A custom experimental sampler designed for ComfyUI, the DistanceSampler utilizes relative distances to enhance image generation quality. By slowing down the initial steps and accelerating towards the end, it aims to achieve greater precision during the crucial early phase of image creation.
- The sampler significantly reduces issues such as body horror and unwanted merging of subjects, resulting in cleaner outputs.
- It operates effectively with fewer steps (recommended 7) while maintaining high-quality results, even with a low CFG scale.
- Variants of the sampler, particularly the "n" and "p" versions, offer improved outcomes by utilizing unconditional prediction and step-by-step comparisons, respectively.
Context
DistanceSampler is an innovative tool within the ComfyUI framework that modifies the Heun sampling technique by focusing on normalized distances. Its primary goal is to enhance the quality of image generation, particularly during the initial stages, where precision is critical.
Key Features & Benefits
This sampler's unique approach allows for a slower, more deliberate beginning, which is crucial for achieving detailed and accurate results. Users benefit from reduced artifacts like body horror, making it easier to generate visually appealing images with fewer steps than conventional samplers.
Advanced Functionalities
DistanceSampler includes specialized variants such as the "n" version, which employs unconditional predictions to refine outputs further, and the "p" version, which enhances results by comparing each step to previous ones. These advanced functionalities help users produce smoother and more coherent images.
Practical Benefits
By improving the control and quality of image generation, DistanceSampler streamlines the workflow in ComfyUI. Its ability to deliver high-quality results with fewer steps enhances efficiency, allowing artists to focus on creativity rather than troubleshooting common issues associated with image generation.
Credits/Acknowledgments
The DistanceSampler was developed by the original author and contributors, with special thanks to Blepping for providing an ancestral version. The tool is available under an open-source license, facilitating community collaboration and further development.