Files
2025-12-25 00:25:26 +08:00

133 lines
4.0 KiB
Markdown

# ComfyUI-LG_SamplingUtils
![Image](./assets/1.jpg)
[中文文档](README_CN.md) | **English**
---
## Overview
**ComfyUI-LG_SamplingUtils** is a comprehensive toolset designed for ComfyUI by LAOGOU-666, providing a series of practical sampling nodes that make operations more intuitive and convenient. This extension focuses on advanced sampling techniques, particularly optimized for Flow Matching models like ZImage and Lumina2.
## Features
This extension includes four powerful nodes:
### 1. 🎈 ZImage Timestep Noise
Adds noise perturbation to timesteps during the sampling process to break model homogenization and produce more diverse outputs with different seeds.
**Key Features:**
- Two modes: `sigma` (traditional diffusion models) and `flow` (Flow Matching models)
- Adjustable noise strength and application range
- Optional mask support for localized effects
- Seed-based reproducibility
**Parameters:**
- `mode`: Choose between `sigma` (multiplicative noise) or `flow` (additive noise)
- `noise_strength`: Control the intensity of noise (0.0-2.0)
- `seed`: Random seed for reproducibility
- `start_percent` / `end_percent`: Define the sampling range where noise is applied
- `mask` (optional): Limit the effect to specific regions
### 2. 🎈 LG Noise Injection
Injects features from a reference image into the generation process through the CFG mechanism, allowing the model to "learn" specific qualities from the reference.
**Key Features:**
- Inject surface details like water droplets, sweat, textures
- Add material properties and reflections
- Mask support for targeted feature injection
- Strength decay over sampling steps
**Parameters:**
- `reference_image`: Image containing desired features
- `strength`: Injection strength (0.1-0.2 subtle, 0.2-0.4 noticeable)
- `start_percent` / `end_percent`: Define when feature injection occurs
- `mask` (optional): White areas receive feature injection
**Variant:** 🎈 **LG Noise Injection (Latent)** - Works directly with latent representations and automatically uses `noise_mask` if present in the latent.
### 3. Model Sampling ZImage
Adjusts sampling parameters for ZImage/Lumina2 models.
**Key Features:**
- Proper timestep scaling for Flow Matching models
- Adjustable noise schedule shift
- Compatible with ZImage, Lumina2, and AuraFlow
**Parameters:**
- `shift`: Noise schedule shift (default 3.0 for ZImage)
- `shift=1.0`: Linear schedule
- `shift>1.0`: Shifts toward high noise, more aggressive early steps
- `multiplier`: Timestep multiplier (1.0 for ZImage/AuraFlow, 1000 for SD3/Flux)
### 4. Sigmas Editor 🎚️
Interactive visual editor for adjusting the sigmas curve in real-time.
**Key Features:**
- Drag-and-drop curve editing
- Real-time visualization
- Fine-tune noise schedules for optimal results
## Examples
<div align="center">
![Example 1](./assets/01.jpg)
![Example 2](./assets/02.jpg)
![Example 3](./assets/03.jpg)
![Example 4](./assets/04.jpg)
![Example 5](./assets/05.jpg)
![Example 6](./assets/06.jpg)
![Example 7](./assets/07.jpg)
![Example 8](./assets/08.jpg)
</div>
## Installation
### Method 1: ComfyUI Manager (Recommended)
1. Open ComfyUI Manager
2. Search for "ComfyUI-LG_SamplingUtils"
3. Click Install
### Method 2: Manual Installation
```bash
cd ComfyUI/custom_nodes
git clone https://github.com/LAOGOU-666/ComfyUI-LG_SamplingUtils.git
```
Restart ComfyUI after installation.
## Contact
- **WeChat**: wenrulaogou2033
- **Bilibili**: [老狗_学习笔记](https://space.bilibili.com/175873218?spm_id_from=333.788.0.0)
## Repository
[https://github.com/LAOGOU-666/ComfyUI-LG_SamplingUtils](https://github.com/LAOGOU-666/ComfyUI-LG_SamplingUtils)
## Support
If you encounter any issues or have suggestions, please open an issue on GitHub.
## Changelog
### v1.0.0
- Initial release
- Added ZImage Timestep Noise node
- Added LG Noise Injection nodes (Image and Latent variants)
- Added Model Sampling ZImage node
- Added Sigmas Editor node
---
**Enjoy creating with ComfyUI-LG_SamplingUtils! 🎈**