscaling (512×682 to 1024×1364) in just 1.29 seconds on RTX 4090, compared to minutes required by traditional SD upscaling methods
• Low VRAM Requirements: ComfyUI PMRF operates efficiently with only 3.3GB VRAM, significantly lower than competing solutions (DifFBIR requires 8GB, Topaz PhotoAI needs 20GB)
• Superior Detail Preservation: ComfyUI PMRF's advanced posterior-mean rectified flow algorithm maintains natural facial features while eliminating blur and noise artifacts
• Memory Issue Fixed: This ComfyUI PMRF version resolves the 1GB VRAM occupation bug present in the original PMRF release
• Mathematically Optimal: ComfyUI PMRF provably approximates the theoretically optimal estimator for photo-realistic restoration tasks

3. How to Use the ComfyUI PMRF Workflow

3.1 Generation Methods with ComfyUI PMRF

Example Setup for ComfyUI PMRF Face Restoration:

1. Prepare inputs: In Load Image node:
   • Upload your degraded/blurry face image
   • Ensure image is properly aligned and cropped to focus on the face
2. Configure ComfyUI PMRF node:
   • Set num_steps (25 for speed, 100 for maximum quality)
   • Set scale (2.0 for 2x upscaling, adjust as needed)
3. Click Queue Prompt button to run the ComfyUI PMRF workflow
4. In Save Image: get your enhanced face restoration output

3.2 Parameter Reference for ComfyUI PMRF

ComfyUI PMRF Core Node: This node performs the posterior-mean rectified flow restoration process.

• scale: Upscaling factor for the output image (2.0 = 2x larger, 1.5 = 1.5x larger, etc.).
• num_steps: Number of rectified flow iterations.
• seed: Random seed for reproducible results.
• control_after_generate: Determines seed behavior for batch processing (randomize/fixed).
• interpolation: Resampling method used during upscaling process (lanczos4 recommended for best quality).

3.3 Advanced Optimization with ComfyUI PMRF

Understanding the Scale Parameter in ComfyUI PMRF: The scale parameter controls the upscaling factor - it's a multiplier for your image dimensions. To calculate the correct scale value for ComfyUI PMRF:

Scale Calculation Formula:

scale = Target Resolution ÷ Input Resolution

Practical Examples for ComfyUI PMRF:

• For 4K output (3840×2160): If your input is 1920×1080, use scale: 2.0 (3840÷1920=2.0)
• For 4K output from 1280×720: Use scale: 3.0 (3840÷1280=3.0)
• For 2K output (2560×1440) from 1280×720: Use scale: 2.0 (2560÷1280=2.0)
• For custom sizes: Always divide your target width by input width to get the scale value

💡 Pro Tip: Iterative Enhancement

For heavily degraded images where single-pass results aren't satisfactory, you can use iterative processing: Take the ComfyUI PMRF output and feed it back as input for another round of restoration. This multi-pass approach can achieve even better results for extremely challenging images.

Technical Background

How ComfyUI PMRF Works

Think of image restoration like fixing a blurry photograph. Traditional methods either make the image too smooth (losing important details) or add weird artifacts when trying to make it look natural. ComfyUI PMRF solves this by using a two-step approach: first, it creates the "best guess" of what the clear image should look like, then it uses advanced mathematics to transport this guess to look perfectly natural - like the difference between a rough sketch and a finished painting.

The Science Behind ComfyUI PMRF

ComfyUI PMRF addresses the fundamental "distortion-perception tradeoff" in image restoration. The key insight is that the optimal way to restore images isn't to guess randomly (like most AI methods do), but to follow a mathematically proven path. ComfyUI PMRF first predicts the "posterior mean" (the statistically best guess), then uses "rectified flow" to optimally transport this prediction to the natural image distribution. This ensures both minimal error and maximum visual quality.

More Information about ComfyUI PMRF

For additional details and development references:

• PMRF original research by Technion—Israel Institute of Technology
• Paper: "Posterior-Mean Rectified Flow: Towards Minimum MSE Photo-Realistic Image Restoration" (ICLR 2025)
• Project Page: https://pmrf-ml.github.io/
• Online Demo: https://huggingface.co/spaces/ohayonguy/PMRF

Acknowledgements

This ComfyUI PMRF workflow is powered by PMRF (Posterior-Mean Rectified Flow), developed by Guy Ohayon, Tomer Michaeli, and Michael Elad from Technion—Israel Institute of Technology. The research was published at ICLR 2025.

The ComfyUI PMRF integration includes bug fixes for memory issues in the original implementation. Full credit goes to the original authors for their groundbreaking work in photo-realistic image restoration.