These nodes enhance the KSampler functionality within ComfyUI, specifically designed to work with the Wan2.2 A14B models that utilize a Mixture of Expert architecture. They automate the transition between high and low noise models based on a calculated signal-to-noise ratio during the diffusion process.
- Supports Mixture of Expert Flow models, improving sampling accuracy.
- Automatically switches to low noise models at the optimal diffusion timestep, enhancing output quality.
- Provides flexibility with adjustable boundary parameters tailored for specific model types.
Context
This tool is a specialized modification of the KSampler within ComfyUI, aimed at optimizing the performance of the Wan2.2 A14B models. Its primary function is to facilitate the efficient use of high and low noise experts in the Mixture of Expert architecture by automating the transition between them during the denoising process.
Key Features & Benefits
The key feature of this tool is its ability to automatically adjust the model being used based on the current diffusion timestep. This means that instead of manually determining when to switch from the high noise expert to the low noise expert, users can rely on the tool to make this transition at the precise moment when the signal-to-noise ratio is optimal, significantly improving the fidelity of the generated outputs.
Advanced Functionalities
An advanced capability of this tool lies in its boundary parameter, which dictates the specific diffusion timestep at which the low noise expert should be engaged. For different model configurations like Wan 2.2 T2V and I2V, specific values (0.875 and 0.900, respectively) are recommended, allowing for tailored performance adjustments. Understanding the relationship between diffusion timesteps and denoising steps is crucial for leveraging this functionality effectively.
Practical Benefits
By automating the model switching process based on diffusion timesteps, this tool enhances the workflow within ComfyUI. It provides users with greater control over the quality of their outputs, minimizes guesswork in model management, and ultimately leads to more efficient and effective image and video generation processes.
Credits/Acknowledgments
This project builds upon code from ComfyUI, which is licensed under GPL 3.0, and thus inherits the same licensing terms. For further details, users are encouraged to refer to the LICENSE file included in the repository.
