This repository provides a ComfyUI extension that allows users to schedule the Classifier-Free Guidance (CFG) scale value over the sampling time in a dynamic manner. It includes various nodes that facilitate the adjustment of CFG values based on a defined schedule, enhancing the control over the diffusion process.
- Supports multiple scheduling methods including cosine, Gaussian, and log-normal distributions for CFG adjustments.
- Offers advanced nodes like
PerpNegSheduledCFGGuiderthat allow for both positive and negative conditioning, enabling more nuanced output control. - Includes utility nodes such as
ConcatSigmasandInvertSigmasthat help manipulate sigma value sequences for improved sampling flexibility.
Context
This extension is designed to enhance the capabilities of ComfyUI by allowing users to implement scheduled adjustments to the CFG scale during the sampling process. By enabling dynamic changes to the guidance scale based on the noise schedule, it aims to improve the quality and precision of generated outputs in diffusion models.
Key Features & Benefits
The extension introduces several nodes that allow for intricate control over the CFG scheduling process. Each node serves a specific purpose, such as dynamically adjusting CFG values, applying negative conditioning, or generating sigma values through various mathematical functions. This level of customization is crucial for users looking to fine-tune their generative models.
Advanced Functionalities
The extension features advanced nodes like PerpNegSheduledCFGGuider, which allows users to apply negative conditioning to the CFG scale. This functionality is particularly beneficial for penalizing unwanted features in the generated outputs, thus providing greater creative control. Additionally, nodes like GaussianScheduler and LogNormalScheduler allow for the generation of sigma values through statistical methods, providing flexibility in how sigma values are applied.
Practical Benefits
By integrating this extension into their workflow, users can achieve more refined control over the sampling process in ComfyUI. The ability to schedule CFG values dynamically leads to improved output quality and efficiency, allowing for more precise adjustments based on the specific needs of their projects. This ultimately enhances the overall user experience and effectiveness of the generative models.
Credits/Acknowledgments
The extension builds upon the foundational work of Clybius, whose source code for the WarmupDecayCFGGuider has been adapted for this tool. The repository is open source, allowing for community contributions and further enhancements.