Latent Utils for ComfyUI

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

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

🌟 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

---

🎚️ 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-frequency details during enhancement | | Frequency Split Sigma | 5.0 | 0.01-20.0 | Controls frequency separation during enhancement | | Noise Threshold | 0.05 | 0.0-1.0 | Minimum magnitude to preserve details during enhancement | | Mask Hardness | 2.0 | 0.01-100.0 | Transition sharpness in noise suppression during enhancement | | Noise Grouping (Pre-Blur) | 0.5 | 0.0-1.0 | Pre-blur strength for noise coherence during enhancement |

---

🖼️ Output Preview

The Latent Frequency Enhancer node outputs two items:

1. Enhanced Latent (enhanced_latent) The processed latent ready for decoding

2. Mask Preview (mask_preview) Visual representation of the processing mask

The HFEPostProcessor node outputs one item:

1. Enhanced Latent (LATENT) The final processed latent after both sampling and enhancement

---

💡 Pro Tips

For Latent Frequency Enhancer Node:

1. Start with defaults for most images, then adjust:

   • Increase Detail Strength for sharper outputs
   • Raise Noise Threshold for noisy generations
   • Lower Frequency Split Sigma for cartoon/anime styles

2. Mask interpretation:

   • If mask shows important details as black → Lower noise threshold
   • If noise remains visible → Increase mask hardness
   • For soft-focus effects → Increase pre-blur sigma

3. Combine with other nodes:

   • Use after KSampler but before VAEDecode
   • Chain with ControlNet for detail preservation
   • Follow with Color Correct nodes for final polish

For HFEPostProcessor Node:

1. Integration with sampling:

   • The Steps parameter should match your basic sampling steps
   • Use HFE Steps to specify how many steps to apply high-frequency enhancement
   • Start with fewer HFE steps (2-4) and adjust as needed

2. Parameter adjustments:

   • Use lower Detail Strength values (1.05-1.15) compared to post-process mode
   • Higher Frequency Split Sigma (5.0+) may work better during sampling
   • The enhancement happens mid-generation, so parameters may differ from post-processing

3. Workflow integration:

   • Replace standard samplers with HFEPostProcessor to combine sampling and enhancement
   • The node handles both sampling and enhancement in one step
   • Use when you want the enhancement to influence the remaining sampling steps

---

📜 Technical Notes

For Latent Frequency Enhancer Node:

• WAN Format Compatible: Automatically handles WAN-style latent tensors
• Memory Efficient: Processes entirely on GPU when available
• Deterministic: Uses stable FFT operations with no random elements

For HFEPostProcessor Node:

• Integrated Sampling: Combines sampling and enhancement in one node
• Step Control: Allows specifying when during sampling enhancement begins
• Compatible: Works with all standard ComfyUI samplers and schedulers
• Efficient: Applies enhancement only at specified steps for optimal performance

---

🙏 Credits

• FFT implementation inspired by academic frequency-domain processing papers
• Sigmoid gating adapted from noise2noise research
• ComfyUI node template by ComfyOrg

Note: This is a research-grade implementation. Results may vary based on model and generation parameters. Always validate outputs visually.

For research and personal use only. Not for commercial deployment without permission.

Contact

• Twitter: @Lrzjason
• Email: [email protected]
• QQ Group: 866612947
• Wechatid: fkdeai
• Civitai: xiaozhijason

Sponsors me for more open source projects:

<div align="center"> <table> <tr> <td align="center"> <p>Buy me a coffee:</p> <img src="https://github.com/lrzjason/Comfyui-In-Context-Lora-Utils/blob/main/image/bmc_qr.png" alt="Buy Me a Coffee QR" width="200" /> </td> <td align="center"> <p>WeChat:</p> <img src="https://github.com/lrzjason/Comfyui-In-Context-Lora-Utils/blob/main/image/wechat.jpg" alt="WeChat QR" width="200" /> </td> </tr> </table> </div>