Comfyui-LatentUtils

Latent Utils for ComfyUI

Effect Preview
Example output showing original (left) and enhanced image (right)

high_freq_mult: 2 is an extreme value to show the difference (demo only).

Node Preview

🌟 Overview

This custom ComfyUI node performs selective latent denoising and detail enhancement using Fourier Transform (FFT) techniques. It intelligently separates image frequencies to:

• Preserve and enhance important high-frequency details
• Suppress background noise and artifacts
• Output a visual preview of the processing mask

Ideal for refining AI-generated images while maintaining sharp features and eliminating graininess.

Supports models:

• Wan vae (Qwen, etc)
• Flux vae (ZImageTurbo, Flux1.dev, etc)
• Other models not tested

---

⚙️ Installation

1. Navigate to your ComfyUI custom nodes directory: cd ComfyUI/custom_nodes
2. Clone this repository: git clone https://github.com/lrzjason/comfyui-latent-frequency-enhancer.git
3. Restart ComfyUI
4. Find the nodes under:
   • latent → enhancement → Latent Frequency Enhancer (lrzjason)
   • sampling → HFEPostProcessor (lrzjason)

---

🔬 How It Works

1. Frequency Separation Uses FFT to split latent into:

   • Low frequencies: Base composition and smooth areas
   • High frequencies: Details, textures, and noise

2. Smart Mask Generation Creates a dynamic mask using:

   • Sigmoid soft-gating function
   • Adjustable noise threshold
   • Pre-blur for noise coherence

3. Selective Enhancement

   • Boosts important high-frequency details
   • Suppresses noise below threshold
   • Smoothly blends components using frequency-aware masking

4. Visual Feedback Outputs a mask preview showing:

   • White areas: Preserved/enhanced details
   • Black areas: Suppressed noise
   • Gray transitions: Smooth blending zones

🚀 HFEPostProcessor: In-Process Enhancement

The HFEPostProcessor node applies high-frequency enhancement during the sampling process rather than after it. This approach:

• Applies enhancement at specific sampling steps (after basic sampling)
• Integrates frequency enhancement directly into the generation pipeline
• Allows for more refined control by specifying which steps to enhance
• Works as a sampler replacement that combines sampling with enhancement

---

Qwen Preview

🎚️ Parameters Explained

Latent Frequency Enhancer Node

Parameter	Default	Range	Description
Detail Strength (HF Mult)	1.15	1.0-2.0	Multiplier for high-frequency details (values >1 enhance details)
Frequency Split Sigma	2.0	0.1-20.0	Controls frequency separation sharpness (higher = more low frequencies preserved)
Noise Threshold	0.05	0.0-1.0	Minimum magnitude to preserve details (higher = more aggressive denoising)
Mask Hardness	2.0	1.0-100.0	Transition sharpness in noise suppression (higher = sharper cutoff)
Noise Grouping (Pre-Blur)	0.5	0.0-1.0	Pre-blur strength for noise coherence (0.0 disables)

HFEPostProcessor Node

Parameter	Default	Range	Description
Model	-	-	The diffusion model to use for sampling
Steps	8	1-10000	Total number of sampled steps
HFE Steps	2	1-100	Number of steps to apply high-frequency enhancement
Latent Image	-	-	The latent image to enhance
Noise Seed	0	0-18446744073709551615	Random seed for noise generation
CFG Scale	1.0	0.0-100.0	Classifier-free guidance scale
Sampler Name	-	Various	Name of the sampler to use
Scheduler	-	Various	Scheduler to use for sampling
Positive	-	-	Positive conditioning
Negative	-	-	Negative conditioning
Detail Strength (HF Mult)	1.05	1.0-2.0	Multiplier for high-frequ